1================================== 2VFIO - "Virtual Function I/O" [1]_ 3================================== 4 5Many modern systems now provide DMA and interrupt remapping facilities 6to help ensure I/O devices behave within the boundaries they've been 7allotted. This includes x86 hardware with AMD-Vi and Intel VT-d, 8POWER systems with Partitionable Endpoints (PEs) and embedded PowerPC 9systems such as Freescale PAMU. The VFIO driver is an IOMMU/device 10agnostic framework for exposing direct device access to userspace, in 11a secure, IOMMU protected environment. In other words, this allows 12safe [2]_, non-privileged, userspace drivers. 13 14Why do we want that? Virtual machines often make use of direct device 15access ("device assignment") when configured for the highest possible 16I/O performance. From a device and host perspective, this simply 17turns the VM into a userspace driver, with the benefits of 18significantly reduced latency, higher bandwidth, and direct use of 19bare-metal device drivers [3]_. 20 21Some applications, particularly in the high performance computing 22field, also benefit from low-overhead, direct device access from 23userspace. Examples include network adapters (often non-TCP/IP based) 24and compute accelerators. Prior to VFIO, these drivers had to either 25go through the full development cycle to become proper upstream 26driver, be maintained out of tree, or make use of the UIO framework, 27which has no notion of IOMMU protection, limited interrupt support, 28and requires root privileges to access things like PCI configuration 29space. 30 31The VFIO driver framework intends to unify these, replacing both the 32KVM PCI specific device assignment code as well as provide a more 33secure, more featureful userspace driver environment than UIO. 34 35Groups, Devices, and IOMMUs 36--------------------------- 37 38Devices are the main target of any I/O driver. Devices typically 39create a programming interface made up of I/O access, interrupts, 40and DMA. Without going into the details of each of these, DMA is 41by far the most critical aspect for maintaining a secure environment 42as allowing a device read-write access to system memory imposes the 43greatest risk to the overall system integrity. 44 45To help mitigate this risk, many modern IOMMUs now incorporate 46isolation properties into what was, in many cases, an interface only 47meant for translation (ie. solving the addressing problems of devices 48with limited address spaces). With this, devices can now be isolated 49from each other and from arbitrary memory access, thus allowing 50things like secure direct assignment of devices into virtual machines. 51 52This isolation is not always at the granularity of a single device 53though. Even when an IOMMU is capable of this, properties of devices, 54interconnects, and IOMMU topologies can each reduce this isolation. 55For instance, an individual device may be part of a larger multi- 56function enclosure. While the IOMMU may be able to distinguish 57between devices within the enclosure, the enclosure may not require 58transactions between devices to reach the IOMMU. Examples of this 59could be anything from a multi-function PCI device with backdoors 60between functions to a non-PCI-ACS (Access Control Services) capable 61bridge allowing redirection without reaching the IOMMU. Topology 62can also play a factor in terms of hiding devices. A PCIe-to-PCI 63bridge masks the devices behind it, making transaction appear as if 64from the bridge itself. Obviously IOMMU design plays a major factor 65as well. 66 67Therefore, while for the most part an IOMMU may have device level 68granularity, any system is susceptible to reduced granularity. The 69IOMMU API therefore supports a notion of IOMMU groups. A group is 70a set of devices which is isolatable from all other devices in the 71system. Groups are therefore the unit of ownership used by VFIO. 72 73While the group is the minimum granularity that must be used to 74ensure secure user access, it's not necessarily the preferred 75granularity. In IOMMUs which make use of page tables, it may be 76possible to share a set of page tables between different groups, 77reducing the overhead both to the platform (reduced TLB thrashing, 78reduced duplicate page tables), and to the user (programming only 79a single set of translations). For this reason, VFIO makes use of 80a container class, which may hold one or more groups. A container 81is created by simply opening the /dev/vfio/vfio character device. 