1 // Copyright © 2019 Intel Corporation 2 // 3 // SPDX-License-Identifier: Apache-2.0 AND BSD-3-Clause 4 5 use super::Error as DeviceError; 6 use super::{ 7 ActivateResult, EpollHelper, EpollHelperError, EpollHelperHandler, VirtioCommon, VirtioDevice, 8 VirtioDeviceType, EPOLL_HELPER_EVENT_LAST, VIRTIO_F_VERSION_1, 9 }; 10 use crate::seccomp_filters::Thread; 11 use crate::thread_helper::spawn_virtio_thread; 12 use crate::GuestMemoryMmap; 13 use crate::{DmaRemapping, VirtioInterrupt, VirtioInterruptType}; 14 use anyhow::anyhow; 15 use seccompiler::SeccompAction; 16 use serde::{Deserialize, Serialize}; 17 use std::collections::BTreeMap; 18 use std::io; 19 use std::mem::size_of; 20 use std::ops::Bound::Included; 21 use std::os::unix::io::AsRawFd; 22 use std::result; 23 use std::sync::atomic::{AtomicBool, Ordering}; 24 use std::sync::{Arc, Barrier, Mutex, RwLock}; 25 use thiserror::Error; 26 use virtio_queue::{DescriptorChain, Queue, QueueT}; 27 use vm_device::dma_mapping::ExternalDmaMapping; 28 use vm_memory::{ 29 Address, ByteValued, Bytes, GuestAddress, GuestAddressSpace, GuestMemoryAtomic, 30 GuestMemoryError, GuestMemoryLoadGuard, 31 }; 32 use vm_migration::{Migratable, MigratableError, Pausable, Snapshot, Snapshottable, Transportable}; 33 use vm_virtio::AccessPlatform; 34 use vmm_sys_util::eventfd::EventFd; 35 36 /// Queues sizes 37 const QUEUE_SIZE: u16 = 256; 38 const NUM_QUEUES: usize = 2; 39 const QUEUE_SIZES: &[u16] = &[QUEUE_SIZE; NUM_QUEUES]; 40 41 /// New descriptors are pending on the request queue. 42 /// "requestq" is meant to be used anytime an action is required to be 43 /// performed on behalf of the guest driver. 44 const REQUEST_Q_EVENT: u16 = EPOLL_HELPER_EVENT_LAST + 1; 45 /// New descriptors are pending on the event queue. 46 /// "eventq" lets the device report any fault or other asynchronous event to 47 /// the guest driver. 48 const _EVENT_Q_EVENT: u16 = EPOLL_HELPER_EVENT_LAST + 2; 49 50 /// PROBE properties size. 51 /// This is the minimal size to provide at least one RESV_MEM property. 52 /// Because virtio-iommu expects one MSI reserved region, we must provide it, 53 /// otherwise the driver in the guest will define a predefined one between 54 /// 0x8000000 and 0x80FFFFF, which is only relevant for ARM architecture, but 55 /// will conflict with x86. 56 const PROBE_PROP_SIZE: u32 = 57 (size_of::<VirtioIommuProbeProperty>() + size_of::<VirtioIommuProbeResvMem>()) as u32; 58 59 /// Virtio IOMMU features 60 #[allow(unused)] 61 const VIRTIO_IOMMU_F_INPUT_RANGE: u32 = 0; 62 #[allow(unused)] 63 const VIRTIO_IOMMU_F_DOMAIN_RANGE: u32 = 1; 64 #[allow(unused)] 65 const VIRTIO_IOMMU_F_MAP_UNMAP: u32 = 2; 66 #[allow(unused)] 67 const VIRTIO_IOMMU_F_BYPASS: u32 = 3; 68 const VIRTIO_IOMMU_F_PROBE: u32 = 4; 69 #[allow(unused)] 70 const VIRTIO_IOMMU_F_MMIO: u32 = 5; 71 const VIRTIO_IOMMU_F_BYPASS_CONFIG: u32 = 6; 72 73 // Support 2MiB and 4KiB page sizes. 