The purpose of virtio and this speciﬁcation is that virtual environments and guests should have a straightforward, eﬃcient, standard and extensible mechanism for virtual devices, rather than boutique per-environment or per-OS mechanisms.
Technical Committee members should send comments on this speciﬁcation to the Technical Committee’s email list. Others should send comments to the Technical Committee by using the "Send A Comment" button on the Technical Committee’s web page at http://www.oasis-open.org/committees/virtio/.
For information on whether any patents have been disclosed that may be essential to implementing this speciﬁcation, and any oﬀers of patent licensing terms, please refer to the Intellectual Property Rights section of the Technical Committee web page (http://www.oasis-open.org/committees/virtio/ipr.php).
Virtual I/O Device (VIRTIO) Version 1.0. 03 December 2013. Committee Speciﬁcation Draft 01 / Public Review Draft 01. http://docs.oasis-open.org/virtio/virtio/v1.0/csprd01/virtio-v1.0-csprd01.html.
Copyright © OASIS Open 2013. All Rights Reserved.
All capitalized terms in the following text have the meanings assigned to them in the OASIS Intellectual Property Rights Policy (the "OASIS IPR Policy"). The full Policy may be found at the OASIS website.
This document and translations of it may be copied and furnished to others, and derivative works that comment on or otherwise explain it or assist in its implementation may be prepared, copied, published, and distributed, in whole or in part, without restriction of any kind, provided that the above copyright notice and this section are included on all such copies and derivative works. However, this document itself may not be modiﬁed in any way, including by removing the copyright notice or references to OASIS, except as needed for the purpose of developing any document or deliverable produced by an OASIS Technical Committee (in which case the rules applicable to copyrights, as set forth in the OASIS IPR Policy, must be followed) or as required to translate it into languages other than English.
The limited permissions granted above are perpetual and will not be revoked by OASIS or its successors or assigns.
This document and the information contained herein is provided on an "AS IS" basis and OASIS DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION HEREIN WILL NOT INFRINGE ANY OWNERSHIP RIGHTS OR ANY IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
OASIS requests that any OASIS Party or any other party that believes it has patent claims that would necessarily be infringed by implementations of this OASIS Committee Speciﬁcation or OASIS Standard, to notify OASIS TC Administrator and provide an indication of its willingness to grant patent licenses to such patent claims in a manner consistent with the IPR Mode of the OASIS Technical Committee that produced this speciﬁcation.
OASIS invites any party to contact the OASIS TC Administrator if it is aware of a claim of ownership of any patent claims that would necessarily be infringed by implementations of this speciﬁcation by a patent holder that is not willing to provide a license to such patent claims in a manner consistent with the IPR Mode of the OASIS Technical Committee that produced this speciﬁcation. OASIS may include such claims on its website, but disclaims any obligation to do so.
OASIS takes no position regarding the validity or scope of any intellectual property or other rights that might be claimed to pertain to the implementation or use of the technology described in this document or the extent to which any license under such rights might or might not be available; neither does it represent that it has made any eﬀort to identify any such rights. Information on OASIS’ procedures with respect to rights in any document or deliverable produced by an OASIS Technical Committee can be found on the OASIS website. Copies of claims of rights made available for publication and any assurances of licenses to be made available, or the result of an attempt made to obtain a general license or permission for the use of such proprietary rights by implementers or users of this OASIS Committee Speciﬁcation or OASIS Standard, can be obtained from the OASIS TC Administrator. OASIS makes no representation that any information or list of intellectual property rights will at any time be complete, or that any claims in such list are, in fact, Essential Claims.
The name "OASIS" is a trademark of OASIS, the owner and developer of this
speciﬁcation, and should be used only to refer to the organization and its oﬃcial
outputs. OASIS welcomes reference to, and implementation and use of, speciﬁcations,
while reserving the right to enforce its marks against misleading uses. Please see
https://www.oasis-open.org/policies-guidelines/trademark for above guidance.
The purpose of virtio and this speciﬁcation is that virtual environments and guests should have a straightforward, eﬃcient, standard and extensible mechanism for virtual devices, rather than boutique per-environment or per-OS mechanisms.
Straightforward: Virtio devices use normal bus mechanisms of interrupts and DMA which should be familiar to any device driver author. There is no exotic page-ﬂipping or COW mechanism: it’s just a normal device. 1
Eﬃcient: Virtio devices consist of rings of descriptors for input and output, which are neatly separated to avoid cache eﬀects from both driver and device writing to the same cache lines.
Standard: Virtio makes no assumptions about the environment in which it operates, beyond supporting the bus attaching the device. Virtio devices are implemented over PCI and other buses, and earlier drafts been implemented on other buses not included in this spec. 2
Extensible: Virtio PCI devices contain feature bits which are acknowledged by the guest operating system during device setup. This allows forwards and backwards compatibility: the device oﬀers all the features it knows about, and the driver acknowledges those it understands and wishes to use.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].
S. Bradner, Key words for use in RFCs to Indicate Requirement Levels,
z/Architecture Principles of Operation,
|[S390 Common I/O]||
ESA/390 Common I/O-Device and Self-Description,
Unless explicitly speciﬁed otherwise, all multi-byte ﬁelds are little-endian. To reinforce this the examples use typenames like "le16" instead of "uint16_t".
The driver MUST update the Device Status ﬁeld in the order below to indicate its progress. This provides a simple low-level diagnostic: it’s most useful to imagine them hooked up to traﬃc lights on the console indicating the status of each device. The driver MUST NOT clear a device status bit.
This ﬁeld is 0 upon reset, otherwise at least one bit should be set:
Each virtio device oﬀers all the features it understands. During device initialization, the driver reads this and tells the device the subset that it accepts. The only way to renegotiate is to reset the device.
This allows for forwards and backwards compatibility: if the device is enhanced with a new feature bit, older drivers will not write that feature bit back to the device and it SHOULD go into backwards compatibility mode. Similarly, if a driver is enhanced with a feature that the device doesn’t support, it see the new feature is not oﬀered and SHOULD go into backwards compatibility mode (or, for poor implementations it MAY set the FAILED Device Status bit).
The driver MUST NOT accept a feature which the device did not oﬀer, and MUST NOT accept a feature which requires another feature which was not accepted.
The device MUST NOT oﬀer a feature which requires another feature which was not oﬀered.
Feature bits are allocated as follows:
For example, feature bit 0 for a network device (i.e. Subsystem Device ID 1) indicates that the device supports checksumming of packets.
In particular, new ﬁelds in the device conﬁguration space are indicated by oﬀering a feature bit, so the driver MUST check that the feature is oﬀered before accessing that part of the conﬁguration space.
Earlier drafts of this speciﬁcation (up to 0.9.X) deﬁned a similar, but diﬀerent interface between the hypervisor and the guest. Since these are widely deployed, this speciﬁcation accommodates optional features to simplify transition from these earlier draft interfaces. Speciﬁcally:
Legacy devices and legacy drivers are not compliant with this speciﬁcation.
To simplify transition from these earlier draft interfaces, it is possible to implement:
Transitional devices and transitional drivers can be compliant with this speciﬁcation (ie. when not operating in legacy mode).
Devices or drivers with no legacy compatibility are referred to as non-transitional devices and drivers, respectively.
Transitional Drivers can detect Legacy Devices by detecting that the feature bit VIRTIO_F_VERSION_1 is not oﬀered. Transitional devices can detect Legacy drivers by detecting that VIRTIO_F_VERSION_1 has not been acknowledged by the driver. In this case device is used through the legacy interface.
To make them easier to locate, speciﬁcation sections documenting these transitional features are explicitly marked with ’Legacy Interface’ in the section title.
Conﬁguration space is generally used for rarely-changing or initialization-time parameters. Drivers MUST NOT assume reads from ﬁelds greater than 32 bits wide are atomic, nor or reads from multiple ﬁelds.
Each transport provides a generation count for the conﬁguration space, which must change whenever there is a possibility that two accesses to the conﬁguration space can see diﬀerent versions of that space.
Thus drivers SHOULD read conﬁguration space ﬁelds like so:
Note that conﬁguration space uses the little-endian format for multi-byte ﬁelds.
Note that future versions of this speciﬁcation will likely extend the conﬁguration space for devices by adding extra ﬁelds at the tail end of some structures in conﬁguration space.
To allow forward compatibility with such extensions, drivers MUST NOT limit structure size and conﬁguration space size. Instead, drivers SHOULD only check that conﬁguration space is *large enough* to contain the ﬁelds required for device operation.
