1 // Copyright © 2021 Intel Corporation 2 // 3 // SPDX-License-Identifier: Apache-2.0 4 // 5 6 use std::any::Any; 7 use std::os::unix::prelude::AsRawFd; 8 use std::ptr::null_mut; 9 use std::sync::{Arc, Barrier, Mutex}; 10 11 use hypervisor::HypervisorVmError; 12 use thiserror::Error; 13 use vfio_bindings::bindings::vfio::*; 14 use vfio_ioctls::VfioIrq; 15 use vfio_user::{Client, Error as VfioUserError}; 16 use vm_allocator::{AddressAllocator, MemorySlotAllocator, SystemAllocator}; 17 use vm_device::dma_mapping::ExternalDmaMapping; 18 use vm_device::interrupt::{InterruptManager, InterruptSourceGroup, MsiIrqGroupConfig}; 19 use vm_device::{BusDevice, Resource}; 20 use vm_memory::bitmap::AtomicBitmap; 21 use vm_memory::{ 22 Address, GuestAddress, GuestAddressSpace, GuestMemory, GuestMemoryRegion, GuestRegionMmap, 23 }; 24 use vm_migration::{Migratable, MigratableError, Pausable, Snapshot, Snapshottable, Transportable}; 25 use vmm_sys_util::eventfd::EventFd; 26 27 use crate::vfio::{UserMemoryRegion, Vfio, VfioCommon, VfioError, VFIO_COMMON_ID}; 28 use crate::{ 29 BarReprogrammingParams, PciBarConfiguration, PciBdf, PciDevice, PciDeviceError, PciSubclass, 30 VfioPciError, 31 }; 32 33 pub struct VfioUserPciDevice { 34 id: String, 35 vm: Arc<dyn hypervisor::Vm>, 36 client: Arc<Mutex<Client>>, 37 common: VfioCommon, 38 memory_slot_allocator: MemorySlotAllocator, 39 } 40 41 #[derive(Error, Debug)] 42 pub enum VfioUserPciDeviceError { 43 #[error("Client error")] 44 Client(#[source] VfioUserError), 45 #[error("Failed to map VFIO PCI region into guest")] 46 MapRegionGuest(#[source] HypervisorVmError), 47 #[error("Failed to DMA map")] 48 DmaMap(#[source] VfioUserError), 49 #[error("Failed to DMA unmap")] 50 DmaUnmap(#[source] VfioUserError), 51 #[error("Failed to initialize legacy interrupts")] 52 InitializeLegacyInterrupts(#[source] VfioPciError), 53 #[error("Failed to create VfioCommon")] 54 CreateVfioCommon(#[source] VfioPciError), 55 } 56 57 #[derive(Copy, Clone)] 58 enum PciVfioUserSubclass { 59 VfioUserSubclass = 0xff, 60 } 61 62 impl PciSubclass for PciVfioUserSubclass { 63 fn get_register_value(&self) -> u8 { 64 *self as u8 65 } 66 } 67 68 impl VfioUserPciDevice { 69 #[allow(clippy::too_many_arguments)] 70 pub fn new( 71 id: String, 72 vm: &Arc<dyn hypervisor::Vm>, 73 client: Arc<Mutex<Client>>, 74 msi_interrupt_manager: Arc<dyn InterruptManager<GroupConfig = MsiIrqGroupConfig>>, 75 legacy_interrupt_group: Option<Arc<dyn InterruptSourceGroup>>, 76 bdf: PciBdf, 77 memory_slot_allocator: MemorySlotAllocator, 78 snapshot: Option<Snapshot>, 79 ) -> Result<Self, VfioUserPciDeviceError> { 80 let resettable = client.lock().unwrap().resettable(); 81 if resettable { 82 client 83 .lock() 84 .unwrap() 85 .reset() 86 .map_err(VfioUserPciDeviceError::Client)?; 87 } 88 89 let vfio_wrapper = VfioUserClientWrapper { 90 client: client.clone(), 91 }; 92 93 let common = VfioCommon::new( 94 msi_interrupt_manager, 95 legacy_interrupt_group, 96 Arc::new(vfio_wrapper) as Arc<dyn Vfio>, 97 &PciVfioUserSubclass::VfioUserSubclass, 98 bdf, 