When MSI is off, each interrupt needs to be bounced through the io
thread when it's set/cleared, so vhost-net causes more context switches and
higher CPU utilization than userspace virtio which handles networking in
the same thread.
We'll need to fix this by adding level irq support in kvm irqfd,
for now disable vhost-net in these configurations.
Added a vhostforce flag to force vhost-net back on.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
- Don't return status from start/stop functions where it's ignored
- report errors to make debugging easier
- assert on unexpected failures
- don't disable notifiers on error so that we'll
retry when guest driver restarts
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Virtqueue notify is currently handled synchronously in userspace virtio. This
prevents the vcpu from executing guest code while hardware emulation code
handles the notify.
On systems that support KVM, the ioeventfd mechanism can be used to make
virtqueue notify a lightweight exit by deferring hardware emulation to the
iothread and allowing the VM to continue execution. This model is similar to
how vhost receives virtqueue notifies.
The result of this change is improved performance for userspace virtio devices.
Virtio-blk throughput increases especially for multithreaded scenarios and
virtio-net transmit throughput increases substantially.
Some virtio devices are known to have guest drivers which expect a notify to be
processed synchronously and spin waiting for completion.
For virtio-net, this also seems to interact with the guest stack in strange
ways so that TCP throughput for small message sizes (~200bytes)
is harmed. Only enable ioeventfd for virtio-blk for now.
Care must be taken not to interfere with vhost-net, which uses host
notifiers. If the set_host_notifier() API is used by a device
virtio-pci will disable virtio-ioeventfd and let the device deal with
host notifiers as it wishes.
Finally, there used to be a limit of 6 KVM io bus devices inside the
kernel. On such a kernel, don't use ioeventfd for virtqueue host
notification since the limit is reached too easily. This ensures that
existing vhost-net setups (which always use ioeventfd) have ioeventfds
available so they can continue to work.
After migration and on VM change state (running/paused) virtio-ioeventfd
will enable/disable itself.
* VIRTIO_CONFIG_S_DRIVER_OK -> enable virtio-ioeventfd
* !VIRTIO_CONFIG_S_DRIVER_OK -> disable virtio-ioeventfd
* virtio_pci_set_host_notifier() -> disable virtio-ioeventfd
* vm_change_state(running=0) -> disable virtio-ioeventfd
* vm_change_state(running=1) -> enable virtio-ioeventfd
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The VirtIOPCIProxy bugs field is currently used to enable workarounds
for older guests. Rename it to flags so that other per-device behavior
can be tracked.
A later patch uses the flags field to remember whether ioeventfd should
be used for virtqueue host notification.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Add "fw_name" to DeviceInfo to use in device path building. In
contrast to "name" "fw_name" should refer to functionality device
provides instead of particular device model like "name" does.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
This patch enables MSI-X for virtfs-9p-pci. It also adds a
compat property to pc-0.13 which turns it of there to stay
compatible to 0.13-stable.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
When using irqfd with vhost-net to inject interrupts,
a single evenfd might inject multiple interrupts.
Implementing this is much easier with a single
per-device callback to set guest notifiers.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Based on a patch from Mark McLoughlin, this patch introduces a new
bottom half packet transmitter that avoids the latency imposed by
the tx_timer approach. Rather than scheduling a timer when a TX
packet comes in, schedule a bottom half to be run from the iothread.
The bottom half handler first attempts to flush the queue with
notification disabled (this is where we could race with a guest
without txburst). If we flush a full burst, reschedule immediately.
If we send short of a full burst, try to re-enable notification.
To avoid a race with TXs that may have occurred, we must then
flush again. If we find some packets to send, the guest it probably
active, so we can reschedule again.
tx_timer and tx_bh are mutually exclusive, so we can re-use the
tx_waiting flag to indicate one or the other needs to be setup.
This allows us to seamlessly migrate between timer and bh TX
handling.
The bottom half handler becomes the new default and we add a new
tx= option to virtio-net-pci. Usage:
-device virtio-net-pci,tx=timer # select timer mitigation vs "bh"
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
If virtio_net_flush_tx() is called with notification disabled, we can
race with the guest, processing packets at the same rate as they
get produced. The trouble is that this means we have no guaranteed
exit condition from the function and can spend minutes in there.
Currently flush_tx is only called with notification on, which seems
to limit us to one pass through the queue per call. An upcoming
patch changes this.