74 const VIRTIO_IOMMU_PAGE_SIZE_MASK: u64 = (2 << 20) | (4 << 10); 75 76 #[derive(Copy, Clone, Debug, Default)] 77 #[repr(packed)] 78 #[allow(dead_code)] 79 struct VirtioIommuRange32 { 80 start: u32, 81 end: u32, 82 } 83 84 #[derive(Copy, Clone, Debug, Default)] 85 #[repr(packed)] 86 #[allow(dead_code)] 87 struct VirtioIommuRange64 { 88 start: u64, 89 end: u64, 90 } 91 92 #[derive(Copy, Clone, Debug, Default)] 93 #[repr(packed)] 94 #[allow(dead_code)] 95 struct VirtioIommuConfig { 96 page_size_mask: u64, 97 input_range: VirtioIommuRange64, 98 domain_range: VirtioIommuRange32, 99 probe_size: u32, 100 bypass: u8, 101 _reserved: [u8; 7], 102 } 103 104 /// Virtio IOMMU request type 105 const VIRTIO_IOMMU_T_ATTACH: u8 = 1; 106 const VIRTIO_IOMMU_T_DETACH: u8 = 2; 107 const VIRTIO_IOMMU_T_MAP: u8 = 3; 108 const VIRTIO_IOMMU_T_UNMAP: u8 = 4; 109 const VIRTIO_IOMMU_T_PROBE: u8 = 5; 110 111 #[derive(Copy, Clone, Debug, Default)] 112 #[repr(packed)] 113 struct VirtioIommuReqHead { 114 type_: u8, 115 _reserved: [u8; 3], 116 } 117 118 /// Virtio IOMMU request status 119 const VIRTIO_IOMMU_S_OK: u8 = 0; 120 #[allow(unused)] 121 const VIRTIO_IOMMU_S_IOERR: u8 = 1; 122 #[allow(unused)] 123 const VIRTIO_IOMMU_S_UNSUPP: u8 = 2; 124 #[allow(unused)] 125 const VIRTIO_IOMMU_S_DEVERR: u8 = 3; 126 #[allow(unused)] 127 const VIRTIO_IOMMU_S_INVAL: u8 = 4; 128 #[allow(unused)] 129 const VIRTIO_IOMMU_S_RANGE: u8 = 5; 130 #[allow(unused)] 131 const VIRTIO_IOMMU_S_NOENT: u8 = 6; 132 #[allow(unused)] 133 const VIRTIO_IOMMU_S_FAULT: u8 = 7; 134 #[allow(unused)] 135 const VIRTIO_IOMMU_S_NOMEM: u8 = 8; 136 137 #[derive(Copy, Clone, Debug, Default)] 138 #[repr(packed)] 139 #[allow(dead_code)] 140 struct VirtioIommuReqTail { 141 status: u8, 142 _reserved: [u8; 3], 143 } 144 145 /// ATTACH request 146 #[derive(Copy, Clone, Debug, Default)] 147 #[repr(packed)] 148 struct VirtioIommuReqAttach { 149 domain: u32, 150 endpoint: u32, 151 flags: u32, 152 _reserved: [u8; 4], 153 } 154 155 const VIRTIO_IOMMU_ATTACH_F_BYPASS: u32 = 1; 156 157 /// DETACH request 158 #[derive(Copy, Clone, Debug, Default)] 159 #[repr(packed)] 160 struct VirtioIommuReqDetach { 161 domain: u32, 162 endpoint: u32, 163 _reserved: [u8; 8], 164 } 165 166 /// Virtio IOMMU request MAP flags 167 #[allow(unused)] 168 const VIRTIO_IOMMU_MAP_F_READ: u32 = 1; 169 #[allow(unused)] 170 const VIRTIO_IOMMU_MAP_F_WRITE: u32 = 1 << 1; 171 #[allow(unused)] 172 const VIRTIO_IOMMU_MAP_F_MMIO: u32 = 1 << 2; 173 #[allow(unused)] 174 const VIRTIO_IOMMU_MAP_F_MASK: u32 = 175 VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE | VIRTIO_IOMMU_MAP_F_MMIO; 176 177 /// MAP request 178 #[derive(Copy, Clone, Debug, Default)] 179 #[repr(packed)] 180 struct VirtioIommuReqMap { 181 domain: u32, 182 virt_start: u64, 183 virt_end: u64, 184 phys_start: u64, 185 _flags: u32, 186 } 187 188 /// UNMAP request 189 #[derive(Copy, Clone, Debug, Default)] 190 #[repr(packed)] 191 struct VirtioIommuReqUnmap { 192 domain: u32, 193 virt_start: u64, 194 virt_end: u64, 195 _reserved: [u8; 4], 196 } 197 198 /// Virtio IOMMU request PROBE types 199 #[allow(unused)] 200 const VIRTIO_IOMMU_PROBE_T_NONE: u16 = 0; 201 const VIRTIO_IOMMU_PROBE_T_RESV_MEM: u16 = 1; 202 #[allow(unused)] 203 const VIRTIO_IOMMU_PROBE_T_MASK: u16 = 0xfff; 204 205 /// PROBE request 206 #[derive(Copy, Clone, Debug, Default)] 207 #[repr(packed)] 208 #[allow(dead_code)] 209 struct VirtioIommuReqProbe { 210 endpoint: u32, 211 _reserved: [u64; 8], 212 } 213 214 #[derive(Copy, Clone, Debug, Default)] 215 #[repr(packed)] 216 #[allow(dead_code)] 217 struct VirtioIommuProbeProperty { 218 type_: u16, 219 length: u16, 220 } 221 222 /// Virtio IOMMU request PROBE property RESV_MEM subtypes 223 #[allow(unused)] 224 const VIRTIO_IOMMU_RESV_MEM_T_RESERVED: u8 = 0; 225 const VIRTIO_IOMMU_RESV_MEM_T_MSI: u8 = 1; 226 227 #[derive(Copy, Clone, Debug, Default)] 228 #[repr(packed)] 229 #[allow(dead_code)] 230 struct VirtioIommuProbeResvMem { 231 subtype: u8, 232 _reserved: [u8; 3], 233 start: u64, 234 end: u64, 235 } 236 237 /// Virtio IOMMU fault flags 238 #[allow(unused)] 239 const VIRTIO_IOMMU_FAULT_F_READ: u32 = 1; 240 #[allow(unused)] 241 const VIRTIO_IOMMU_FAULT_F_WRITE: u32 = 1 << 1; 242 #[allow(unused)] 243 const VIRTIO_IOMMU_FAULT_F_EXEC: u32 = 1 << 2; 244 #[allow(unused)] 245 const VIRTIO_IOMMU_FAULT_F_ADDRESS: u32 = 1 << 8; 246 247 /// Virtio IOMMU fault reasons 248 #[allow(unused)] 249 const VIRTIO_IOMMU_FAULT_R_UNKNOWN: u32 = 0; 250 #[allow(unused)] 251 const VIRTIO_IOMMU_FAULT_R_DOMAIN: u32 = 1; 252 #[allow(unused)] 253 const VIRTIO_IOMMU_FAULT_R_MAPPING: u32 = 2; 254 255 /// Fault reporting through eventq 256 #[allow(unused)] 257 #[derive(Copy, Clone, Debug, Default)] 258 #[repr(packed)] 259 struct VirtioIommuFault { 260 reason: u8, 261 reserved: [u8; 3], 262 flags: u32, 263 endpoint: u32, 264 reserved2: [u8; 4], 265 address: u64, 266 } 267 268 // SAFETY: data structure only contain integers and have no implicit padding 269 unsafe impl ByteValued for VirtioIommuRange32 {} 270 // SAFETY: data structure only contain integers and have no implicit padding 271 unsafe impl ByteValued for VirtioIommuRange64 {} 272 // SAFETY: data structure only contain integers and have no implicit padding 273 unsafe impl ByteValued for VirtioIommuConfig {} 274 // SAFETY: data structure only contain integers and have no implicit padding 275 unsafe impl ByteValued for VirtioIommuReqHead {} 276 // SAFETY: data structure only contain integers and have no implicit padding 277 unsafe impl ByteValued for VirtioIommuReqTail {} 278 // SAFETY: data structure only contain integers and have no implicit padding 279 unsafe impl ByteValued for VirtioIommuReqAttach {} 280 // SAFETY: data structure only contain integers and have no implicit padding 281 unsafe impl ByteValued for VirtioIommuReqDetach {} 282 // SAFETY: data structure only contain integers and have no implicit padding 283 unsafe impl ByteValued for VirtioIommuReqMap {} 284 // SAFETY: data structure only contain integers and have no implicit padding 285 unsafe impl ByteValued for VirtioIommuReqUnmap {} 286 // SAFETY: data structure only contain integers and have no implicit padding 287 unsafe impl ByteValued for VirtioIommuReqProbe {} 288 // SAFETY: data structure only contain integers and have no implicit padding 289 unsafe impl ByteValued for VirtioIommuProbeProperty {} 290 // SAFETY: data structure only contain integers and have no implicit padding 291 unsafe impl ByteValued for VirtioIommuProbeResvMem {} 292 // SAFETY: data structure only contain integers and have no implicit padding 293 unsafe impl ByteValued for VirtioIommuFault {} 294 295 #[derive(Error, Debug)] 296 enum Error { 297 #[error("Guest gave us bad memory addresses: {0}")] 298 GuestMemory(GuestMemoryError), 299 #[error("Guest gave us a write only descriptor that protocol says to read from")] 300 UnexpectedWriteOnlyDescriptor, 301 #[error("Guest gave us a read only descriptor that protocol says to write to")] 302 UnexpectedReadOnlyDescriptor, 303 #[error("Guest gave us too few descriptors in a descriptor chain")] 304 DescriptorChainTooShort, 305 #[error("Guest gave us a buffer that was too short to use")] 306 BufferLengthTooSmall, 307 #[error("Guest sent us invalid request")] 308 InvalidRequest, 309 #[error("Guest sent us invalid ATTACH request")] 310 InvalidAttachRequest, 311 #[error("Guest sent us invalid DETACH request")] 312 InvalidDetachRequest, 313 #[error("Guest sent us invalid MAP request")] 314 InvalidMapRequest, 315 #[error("Invalid to map because the domain is in bypass mode")] 316 InvalidMapRequestBypassDomain, 317 #[error("Invalid to map because the domain is missing")] 318 InvalidMapRequestMissingDomain, 319 #[error("Guest sent us invalid UNMAP request")] 320 InvalidUnmapRequest, 321 #[error("Invalid to unmap because the domain is in bypass mode")] 322 InvalidUnmapRequestBypassDomain, 323 #[error("Invalid to unmap because the domain is missing")] 324 InvalidUnmapRequestMissingDomain, 325 #[error("Guest sent us invalid PROBE request")] 326 InvalidProbeRequest, 327 #[error("Failed to performing external mapping: {0}")] 328 ExternalMapping(io::Error), 329 #[error("Failed to performing external unmapping: {0}")] 330 ExternalUnmapping(io::Error), 331 #[error("Failed adding used index: {0}")] 332 QueueAddUsed(virtio_queue::Error), 333 } 334 335 struct Request {} 336 337 impl Request { 338 // Parse the available vring buffer. Based on the hashmap table of external 339 // mappings required from various devices such as VFIO or vhost-user ones, 340 // this function might update the hashmap table of external mappings per 341 // domain. 342 // Basically, the VMM knows about the device_id <=> mapping relationship 343 // before running the VM, but at runtime, a new domain <=> mapping hashmap 344 // is created based on the information provided from the guest driver for 345 // virtio-iommu (giving the link device_id <=> domain). 346 fn parse( 347 desc_chain: &mut DescriptorChain<GuestMemoryLoadGuard<GuestMemoryMmap>>, 348 mapping: &Arc<IommuMapping>, 349 ext_mapping: &BTreeMap<u32, Arc<dyn ExternalDmaMapping>>, 350 msi_iova_space: (u64, u64), 351 ) -> result::Result<usize, Error> { 352 let desc = desc_chain 353 .next() 354 .ok_or(Error::DescriptorChainTooShort) 355 .inspect_err(|_| { 356 error!("Missing head descriptor"); 357 })?; 358 359 // The descriptor contains the request type which MUST be readable. 