82 83On its own, the container provides little functionality, with all 84but a couple version and extension query interfaces locked away. 85The user needs to add a group into the container for the next level 86of functionality. To do this, the user first needs to identify the 87group associated with the desired device. This can be done using 88the sysfs links described in the example below. By unbinding the 89device from the host driver and binding it to a VFIO driver, a new 90VFIO 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. 92If the IOMMU group contains multiple devices, each will need to 93be bound to a VFIO driver before operations on the VFIO group 94are allowed (it's also sufficient to only unbind the device from 95host drivers if a VFIO driver is unavailable; this will make the 96group available, but not that particular device). TBD - interface 97for disabling driver probing/locking a device. 98 99Once the group is ready, it may be added to the container by opening 100the VFIO group character device (/dev/vfio/$GROUP) and using the 101VFIO_GROUP_SET_CONTAINER ioctl, passing the file descriptor of the 102previously opened container file. If desired and if the IOMMU driver 103supports sharing the IOMMU context between groups, multiple groups may 104be set to the same container. If a group fails to set to a container 105with existing groups, a new empty container will need to be used 106instead. 107 108With a group (or groups) attached to a container, the remaining 109ioctls become available, enabling access to the VFIO IOMMU interfaces. 110Additionally, it now becomes possible to get file descriptors for each 111device within a group using an ioctl on the VFIO group file descriptor. 112 113The VFIO device API includes ioctls for describing the device, the I/O 114regions and their read/write/mmap offsets on the device descriptor, as 115well as mechanisms for describing and registering interrupt 116notifications. 117 118VFIO Usage Example 119------------------ 120 121Assume user wants to access PCI device 0000:06:0d.0:: 122 123 $ readlink /sys/bus/pci/devices/0000:06:0d.0/iommu_group 124 ../../../../kernel/iommu_groups/26 125 126This device is therefore in IOMMU group 26. This device is on the 127pci bus, therefore the user will make use of vfio-pci to manage the 128group:: 129 130 # modprobe vfio-pci 131 132Binding this device to the vfio-pci driver creates the VFIO group 133character devices for this group:: 134 135 $ 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 139 140Now we need to look at what other devices are in the group to free 141it for use by VFIO:: 142 143 $ ls -l /sys/bus/pci/devices/0000:06:0d.0/iommu_group/devices 144 total 0 145 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.1 151 152This device is behind a PCIe-to-PCI bridge [4]_, therefore we also 153need to add device 0000:06:0d.1 to the group following the same 154procedure as above. Device 0000:00:1e.0 is a bridge that does 155not currently have a host driver, therefore it's not required to 156bind this device to the vfio-pci driver (vfio-pci does not currently 157support PCI bridges). 158 159The final step is to provide the user with access to the group if 160unprivileged operation is desired (note that /dev/vfio/vfio provides 161no capabilities on its own and is therefore expected to be set to 162mode 0666 by the system):: 163 164 # chown user:user /dev/vfio/26 165 166The user now has full access to all the devices and the iommu for this 167group and can access them as follows:: 168 169 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) }; 175 176 /* Create a new container */ 177 container = open("/dev/vfio/vfio", O_RDWR); 178 179 if (ioctl(container, VFIO_GET_API_VERSION) != VFIO_API_VERSION) 180 /* Unknown API version */ 181 182 if (!ioctl(container, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) 183 /* Doesn't support the IOMMU driver we want. */ 184 185 /* Open the group */ 186 group = open("/dev/vfio/26", O_RDWR); 187 188 /* Test the group is viable and available */ 189 ioctl(group, VFIO_GROUP_GET_STATUS, &group_status); 190 191 if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) 192 /* Group is not viable (ie, not all devices bound for vfio) */ 193 194 /* Add the group to the container */ 195 ioctl(group, VFIO_GROUP_SET_CONTAINER, &container); 196 197 /* Enable the IOMMU model we want */ 198 ioctl(container, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU); 199 200 /* Get addition