For example, if the speciﬁcation states that conﬁguration space ’includes a single 8-bit ﬁeld’ drivers should understand this to mean that the conﬁguration space might also include an arbitrary amount of tail padding, and accept any conﬁguration space size equal to or greater than the speciﬁed 8-bit size.
Note that for legacy interfaces, conﬁguration space is generally the guest’s native endian, rather than PCI’s little-endian.
Legacy devices did not have a conﬁguration generation ﬁeld, thus are susceptible to race conditions if conﬁguration is updated. This eﬀects the block capacity and network mac ﬁelds; best practice is to read these ﬁelds multiple times until two reads generate a consistent result.
The mechanism for bulk data transport on virtio devices is pretentiously called a virtqueue. Each device can have zero or more virtqueues: for example, the simplest network device has one for transmit and one for receive. Each queue has a 16-bit queue size parameter, which sets the number of entries and implies the total size of the queue.
Each virtqueue consists of three parts:
where each part is physically-contiguous in guest memory, and has diﬀerent alignment requirements.
The memory aligment and size requirements, in bytes, of each part of the virtqueue are summarized in the following table:
|Descriptor Table||16||16∗(Queue Size)|
|Available Ring||2||6 + 2∗(Queue Size)|
|Used Ring||4||6 + 4∗(Queue Size)|
The Alignment column gives the miminum alignment: for each part of the virtqueue, the physical address of the ﬁrst byte MUST be a multiple of the speciﬁed alignment value.
The Size column gives the total number of bytes required for each part of the virtqueue.
Queue Size corresponds to the maximum number of buﬀers in the virtqueue. For example, if Queue Size is 4 then at most 4 buﬀers can be queued at any given time. Queue Size value is always a power of 2. The maximum Queue Size value is 32768. This value is speciﬁed in a bus-speciﬁc way.
When the driver wants to send a buﬀer to the device, it ﬁlls in a slot in the descriptor table (or chains several together), and writes the descriptor index into the available ring. It then notiﬁes the device. When the device has ﬁnished a buﬀer, it writes the descriptor index into the used ring, and sends an interrupt.
For Legacy Interfaces, several additional restrictions are placed on the virtqueue layout:
Each virtqueue occupies two or more physically-contiguous pages (usually deﬁned as 4096 bytes, but depending on the transport) and consists of three parts:
|Descriptor Table||Available Ring (…padding…)||Used Ring|
The bus-speciﬁc Queue Size ﬁeld controls the total number of bytes required for the virtqueue according to the following formula:
This wastes some space with padding. The legacy virtqueue layout structure therefore looks like this:
Note that the endian of ﬁelds and in the virtqueue is the native endian of the guest, not little-endian as speciﬁed by this standard. It is assumed that the host is already aware of the guest endian.
The device MUST NOT make assumptions about the particular arrangement of descriptors: the message framing is independent of the contents of the buﬀers. For example, a network transmit buﬀer consists of a 12 byte header followed by the network packet. This could be most simply placed in the descriptor table as a 12 byte output descriptor followed by a 1514 byte output descriptor, but it could also consist of a single 1526 byte output descriptor in the case where the header and packet are adjacent, or even three or more descriptors (possibly with loss of eﬃciency in that case).
Note that, some implementations may have large-but-reasonable restrictions on total descriptor size (such as based on IOV_MAX in the host OS). This has not been a problem in practice: little sympathy will be given to drivers which create unreasonably-sized descriptors such as by dividing a network packet into 1500 single-byte descriptors!
Regrettably, initial driver implementations used simple layouts, and devices came to rely on it, despite this speciﬁcation wording. In addition, the speciﬁcation for virtio_blk SCSI commands required intuiting ﬁeld lengths from frame boundaries (see 18.104.22.168 Legacy Interface: Device Operation)
It is thus recommended that when using legacy interfaces, transitional drivers be conservative in their assumptions, unless the VIRTIO_F_ANY_LAYOUT feature is accepted.
The descriptor table refers to the buﬀers the driver is using for the device. The addresses are physical addresses, and the buﬀers can be chained via the next ﬁeld. Each descriptor describes a buﬀer which is read-only or write-only, but a chain of descriptors can contain both read-only and write-only buﬀers.
The actual contents of the memory oﬀered to the device depends on the device type. Most common is to begin the data with a header (containing little-endian ﬁelds) for the device to read, and postﬁx it with a status tailer for the device to write.
Drivers MUST NOT add a descriptor chain over than 232 bytes long in total; this implies that loops in the descriptor chain are forbidden!
The number of descriptors in the table is deﬁned by the queue size for this virtqueue.
Some devices beneﬁt by concurrently dispatching a large number of large requests. The VIRTIO_RING_F_INDIRECT_DESC feature allows this (see 7 virtio_ring.h). To increase ring capacity the driver can store a table of indirect descriptors anywhere in memory, and insert a descriptor in main virtqueue (with ﬂags&VRING_DESC_F_INDIRECT on) that refers to memory buﬀer containing this indirect descriptor table; ﬁelds addr and len refer to the indirect table address and length in bytes, respectively.
The driver MUST NOT set the VRING_DESC_F_INDIRECT ﬂag unless the VIRTIO_RING_F_INDIRECT_DESC feature was negotiated.
The indirect table layout structure looks like this (len is the length of the descriptor that refers to this table, which is a variable, so this code won’t compile):
The ﬁrst indirect descriptor is located at start of the indirect descriptor table (index 0), additional indirect descriptors are chained by next ﬁeld. An indirect descriptor without next ﬁeld (with ﬂags&VRING_DESC_F_NEXT oﬀ) signals the end of the descriptor. An indirect descriptor can not refer to another indirect descriptor table (ﬂags&VRING_DESC_F_INDIRECT MUST be oﬀ). A single indirect descriptor table can include both read-only and write-only descriptors; the device MUST ignore the write-only ﬂag (ﬂags&VRING_DESC_F_WRITE) in the descriptor that refers to it.
The available ring refers to what descriptor chains the driver is oﬀering the device: each ring entry refers to the head of a descriptor chain. It is only written by the driver and read by the device.
The “idx” ﬁeld indicates where we would put the next descriptor entry in the ring (modulo the queue size). This starts at 0, and increases.
If the VIRTIO_RING_F_INDIRECT_DESC feature bit is not negotiated, the “ﬂags” ﬁeld oﬀers a crude interrupt control mechanism. The driver MUST set this to 0 or 1: 1 indicates that the device SHOULD NOT send an interrupt when it consumes a descriptor chain from the available ring. The device MUST ignore the used_event value in this case.
Otherwise, if the VIRTIO_RING_F_EVENT_IDX feature bit is negotiated, the driver MUST set the "ﬂags" ﬁeld to 0, and use the “used_event” ﬁeld in the used ring instead. The driver can ask the device to delay interrupts until an entry with an index speciﬁed by the “used_event” ﬁeld is written in the used ring (equivalently, until the idx ﬁeld in the used ring will reach the value used_event + 1).
The driver MUST handle spurious interrupts: either form of interrupt suppression is merely an optimization; it may not suppress interrupts entirely.
The used ring is where the device returns buﬀers once it is done with them: it is only written to by the device, and read by the driver.
Each entry in the ring is a pair: the head entry of the descriptor chain describing the buﬀer (this matches an entry placed in the available ring by the guest earlier), and the total of bytes written into the buﬀer. The latter is extremely useful for drivers using untrusted buﬀers: if you do not know exactly how much has been written by the device, you usually have to zero the buﬀer to ensure no data leakage occurs.
If the VIRTIO_RING_F_INDIRECT_DESC feature bit is not negotiated, the “ﬂags” ﬁeld oﬀers a crude interrupt control mechanism. The driver MUST initialize this to 0, the device MUST set this to 0 or 1: 1 indicates that the driver SHOULD NOT send an notiﬁcation when it adds a descriptor chain to the available ring. The driver MUST ignore the used_event value in this case.
Otherwise, if the VIRTIO_RING_F_EVENT_IDX feature bit is negotiated, the device MUST leave the "ﬂags" ﬁeld at 0, and use the “avail_event” ﬁeld in the used ring instead. The device can ask the driver to delay notiﬁcations until an entry with an index speciﬁed by the “avail_event” ﬁeld is written in the available ring (equivalently, until the idx ﬁeld in the used ring will reach the value avail_event + 1).
