1 ==================================
2 VFIO - "Virtual Function I/O" [1]_
3 ==================================
5 Many modern systems now provide DMA and interrupt remapping facilities
6 to help ensure I/O devices behave within the boundaries they've been
7 allotted. This includes x86 hardware with AMD-Vi and Intel VT-d,
8 POWER systems with Partitionable Endpoints (PEs) and embedded PowerPC
9 systems such as Freescale PAMU. The VFIO driver is an IOMMU/device
10 agnostic framework for exposing direct device access to userspace, in
11 a secure, IOMMU protected environment. In other words, this allows
12 safe [2]_, non-privileged, userspace drivers.
14 Why do we want that? Virtual machines often make use of direct device
15 access ("device assignment") when configured for the highest possible
16 I/O performance. From a device and host perspective, this simply
17 turns the VM into a userspace driver, with the benefits of
18 significantly reduced latency, higher bandwidth, and direct use of
19 bare-metal device drivers [3]_.
21 Some applications, particularly in the high performance computing
22 field, also benefit from low-overhead, direct device access from
23 userspace. Examples include network adapters (often non-TCP/IP based)
24 and compute accelerators. Prior to VFIO, these drivers had to either
25 go through the full development cycle to become proper upstream
26 driver, be maintained out of tree, or make use of the UIO framework,
27 which has no notion of IOMMU protection, limited interrupt support,
28 and requires root privileges to access things like PCI configuration
31 The VFIO driver framework intends to unify these, replacing both the
32 KVM PCI specific device assignment code as well as provide a more
33 secure, more featureful userspace driver environment than UIO.
35 Groups, Devices, and IOMMUs
36 ---------------------------
38 Devices are the main target of any I/O driver. Devices typically
39 create a programming interface made up of I/O access, interrupts,
40 and DMA. Without going into the details of each of these, DMA is
41 by far the most critical aspect for maintaining a secure environment
42 as allowing a device read-write access to system memory imposes the
43 greatest risk to the overall system integrity.
45 To help mitigate this risk, many modern IOMMUs now incorporate
46 isolation properties into what was, in many cases, an interface only
47 meant for translation (ie. solving the addressing problems of devices
48 with limited address spaces). With this, devices can now be isolated
49 from each other and from arbitrary memory access, thus allowing
50 things like secure direct assignment of devices into virtual machines.
52 This isolation is not always at the granularity of a single device
53 though. Even when an IOMMU is capable of this, properties of devices,
54 interconnects, and IOMMU topologies can each reduce this isolation.
55 For instance, an individual device may be part of a larger multi-
56 function enclosure. While the IOMMU may be able to distinguish
57 between devices within the enclosure, the enclosure may not require
58 transactions between devices to reach the IOMMU. Examples of this
59 could be anything from a multi-function PCI device with backdoors
60 between functions to a non-PCI-ACS (Access Control Services) capable
61 bridge allowing redirection without reaching the IOMMU. Topology
62 can also play a factor in terms of hiding devices. A PCIe-to-PCI
63 bridge masks the devices behind it, making transaction appear as if
64 from the bridge itself. Obviously IOMMU design plays a major factor
67 Therefore, while for the most part an IOMMU may have device level
68 granularity, any system is susceptible to reduced granularity. The
69 IOMMU API therefore supports a notion of IOMMU groups. A group is
70 a set of devices which is isolatable from all other devices in the
71 system. Groups are therefore the unit of ownership used by VFIO.
73 While the group is the minimum granularity that must be used to
74 ensure secure user access, it's not necessarily the preferred
75 granularity. In IOMMUs which make use of page tables, it may be
76 possible to share a set of page tables between different groups,
77 reducing the overhead both to the platform (reduced TLB thrashing,
78 reduced duplicate page tables), and to the user (programming only
79 a single set of translations). For this reason, VFIO makes use of
80 a container class, which may hold one or more groups. A container
81 is created by simply opening the /dev/vfio/vfio character device.
83 On its own, the container provides little functionality, with all
84 but a couple version and extension query interfaces locked away.