Also add an option to set this value on the command line as different
workloads may wish to use different values. We can't necessarily
support any random value, so this is a developer option: x-txburst=
Usage:
-device virtio-net-pci,x-txburst=64 # 64 packets per tx flush
One pass through the queue (256) seems to be a good default value
for this, balancing latency with throughput. We use a signed int
for x-txburst because 2^31 packets in a burst would take many, many
minutes to process and it allows us to easily return a negative
value value from virtio_net_flush_tx() to indicate a back-off
or error condition.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Add an option to make the TX mitigation timer adjustable as a device
option. The 150us hard coded default used currently is reasonable,
but may not be suitable for all workloads, this gives us a way to
adjust it using a single binary. We can't support any random option
though, so use the "x-" prefix to indicate this is a developer
option. Usage:
-device virtio-net-pci,x-txtimer=500000,... # .5ms timeout
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Changing block.h or blockdev.h resulted in recompiling most objects.
Move DriveInfo typedef and BlockInterfaceType enum definitions
to qemu-common.h and rearrange blockdev.h use to decrease churn.
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
Free malloc'ed memory, unregister from savevm and clean up virtio-common
bits on device hot-unplug.
This was found performing a migration after device hot-unplug.
Reported-by: <lihuang@redhat.com>
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Otherwise we can't migrate after we've removed a virtio block device.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Move the check from virtio_blk_init_pci(), where it protects only
virtio-blk-pci, to virtio_blk_init(). Without that, virtio-blk-s390
initializes without a drive. I figure that can lead to null pointer
dereferences.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
It can't actually fail now, but the next commit will change that.
s390_virtio_blk_init() already checks for failure, but
virtio_blk_init_pci() doesn't. Fix that.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Don't overwrite pci header type.
Otherwise, multi function bit which pci_init_header_type() sets
appropriately is lost.
Anyway PCI_HEADER_TYPE_NORMAL is zero, so it is unnecessary to zero
which is already zero cleared.
how to test:
run qemu and issue info pci to see whether a device in question is
normal device, not pci-to-pci bridge.
This is handy because guest os isn't required.
tested changes:
The following files are covered by using following commands.
sparc64-softmmu
apb_pci.c, vga-pci.c, cmd646.c, ne2k_pci.c, sun4u.c
ppc-softmmu
grackle_pci.c, cmd646.c, ne2k_pci.c, vga-pci.c, macio.c
ppc-softmmu -M mac99
unin_pci.c(uni-north, uni-north-agp)
ppc64-softmmu
pci-ohci, ne2k_pci, vga-pci, unin_pci.c(u3-agp)
x86_64-softmmu
acpi_piix4.c, ide/piix.c, piix_pci.c
-vga vmware vmware_vga.c
-watchdog i6300esb wdt_i6300esb.c
-usb usb-uhci.c
-sound ac97 ac97.c
-nic model=rtl8139 rtl8139.c
-nic model=pcnet pcnet.c
-balloon virtio virtio-pci.c:
untested changes:
The following changes aren't tested.
prep_pci.c: ppc-softmmu -M prep should cover, but core dumped.
unin_pci.c(uni-north-pci): the caller is commented out.
openpic.c: the caller is commented out in ppc_prep.c
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
Make the property point to BlockDriverState, cutting out the DriveInfo
middleman. This prepares the ground for block devices that don't have
a DriveInfo.
Currently all user-defined ones have a DriveInfo, because the only way
to define one is -drive & friends (they go through drive_init()).
DriveInfo is closely tied to -drive, and like -drive, it mixes
information about host and guest part of the block device. I'm
working towards a new way to define block devices, with clean
host/guest separation, and I need to get DriveInfo out of the way for
that.
Fortunately, the device models are perfectly happy with
BlockDriverState, except for two places: ide_drive_initfn() and
scsi_disk_initfn() need to check the DriveInfo for a serial number set
with legacy -drive serial=... Use drive_get_by_blockdev() there.
Device model code should now use DriveInfo only when explicitly
dealing with drives defined the old way, i.e. without -device.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
We automatically delete blockdev host parts on unplug of the guest
device. Too much magic, but we can't change that now.
The delete happens early in the guest device teardown, before the
connection to the host part is severed. Thus, the guest part's
pointer to the host part dangles for a brief time. No actual harm
comes from this, but we'll catch such dangling pointers a few commits
down the road. Clean up the dangling pointers by delaying the
automatic deletion until the guest part's pointer is gone.