The device MUST handle spurious notiﬁcation: either form of notiﬁcation suppression is merely an optimization; it may not suppress them entirely.
The Linux Kernel Source code contains the deﬁnitions above and helper routines in a more usable form, in include/linux/virtio_ring.h. This was explicitly licensed by IBM and Red Hat under the (3-clause) BSD license so that it can be freely used by all other projects, and is reproduced (with slight variation to remove Linux assumptions) in 7 virtio_ring.h.
The driver MUST follow this sequence to initialize a device:
If any of these steps go irrecoverably wrong, the driver SHOULD set the FAILED status bit to indicate that it has given up on the device (it can reset the device later to restart if desired). The driver MUST not continue initialization in that case.
The device MUST NOT consume buﬀers before DRIVER_OK, and the driver MUST NOT notify the device before it sets DRIVER_OK.
Devices SHOULD support all valid combinations of features, but we know that implementations may well make assuptions that they will only be used by fully-optimized drivers. The resetting of the FEATURES_OK ﬂag provides a semi-graceful failure mode for this case.
Legacy devices do not support the FEATURES_OK status bit, and thus did not have a graceful way for the device to indicate unsupported feature combinations. It also did not provide a clear mechanism to end feature negotiation, which meant that devices ﬁnalized features on ﬁrst-use, and no features could be introduced which radically changed the initial operation of the device.
Legacy device implementations often used the device before setting the DRIVER_OK bit.
The result was the steps 5 and 6 were omitted, and steps 7 and 8 were conﬂated.
There are two parts to device operation: supplying new buﬀers to the device, and processing used buﬀers from the device. As an example, the simplest virtio network device has two virtqueues: the transmit virtqueue and the receive virtqueue. The driver adds outgoing (read-only) packets to the transmit virtqueue, and then frees them after they are used. Similarly, incoming (write-only) buﬀers are added to the receive virtqueue, and processed after they are used.
The driver oﬀers buﬀers to one of the device’s virtqueues as follows:
Note that the above code does not take precautions against the available ring buﬀer wrapping around: this is not possible since the ring buﬀer is the same size as the descriptor table, so step (1) will prevent such a condition.
In addition, the maximum queue size is 32768 (it must be a power of 2 which ﬁts in 16 bits), so the 16-bit “idx” value can always distinguish between a full and empty buﬀer.
Here is a description of each stage in more detail.
A buﬀer consists of zero or more read-only physically-contiguous elements followed by zero or more physically-contiguous write-only elements (it must have at least one element). This algorithm maps it into the descriptor table to form a descriptor chain:
for each buﬀer element, b:
In practice, the d.next ﬁelds are usually used to chain free descriptors, and a separate count kept to check there are enough free descriptors before beginning the mappings.
The head of the buﬀer we mapped is the ﬁrst d in the algorithm above (the descriptor chain head). A naive implementation would do the following (with the appropriate conversion to-and-from little-endian assumed):
However, in general we can add many descriptor chains before we update the “idx” ﬁeld (at which point they become visible to the device), so we keep a counter of how many we’ve added:
Once the index ﬁeld of the virtqueue is updated, the device will be able to access the descriptor chains we’ve created and the memory they refer to. This is why a memory barrier is generally used before the index update, to ensure it sees the most up-to-date copy.
The index ﬁeld always increments, and we let it wrap naturally at 65536:
The actual method of device notiﬁcation is bus-speciﬁc, but generally it can be expensive. So the device can suppress such notiﬁcations if it doesn’t need them. The driver has to be careful to expose the new index value before checking if notiﬁcations are suppressed: it’s OK to notify gratuitously, but not to omit a required notiﬁcation. So again, we use a memory barrier here before reading the ﬂags or the avail_event ﬁeld.
If the VIRTIO_F_RING_EVENT_IDX feature is not negotiated, and if the VRING_USED_F_NOTIFY ﬂag is not set, we go ahead and notify the device.
If the VIRTIO_F_RING_EVENT_IDX feature is negotiated, we read the avail_event ﬁeld in the available ring structure. If the available index crossed_the avail_event ﬁeld value since the last notiﬁcation, we go ahead and write to the PCI conﬁguration space. The avail_event ﬁeld wraps naturally at 65536 as well, iving the following algorithm for calculating whether a device needs notiﬁcation:
Once the device has used a buﬀer (read from or written to it, or parts of both, depending on the nature of the virtqueue and the device), it sends an interrupt, following an algorithm very similar to the algorithm used for the driver to send the device a buﬀer:
For each ring, the driver should then disable interrupts by writing VRING_AVAIL_F_NO_INTERRUPT ﬂag in avail structure, if required. It can then process used ring entries ﬁnally enabling interrupts by clearing the VRING_AVAIL_F_NO_INTERRUPT ﬂag or updating the EVENT_IDX ﬁeld in the available structure. The driver should then execute a memory barrier, and then recheck the ring empty condition. This is necessary to handle the case where after the last check and before enabling interrupts, an interrupt has been suppressed by the device:
For devices where the conﬁguration information can be changed, an interrupt is delivered when a conﬁguration change occurs.
Virtio devices are commonly implemented as PCI devices.
Any PCI device with Vendor ID 0x1AF4, and Device ID 0x1000 through 0x103F inclusive is a virtio device3 .
The Subsystem Device ID indicates which virtio device is supported by the device. The Subsystem Vendor ID SHOULD reﬂect the PCI Vendor ID of the environment (it’s currently only used for informational purposes by the driver).
All drivers MUST match devices with any Revision ID, this is to allow devices to be versioned without breaking drivers.
Transitional devices must have a Revision ID of 0 to match legacy drivers.
Non-transitional devices must have a Revision ID of 1 or higher.
Both transitional and non-transitional drivers must match any Revision ID value.
To conﬁgure the device, use I/O and/or memory regions and/or PCI conﬁguration space of the PCI device. These contain the virtio header registers, the notiﬁcation register, the ISR status register and device speciﬁc registers, as speciﬁed by Virtio Structure PCI Capabilities.
There may be diﬀerent widths of accesses to the I/O region; the “natural” access method for each ﬁeld must be used (i.e. 32-bit accesses for 32-bit ﬁelds, etc).
PCI Device Conﬁguration Layout includes the common conﬁguration, ISR, notiﬁcation and device speciﬁc conﬁguration structures.
All multi-byte ﬁelds are little-endian.
Common conﬁguration structure layout is documented below:
The driver MUST conﬁgure the other virtqueue ﬁelds before enabling the virtqueue.
ISR status structure includes a single 8-bit ISR status ﬁeld.
Notiﬁcation structure is always a multiple of 2 bytes in size. It includes 2-byte Queue Notify ﬁelds for each virtqueue of the device. Note that multiple virtqueues can use the same Queue Notify ﬁeld, if necessary: see notify_oﬀ_multiplier below.
Device speciﬁc structure is optional.
Transitional devices should present part of conﬁguration registers in a legacy conﬁguration structure in BAR0 in the ﬁrst I/O region of the PCI device, as documented below.
There may be diﬀerent widths of accesses to the I/O region; the “natural” access method for each ﬁeld in the virtio header must be used (i.e. 32-bit accesses for 32-bit ﬁelds, etc), but when accessed through the legacy interface the device-speciﬁc region can be accessed using any width accesses, and should obtain the same results.
Note that this is possible because while the virtio header is PCI (i.e. little) endian, when using the legacy interface the device-speciﬁc region is encoded in the native endian of the guest (where such distinction is applicable).
When used through the legacy interface, the virtio header looks as follows:
|Read / Write||R||R+W||R+W||R||R+W||R+W||R+W||R|
|Purpose||Device Features bits 0:31||Driver Features bits 0:31||Queue Size||Queue Select||Queue Notify||Queue Address||Device Status|| ISR
If MSI-X is enabled for the device, two additional ﬁelds immediately follow this header:
|Purpose (MSI-X)||Conﬁguration Vector||Queue Vector|
Note: When MSI-X capability is enabled, device speciﬁc conﬁguration starts at byte oﬀset 24 in virtio header structure. When MSI-X capability is not enabled, device speciﬁc conﬁguration starts at byte oﬀset 20 in virtio header. ie. once you enable MSI-X on the device, the other ﬁelds move. If you turn it oﬀ again, they move back!