99 vm_migration::snapshot_from_id(snapshot.as_ref(), VFIO_COMMON_ID), 100 None, 101 ) 102 .map_err(VfioUserPciDeviceError::CreateVfioCommon)?; 103 104 Ok(Self { 105 id, 106 vm: vm.clone(), 107 client, 108 common, 109 memory_slot_allocator, 110 }) 111 } 112 113 pub fn map_mmio_regions(&mut self) -> Result<(), VfioUserPciDeviceError> { 114 for mmio_region in &mut self.common.mmio_regions { 115 let region_flags = self 116 .client 117 .lock() 118 .unwrap() 119 .region(mmio_region.index) 120 .unwrap() 121 .flags; 122 let file_offset = self 123 .client 124 .lock() 125 .unwrap() 126 .region(mmio_region.index) 127 .unwrap() 128 .file_offset 129 .clone(); 130 131 let sparse_areas = self 132 .client 133 .lock() 134 .unwrap() 135 .region(mmio_region.index) 136 .unwrap() 137 .sparse_areas 138 .clone(); 139 140 if region_flags & VFIO_REGION_INFO_FLAG_MMAP != 0 { 141 let mut prot = 0; 142 if region_flags & VFIO_REGION_INFO_FLAG_READ != 0 { 143 prot |= libc::PROT_READ; 144 } 145 if region_flags & VFIO_REGION_INFO_FLAG_WRITE != 0 { 146 prot |= libc::PROT_WRITE; 147 } 148 149 let mmaps = if sparse_areas.is_empty() { 150 vec![vfio_region_sparse_mmap_area { 151 offset: 0, 152 size: mmio_region.length, 153 }] 154 } else { 155 sparse_areas 156 }; 157 158 for s in mmaps.iter() { 159 // SAFETY: FFI call with correct arguments 160 let host_addr = unsafe { 161 libc::mmap( 162 null_mut(), 163 s.size as usize, 164 prot, 165 libc::MAP_SHARED, 166 file_offset.as_ref().unwrap().file().as_raw_fd(), 167 file_offset.as_ref().unwrap().start() as libc::off_t 168 + s.offset as libc::off_t, 169 ) 170 }; 171 172 if std::ptr::eq(host_addr, libc::MAP_FAILED) { 173 error!( 174 "Could not mmap regions, error:{}", 175 std::io::Error::last_os_error() 176 ); 177 continue; 178 } 179 180 let user_memory_region = UserMemoryRegion { 181 slot: self.memory_slot_allocator.next_memory_slot(), 182 start: mmio_region.start.0 + s.offset, 183 size: s.size, 184 host_addr: host_addr as u64, 185 }; 186 187 mmio_region.user_memory_regions.push(user_memory_region); 188 189 let mem_region = self.vm.make_user_memory_region( 190 user_memory_region.slot, 191 user_memory_region.start, 192 user_memory_region.size, 193 user_memory_region.host_addr, 194 false, 195 false, 196 ); 197 198 self.vm 199 .create_user_memory_region(mem_region) 200 .map_err(VfioUserPciDeviceError::MapRegionGuest)?; 201 } 202 } 203 } 204 205 Ok(()) 206 } 207 208 pub fn unmap_mmio_regions(&mut self) { 209 for mmio_region in self.common.mmio_regions.iter() { 210 for user_memory_region in mmio_region.user_memory_regions.iter() { 211 // Remove region 212 let r = self.vm.make_user_memory_region( 213 user_memory_region.slot, 214 user_memory_region.start, 215 user_memory_region.size, 216 user_memory_region.host_addr, 217 false, 218 false, 219 ); 220 221 if let Err(e) = self.vm.remove_user_memory_region(r) { 222 error!("Could not remove the userspace memory region: {}", e); 223 } 224 225 self.memory_slot_allocator 226 .free_memory_slot(user_memory_region.slot); 227 228 // Remove mmaps 229 // SAFETY: FFI call with correct arguments 230 let ret = unsafe { 231 libc::munmap( 232 user_memory_region.host_addr as *mut libc::c_void, 233 user_memory_region.size as usize, 234 ) 235 }; 236 if ret != 0 { 237 error!