85 The user needs to add a group into the container for the next level
86 of functionality. To do this, the user first needs to identify the
87 group associated with the desired device. This can be done using
88 the sysfs links described in the example below. By unbinding the
89 device from the host driver and binding it to a VFIO driver, a new
90 VFIO group will appear for the group as /dev/vfio/$GROUP, where
91 $GROUP is the IOMMU group number of which the device is a member.
92 If the IOMMU group contains multiple devices, each will need to
93 be bound to a VFIO driver before operations on the VFIO group
94 are allowed (it's also sufficient to only unbind the device from
95 host drivers if a VFIO driver is unavailable; this will make the
96 group available, but not that particular device). TBD - interface
97 for disabling driver probing/locking a device.
99 Once the group is ready, it may be added to the container by opening
100 the VFIO group character device (/dev/vfio/$GROUP) and using the
101 VFIO_GROUP_SET_CONTAINER ioctl, passing the file descriptor of the
102 previously opened container file. If desired and if the IOMMU driver
103 supports sharing the IOMMU context between groups, multiple groups may
104 be set to the same container. If a group fails to set to a container
105 with existing groups, a new empty container will need to be used
108 With a group (or groups) attached to a container, the remaining
109 ioctls become available, enabling access to the VFIO IOMMU interfaces.
110 Additionally, it now becomes possible to get file descriptors for each
111 device within a group using an ioctl on the VFIO group file descriptor.
113 The VFIO device API includes ioctls for describing the device, the I/O
114 regions and their read/write/mmap offsets on the device descriptor, as
115 well as mechanisms for describing and registering interrupt
121 Assume user wants to access PCI device 0000:06:0d.0::
123 $ readlink /sys/bus/pci/devices/0000:06:0d.0/iommu_group
124 ../../../../kernel/iommu_groups/26
126 This device is therefore in IOMMU group 26. This device is on the
127 pci bus, therefore the user will make use of vfio-pci to manage the
132 Binding this device to the vfio-pci driver creates the VFIO group
133 character devices for this group::
135 $ lspci -n -s 0000:06:0d.0
136 06:0d.0 0401: 1102:0002 (rev 08)
137 # echo 0000:06:0d.0 > /sys/bus/pci/devices/0000:06:0d.0/driver/unbind
138 # echo 1102 0002 > /sys/bus/pci/drivers/vfio-pci/new_id
140 Now we need to look at what other devices are in the group to free
143 $ ls -l /sys/bus/pci/devices/0000:06:0d.0/iommu_group/devices
145 lrwxrwxrwx. 1 root root 0 Apr 23 16:13 0000:00:1e.0 ->
146 ../../../../devices/pci0000:00/0000:00:1e.0
147 lrwxrwxrwx. 1 root root 0 Apr 23 16:13 0000:06:0d.0 ->
148 ../../../../devices/pci0000:00/0000:00:1e.0/0000:06:0d.0
149 lrwxrwxrwx. 1 root root 0 Apr 23 16:13 0000:06:0d.1 ->
150 ../../../../devices/pci0000:00/0000:00:1e.0/0000:06:0d.1
152 This device is behind a PCIe-to-PCI bridge [4]_, therefore we also
153 need to add device 0000:06:0d.1 to the group following the same
154 procedure as above. Device 0000:00:1e.0 is a bridge that does
155 not currently have a host driver, therefore it's not required to
156 bind this device to the vfio-pci driver (vfio-pci does not currently
157 support PCI bridges).