Immediately following these general headers, there may be device-speciﬁc headers:
|Bits||Device Speciﬁc|| |
|Read / Write||Device Speciﬁc|
Note that only Feature Bits 0 to 31 are accessible through the Legacy Interface. When used through the Legacy Interface, Transitional Devices must assume that Feature Bits 32 to 63 are not acknowledged by Driver.
As legacy devices had no conﬁguration generation ﬁeld, see 2.3.2 Legacy Interface: Conﬁguration Space for workarounds.
This documents PCI-speciﬁc steps executed during Device Initialization. As the ﬁrst step, driver must detect device conﬁguration layout to locate conﬁguration ﬁelds in memory, I/O or conﬁguration space of the device.
Virtio Device Conﬁguration Layout includes virtio conﬁguration header, Notiﬁcation and ISR Status and device conﬁguration structures. Each structure can be mapped by a Base Address register (BAR) belonging to the function, located beginning at 10h in Conﬁguration Space, or accessed though PCI conﬁguration space.
Actual location of each structure is speciﬁed using vendor-speciﬁc PCI capability located on capability list in PCI conﬁguration space of the device. This virtio structure capability uses little-endian format; all bits are read-only:
This structure can optionally followed by extra data, depending on other ﬁelds, as documented below.
Note that future versions of this speciﬁcation will likely extend devices by adding extra ﬁelds at the tail end of some structures.
To allow forward compatibility with such extensions, drivers must not limit structure size. Instead, drivers should only check that structures are *large enough* to contain the ﬁelds required for device operation.
For example, if the speciﬁcation states ’structure includes a single 8-bit ﬁeld’ drivers should understand this to mean that the structure can also include an arbitrary amount of tail padding, and accept any structure size equal to or greater than the speciﬁed 8-bit size.
The ﬁelds are interpreted as follows:
Any other value - reserved for future use. Drivers MUST ignore any vendor-speciﬁc capability structure which has a reserved cfg_type value.
More than one capability can identify the same structure - this makes it possible for the device to expose multiple interfaces to drivers. The order of the capabilities in the capability list speciﬁes the order of preference suggested by the device; drivers SHOULD use the ﬁrst interface that they can support. For example, on some hypervisors, notiﬁcations using IO accesses are faster than memory accesses. In this case, hypervisor can expose two capabilities with cfg_type set to VIRTIO_PCI_CAP_NOTIFY_CFG: the ﬁrst one addressing an I/O BAR, the second one addressing a memory BAR. Driver will use the I/O BAR if I/O resources are available, and fall back on memory BAR when I/O resources are unavailable.
Any other value is reserved for future use. Drivers MUST ignore any vendor-speciﬁc capability structure which has a reserved bar value.
If cfg_type is VIRTIO_PCI_CAP_NOTIFY_CFG this structure is immediately followed by additional ﬁelds:
Virtqueue oﬀset multiplier, in bytes. Must be even and either a power of two, or 0. Value 0x1 is reserved. For a given virtqueue, the address to use for notiﬁcations is calculated as follows:
queue_notify_oﬀ * notify_oﬀ_multiplier + oﬀset
If notify_oﬀ_multiplier is 0, all virtqueues use the same address in the Notiﬁcations structure!
If cfg_type is VIRTIO_PCI_CAP_PCI_CFG the ﬁelds bar, oﬀset and length are RW and this structure is immediately followed by an additional ﬁeld:
This RW ﬁeld allows an indirect access to any BAR on the device using PCI conﬁguration accesses.
The BAR to access is selected using the bar ﬁeld. The length of the access is speciﬁed by the length ﬁeld, which can be set to 1, 2 and 4. The oﬀset within the BAR is speciﬁed by the oﬀset ﬁeld, which must be aligned to length bytes.
After this ﬁeld is written by driver, the ﬁrst length bytes in pci_cfg_data are written at the selected oﬀset in the selected BAR.
When this ﬁeld is read by driver, length bytes at the selected oﬀset in the selected BAR are read into pci_cfg_data.
Legacy devices did not have the Virtio PCI Capability in their capability list.
Transitional devices should expose the Legacy Interface in I/O space in BAR0.
Transitional drivers should look for the Virtio PCI Capabilities on the capability list. If these are not present, driver should assume a legacy device.
Non-transitional drivers should look for the Virtio PCI Capabilities on the capability list. If these are not present, driver should assume a legacy device, and fail gracefully.
Non-transitional devices, on a platform where a legacy driver for a legacy device with the same ID might have previously existed, must take the following steps to fail gracefully when a legacy driver attempts to drive them:
Writing a valid MSI-X Table entry number, 0 to 0x7FF, to one of Conﬁguration/Queue Vector registers, maps interrupts triggered by the conﬁguration change/selected queue events respectively to the corresponding MSI-X vector. To disable interrupts for a speciﬁc event type, unmap it by writing a special NO_VECTOR value:
Reading these registers returns vector mapped to a given event, or NO_VECTOR if unmapped. All queue and conﬁguration change events are unmapped by default.
Note that mapping an event to vector might require allocating internal device resources, and might fail. Devices MUST report such failures by returning the NO_VECTOR value when the relevant Vector ﬁeld is read. After mapping an event to vector, the driver MUST verify success by reading the Vector ﬁeld value: on success, the previously written value is returned, and on failure, NO_VECTOR is returned. If a mapping failure is detected, the driver can retry mapping with fewer vectors, or disable MSI-X.
The driver does this as follows, for each virtqueue a device has:
Device notiﬁcation occurs by writing the 16-bit virtqueue index of this virtqueue to the Queue Notify ﬁeld.
If an interrupt is necessary:
The driver interrupt handler should:
Some virtio PCI devices can change the device conﬁguration state, as reﬂected in the virtio header in the PCI conﬁguration space. In this case:
Virtual environments without PCI support (a common situation in embedded devices models) might use simple memory mapped device ("virtio-mmio") instead of the PCI device.
The memory mapped virtio device behaviour is based on the PCI device speciﬁcation. Therefore most of operations like device initialization, queues conﬁguration and buﬀer transfers are nearly identical. Existing diﬀerences are described in the following sections.
Unlike PCI, MMIO provides no generic device discovery. For systems using Flattened Device Trees the suggested format is:
MMIO virtio devices provides a set of memory mapped control registers, all 32 bits wide, followed by device-speciﬁc conﬁguration space. The following table presents their names, oﬀset from the base address, and whether they are read-only (R) or write-only (W) from the driver’s perspective:
– Bit 0 | Used Ring Update This interrupt is asserted when the device has updated the Used Ring in at least one of the active virtual queues.
– Bit 1 | Conﬁguration change This interrupt is asserted when conﬁguration of the device has changed.
All register values are organized as Little Endian.
Accessing memory locations not explicitly described above (or - in case of the conﬁguration space - described in the device speciﬁcation), writing to the registers described as "R" and reading from registers described as "W" is not permitted and can cause undeﬁned behavior.
The driver must start the device initialization by reading and checking values from the MagicValue and the Version registers. If both values are valid, it must read the DeviceID register and if its value is zero (0x0) must abort initialization and must not access any other register.
Further initialization must follow the procedure described in 3.1 Device Initialization.
The device is notiﬁed about new buﬀers available in a queue by writing the queue index to the QueueNum register.
The memory mapped virtio device is using single, dedicated interrupt signal, which is raised when at least one of the interrupts described in the InterruptStatus register description is asserted. After receiving an interrupt, the driver must read the InterruptStatus register to check what caused the interrupt (see the register description). After the interrupt is handled, the driver must acknowledge it by writing a bit mask corresponding to the serviced interrupt to the InterruptACK register.
As documented in the InterruptStatus register description, the device may notify the driver about a new used buﬀer being available in the queue or about a change in the device conﬁguration.
The legacy MMIO transport used page-based addressing, resulting in a slightly diﬀerent control register layout, the device initialization and the virtual queue conﬁguration procedure.
The following list presents control registers layout, omitting descriptions of registers which did not change their function nor behaviour:
* Oﬀset from the device base address | Direction | Name Description
* 0x000 | R | MagicValue
* 0x004 | R | Version Device version number. Legacy devices must return value 0x1.
* 0x008 | R | DeviceID
* 0x00c | R | VendorID
* 0x010 | R | HostFeatures
* 0x014 | W | HostFeaturesSel
* 0x020 | W | GuestFeatures
* 0x024 | W | GuestFeaturesSel
* 0x028 | W | GuestPageSize Guest page size. The driver must write the guest page size in bytes to the register during initialization, before any queues are used. This value must be a power of 2 and is used by the device to calculate the Guest address of the ﬁrst queue page (see QueuePFN).