IOMMU info */ 201 ioctl(container, VFIO_IOMMU_GET_INFO, &iommu_info); 202 203 /* 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; 209 210 ioctl(container, VFIO_IOMMU_MAP_DMA, &dma_map); 211 212 /* Get a file descriptor for the device */ 213 device = ioctl(group, VFIO_GROUP_GET_DEVICE_FD, "0000:06:0d.0"); 214 215 /* Test and setup the device */ 216 ioctl(device, VFIO_DEVICE_GET_INFO, &device_info); 217 218 for (i = 0; i < device_info.num_regions; i++) { 219 struct vfio_region_info reg = { .argsz = sizeof(reg) }; 220 221 reg.index = i; 222 223 ioctl(device, VFIO_DEVICE_GET_REGION_INFO, ®); 224 225 /* Setup mappings... read/write offsets, mmaps 226 * For PCI devices, config space is a region */ 227 } 228 229 for (i = 0; i < device_info.num_irqs; i++) { 230 struct vfio_irq_info irq = { .argsz = sizeof(irq) }; 231 232 irq.index = i; 233 234 ioctl(device, VFIO_DEVICE_GET_IRQ_INFO, &irq); 235 236 /* Setup IRQs... eventfds, VFIO_DEVICE_SET_IRQS */ 237 } 238 239 /* Gratuitous device reset and go... */ 240 ioctl(device, VFIO_DEVICE_RESET); 241 242VFIO User API 243------------------------------------------------------------------------------- 244 245Please see include/linux/vfio.h for complete API documentation. 246 247VFIO bus driver API 248------------------------------------------------------------------------------- 249 250VFIO bus drivers, such as vfio-pci make use of only a few interfaces 251into VFIO core. When devices are bound and unbound to the driver, 252the driver should call vfio_register_group_dev() and 253vfio_unregister_group_dev() respectively:: 254 255 void vfio_init_group_dev(struct vfio_device *device, 256 struct device *dev, 257 const struct vfio_device_ops *ops); 258 void vfio_uninit_group_dev(struct vfio_device *device); 259 int vfio_register_group_dev(struct vfio_device *device); 260 void vfio_unregister_group_dev(struct vfio_device *device); 261 262The driver should embed the vfio_device in its own structure and call 263vfio_init_group_dev() to pre-configure it before going to registration 264and call vfio_uninit_group_dev() after completing the un-registration. 265vfio_register_group_dev() indicates to the core to begin tracking the 266iommu_group of the specified dev and register the dev as owned by a VFIO bus 267driver. Once vfio_register_group_dev() returns it is possible for userspace to 268start accessing the driver, thus the driver should ensure it is completely 269ready before calling it. The driver provides an ops structure for callbacks 270similar to a file operations structure:: 271 272 struct vfio_device_ops { 273 int (*open)(struct vfio_device *vdev); 274 void (*release)(struct vfio_device *vdev); 275 ssize_t (*read)(struct vfio_device *vdev, char __user *buf, 276 size_t count, loff_t *ppos); 277 ssize_t (*write)(struct vfio_device *vdev, 278 const char __user *buf, 279 size_t size, loff_t *ppos); 280 long (*ioctl)(struct vfio_device *vdev, unsigned int cmd, 281 unsigned long arg); 282 int (*mmap)(struct vfio_device *vdev, 283 struct vm_area_struct *vma); 284 }; 285 286Each function is passed the vdev that was originally registered 287in the vfio_register_group_dev() call above. This allows the bus driver 288to obtain its private data using container_of(). The open/release 289callbacks are issued when a new file descriptor is created for a 290device (via VFIO_GROUP_GET_DEVICE_FD). The ioctl interface provides 291a direct pass through for VFIO_DEVICE_* ioctls. The read/write/mmap 292interfaces implement the device region access defined by the device's 293own VFIO_DEVICE_GET_REGION_INFO ioctl. 294 295 296PPC64 sPAPR implementation note 297------------------------------- 298 299This implementation has some specifics: 300 3011) On older systems (POWER7 with P5IOC2/IODA1) only one IOMMU group per 302 container is supported as an IOMMU table is allocated at the boot time, 303 one table per a IOMMU group which is a Partitionable Endpoint (PE) 304 (PE is often a PCI domain but not always). 305 306 Newer systems (POWER8 with IODA2) have improved hardware design which allows 307 to remove this limitation and have multiple IOMMU groups per a VFIO 308 container. 309 3102) The hardware supports so called DMA windows - the PCI address range 311 within which DMA transfer is allowed, any attempt to access address space 312 out of the window leads to the whole PE isolation. 