159 The final step is to provide the user with access to the group if
160 unprivileged operation is desired (note that /dev/vfio/vfio provides
161 no capabilities on its own and is therefore expected to be set to
162 mode 0666 by the system)::
164 # chown user:user /dev/vfio/26
166 The user now has full access to all the devices and the iommu for this
167 group and can access them as follows::
169 int container, group, device, i;
170 struct vfio_group_status group_status =
171 { .argsz = sizeof(group_status) };
172 struct vfio_iommu_type1_info iommu_info = { .argsz = sizeof(iommu_info) };
173 struct vfio_iommu_type1_dma_map dma_map = { .argsz = sizeof(dma_map) };
174 struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
176 /* Create a new container */
177 container = open("/dev/vfio/vfio", O_RDWR);
179 if (ioctl(container, VFIO_GET_API_VERSION) != VFIO_API_VERSION)
180 /* Unknown API version */
182 if (!ioctl(container, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU))
183 /* Doesn't support the IOMMU driver we want. */
186 group = open("/dev/vfio/26", O_RDWR);
188 /* Test the group is viable and available */
189 ioctl(group, VFIO_GROUP_GET_STATUS, &group_status);
191 if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE))
192 /* Group is not viable (ie, not all devices bound for vfio) */
194 /* Add the group to the container */
195 ioctl(group, VFIO_GROUP_SET_CONTAINER, &container);
197 /* Enable the IOMMU model we want */
198 ioctl(container, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU);
200 /* Get addition IOMMU info */
201 ioctl(container, VFIO_IOMMU_GET_INFO, &iommu_info);
203 /* Allocate some space and setup a DMA mapping */
204 dma_map.vaddr = mmap(0, 1024 * 1024, PROT_READ | PROT_WRITE,
205 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
206 dma_map.size = 1024 * 1024;
207 dma_map.iova = 0; /* 1MB starting at 0x0 from device view */
208 dma_map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
210 ioctl(container, VFIO_IOMMU_MAP_DMA, &dma_map);
212 /* Get a file descriptor for the device */
213 device = ioctl(group, VFIO_GROUP_GET_DEVICE_FD, "0000:06:0d.0");
215 /* Test and setup the device */
216 ioctl(device, VFIO_DEVICE_GET_INFO, &device_info);
218 for (i = 0; i < device_info.num_regions; i++) {
219 struct vfio_region_info reg = { .argsz = sizeof(reg) };
223 ioctl(device, VFIO_DEVICE_GET_REGION_INFO, ®);
225 /* Setup mappings... read/write offsets, mmaps
226 * For PCI devices, config space is a region */
229 for (i = 0; i < device_info.num_irqs; i++) {
230 struct vfio_irq_info irq = { .argsz = sizeof(irq) };
234 ioctl(device, VFIO_DEVICE_GET_IRQ_INFO, &irq);
236 /* Setup IRQs... eventfds, VFIO_DEVICE_SET_IRQS */
239 /* Gratuitous device reset and go... */
240 ioctl(device, VFIO_DEVICE_RESET);
243 -------------------------------------------------------------------------------
245 Please see include/uapi/linux/vfio.h for complete API documentation.
248 -------------------------------------------------------------------------------
250 VFIO bus drivers, such as vfio-pci make use of only a few interfaces
251 into VFIO core. When devices are bound and unbound to the driver,
252 Following interfaces are called when devices are bound to and
253 unbound from the driver::
255 int vfio_register_group_dev(struct vfio_device *device);
256 int vfio_register_emulated_iommu_dev(struct vfio_device *device);
257 void vfio_unregister_group_dev(struct vfio_device *device);
259 The driver should embed the vfio_device in its own structure and use
260 vfio_alloc_device() to allocate the structure, and can register
261 @init/@release callbacks to manage any private state wrapping the
264 vfio_alloc_device(dev_struct, member, dev, ops);
265 void vfio_put_device(struct vfio_device *device);
267 vfio_register_group_dev() indicates to the core to begin tracking the
268 iommu_group of the specified dev and register the dev as owned by a VFIO bus
269 driver. Once vfio_register_group_dev() returns it is possible for userspace to
270 start accessing the driver, thus the driver should ensure it is completely
271 ready before calling it. The driver provides an ops structure for callbacks
272 similar to a file operations structure::
274 struct vfio_device_ops {
276 int (*init)(struct vfio_device *vdev);
277 void (*release)(struct vfio_device *vdev);
278 int (*bind_iommufd)(struct vfio_device *vdev,