* 0x030 | W | QueueSel Virtual queue index (ﬁrst queue is 0). Writing to this register selects the virtual queue that the following operations on the QueueNumMAx, QueueNum, QueueAlign and QueuePFN registers apply to.
* 0x034 | R | QueueNumMax Maximum virtual queue size. Reading from the register returns the maximum size of the queue the device is ready to process or zero (0x0) if the queue is not available. This applies to the queue selected by writing to the QueueSel and is allowed only when the QueuePFN is set to zero (0x0), so when the queue is not actively used.
* 0x038 | W | QueueNum Virtual queue size. Queue size is the number of elements in the queue, therefore size of the descriptor table and both available and used rings. Writing to this register notiﬁes the device what size of the queue the Driver will use. This applies to the queue selected by writing to the QueueSel register.
* 0x03c | W | QueueAlign Used Ring alignment in the virtual queue. Writing to this register notiﬁes the device about alignment boundary of the Used Ring in bytes. This value must be a power of 2 and applies to the queue selected by writing to the QueueSel register.
* 0x040 | RW | QueuePFN Guest physical page number of the virtual queue. Writing to this register notiﬁes the device about location of the virtual queue in the Guest’s physical address space. This value is the index number of a page starting with the queue Descriptor Table. Value zero (0x0) means physical address zero (0x00000000) and is illegal. When the Driver stops using the queue it must write zero (0x0) to this register. Reading from this register returns the currently used page number of the queue, therefore a value other than zero (0x0) means that the queue is in use. Both read and write accesses apply to the queue selected by writing to the QueueSel register.
* 0x050 | W | QueueNotify
* 0x060 | R | InterruptStatus
* 0x064 | W | InterruptACK
* 0x070 | RW | Status Device status. Reading from this register returns the current device status ﬂags. Writing non-zero values to this register sets the status ﬂags, indicating the OS/driver progress. Writing zero (0x0) to this register triggers a device reset. This should include setting QueuePFN to zero (0x0) for all queues in the device. Also see 3.1 Device Initialization.
* 0x100+ | RW | Conﬁg
The virtual queue page size is deﬁned by writing to the GuestPageSize register, as written by the guest. This must be done before the virtual queues are conﬁgured.
The virtual queue layout follows p. 2.4.1 Legacy Interfaces: A Note on Virtqueue Layout, with the alignment deﬁned in the QueueAlign register.
The virtual queue is conﬁgured as follows:
Notiﬁcation mechanisms did not change.
S/390 based virtual machines support neither PCI nor MMIO, so a diﬀerent transport is needed there.
virtio-ccw uses the standard channel I/O based mechanism used for the majority of devices on S/390. A virtual channel device with a special control unit type acts as proxy to the virtio device (similar to the way virtio-pci uses a PCI device) and conﬁguration and operation of the virtio device is accomplished (mostly) via channel commands. This means virtio devices are discoverable via standard operating system algorithms, and adding virtio support is mainly a question of supporting a new control unit type.
As the S/390 is a big endian machine, the data structures transmitted via channel commands are big-endian: this is made clear by use of the types be16, be32 and be64.
As a proxy device, virtio-ccw uses a channel-attached I/O control unit with a special control unit type (0x3832) and a control unit model corresponding to the attached virtio device’s subsystem device ID, accessed via a virtual I/O subchannel and a virtual channel path of type 0x32. This proxy device is discoverable via normal channel subsystem device discovery (usually a STORE SUBCHANNEL loop) and answers to the basic channel commands, most importantly SENSE ID.
For a virtio-ccw proxy device, SENSE ID will return the following information:
|1-2||control unit type||0x3832|
|3||control unit model||<virtio device id>|
|4-5||device type||zeroes (unset)|
|6||device model||zeroes (unset)|
|7-255||extended SenseId data||zeroes (unset)|
A driver for virtio-ccw devices MUST check for a control unit type of 0x3832 and MUST ignore the device type and model.
In addition to the basic channel commands, virtio-ccw deﬁnes a set of channel commands related to conﬁguration and operation of virtio:
The virtio-ccw device acts like a normal channel device, as speciﬁed in [S390 PoP] and [S390 Common I/O]. In particular:
virtio-ccw uses several channel commands to set up a device.
CCW_CMD_SET_VIRTIO_REV is issued by the driver to set the revision of the virtio-ccw transport it intends to drive the device with. It uses the following communication structure:
revision contains the desired revision id, length the length of the data portion and data revision-dependent additional desired options.
The following values are supported:
|0||0||<empty>||legacy interface; transitional devices only|
|2-n||reserved for later revisions|
Note that a change in the virtio standard does not neccessarily correspond to a change in the virtio-ccw revision.
A device must post a unit check with command reject for any revision it does not support. For any invalid combination of revision, length and data, it must post a unit check with command reject as well. A non-transitional device must reject revision id 0.
A driver should start with trying to set the highest revision it supports and continue with lower revisions if it gets a command reject.
A driver must not issue any other virtio-ccw speciﬁc channel commands prior to setting the revision.
A device must answer with command reject to any virtio-ccw speciﬁc channel command that is not contained in the revision selected by the driver.
After a revision has been successfully selected by the driver, it must not attempt to select a diﬀerent revision. A device must answer to any such attempt with a command reject.
A device must treat the revision as unset from the time the associated subchannel has been enabled until a revision has been successfully set by the driver. This implies that revisions are not persistent across disabling and enabling of the associated subchannel.
A legacy driver will not issue the CCW_CMD_SET_VIRTIO_REV prior to issueing other virtio-ccw speciﬁc channel commands. A non-transitional device therefore must answer any such attempts with a command reject. A transitional device must assume in this case that the driver is a legacy driver and continue as if the driver selected revision 0. This implies that the device must reject any command not valid for revision 0, including a subsequent CCW_CMD_SET_VIRTIO_REV.
CCW_CMD_READ_VQ_CONF is issued by the driver to obtain information about a queue. It uses the following structure for communicating:
The requested number of buﬀers for queue index is returned in max_num.
Afterwards, CCW_CMD_SET_VQ is issued by the driver to inform the device about the location used for its queue. The transmitted structure is
desc, avail and used contain the guest addresses for the descriptor table, available ring and used ring for queue index, respectively. The actual virtqueue size (number of allocated buﬀers) is transmitted in num. res0 is reserved and must be ignored by the device.
queue contains the guest address for queue index, num the number of buﬀers and align the alignment.
The virtqueue is physically contiguous, with padded added to make the used ring meet the align value:
|Descriptor Table||Available Ring (…padding…)||Used Ring|
The calculation for total size is as follows:
The driver can change the status of a device via the CCW_CMD_WRITE_STATUS command, which transmits an 8 bit status value.
Feature bits are arranged in an array of 32 bit values, making for a total of 8192 feature bits. Feature bits are in little-endian byte order.
The CCW commands dealing with features use the following communication block:
features are the 32 bits of features currently accessed, while index describes which of the feature bit values is to be accessed.
The guest may obtain the device’s device feature set via the CCW_CMD_READ_FEAT command. The device stores the features at index to features.
For communicating its supported features to the device, the driver may use the CCW_CMD_WRITE_FEAT command, denoting a features/index combination.
The device’s conﬁguration space is located in host memory. It is the same size as the standard PCI conﬁguration space.
To obtain information from the conﬁguration space, the driver may use CCW_CMD_READ_CONF, specifying the guest memory for the device to write to.
For changing conﬁguration information, the driver may use CCW_CMD_WRITE_CONF, specifying the guest memory for the device to read from.
In both cases, the complete conﬁguration space is transmitted. This allows the driver to compare the new conﬁguration space with the old version, and keep a generation count internally whenever it changes.
In order to set up the indicator bits for host->guest notiﬁcation, the driver uses diﬀerent channel commands depending on whether it wishes to use traditional I/O interrupts tied to a subchannel or adapter I/O interrupts for virtqueue notiﬁcations. For any given device, the two mechanisms are mutually exclusive.
For the conﬁguration change indicators, only a mechanism using traditional I/O interrupts is provided, regardless of whether traditional or adapter I/O interrupts are used for virtqueue notiﬁcations.
To communicate the location of the indicator bits for host->guest notiﬁcation, the driver uses the CCW_CMD_SET_IND command, pointing to a location containing the guest address of the indicators in a 64 bit value.
