ChangeLog/4.2: Difference between revisions

From QEMU
No edit summary
 
(5 intermediate revisions by 3 users not shown)
Line 3: Line 3:
=== Incompatible changes ===
=== Incompatible changes ===


* The minimum version of the glib library is now 2.48
* On s390x, using KVM now explicitly requires a host kernel version of at least 3.15 (which includes the 'flic' KVM device). This had been broken since QEMU 2.10 already.
* On s390x, using KVM now explicitly requires a host kernel version of at least 3.15 (which includes the 'flic' KVM device). This had been broken since QEMU 2.10 already.
* On s390x, migration of KVM guests with >8TB of memory from pre-4.2 QEMUs will no longer work. Migration of KVM guests with up to 8TB of memory continues to work, as will migration of >8TB guests from QEMU 4.2 on.
* On s390x, migration of KVM guests with >8TB of memory from pre-4.2 QEMUs will no longer work. Migration of KVM guests with up to 8TB of memory continues to work, as will migration of >8TB guests from QEMU 4.2 on.
Line 57: Line 56:
* pseries machine type no longer permits NUMA nodes with no CPU and no memory - this was already unlikely to work due to problems expressing this configuration to the guest
* pseries machine type no longer permits NUMA nodes with no CPU and no memory - this was already unlikely to work due to problems expressing this configuration to the guest
* On pseries-4.2 and later machine types, "vsmt" is set to be equal to smp_threads by default (previously it was different in some circumstances for compatibility with old KVM versions)
* On pseries-4.2 and later machine types, "vsmt" is set to be equal to smp_threads by default (previously it was different in some circumstances for compatibility with old KVM versions)
* New defaults mean that with the pseries machine type and host kernels prior to 4.13 you may get errors about VSMT mode.  You can either update your host kernel or work around this by setting <code>-machine vsmt=8</code>.


=== RISC-V ===
=== RISC-V ===
Line 77: Line 77:
=== x86 ===
=== x86 ===
* VMX features can be enabled/disabled via the "-cpu" flag.
* VMX features can be enabled/disabled via the "-cpu" flag.
* When nested virtualization is enabled with an option like "-cpu Haswell,+vmx", the set of VMX features will also be constrained to what was available on the corresponding CPU model.
* New "microvm" machine type that has virtio-mmio instead of PCI, and no ACPI support (so no hotplug too).  The new machine type is meant as a baseline for performance optimizations of QEMU, firmware and guests.  While inspired by Firecracker it is not entirely compatible with it (for example it does not have Firecracker's userspace IP stack and MicroVM Metadata Service).
* New "microvm" machine type that has virtio-mmio instead of PCI, and no ACPI support (so no hotplug too).  The new machine type is meant as a baseline for performance optimizations of QEMU, firmware and guests.  While inspired by Firecracker it is not entirely compatible with it (for example it does not have Firecracker's userspace IP stack and MicroVM Metadata Service).
* Support for AVX512 BFloat16extensions.
* Support for AVX512 BFloat16 extensions.
* New CPU model Denverton (a server-class Atom-based SoC).
* New CPU models: Denverton (a server-class Atom-based SoC), Snowridge, Dhyana
* New CPU features clzero and xsaveerptr
* New CPU features clzero and xsaveerptr
* Latest version of all CPU models now have TSX (HLE and RTM) disabled by default
* Support for new IA32_ARCH_CAPABILITIES flags: MDS_NO, TSX_NO, TSX_CTRL
* macOS Hypervisor.framework support ("-accel hvf") is now considered stable


=== Xtensa ===
=== Xtensa ===
Line 152: Line 156:


=== Monitor ===
=== Monitor ===
* Fix not to leave the monitor wedged when the client disconnects while the queue is full, and the event timing is unlucky.
* Fix not to leave the monitor wedged when the client disconnects while the queue is full, and the event timing is unlucky.
==== QMP ====
* query-machines now report the default CPU type for each machine


=== Migration ===
=== Migration ===

Latest revision as of 11:24, 3 December 2019

System emulation

Incompatible changes

  • On s390x, using KVM now explicitly requires a host kernel version of at least 3.15 (which includes the 'flic' KVM device). This had been broken since QEMU 2.10 already.
  • On s390x, migration of KVM guests with >8TB of memory from pre-4.2 QEMUs will no longer work. Migration of KVM guests with up to 8TB of memory continues to work, as will migration of >8TB guests from QEMU 4.2 on.
  • The "autoload" parameter of the "block-dirty-bitmap-add" QMP command has been removed (it's no longer needed).