279 struct iommufd_ctx *ictx, u32 *out_device_id);
280 void (*unbind_iommufd)(struct vfio_device *vdev);
281 int (*attach_ioas)(struct vfio_device *vdev, u32 *pt_id);
282 int (*open_device)(struct vfio_device *vdev);
283 void (*close_device)(struct vfio_device *vdev);
284 ssize_t (*read)(struct vfio_device *vdev, char __user *buf,
285 size_t count, loff_t *ppos);
286 ssize_t (*write)(struct vfio_device *vdev, const char __user *buf,
287 size_t count, loff_t *size);
288 long (*ioctl)(struct vfio_device *vdev, unsigned int cmd,
290 int (*mmap)(struct vfio_device *vdev, struct vm_area_struct *vma);
291 void (*request)(struct vfio_device *vdev, unsigned int count);
292 int (*match)(struct vfio_device *vdev, char *buf);
293 void (*dma_unmap)(struct vfio_device *vdev, u64 iova, u64 length);
294 int (*device_feature)(struct vfio_device *device, u32 flags,
295 void __user *arg, size_t argsz);
298 Each function is passed the vdev that was originally registered
299 in the vfio_register_group_dev() or vfio_register_emulated_iommu_dev()
300 call above. This allows the bus driver to obtain its private data using
305 - The init/release callbacks are issued when vfio_device is initialized
308 - The open/close device callbacks are issued when the first
309 instance of a file descriptor for the device is created (eg.
310 via VFIO_GROUP_GET_DEVICE_FD) for a user session.
312 - The ioctl callback provides a direct pass through for some VFIO_DEVICE_*
315 - The [un]bind_iommufd callbacks are issued when the device is bound to
316 and unbound from iommufd.
318 - The attach_ioas callback is issued when the device is attached to an
319 IOAS managed by the bound iommufd. The attached IOAS is automatically
320 detached when the device is unbound from iommufd.
322 - The read/write/mmap callbacks implement the device region access defined
323 by the device's own VFIO_DEVICE_GET_REGION_INFO ioctl.
325 - The request callback is issued when device is going to be unregistered,
326 such as when trying to unbind the device from the vfio bus driver.
328 - The dma_unmap callback is issued when a range of iovas are unmapped
329 in the container or IOAS attached by the device. Drivers which make
330 use of the vfio page pinning interface must implement this callback in
331 order to unpin pages within the dma_unmap range. Drivers must tolerate
332 this callback even before calls to open_device().
334 PPC64 sPAPR implementation note
335 -------------------------------
337 This implementation has some specifics:
339 1) On older systems (POWER7 with P5IOC2/IODA1) only one IOMMU group per
340 container is supported as an IOMMU table is allocated at the boot time,
341 one table per a IOMMU group which is a Partitionable Endpoint (PE)
342 (PE is often a PCI domain but not always).
344 Newer systems (POWER8 with IODA2) have improved hardware design which allows
345 to remove this limitation and have multiple IOMMU groups per a VFIO
348 2) The hardware supports so called DMA windows - the PCI address range
349 within which DMA transfer is allowed, any attempt to access address space
350 out of the window leads to the whole PE isolation.
352 3) PPC64 guests are paravirtualized but not fully emulated. There is an API
353 to map/unmap pages for DMA, and it normally maps 1..32 pages per call and
354 currently there is no way to reduce the number of calls. In order to make
355 things faster, the map/unmap handling has been implemented in real mode
356 which provides an excellent performance which has limitations such as
357 inability to do locked pages accounting in real time.
359 4) According to sPAPR specification, A Partitionable Endpoint (PE) is an I/O
360 subtree that can be treated as a unit for the purposes of partitioning and
361 error recovery. A PE may be a single or multi-function IOA (IO Adapter), a
362 function of a multi-function IOA, or multiple IOAs (possibly including
363 switch and bridge structures above the multiple IOAs). PPC64 guests detect
364 PCI errors and recover from them via EEH RTAS services, which works on the
365 basis of additional ioctl commands.
367 So 4 additional ioctls have been added:
369 VFIO_IOMMU_SPAPR_TCE_GET_INFO
370 returns the size and the start of the DMA window on the PCI bus.
373 enables the container. The locked pages accounting
374 is done at this point. This lets user first to know what
375 the DMA window is and adjust rlimit before doing any real job.