If the driver has already set up two-staged queue indicators via the CCW_CMD_SET_IND_ADAPTER command, the device MUST post a unit check with command reject to any subsequent CCW_CMD_SET_IND command.
To communicate the location of the indicator bits used in the conﬁguration change host->guest notiﬁcation, the driver issues the CCW_CMD_SET_CONF_IND command, pointing to a location containing the guest address of the indicators in a 64 bit value.
To communicate the location of the summary and queue indicator bits, the driver uses the CCW_CMD_SET_IND_ADAPTER command with the following payload:
summary_indicator contains the guest address of the 8 bit summary indicator. indicator contains the guest address of an area wherin the indicators for the devices are contained, starting at bit_nr, one bit per virtqueue of the device. Bit numbers start at the left. isc contains the I/O interruption subclass to be used for the adapter I/O interrupt. It may be diﬀerent from the isc used by the proxy virtio-ccw device’s subchannel.
If the driver has already set up classic queue indicators via the CCW_CMD_SET_IND command, the device MUST post a unit check with command reject to any subsequent CCW_CMD_SET_IND_ADAPTER command.
There are two modes of operation regarding host->guest notifcation, classic I/O interrupts and adapter I/O interrupts. The mode to be used is determined by the driver by using CCW_CMD_SET_IND respectively CCW_CMD_SET_IND_ADAPTER to set up queue indicators.
For conﬁguration changes, the driver will always use classic I/O interrupts.
For notifying the driver of virtqueue buﬀers, the device sets the corresponding bit in the guest-provided indicators. If an interrupt is not already pending for the subchannel, the device generates an unsolicited I/O interrupt.
If the device wants to notify the driver about conﬁguration changes, it sets bit 0 in the conﬁguration indicators and generates an unsolicited I/O interrupt, if needed. This also applies if adapter I/O interrupts are used for queue notiﬁcations.
For notifying the driver of virtqueue buﬀers, the device sets the bit in the guest-provided indicator area at the corresponding oﬀset. The guest-provided summary indicator is also set. An adapter I/O interrupt for the corresponding interruption subclass is generated. The device SHOULD only generate an adapter I/O interrupt if the summary indicator had not been set prior to notiﬁcation. The driver MUST clear the summary indicator after receiving an adapter I/O interrupt before it processes the queue indicators.
For notifying the device of virtqueue buﬀers, the driver unfortunately can’t use a channel command (the asynchronous characteristics of channel I/O interact badly with the host block I/O backend). Instead, it uses a diagnose 0x500 call with subcode 3 specifying the queue, as follows:
|GPR||Input Value||Output Value|
|2||Subchannel ID||Host Cookie|
Host cookie is an optional per-virtqueue 64 bit value that can be used by the hypervisor to speed up the notiﬁcation execution. For each notiﬁcation, the output value is returned in GPR2 and should be passed in GPR4 for the next notiﬁcation:
For the early printk mechanism, diagnose 0x500 with subcode 0 is used.
In order to reset a device, a driver may send the CCW_CMD_VDEV_RESET command.
The following device IDs are used to identify diﬀerent types of virtio devices. Some device IDs are reserved for devices which are not currently deﬁned in this standard.
Discovering what devices are available and their type is bus-dependent.
|Device ID||Virtio Device|
The virtio network device is a virtual ethernet card, and is the most complex of the devices supported so far by virtio. It has enhanced rapidly and demonstrates clearly how support for new features should be added to an existing device. Empty buﬀers are placed in one virtqueue for receiving packets, and outgoing packets are enqueued into another for transmission in that order. A third command queue is used to control advanced ﬁltering features.
Virtqueue 2 only exists if VIRTIO_NET_F_CTRL_VQ set.
This was supposed to indicate segmentation oﬄoad support, but upon further investigation it became clear that multiple bits were required.
Two conﬁguration ﬁelds are currently deﬁned. The mac address ﬁeld always exists (though is only valid if VIRTIO_NET_F_MAC is set), and the status ﬁeld only exists if VIRTIO_NET_F_STATUS is set. Two read-only bits are currently deﬁned for the status ﬁeld: VIRTIO_NET_S_LINK_UP and VIRTIO_NET_S_ANNOUNCE.
For legacy devices, the status ﬁeld in struct virtio_net_conﬁg is the native endian of the guest rather than (necessarily) little-endian.
Packets are transmitted by placing them in the transmitq, and buﬀers for incoming packets are placed in the receiveq. In each case, the packet itself is preceeded by a header:
The controlq is used to control device features such as ﬁltering.
For legacy devices, the ﬁelds in struct virtio_net_hdr are the native endian of the guest rather than (necessarily) little-endian.
Transmitting a single packet is simple, but varies depending on the diﬀerent features the driver negotiated.
For example, consider a partially checksummed TCP (IPv4) packet. It will have a 14 byte ethernet header and 20 byte IP header followed by the TCP header (with the TCP checksum ﬁeld 16 bytes into that header). csum_start will be 14+20 = 34 (the TCP checksum includes the header), and csum_oﬀset will be 16. The value in the TCP checksum ﬁeld should be initialized to the sum of the TCP pseudo header, so that replacing it by the ones’ complement checksum of the TCP header and body will give the correct result.
The normal behavior in this interrupt handler is to retrieve and new descriptors from the used ring and free the corresponding headers and packets.
It is generally a good idea to keep the receive virtqueue as fully populated as possible: if it runs out, network performance will suﬀer.
If the VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6 or VIRTIO_NET_F_GUEST_UFO features are used, the Driver will need to accept packets of up to 65550 bytes long (the maximum size of a TCP or UDP packet, plus the 14 byte ethernet header), otherwise 1514. bytes. So unless VIRTIO_NET_F_MRG_RXBUF is negotiated, every buﬀer in the receive queue needs to be at least this length.11
If VIRTIO_NET_F_MRG_RXBUF is negotiated, each buﬀer must be at least the size of the struct virtio_net_hdr.
Processing packet involves:
The driver uses the control virtqueue (if VIRTIO_NET_F_VTRL_VQ is negotiated) to send commands to manipulate various features of the device which would not easily map into the conﬁguration space.
All commands are of the following form:
The class, command and command-speciﬁc-data are set by the driver, and the device sets the ack byte. There is little it can do except issue a diagnostic if the ack byte is not VIRTIO_NET_OK.
Note that in general, these commands are best-eﬀort: unwanted packets may still arrive.
The class VIRTIO_NET_CTRL_RX has two commands: VIRTIO_NET_CTRL_RX_PROMISC turns promiscuous mode on and oﬀ, and VIRTIO_NET_CTRL_RX_ALLMULTI turns all-multicast receive on and oﬀ. The command-speciﬁc-data is one byte containing 0 (oﬀ) or 1 (on).
The device can ﬁlter incoming packets by any number of destination MAC addresses.12 This table is set using the class VIRTIO_NET_CTRL_MAC and the command VIRTIO_NET_CTRL_MAC_TABLE_SET. The command-speciﬁc-data is two variable length tables of 6-byte MAC addresses. The ﬁrst table contains unicast addresses, and the second contains multicast addresses.
Both the VIRTIO_NET_CTRL_VLAN_ADD and VIRTIO_NET_CTRL_VLAN_DEL command take a little-endian 16-bit VLAN id as the command-speciﬁc-data.
The Driver needs to check VIRTIO_NET_S_ANNOUNCE bit in status ﬁeld when it notices the changes of device conﬁguration. The command VIRTIO_NET_CTRL_ANNOUNCE_ACK is used to indicate that driver has recevied the notiﬁcation and device would clear the VIRTIO_NET_S_ANNOUNCE bit in the status ﬁled after it received this command.
Processing this notiﬁcation involves:
The class VIRTIO_NET_CTRL_GUEST_OFFLOADS has one command: VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET applies the new oﬄoads conﬁguration.
le64 value passed as command data is a bitmask, bits set deﬁne oﬄoads to be enabled, bits cleared - oﬄoads to be disabled.
There is a corresponding device feature for each oﬄoad. Upon feature negotiation corresponding oﬄoad gets enabled to preserve backward compartibility.
Corresponding feature must be negotiated at startup in order to allow dynamic change of speciﬁc oﬄoad state.
The virtio block device is a simple virtual block device (ie. disk). Read and write requests (and other exotic requests) are placed in the queue, and serviced (probably out of order) by the device except where noted.
VIRTIO_BLK_F_FLUSH was also called VIRTIO_BLK_F_WCE: Legacy drivers should only negotiate this feature if they are capable of sending VIRTIO_BLK_T_FLUSH commands.