( 238 "Could not unmap region {}, error:{}", 239 mmio_region.index, 240 std::io::Error::last_os_error() 241 ); 242 } 243 } 244 } 245 } 246 247 pub fn dma_map( 248 &mut self, 249 region: &GuestRegionMmap<AtomicBitmap>, 250 ) -> Result<(), VfioUserPciDeviceError> { 251 let (fd, offset) = match region.file_offset() { 252 Some(_file_offset) => (_file_offset.file().as_raw_fd(), _file_offset.start()), 253 None => return Ok(()), 254 }; 255 256 self.client 257 .lock() 258 .unwrap() 259 .dma_map(offset, region.start_addr().raw_value(), region.len(), fd) 260 .map_err(VfioUserPciDeviceError::DmaMap) 261 } 262 263 pub fn dma_unmap( 264 &mut self, 265 region: &GuestRegionMmap<AtomicBitmap>, 266 ) -> Result<(), VfioUserPciDeviceError> { 267 self.client 268 .lock() 269 .unwrap() 270 .dma_unmap(region.start_addr().raw_value(), region.len()) 271 .map_err(VfioUserPciDeviceError::DmaUnmap) 272 } 273 } 274 275 impl BusDevice for VfioUserPciDevice { 276 fn read(&mut self, base: u64, offset: u64, data: &mut [u8]) { 277 self.read_bar(base, offset, data) 278 } 279 280 fn write(&mut self, base: u64, offset: u64, data: &[u8]) -> Option<Arc<Barrier>> { 281 self.write_bar(base, offset, data) 282 } 283 } 284 285 #[repr(u32)] 286 #[derive(Debug, PartialEq, Eq, PartialOrd, Ord)] 287 #[allow(dead_code)] 288 enum Regions { 289 Bar0, 290 Bar1, 291 Bar2, 292 Bar3, 293 Bar4, 294 Bar5, 295 Rom, 296 Config, 297 Vga, 298 Migration, 299 } 300 301 struct VfioUserClientWrapper { 302 client: Arc<Mutex<Client>>, 303 } 304 305 impl Vfio for VfioUserClientWrapper { 306 fn region_read(&self, index: u32, offset: u64, data: &mut [u8]) { 307 self.client 308 .lock() 309 .unwrap() 310 .region_read(index, offset, data) 311 .ok(); 312 } 313 314 fn region_write(&self, index: u32, offset: u64, data: &[u8]) { 315 self.client 316 .lock() 317 .unwrap() 318 .region_write(index, offset, data) 319 .ok(); 320 } 321 322 fn get_irq_info(&self, irq_index: u32) -> Option<VfioIrq> { 323 self.client 324 .lock() 325 .unwrap() 326 .get_irq_info(irq_index) 327 .ok() 328 .map(|i| VfioIrq { 329 index: i.index, 330 flags: i.flags, 331 count: i.count, 332 }) 333 } 334 335 fn enable_irq(&self, irq_index: u32, event_fds: Vec<&EventFd>) -> Result<(), VfioError> { 336 info!( 337 "Enabling IRQ {:x} number of fds = {:?}", 338 irq_index, 339 event_fds.len() 340 ); 341 let fds: Vec<i32> = event_fds.iter().map(|e| e.as_raw_fd()).collect(); 342 343 // Batch into blocks of 16 fds as sendmsg() has a size limit 344 let mut sent_fds = 0; 345 let num_fds = event_fds.len() as u32; 346 while sent_fds < num_fds { 347 let remaining_fds = num_fds - sent_fds; 348 let count = if remaining_fds > 16 { 349 16 350 } else { 351 remaining_fds 352 }; 353 354 self.client 355 .lock() 356 .unwrap() 357 .set_irqs( 358 irq_index, 359 VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER, 360 sent_fds, 361 count, 362 &fds[sent_fds as usize..