313 3143) PPC64 guests are paravirtualized but not fully emulated. There is an API 315 to map/unmap pages for DMA, and it normally maps 1..32 pages per call and 316 currently there is no way to reduce the number of calls. In order to make 317 things faster, the map/unmap handling has been implemented in real mode 318 which provides an excellent performance which has limitations such as 319 inability to do locked pages accounting in real time. 320 3214) According to sPAPR specification, A Partitionable Endpoint (PE) is an I/O 322 subtree that can be treated as a unit for the purposes of partitioning and 323 error recovery. A PE may be a single or multi-function IOA (IO Adapter), a 324 function of a multi-function IOA, or multiple IOAs (possibly including 325 switch and bridge structures above the multiple IOAs). PPC64 guests detect 326 PCI errors and recover from them via EEH RTAS services, which works on the 327 basis of additional ioctl commands. 328 329 So 4 additional ioctls have been added: 330 331 VFIO_IOMMU_SPAPR_TCE_GET_INFO 332 returns the size and the start of the DMA window on the PCI bus. 333 334 VFIO_IOMMU_ENABLE 335 enables the container. The locked pages accounting 336 is done at this point. This lets user first to know what 337 the DMA window is and adjust rlimit before doing any real job. 338 339 VFIO_IOMMU_DISABLE 340 disables the container. 341 342 VFIO_EEH_PE_OP 343 provides an API for EEH setup, error detection and recovery. 344 345 The code flow from the example above should be slightly changed:: 346 347 struct vfio_eeh_pe_op pe_op = { .argsz = sizeof(pe_op), .flags = 0 }; 348 349 ..... 350 /* Add the group to the container */ 351 ioctl(group, VFIO_GROUP_SET_CONTAINER, &container); 352 353 /* Enable the IOMMU model we want */ 354 ioctl(container, VFIO_SET_IOMMU, VFIO_SPAPR_TCE_IOMMU) 355 356 /* Get addition sPAPR IOMMU info */ 357 vfio_iommu_spapr_tce_info spapr_iommu_info; 358 ioctl(container, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &spapr_iommu_info); 359 360 if (ioctl(container, VFIO_IOMMU_ENABLE)) 361 /* Cannot enable container, may be low rlimit */ 362 363 /* Allocate some space and setup a DMA mapping */ 364 dma_map.vaddr = mmap(0, 1024 * 1024, PROT_READ | PROT_WRITE, 365 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0); 366 367 dma_map.size = 1024 * 1024; 368 dma_map.iova = 0; /* 1MB starting at 0x0 from device view */ 369 dma_map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE; 370 371 /* Check here is .iova/.size are within DMA window from spapr_iommu_info */ 372 ioctl(container, VFIO_IOMMU_MAP_DMA, &dma_map); 373 374 /* Get a file descriptor for the device */ 375 device = ioctl(group, VFIO_GROUP_GET_DEVICE_FD, "0000:06:0d.0"); 376 377 .... 378 379 /* Gratuitous device reset and go... */ 380 ioctl(device, VFIO_DEVICE_RESET); 381 382 /* Make sure EEH is supported */ 383 ioctl(container, VFIO_CHECK_EXTENSION, VFIO_EEH); 384 385 /* Enable the EEH functionality on the device */ 386 pe_op.op = VFIO_EEH_PE_ENABLE; 387 ioctl(container, VFIO_EEH_PE_OP, &pe_op); 388 389 /* You're suggested to create additional data struct to represent 390 * PE, and put child devices belonging to same IOMMU group to the 391 * PE instance for later reference. 392 */ 393 394 /* Check the PE's state and make sure it's in functional state */ 395 pe_op.op = VFIO_EEH_PE_GET_STATE; 396 ioctl(container, VFIO_EEH_PE_OP, &pe_op); 397 398 /* Save device state using pci_save_state(). 399 * EEH should be enabled on the specified device. 400 */ 401 402 .... 403 404 /* Inject EEH error, which is expected to be caused by 32-bits 405 * config load. 406 */ 407 pe_op.op = VFIO_EEH_PE_INJECT_ERR; 408 pe_op.err.type = EEH_ERR_TYPE_32; 409 pe_op.err.func = EEH_ERR_FUNC_LD_CFG_ADDR; 410 pe_op.err.addr = 0ul; 411 pe_op.err.mask = 0ul; 412 ioctl(container, VFIO_EEH_PE_OP, &pe_op); 413 414 .... 415 416 /* When 0xFF's returned from reading PCI config space or IO BARs 417 * of the PCI device. Check the PE's state to see if that has been 418 * frozen. 419 */ 420 ioctl(container, VFIO_EEH_PE_OP, &pe_op); 421 422 /* Waiting for pending PCI transactions to be completed and don't 423 * produce any more PCI traffic from/to the affected PE until 424 * recovery is finished. 425 */ 426 427 /* Enable IO for the affected PE and collect logs. Usually, the 428 * standard part of PCI config space, AER registers are dumped 429 * as logs for further analysis. 430 */ 431 pe_op.op = VFIO_EEH_PE_UNFREEZE_IO; 432 ioctl(container, VFIO_EEH_PE_OP, &pe_op); 433 434 /* 435 * Issue PE reset: hot or fundamental reset. Usually, hot reset 436 * is enough. However, the firmware of some PCI adapters would 437 * require fundamental reset. 