The capacity of the device (expressed in 512-byte sectors) is always present. The availability of the others all depend on various feature bits as indicated above.
The reserved ﬁeld used to be called writeback. If the VIRTIO_BLK_F_CONFIG_WCE feature is oﬀered, the cache mode should be read from the writeback ﬁeld of the conﬁguration if available; the driver can also write to the ﬁeld in order to toggle the cache between writethrough (0) and writeback (1) mode. If the feature is not available, the driver can instead look at the result of negotiating VIRTIO_BLK_F_FLUSH: the cache will be in writeback mode after reset if and only if VIRTIO_BLK_F_FLUSH is negotiated.
Some older legacy devices did not operate in writethrough mode even after a driver announced lack of support for VIRTIO_BLK_F_FLUSH.
The driver queues requests to the virtqueue, and they are used by the device (not necessarily in order). Each request is of form:
The type of the request is either a read (VIRTIO_BLK_T_IN), a write (VIRTIO_BLK_T_OUT), or a ﬂush (VIRTIO_BLK_T_FLUSH or VIRTIO_BLK_T_FLUSH_OUT13 ).
The sector number indicates the oﬀset (multiplied by 512) where the read or write is to occur. This ﬁeld is unused and set to 0 for scsi packet commands and for ﬂush commands.
The ﬁnal status byte is written by the device: either VIRTIO_BLK_S_OK for success, VIRTIO_BLK_S_IOERR for device or driver error or VIRTIO_BLK_S_UNSUPP for a request unsupported by device:
Any writes completed before the submission of the ﬂush command should be committed to non-volatile storage by the device.
For legacy devices, the ﬁelds in struct virtio_blk_req are the native endian of the guest rather than (necessarily) little-endian.
The ’reserved’ ﬁeld was previously called ioprio. The ioprio ﬁeld is a hint about the relative priorities of requests to the device: higher numbers indicate more important requests.
If the device has VIRTIO_BLK_F_BARRIER feature the high bit (VIRTIO_BLK_T_BARRIER) indicates that this request acts as a barrier and that all preceeding requests must be complete before this one, and all following requests must not be started until this is complete. Note that a barrier does not ﬂush caches in the underlying backend device in host, and thus does not serve as data consistency guarantee. Driver must use FLUSH request to ﬂush the host cache.
If the device has VIRTIO_BLK_F_SCSI feature, it can also support scsi packet command requests, each of these requests is of form:
A request type can also be a scsi packet command (VIRTIO_BLK_T_SCSI_CMD or VIRTIO_BLK_T_SCSI_CMD_OUT). The two types are equivalent, the device does not distinguish between them:
The cmd ﬁeld is only present for scsi packet command requests, and indicates the command to perform. This ﬁeld must reside in a single, separate read-only buﬀer; command length can be derived from the length of this buﬀer.
Note that these ﬁrst three (four for scsi packet commands) ﬁelds are always read-only: the data ﬁeld is either read-only or write-only, depending on the request. The size of the read or write can be derived from the total size of the request buﬀers.
The sense ﬁeld is only present for scsi packet command requests, and indicates the buﬀer for scsi sense data.
The data_len ﬁeld is only present for scsi packet command requests, this ﬁeld is deprecated, and should be ignored by the driver. Historically, devices copied data length there.
The sense_len ﬁeld is only present for scsi packet command requests and indicates the number of bytes actually written to the sense buﬀer.
The residual ﬁeld is only present for scsi packet command requests and indicates the residual size, calculated as data length - number of bytes actually transferred.
Historically, devices assumed that the ﬁelds type, ioprio and sector reside in a single, separate read-only buﬀer; the ﬁelds errors, data_len, sense_len and residual reside in a single, separate write-only buﬀer; the sense ﬁeld in a separate write-only buﬀer of size 96 bytes, by itself; the ﬁelds errors, data_len, sense_len and residual in a single write-only buﬀer; and the status ﬁeld is a separate read-only buﬀer of size 1 byte, by itself.
The virtio console device is a simple device for data input and output. A device may have one or more ports. Each port has a pair of input and output virtqueues. Moreover, a device has a pair of control IO virtqueues. The control virtqueues are used to communicate information between the device and the driver about ports being opened and closed on either side of the connection, indication from the device about whether a particular port is a console port, adding new ports, port hot-plug/unplug, etc., and indication from the driver about whether a port or a device was successfully added, port open/close, etc.. For data IO, one or more empty buﬀers are placed in the receive queue for incoming data and outgoing characters are placed in the transmit queue.
Ports 2 onwards only exist if VIRTIO_CONSOLE_F_MULTIPORT is set.
VIRTIO_CONSOLE_F_SIZE (0) Conﬁguration cols and rows ﬁelds are valid.
VIRTIO_CONSOLE_F_MULTIPORT(1) Device has support for multiple ports; conﬁguration ﬁelds nr_ports and max_nr_ports are valid and control virtqueues will be used.
The size of the console is supplied in the conﬁguration space if the VIRTIO_CONSOLE_F_SIZE feature is set. Furthermore, if the VIRTIO_CONSOLE_F_MULTIPORT feature is set, the maximum number of ports supported by the device can be fetched.
For legacy devices, the ﬁelds in struct virtio_console_conﬁg are the native endian of the guest rather than (necessarily) little-endian.
For legacy devices, the ﬁelds in struct virtio_console_control are the native endian of the guest rather than (necessarily) little-endian.
The virtio entropy device supplies high-quality randomness for guest use.
None currently deﬁned
None currently deﬁned.
When the driver requires random bytes, it places the descriptor of one or more buﬀers in the queue. It will be completely ﬁlled by random data by the device.
The virtio memory balloon device is a primitive device for managing guest memory: the device asks for a certain amount of memory, and the driver supplies it (or withdraws it, if the device has more than it asks for). This allows the guest to adapt to changes in allowance of underlying physical memory. If the feature is negotiated, the device can also be used to communicate guest memory statistics to the host.
Virtqueue 2 only exists if VIRTIO_BALLON_F_STATS_VQ set.
Both ﬁelds of this conﬁguration are always available.
Note that these ﬁelds are always little endian, despite convention that legacy device ﬁelds are guest endian.
Device operation begins immediately.
The device is driven by the receipt of a conﬁguration change interrupt.
The stats virtqueue is atypical because communication is driven by the device (not the driver). The channel becomes active at driver initialization time when the driver adds an empty buﬀer and notiﬁes the device. A request for memory statistics proceeds as follows:
Each statistic consists of a 16 bit tag and a 64 bit value. All statistics are optional and the driver may choose which ones to supply. To guarantee backwards compatibility, unsupported statistics should be omitted.
The virtio SCSI host device groups together one or more virtual logical units (such as disks), and allows communicating to them using the SCSI protocol. An instance of the device represents a SCSI host to which many targets and LUNs are attached.
The virtio SCSI device services two kinds of requests:
The device is also able to send out notiﬁcations about added and removed logical units. Together, these capabilities provide a SCSI transport protocol that uses virtqueues as the transfer medium. In the transport protocol, the virtio driver acts as the initiator, while the virtio SCSI host provides one or more targets that receive and process the requests.
All ﬁelds of this conﬁguration are always available. sense_size and cdb_size are writable by the driver.
For legacy devices, the ﬁelds in struct virtio_scsi_conﬁg are the native endian of the guest rather than (necessarily) little-endian.
The initialization routine should ﬁrst of all discover the device’s virtqueues.
If the driver uses the eventq, it should then place at least a buﬀer in the eventq.
The driver can immediately issue requests (for example, INQUIRY or REPORT LUNS) or task management functions (for example, I_T RESET).
Device operation consists of operating request queues, the control queue and the event queue.
The driver queues requests to an arbitrary request queue, and they are used by the device on that same queue. It is the responsibility of the driver to ensure strict request ordering for commands placed on diﬀerent queues, because they will be consumed with no order constraints.
Requests have the following format:
The lun ﬁeld addresses a target and logical unit in the virtio-scsi device’s SCSI domain. The only supported format for the LUN ﬁeld is: ﬁrst byte set to 1, second byte set to target, third and fourth byte representing a single level LUN structure, followed by four zero bytes. With this representation, a virtio-scsi device can serve up to 256 targets and 16384 LUNs per target.