(sent_fds + count) as usize], 363 ) 364 .map_err(VfioError::VfioUser)?; 365 366 sent_fds += count; 367 } 368 369 Ok(()) 370 } 371 372 fn disable_irq(&self, irq_index: u32) -> Result<(), VfioError> { 373 info!("Disabling IRQ {:x}", irq_index); 374 self.client 375 .lock() 376 .unwrap() 377 .set_irqs( 378 irq_index, 379 VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER, 380 0, 381 0, 382 &[], 383 ) 384 .map_err(VfioError::VfioUser) 385 } 386 387 fn unmask_irq(&self, irq_index: u32) -> Result<(), VfioError> { 388 info!("Unmasking IRQ {:x}", irq_index); 389 self.client 390 .lock() 391 .unwrap() 392 .set_irqs( 393 irq_index, 394 VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK, 395 0, 396 1, 397 &[], 398 ) 399 .map_err(VfioError::VfioUser) 400 } 401 } 402 403 impl PciDevice for VfioUserPciDevice { 404 fn allocate_bars( 405 &mut self, 406 allocator: &Arc<Mutex<SystemAllocator>>, 407 mmio32_allocator: &mut AddressAllocator, 408 mmio64_allocator: &mut AddressAllocator, 409 resources: Option<Vec<Resource>>, 410 ) -> Result<Vec<PciBarConfiguration>, PciDeviceError> { 411 self.common 412 .allocate_bars(allocator, mmio32_allocator, mmio64_allocator, resources) 413 } 414 415 fn free_bars( 416 &mut self, 417 allocator: &mut SystemAllocator, 418 mmio32_allocator: &mut AddressAllocator, 419 mmio64_allocator: &mut AddressAllocator, 420 ) -> Result<(), PciDeviceError> { 421 self.common 422 .free_bars(allocator, mmio32_allocator, mmio64_allocator) 423 } 424 425 fn as_any_mut(&mut self) -> &mut dyn Any { 426 self 427 } 428 429 fn write_config_register( 430 &mut self, 431 reg_idx: usize, 432 offset: u64, 433 data: &[u8], 434 ) -> (Vec<BarReprogrammingParams>, Option<Arc<Barrier>>) { 435 self.common.write_config_register(reg_idx, offset, data) 436 } 437 438 fn read_config_register(&mut self, reg_idx: usize) -> u32 { 439 self.common.read_config_register(reg_idx) 440 } 441 442 fn read_bar(&mut self, base: u64, offset: u64, data: &mut [u8]) { 443 self.common.read_bar(base, offset, data) 444 } 445 446 fn write_bar(&mut self, base: u64, offset: u64, data: &[u8]) -> Option<Arc<Barrier>> { 447 self.common.write_bar(base, offset, data) 448 } 449 450 fn move_bar(&mut self, old_base: u64, new_base: u64) -> Result<(), std::io::Error> { 451 info!("Moving BAR 0x{:x} -> 0x{:x}", old_base, new_base); 452 for mmio_region in self.common.mmio_regions.iter_mut() { 453 if mmio_region.start.raw_value() == old_base { 454 mmio_region.start = GuestAddress(new_base); 455 456 for user_memory_region in mmio_region.user_memory_regions.iter_mut() { 457 // Remove old region 458 let old_region = self.vm.make_user_memory_region( 459 user_memory_region.slot, 460 user_memory_region.start, 461 user_memory_region.size, 462 user_memory_region.host_addr, 463 false, 464 false, 465 ); 466 467 self.vm 468 .remove_user_memory_region(old_region) 469 .map_err(std::io::Error::other)?; 470 471 // Update the user memory region with the correct start address. 