Device usb-storage deliberately makes two qdev properties refer to the
same drive, because it automatically creates a second device. Again,
too much magic we can't change now. Multiple references worked okay
before, but now free_drive() dies for the second one. Zap the extra
reference.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The comment suggests we're checking for the driver in the ready
state and bus master disabled, but the code is checking that it's
not in the ready state.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Found-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch adds required infrastructure for the new security model.
- A new configure option for attr/xattr.
- if CONFIG_VIRTFS will be defined if both CONFIG_LINUX and CONFIG_ATTR defined.
- Defines routines related to both security models.
Signed-off-by: Venkateswararao Jujjuri <jvrao@linux.vnet.ibm.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Anything that moves hundreds of lines out of vl.c can't be all bad.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch doesn't implement the 9p protocol handling
code. It adds a simple device which dump the protocol data.
[jvrao@linux.vnet.ibm.com: Little-Endian to host format conversion]
[aneesh.kumar@linux.vnet.ibm.com: Multiple-mounts support]
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Support host/guest notifiers in virtio-pci.
The last one only with kvm, that's okay
because vhost relies on kvm anyway.
Note on kvm usage: kvm ioeventfd API
is implemented on non-kvm systems as well,
this is the reason we don't need if (kvm_enabled())
around it.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
vhost net backend needs to be notified when
frontend status changes. Add a callback,
similar to set_features.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Older Linux guests don't activate the bus master enable bit. So for those we
can just try to be clever and track if they set the DEVICE_OK bit even though
bus mastering is still disabled.
Under that condition we can disable the windows safety check. With that logic
in place both guests should work just fine. Without PCI hotplug breaks
virtio-net in Linux < 2.6.34 guests.
Signed-off-by: Alexander Graf <agraf@suse.de>
CC: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
error_report() terminates the message with a newline. Strip it it
from its arguments.
This fixes a few error messages lacking a newline:
net_handle_fd_param()'s "No file descriptor named %s found", and
tap_open()'s "vnet_hdr=1 requested, but no kernel support for
IFF_VNET_HDR available" (all three versions).
There's one place that passes arguments without newlines
intentionally: load_vmstate(). Fix it up.
Use the named constant instead of -1.
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Reported-by: "Michael S. Tsirkin" <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Michael noted we don't allow disabling of MSI for the virtio-serial-pci
device. Fix that.
Signed-off-by: Amit Shah <amit.shah@redhat.com>
CC: "Michael S. Tsirkin" <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Add three new qdev properties to export block topology information to
the guest. This is needed to get optimal I/O alignment for RAID arrays
or SSDs.
The options are:
- physical_block_size to specify the physical block size of the device,
this is going to increase from 512 bytes to 4096 kilobytes for many
modern storage devices
- min_io_size to specify the minimal I/O size without performance impact,
this is typically set to the RAID chunk size for arrays.
- opt_io_size to specify the optimal sustained I/O size, this is
typically the RAID stripe width for arrays.
I decided to not auto-probe these values from blkid which might easily
be possible as I don't know how to deal with these issues on migration.
Note that we specificly only set the physical_block_size, and not the
logial one which is the unit all I/O is described in. The reason for
that is that IDE does not support increasing the logical block size and
at last for now I want to stick to one meachnisms in queue and allow
for easy switching of transports for a given backing image which would
not be possible if scsi and virtio use real 4k sectors, while ide only
uses the physical block exponent.
To make this more common for the different block drivers introduce a
new BlockConf structure holding all common block properties and a
DEFINE_BLOCK_PROPERTIES macro to add them all together, mirroring
what is done for network drivers. Also switch over all block drivers
to use it, except for the floppy driver which has weird driveA/driveB
properties and probably won't require any advanced block options ever.
Example usage for a virtio device with 4k physical block size and
8k optimal I/O size:
-drive file=scratch.img,media=disk,cache=none,id=scratch \
-device virtio-blk-pci,drive=scratch,physical_block_size=4096,opt_io_size=8192
aliguori: updated patch to take into account BLOCK events
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This commit enables the use of MSI interrupts for virtqueue
notifications for ports. We use nr_ports + 1 (for control channel) msi
entries for the ports, as only the in_vq operations need an interrupt on
the guest.
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This commit converts the virtio-console device to create a new
virtio-serial bus that can host console and generic serial ports. The
file hosting this code is now called virtio-serial-bus.c.
The virtio console is now a very simple qdev device that sits on the
virtio-serial-bus and communicates between the bus and qemu's chardevs.