The id ﬁeld is the command identiﬁer (“tag”).
task_attr, prio and crn should be left to zero. task_attr deﬁnes the task attribute as in the table above, but all task attributes may be mapped to SIMPLE by the device; crn may also be provided by clients, but is generally expected to be 0. The maximum CRN value deﬁned by the protocol is 255, since CRN is stored in an 8-bit integer.
All of these ﬁelds are deﬁned in SAM. They are always read-only, as are the cdb and dataout ﬁeld. The cdb_size is taken from the conﬁguration space.
sense and subsequent ﬁelds are always write-only. The sense_len ﬁeld indicates the number of bytes actually written to the sense buﬀer. The residual ﬁeld indicates the residual size, calculated as “data_length - number_of_transferred_bytes”, for read or write operations. For bidirectional commands, the number_of_transferred_bytes includes both read and written bytes. A residual ﬁeld that is less than the size of datain means that the dataout ﬁeld was processed entirely. A residual ﬁeld that exceeds the size of datain means that the dataout ﬁeld was processed partially and the datain ﬁeld was not processed at all.
The status byte is written by the device to be the status code as deﬁned in SAM.
The response byte is written by the device to be one of the following:
The controlq is used for other SCSI transport operations. Requests have the following format:
The type identiﬁes the remaining ﬁelds.
The following commands are deﬁned:
Task management function
The type is VIRTIO_SCSI_T_TMF; the subtype ﬁeld deﬁnes. All ﬁelds except response are ﬁlled by the driver. The subtype ﬁeld must always be speciﬁed and identiﬁes the requested task management function.
Other ﬁelds may be irrelevant for the requested TMF; if so, they are ignored but they should still be present. The lun ﬁeld is in the same format speciﬁed for request queues; the single level LUN is ignored when the task management function addresses a whole I_T nexus. When relevant, the value of the id ﬁeld is matched against the id values passed on the requestq.
The outcome of the task management function is written by the device in the response ﬁeld. The command-speciﬁc response values map 1-to-1 with those deﬁned in SAM.
Asynchronous notiﬁcation query
By sending this command, the driver asks the device which events the given LUN can report, as described in paragraphs 6.6 and A.6 of the SCSI MMC speciﬁcation. The driver writes the events it is interested in into the event_requested; the device responds by writing the events that it supports into event_actual.
The type is VIRTIO_SCSI_T_AN_QUERY. The lun and event_requested ﬁelds are written by the driver. The event_actual and response ﬁelds are written by the device.
No command-speciﬁc values are deﬁned for the response byte.
Asynchronous notiﬁcation subscription
By sending this command, the driver asks the speciﬁed LUN to report events for its physical interface, again as described in the SCSI MMC speciﬁcation. The driver writes the events it is interested in into the event_requested; the device responds by writing the events that it supports into event_actual.
Event types are the same as for the asynchronous notiﬁcation query message.
The type is VIRTIO_SCSI_T_AN_SUBSCRIBE. The lun and event_requested ﬁelds are written by the driver. The event_actual and response ﬁelds are written by the device.
No command-speciﬁc values are deﬁned for the response byte.
The eventq is used by the device to report information on logical units that are attached to it. The driver should always leave a few buﬀers ready in the eventq. In general, the device will not queue events to cope with an empty eventq, and will end up dropping events if it ﬁnds no buﬀer ready. However, when reporting events for many LUNs (e.g. when a whole target disappears), the device can throttle events to avoid dropping them. For this reason, placing 10-15 buﬀers on the event queue should be enough.
Buﬀers are placed in the eventq and ﬁlled by the device when interesting events occur. The buﬀers should be strictly write-only (device-ﬁlled) and the size of the buﬀers should be at least the value given in the device’s conﬁguration information.
Buﬀers returned by the device on the eventq will be referred to as "events" in the rest of this section. Events have the following format:
If bit 31 is set in the event ﬁeld, the device failed to report an event due to missing buﬀers. In this case, the driver should poll the logical units for unit attention conditions, and/or do whatever form of bus scan is appropriate for the guest operating system.
The meaning of the reason ﬁeld depends on the contents of the event ﬁeld. The following events are deﬁned:
This event is ﬁred in the following cases:
By sending this event, the device signals that a logical unit on a target has been reset, including the case of a new device appearing or disappearing on the bus.The device ﬁlls in all ﬁelds. The event ﬁeld is set to VIRTIO_SCSI_T_TRANSPORT_RESET. The lun ﬁeld addresses a logical unit in the SCSI host.
The reason value is one of the three #deﬁne values appearing above:
The “removed” and “rescan” events, when sent for LUN 0, may apply to the entire target. After receiving them the driver should ask the initiator to rescan the target, in order to detect the case when an entire target has appeared or disappeared. These two events will never be reported unless the VIRTIO_SCSI_F_HOTPLUG feature was negotiated between the device and the driver.
Events will also be reported via sense codes (this obviously does not apply to newly appeared buses or targets, since the application has never discovered them):
The preferred way to detect transport reset is always to use events, because sense codes are only seen by the driver when it sends a SCSI command to the logical unit or target. However, in case events are dropped, the initiator will still be able to synchronize with the actual state of the controller if the driver asks the initiator to rescan of the SCSI bus. During the rescan, the initiator will be able to observe the above sense codes, and it will process them as if it the driver had received the equivalent event.
By sending this event, the device signals that an asynchronous event was ﬁred from a physical interface.
All ﬁelds are written by the device. The event ﬁeld is set to VIRTIO_SCSI_T_ASYNC_NOTIFY. The lun ﬁeld addresses a logical unit in the SCSI host. The reason ﬁeld is a subset of the events that the driver has subscribed to via the "Asynchronous notiﬁcation subscription" command.
When dropped events are reported, the driver should poll for asynchronous events manually using SCSI commands.
LUN parameter change
By sending this event, the device signals that the conﬁguration parameters (for example the capacity) of a logical unit have changed. The event ﬁeld is set to VIRTIO_SCSI_T_PARAM_CHANGE. The lun ﬁeld addresses a logical unit in the SCSI host.
The same event is also reported as a unit attention condition. The reason ﬁeld contains the additional sense code and additional sense code qualiﬁer, respectively in bits 0..7 and 8..15. For example, a change in capacity will be reported as asc 0x2a, ascq 0x09 (CAPACITY DATA HAS CHANGED).
For MMC devices (inquiry type 5) there would be some overlap between this event and the asynchronous notiﬁcation event. For simplicity, as of this version of the speciﬁcation the host must never report this event for MMC devices.
In addition, bit 30 is used by qemu’s implementation to check for experimental early versions of virtio which did not perform correct feature negotiation, and should not be used.
Legacy or transitional devices may oﬀer the following:
It is possible that a very simple device will operate entirely through its conﬁguration space, but most will need at least one virtqueue in which it will place requests. A device with both input and output (eg. console and network devices described here) need two queues: one which the driver ﬁlls with buﬀers to receive input, and one which the driver places buﬀers to transmit output.
Conﬁguration space should only be used for initialization-time parameters. It is a limited resource with no synchronization between writable ﬁelds, so for most uses it is better to use a virtqueue to update conﬁguration information (the network device does this for ﬁltering, otherwise the table in the conﬁg space could potentially be very large).
Devices must not assume that conﬁguration ﬁelds over 32 bits wide are atomically writable.
Device numbers can be reserved by the OASIS committee: email email@example.com to secure a unique one.
Meanwhile for experimental drivers, use 65535 and work backwards.
Using the optional MSI-X capability devices can speed up interrupt processing by removing the need to read ISR Status register by guest driver (which might be an expensive operation), reducing interrupt sharing between devices and queues within the device, and handling interrupts from multiple CPUs. However, some systems impose a limit (which might be as low as 256) on the total number of MSI-X vectors that can be allocated to all devices. Devices and/or drivers should take this into account, limiting the number of vectors used unless the device is expected to cause a high volume of interrupts. Devices can control the number of vectors used by limiting the MSI-X Table Size or not presenting MSI-X capability in PCI conﬁguration space. Drivers can control this by mapping events to as small number of vectors as possible, or disabling MSI-X capability altogether.
Any change to conﬁguration space, or new virtqueues, or behavioural changes, should be indicated by negotiation of a new feature bit. This establishes clarity18 and avoids future expansion problems.
Clusters of functionality which are always implemented together can use a single bit, but if one feature makes sense without the others they should not be gratuitously grouped together to conserve feature bits.
|[Rev number ]||[Rev Date ]||[Modiﬁed By ]||[Summary of Changes ]|