438 */ 439 pe_op.op = VFIO_EEH_PE_RESET_HOT; 440 ioctl(container, VFIO_EEH_PE_OP, &pe_op); 441 pe_op.op = VFIO_EEH_PE_RESET_DEACTIVATE; 442 ioctl(container, VFIO_EEH_PE_OP, &pe_op); 443 444 /* Configure the PCI bridges for the affected PE */ 445 pe_op.op = VFIO_EEH_PE_CONFIGURE; 446 ioctl(container, VFIO_EEH_PE_OP, &pe_op); 447 448 /* Restored state we saved at initialization time. pci_restore_state() 449 * is good enough as an example. 450 */ 451 452 /* Hopefully, error is recovered successfully. Now, you can resume to 453 * start PCI traffic to/from the affected PE. 454 */ 455 456 .... 457 4585) There is v2 of SPAPR TCE IOMMU. It deprecates VFIO_IOMMU_ENABLE/ 459 VFIO_IOMMU_DISABLE and implements 2 new ioctls: 460 VFIO_IOMMU_SPAPR_REGISTER_MEMORY and VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY 461 (which are unsupported in v1 IOMMU). 462 463 PPC64 paravirtualized guests generate a lot of map/unmap requests, 464 and the handling of those includes pinning/unpinning pages and updating 465 mm::locked_vm counter to make sure we do not exceed the rlimit. 466 The v2 IOMMU splits accounting and pinning into separate operations: 467 468 - VFIO_IOMMU_SPAPR_REGISTER_MEMORY/VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY ioctls 469 receive a user space address and size of the block to be pinned. 470 Bisecting is not supported and VFIO_IOMMU_UNREGISTER_MEMORY is expected to 471 be called with the exact address and size used for registering 472 the memory block. The userspace is not expected to call these often. 473 The ranges are stored in a linked list in a VFIO container. 474 475 - VFIO_IOMMU_MAP_DMA/VFIO_IOMMU_UNMAP_DMA ioctls only update the actual 476 IOMMU table and do not do pinning; instead these check that the userspace 477 address is from pre-registered range. 478 479 This separation helps in optimizing DMA for guests. 480 4816) sPAPR specification allows guests to have an additional DMA window(s) on 482 a PCI bus with a variable page size. Two ioctls have been added to support 483 this: VFIO_IOMMU_SPAPR_TCE_CREATE and VFIO_IOMMU_SPAPR_TCE_REMOVE. 484 The platform has to support the functionality or error will be returned to 485 the userspace. The existing hardware supports up to 2 DMA windows, one is 486 2GB long, uses 4K pages and called "default 32bit window"; the other can 487 be as big as entire RAM, use different page size, it is optional - guests 488 create those in run-time if the guest driver supports 64bit DMA. 489 490 VFIO_IOMMU_SPAPR_TCE_CREATE receives a page shift, a DMA window size and 491 a number of TCE table levels (if a TCE table is going to be big enough and 492 the kernel may not be able to allocate enough of physically contiguous 493 memory). It creates a new window in the available slot and returns the bus 494 address where the new window starts. Due to hardware limitation, the user 495 space cannot choose the location of DMA windows. 496 497 VFIO_IOMMU_SPAPR_TCE_REMOVE receives the bus start address of the window 498 and removes it. 499 500------------------------------------------------------------------------------- 501 502.. [1] VFIO was originally an acronym for "Virtual Function I/O" in its 503 initial implementation by Tom Lyon while as Cisco. We've since 504 outgrown the acronym, but it's catchy. 505 506.. [2] "safe" also depends upon a device being "well behaved". It's 507 possible for multi-function devices to have backdoors between 508 functions and even for single function devices to have alternative 509 access to things like PCI config space through MMIO registers. To 510 guard against the former we can include additional precautions in the 511 IOMMU driver to group multi-function PCI devices together 512 (iommu=group_mf). The latter we can't prevent, but the IOMMU should 513 still provide isolation. For PCI, SR-IOV Virtual Functions are the 514 best indicator of "well behaved", as these are designed for 515 virtualization usage models. 516 517.. [3] As always there are trade-offs to virtual machine device 518 assignment that are beyond the scope of VFIO. It's expected that 519 future IOMMU technologies will reduce some, but maybe not all, of 520 these trade-offs. 521 522.. [4] In this case the device is below a PCI bridge, so transactions 523 from either function of the device are indistinguishable to the iommu:: 524 525 -[0000:00]-+-1e.0-[06]--+-0d.0 526 \-0d.1 527 528 00:1e.0 PCI bridge: Intel Corporation 82801 PCI Bridge (rev 90) 529