472 if new_base > old_base { 473 user_memory_region.start += new_base - old_base; 474 } else { 475 user_memory_region.start -= old_base - new_base; 476 } 477 478 // Insert new region 479 let new_region = self.vm.make_user_memory_region( 480 user_memory_region.slot, 481 user_memory_region.start, 482 user_memory_region.size, 483 user_memory_region.host_addr, 484 false, 485 false, 486 ); 487 488 self.vm 489 .create_user_memory_region(new_region) 490 .map_err(std::io::Error::other)?; 491 } 492 info!("Moved bar 0x{:x} -> 0x{:x}", old_base, new_base); 493 } 494 } 495 496 Ok(()) 497 } 498 499 fn id(&self) -> Option<String> { 500 Some(self.id.clone()) 501 } 502 } 503 504 impl Drop for VfioUserPciDevice { 505 fn drop(&mut self) { 506 self.unmap_mmio_regions(); 507 508 if let Some(msix) = &self.common.interrupt.msix { 509 if msix.bar.enabled() { 510 self.common.disable_msix(); 511 } 512 } 513 514 if let Some(msi) = &self.common.interrupt.msi { 515 if msi.cfg.enabled() { 516 self.common.disable_msi() 517 } 518 } 519 520 if self.common.interrupt.intx_in_use() { 521 self.common.disable_intx(); 522 } 523 524 if let Err(e) = self.client.lock().unwrap().shutdown() { 525 error!("Failed shutting down vfio-user client: {}", e); 526 } 527 } 528 } 529 530 impl Pausable for VfioUserPciDevice {} 531 532 impl Snapshottable for VfioUserPciDevice { 533 fn id(&self) -> String { 534 self.id.clone() 535 } 536 537 fn snapshot(&mut self) -> std::result::Result<Snapshot, MigratableError> { 538 let mut vfio_pci_dev_snapshot = Snapshot::default(); 539 540 // Snapshot VfioCommon 541 vfio_pci_dev_snapshot.add_snapshot(self.common.id(), self.common.snapshot()?); 542 543 Ok(vfio_pci_dev_snapshot) 544 } 545 } 546 impl Transportable for VfioUserPciDevice {} 547 impl Migratable for VfioUserPciDevice {} 548 549 pub struct VfioUserDmaMapping<M: GuestAddressSpace> { 550 client: Arc<Mutex<Client>>, 551 memory: Arc<M>, 552 } 553 554 impl<M: GuestAddressSpace> VfioUserDmaMapping<M> { 555 pub fn new(client: Arc<Mutex<Client>>, memory: Arc<M>) -> Self { 556 Self { client, memory } 557 } 558 } 559 560 impl<M: GuestAddressSpace + Sync + Send> ExternalDmaMapping for VfioUserDmaMapping<M> { 561 fn map(&self, iova: u64, gpa: u64, size: u64) -> std::result::Result<(), std::io::Error> { 562 let mem = self.memory.memory(); 563 let guest_addr = GuestAddress(gpa); 564 let region = mem.find_region(guest_addr); 565 566 if let Some(region) = region { 567 let file_offset = region.file_offset().unwrap(); 568 let offset = (GuestAddress(gpa).checked_offset_from(region.start_addr())).unwrap() 569 + file_offset.start(); 570 571 self.client 572 .lock() 573 .unwrap() 574 .dma_map(offset, iova, size, file_offset.file().as_raw_fd()) 575 .map_err(|e| std::io::Error::other(format!("Error mapping region: {e}"))) 576 } else { 577 Err(std::io::Error::other(format!( 578 "Region not found for 0x{gpa:x}" 579 ))) 580 } 581 } 582 583 fn unmap(&self, iova: u64, size: u64) -> std::result::Result<(), std::io::Error> { 584 self.client 585 .lock() 586 .unwrap() 587 .dma_unmap(iova, size) 588 .map_err(|e| std::io::Error::other(format!("Error unmapping region: {e}"))) 589 } 590 } 591