378 disables the container.
381 provides an API for EEH setup, error detection and recovery.
383 The code flow from the example above should be slightly changed::
385 struct vfio_eeh_pe_op pe_op = { .argsz = sizeof(pe_op), .flags = 0 };
388 /* Add the group to the container */
389 ioctl(group, VFIO_GROUP_SET_CONTAINER, &container);
391 /* Enable the IOMMU model we want */
392 ioctl(container, VFIO_SET_IOMMU, VFIO_SPAPR_TCE_IOMMU)
394 /* Get addition sPAPR IOMMU info */
395 vfio_iommu_spapr_tce_info spapr_iommu_info;
396 ioctl(container, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &spapr_iommu_info);
398 if (ioctl(container, VFIO_IOMMU_ENABLE))
399 /* Cannot enable container, may be low rlimit */
401 /* Allocate some space and setup a DMA mapping */
402 dma_map.vaddr = mmap(0, 1024 * 1024, PROT_READ | PROT_WRITE,
403 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
405 dma_map.size = 1024 * 1024;
406 dma_map.iova = 0; /* 1MB starting at 0x0 from device view */
407 dma_map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
409 /* Check here is .iova/.size are within DMA window from spapr_iommu_info */
410 ioctl(container, VFIO_IOMMU_MAP_DMA, &dma_map);
412 /* Get a file descriptor for the device */
413 device = ioctl(group, VFIO_GROUP_GET_DEVICE_FD, "0000:06:0d.0");
417 /* Gratuitous device reset and go... */
418 ioctl(device, VFIO_DEVICE_RESET);
420 /* Make sure EEH is supported */
421 ioctl(container, VFIO_CHECK_EXTENSION, VFIO_EEH);
423 /* Enable the EEH functionality on the device */
424 pe_op.op = VFIO_EEH_PE_ENABLE;
425 ioctl(container, VFIO_EEH_PE_OP, &pe_op);
427 /* You're suggested to create additional data struct to represent
428 * PE, and put child devices belonging to same IOMMU group to the
429 * PE instance for later reference.
432 /* Check the PE's state and make sure it's in functional state */
433 pe_op.op = VFIO_EEH_PE_GET_STATE;
434 ioctl(container, VFIO_EEH_PE_OP, &pe_op);
436 /* Save device state using pci_save_state().
437 * EEH should be enabled on the specified device.
442 /* Inject EEH error, which is expected to be caused by 32-bits
445 pe_op.op = VFIO_EEH_PE_INJECT_ERR;
446 pe_op.err.type = EEH_ERR_TYPE_32;
447 pe_op.err.func = EEH_ERR_FUNC_LD_CFG_ADDR;
448 pe_op.err.addr = 0ul;
449 pe_op.err.mask = 0ul;
450 ioctl(container, VFIO_EEH_PE_OP, &pe_op);
454 /* When 0xFF's returned from reading PCI config space or IO BARs
455 * of the PCI device. Check the PE's state to see if that has been
458 ioctl(container, VFIO_EEH_PE_OP, &pe_op);
460 /* Waiting for pending PCI transactions to be completed and don't
461 * produce any more PCI traffic from/to the affected PE until
462 * recovery is finished.
465 /* Enable IO for the affected PE and collect logs. Usually, the
466 * standard part of PCI config space, AER registers are dumped
467 * as logs for further analysis.
469 pe_op.op = VFIO_EEH_PE_UNFREEZE_IO;
470 ioctl(container, VFIO_EEH_PE_OP, &pe_op);
473 * Issue PE reset: hot or fundamental reset. Usually, hot reset
474 * is enough. However, the firmware of some PCI adapters would
475 * require fundamental reset.
477 pe_op.op = VFIO_EEH_PE_RESET_HOT;
478 ioctl(container, VFIO_EEH_PE_OP, &pe_op);
479 pe_op.op = VFIO_EEH_PE_RESET_DEACTIVATE;
480 ioctl(container, VFIO_EEH_PE_OP, &pe_op);
482 /* Configure the PCI bridges for the affected PE */
483 pe_op.op = VFIO_EEH_PE_CONFIGURE;
484 ioctl(container, VFIO_EEH_PE_OP, &pe_op);
486 /* Restored state we saved at initialization time. pci_restore_state()
487 * is good enough as an example.