This commit also includes a few changes to the virtio backing code for
pci and s390 to spawn the virtio-serial bus.
As a result of the qdev conversion, we get rid of a lot of legacy code.
The old-style way of instantiating a virtio console using
-virtioconsole ...
is maintained, but the new, preferred way is to use
-device virtio-serial -device virtconsole,chardev=...
With this commit, multiple devices as well as multiple ports with a
single device can be supported.
For multiple ports support, each port gets an IO vq pair. Since the
guest needs to know in advance how many vqs a particular device will
need, we have to set this number as a property of the virtio-serial
device and also as a config option.
In addition, we also spawn a pair of control IO vqs. This is an internal
channel meant for guest-host communication for things like port
open/close, sending port properties over to the guest, etc.
This commit is a part of a series of other commits to get the full
implementation of multiport support. Future commits will add other
support as well as ride on the savevm version that we bump up here.
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Since patch ed757e140c0ada220f213036e4497315d24ca8bct, virtio will
sometimes clear all status registers on bus master disable, which loses
information such as VIRTIO_CONFIG_S_FAILED bit. This is a result of
a patch being misapplied: code uses ! instead of ~ for bit
operations as in Yan's original patch. This obviously does not make
sense.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Add feature bits as properties to virtio. This makes it possible to e.g. define
machine without indirect buffer support, which is required for 0.10
compatibility, or without hardware checksum support, which is required for 0.11
compatibility. Since default values for optional features are now set by qdev,
get_features callback has been modified: it sets non-optional bits, and clears
bits not supported by host.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Rename features->guest_features. This is
what they are, avoid confusion with
host features which we also need to keep around.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch adds a romfile property to the pci bus. It allows to specify
a romfile to load into the rom bar of the pci device. The default value
comes from a new field in PCIDeviceInfo. The property allows to change
the file and also to disable the rom loading using an empty string.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Currently, we preload option roms into the option rom space in memory. This
prevents DDIM from functioning correctly which severely limits the number
of roms we can support.
This patch introduces a pci_add_option_rom() which registers the
PCI_ROM_ADDRESS bar which points to our option rom. It also converts over
the cirrus vga adapter, the rtl8139, virtio, and the e1000 to use this
new mechanism.
The result is that PXE boot functions even with three unique types of cards.
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
migrating between hosts which have different features
might break silently, if the migration destination
does not support some features supported by source.
Prevent this from happening by comparing acked feature
bits with the mask supported by the device.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
msix state is managed by OS, not the
driver, so it's wrong to touch it
on io from driver.
Mark all vectors unused instead.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This patch is preliminary for 64 bit BAR support.
Introduce dedicated type, pcibus_t, to represent pci bus address/size
instead of uint32_t.
Later this type will be changed to uint64_t.
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
make constants for pci base address match pci_regs.h by
renaming PCI_ADDRESS_SPACE_xxx to PCI_BASE_ADDRESS_SPACE_xxx.
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Making pci device cleanup msix automatically makes pci.c depend on
msix.c, which is IMO messy. Since devices do msix_init it's easy and
natural for them to also do msix_uninit.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Cc: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Currently only one virtio_console device is supported. Trying to add
multiple devices fails and such failure should be reported back to the
qdev init functions.
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
it's safe to call msix_write_config if msix
is disabled, so call it unconditionally on
pci config write.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Since cpu_register_phys_memory does not require size to be a multiple of
target page size, simply make msix page size 0x1000. Do this in msix,
reverting part of 5e520a7d50, as we no
longer have to pass target page around.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
virtio pci registers its own reset handler, but fails to unregister it,
which will lead to crashes after device removal. Solve this problem by
switching to qdev reset handler, which is automatically unregistered.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Changes:
* drive_uninit() wants a DriveInfo now.
* drive_uninit() also calls bdrv_delete(),
so callers don't need to do that.
* drive_uninit() calls are moved over to the ->exit()
callbacks, destroy_bdrvs() is zapped.
* setting bdrv->private is not needed any more as the
only user (destroy_bdrvs) is gone.
* usb-storage needs no drive_uninit, scsi-disk will
handle that.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
mst@redhat.com
Stefan Hajnoczi
Gleb Natapov
Gerd Hoffmann
Alex Williamson
Blue Swirl
Amit Shah
Markus Armbruster
Isaku Yamahata
Venkateswararao Jujjuri (JV)
Anthony Liguori
Alexander Graf
Christoph Hellwig