New deprecated options and features

  • The "ide-drive" device is deprecated. Users should use 'ide-hd' or "ide-cd" as appropriate to get an IDE hard disk or CD-ROM as needed.
  • The "change" QMP command has been deprecated. Use "blockdev-change-medium" or "change-vnc-password" instead.
  • The "migrate_set_downtime", "migrate_set_speed" and "migrate-set-cache-size" QMP commands have been deprecated. Use "migrate-set-parameters" instead.
  • The "query-migrate-cache-size" QMP command has been deprecated. Use "query-migrate-parameters" instead.
  • The "scsi-disk" device is deprecated. Users should use 'scsi-hd' or "scsi-cd" as appropriate to get a SCSI hard disk or CD-ROM as needed.

Consult the "Deprecated Features" appendix for the full list of historically deprecated features/options.

Alpha

68k

  • Added new "next-cube" machine for emulating a classic NeXTcube (still incomplete, can only boot to the firmware prompt)
  • Added new "q800" machine for emulating Macintosh Quadra 800 (can only boot Linux, MacOS ROM is not yet supported)

Arm

  • Added Aspeed AST2600 SoC and evaluation board model support
  • QEMU's Arm semihosting implementation now supports v2.0 of the specification, including the "STDOUT_STDERR" and "EXIT_EXTENDED" extensions
  • We now correctly support more than 256 CPUs when using KVM
  • The virt board now supports memory hotplugging, when used with a UEFI guest BIOS and ACPI.
  • The raspberry PI models now implement the bcm2835 system timer device and a dummy thermal sensor.
  • Performance of the TCG emulation of ARM cores has been improved.
  • SVE is supported in KVM guests (requires SVE hardware and kernel support)

HPPA

Microblaze

MIPS

  • Fixed bug involving emulation of ST.W instruction in system mode only (there was no bug in Linux user mode).

Nios2

OpenRISC

PowerPC

  • Behavior for unplug of multifunction PCI devices has been changed to match x86. Attempting to unplug any function will cause an unplug of the whole slot.
  • There are now separate machine types for POWER8 based and POWER9 based non-virtualized machines (powernv8 & powernv9).
  • Default RAM size for powernv machines has been increased to 1.75G which should allow more things to work with the default parameters.
  • RTAS (runtime firmware) code is now supplied by SLOF (boot time firmware) instead of by qemu itself. This won't affect users in most cases, but will affect users of -bios which bypasses SLOF, which includes kvm-unit-tests. You'll need a sufficiently recent version of kvm-unit-tests to handle this.
  • mffsce, mffscrn and mffscrni POWER9 instructions now supported in TCG
  • powernv machine now includes models for the Homer and OCC SRAM system devices
  • "info pic" HMP command now reports whether the irq chip is emulated in kernel or in qemu
  • pseries machine type no longer permits NUMA nodes with no CPU and no memory - this was already unlikely to work due to problems expressing this configuration to the guest
  • On pseries-4.2 and later machine types, "vsmt" is set to be equal to smp_threads by default (previously it was different in some circumstances for compatibility with old KVM versions)
  • New defaults mean that with the pseries machine type and host kernels prior to 4.13 you may get errors about VSMT mode. You can either update your host kernel or work around this by setting -machine vsmt=8.

RISC-V

  • The `-initrd` argument is now supported.
  • The debugger can now see all architectural state regardless of the currently executing privilege mode on each hart.
  • The sifive_u board now contains additional memory regions that more closely match the board.

s390

  • TCG now implements IEP (Instruction Execution Protection), a CPU feature introduced with the z14. Until the "qemu" CPU model is lifted to a z14, it can be enabled using the "max" CPU model ("-cpu max").

SH

SPARC

  • The sun4u IOMMU now support the "invert endianness" bit

TileGX

Tricore

x86

  • VMX features can be enabled/disabled via the "-cpu" flag.
  • When nested virtualization is enabled with an option like "-cpu Haswell,+vmx", the set of VMX features will also be constrained to what was available on the corresponding CPU model.
  • New "microvm" machine type that has virtio-mmio instead of PCI, and no ACPI support (so no hotplug too). The new machine type is meant as a baseline for performance optimizations of QEMU, firmware and guests. While inspired by Firecracker it is not entirely compatible with it (for example it does not have Firecracker's userspace IP stack and MicroVM Metadata Service).
  • Support for AVX512 BFloat16 extensions.
  • New CPU models: Denverton (a server-class Atom-based SoC), Snowridge, Dhyana
  • New CPU features clzero and xsaveerptr
  • Latest version of all CPU models now have TSX (HLE and RTM) disabled by default
  • Support for new IA32_ARCH_CAPABILITIES flags: MDS_NO, TSX_NO, TSX_CTRL
  • macOS Hypervisor.framework support ("-accel hvf") is now considered stable

Xtensa

  • Add new "virt" machine.