490 /* Hopefully, error is recovered successfully. Now, you can resume to
491 * start PCI traffic to/from the affected PE.
496 5) There is v2 of SPAPR TCE IOMMU. It deprecates VFIO_IOMMU_ENABLE/
497 VFIO_IOMMU_DISABLE and implements 2 new ioctls:
498 VFIO_IOMMU_SPAPR_REGISTER_MEMORY and VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY
499 (which are unsupported in v1 IOMMU).
501 PPC64 paravirtualized guests generate a lot of map/unmap requests,
502 and the handling of those includes pinning/unpinning pages and updating
503 mm::locked_vm counter to make sure we do not exceed the rlimit.
504 The v2 IOMMU splits accounting and pinning into separate operations:
506 - VFIO_IOMMU_SPAPR_REGISTER_MEMORY/VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY ioctls
507 receive a user space address and size of the block to be pinned.
508 Bisecting is not supported and VFIO_IOMMU_UNREGISTER_MEMORY is expected to
509 be called with the exact address and size used for registering
510 the memory block. The userspace is not expected to call these often.
511 The ranges are stored in a linked list in a VFIO container.
513 - VFIO_IOMMU_MAP_DMA/VFIO_IOMMU_UNMAP_DMA ioctls only update the actual
514 IOMMU table and do not do pinning; instead these check that the userspace
515 address is from pre-registered range.
517 This separation helps in optimizing DMA for guests.
519 6) sPAPR specification allows guests to have an additional DMA window(s) on
520 a PCI bus with a variable page size. Two ioctls have been added to support
521 this: VFIO_IOMMU_SPAPR_TCE_CREATE and VFIO_IOMMU_SPAPR_TCE_REMOVE.
522 The platform has to support the functionality or error will be returned to
523 the userspace. The existing hardware supports up to 2 DMA windows, one is
524 2GB long, uses 4K pages and called "default 32bit window"; the other can
525 be as big as entire RAM, use different page size, it is optional - guests
526 create those in run-time if the guest driver supports 64bit DMA.
528 VFIO_IOMMU_SPAPR_TCE_CREATE receives a page shift, a DMA window size and
529 a number of TCE table levels (if a TCE table is going to be big enough and
530 the kernel may not be able to allocate enough of physically contiguous
531 memory). It creates a new window in the available slot and returns the bus
532 address where the new window starts. Due to hardware limitation, the user
533 space cannot choose the location of DMA windows.
535 VFIO_IOMMU_SPAPR_TCE_REMOVE receives the bus start address of the window
538 -------------------------------------------------------------------------------
540 .. [1] VFIO was originally an acronym for "Virtual Function I/O" in its
541 initial implementation by Tom Lyon while as Cisco. We've since
542 outgrown the acronym, but it's catchy.
544 .. [2] "safe" also depends upon a device being "well behaved". It's
545 possible for multi-function devices to have backdoors between
546 functions and even for single function devices to have alternative
547 access to things like PCI config space through MMIO registers. To
548 guard against the former we can include additional precautions in the
549 IOMMU driver to group multi-function PCI devices together
550 (iommu=group_mf). The latter we can't prevent, but the IOMMU should
551 still provide isolation. For PCI, SR-IOV Virtual Functions are the
552 best indicator of "well behaved", as these are designed for
553 virtualization usage models.
555 .. [3] As always there are trade-offs to virtual machine device
556 assignment that are beyond the scope of VFIO. It's expected that
557 future IOMMU technologies will reduce some, but maybe not all, of
560 .. [4] In this case the device is below a PCI bridge, so transactions
561 from either function of the device are indistinguishable to the iommu::
563 -[0000:00]-+-1e.0-[06]--+-0d.0
566 00:1e.0 PCI bridge: Intel Corporation 82801 PCI Bridge (rev 90)