Device emulation and assignment

ACPI

  • ARM machines can use the ACPI generic event device for the system powerdown event

Audio

  • Audio devices support an "audiodev" property that can be used to choose a specific backend to connect to the device.

Block devices

  • Unmap operations are now accounted and visible in the output of "query-blockstats"/"info blockstats"

Graphics

  • Support for more ati-vga registers

Input devices

IPMI

  • UUID handling for the built-in IPMI BMC has changed. Before the UUID was set from the qemu UUID, if that was set. However, in a real system, the UUID of a BMC will be independent of the system UUID. So now the UUID must be explicitly set for a BMC if you want one. Otherwise the BMC will not have a UUID. To set the UUID of a BMC, use the new property guid=11223344-5566-7788-99aa-bbccddeeff00 for the ipmi-bmc-sim device.
  • New PCI interfaces for IPMI KCS and BT devices are available. The devices pci-ipmi-kcs and pci-ipmi-bt are used to choose these devices.

Network devices

  • Support for network failover for easier migration of VFIO devices.

NVDIMM

PCI/PCIe

SCSI

SMBIOS

TPM

USB

VFIO

  • vfio-pci supports the "failover_pair_id" property for easier migration of VFIO devices.

virtio

  • virtio-mmio now supports the virtio standard compatible v2 personality (still defaulting to the legacy v1 personality for compatibility reasons)
  • Support for virtio 1.1 packed virtqueues.

Xen

fw_cfg

9pfs

  • the "local" backend now has a new 'multidevs' option to deal with cross-device setups (ie. when the shared directory spans over multiple devices on the host)

Semihosting

Audio

Character devices

  • Any backend that supports connection as a socket client (NBD, chardev, network device, monitor...) can now request to use TCP keep-alive with the server.

Crypto subsystem

  • With new enough gcrypt or nettle libraries, QEMU can now use the library's own XTS cipher mode code. This provides a potentially huge performance boost for AES-XTS encryption, which benefits the LUKS disk encryption block driver I/O performance.

GUI

Host support

Memory backends

Monitor

  • Fix not to leave the monitor wedged when the client disconnects while the queue is full, and the event timing is unlucky.

QMP

  • query-machines now report the default CPU type for each machine

Migration

  • The 'validate-uuid' migration compatibility checks the UUID matches on migration, preventing accidental migration of the wrong VM to the wrong destination. (Yury Kotov)

Network

  • The user mode host network backend now allows to set a guest-visible DNS address which is not in the virtual network, unless restrict mode is enabled.

Block device backends and tools

  • Block drivers can now support BDRV_REQ_PREFETCH for more efficient handling of copy-on-read requests; the NBD driver has wired this up to NBD_CMD_CACHE.
  • The LUKS block driver supports falloc/full preallocation.
  • The NBD server now advertises NBD_FLAG_CAN_MULTI_CONN on shared readonly connections.
  • The NBD server and client support the new NBD_CMD_FLAG_FAST_ZERO as an optimization during the copying of sparse images.
  • The backup block job now creates a filter node, just like mirror and commit do. Clients that care about the structure of the block graph need to take this into account.
  • Several assertion failures relating to the use of an NBD server with iothreads have been fixed.
  • Creation of an NBD client now allows an optional 'reconnect-delay' parameter, which can be used to reconnect rather than permanently fail in the face of a transient server hiccup.
  • NBD client and server now permit export names up to 4k in length (rather than 256 bytes), while at the same time being more careful to avoid over-long strings in violation of the NBD protocol.

Tracing

Miscellaneous

  • bugs in gdbstub handling of F and ! packets have been fixed
  • The arguments to -kernel and -initrd are now mmap-ed for some targets. This allows sharing memory when many VMs are loaded with the same arguments.

User-mode emulation

Alpha

Xtensa

  • call0 ABI may be selected for the user emulation with command line option -xtensa-abi-call0 (or with QEMU_XTENSA_ABI_CALL0 environment variable).

TCG

  • tcg tests are now in $BUILDDIR/tests/tcg/$TARGET/ (rather than $BUILDDIR/$TARGET/tests/)
  • new TCG Plugins feature allows instrumentation experiments to be written

Guest agent

Build Information

Python

  • The minimum supported version of Python is now 3.5.

GIT submodules

Container Based Builds

  • The docker.py tooling now supports podman containers [1] as an alternative to docker
  • the docker.py now requires python3 to run
  • a number of the cross compiler containers have been updated to Buster

Build Dependencies

  • The minimum version of the glib library is now 2.48

Testing

Windows

Known issues