360 if desc.is_write_only() { 361 return Err(Error::UnexpectedWriteOnlyDescriptor); 362 } 363 364 if (desc.len() as usize) < size_of::<VirtioIommuReqHead>() { 365 return Err(Error::InvalidRequest); 366 } 367 368 let req_head: VirtioIommuReqHead = desc_chain 369 .memory() 370 .read_obj(desc.addr()) 371 .map_err(Error::GuestMemory)?; 372 let req_offset = size_of::<VirtioIommuReqHead>(); 373 let desc_size_left = (desc.len() as usize) - req_offset; 374 let req_addr = if let Some(addr) = desc.addr().checked_add(req_offset as u64) { 375 addr 376 } else { 377 return Err(Error::InvalidRequest); 378 }; 379 380 let (msi_iova_start, msi_iova_end) = msi_iova_space; 381 382 // Create the reply 383 let mut reply: Vec<u8> = Vec::new(); 384 let mut status = VIRTIO_IOMMU_S_OK; 385 let mut hdr_len = 0; 386 387 let result = (|| { 388 match req_head.type_ { 389 VIRTIO_IOMMU_T_ATTACH => { 390 if desc_size_left != size_of::<VirtioIommuReqAttach>() { 391 status = VIRTIO_IOMMU_S_INVAL; 392 return Err(Error::InvalidAttachRequest); 393 } 394 395 let req: VirtioIommuReqAttach = desc_chain 396 .memory() 397 .read_obj(req_addr as GuestAddress) 398 .map_err(Error::GuestMemory)?; 399 debug!("Attach request {:?}", req); 400 401 // Copy the value to use it as a proper reference. 402 let domain_id = req.domain; 403 let endpoint = req.endpoint; 404 let bypass = 405 (req.flags & VIRTIO_IOMMU_ATTACH_F_BYPASS) == VIRTIO_IOMMU_ATTACH_F_BYPASS; 406 407 let mut old_domain_id = domain_id; 408 if let Some(&id) = mapping.endpoints.read().unwrap().get(&endpoint) { 409 old_domain_id = id; 410 } 411 412 if old_domain_id != domain_id { 413 detach_endpoint_from_domain(endpoint, old_domain_id, mapping, ext_mapping)?; 414 } 415 416 // Add endpoint associated with specific domain 417 mapping 418 .endpoints 419 .write() 420 .unwrap() 421 .insert(endpoint, domain_id); 422 423 // If any other mappings exist in the domain for other containers, 424 // make sure to issue these mappings for the new endpoint/container 425 if let Some(domain_mappings) = &mapping.domains.read().unwrap().get(&domain_id) 426 { 427 if let Some(ext_map) = ext_mapping.get(&endpoint) { 428 for (virt_start, addr_map) in &domain_mappings.mappings { 429 ext_map 430 .map(*virt_start, addr_map.gpa, addr_map.size) 431 .map_err(Error::ExternalUnmapping)?; 432 } 433 } 434 } 435 436 // Add new domain with no mapping if the entry didn't exist yet 437 let mut domains = mapping.domains.write().unwrap(); 438 let domain = Domain { 439 mappings: BTreeMap::new(), 440 bypass, 441 }; 442 domains.entry(domain_id).or_insert_with(|| domain); 443 } 444 VIRTIO_IOMMU_T_DETACH => { 445 if desc_size_left != size_of::<VirtioIommuReqDetach>() { 446 status = VIRTIO_IOMMU_S_INVAL; 447 return Err(Error::InvalidDetachRequest); 448 } 449 450 let req: VirtioIommuReqDetach = desc_chain 451 .memory() 452 .read_obj(req_addr as GuestAddress) 453 .map_err(Error::GuestMemory)?; 454 debug!("Detach request {:?}", req); 455 456 // Copy the value to use it as a proper reference. 457 let domain_id = req.domain; 458 let endpoint = req.endpoint; 459 460 // Remove endpoint associated with specific domain 461 detach_endpoint_from_domain(endpoint, domain_id, mapping, ext_mapping)?; 462 } 463 VIRTIO_IOMMU_T_MAP => { 464 if desc_size_left != size_of::<VirtioIommuReqMap>() { 465 status = VIRTIO_IOMMU_S_INVAL; 466 return Err(Error::InvalidMapRequest); 467 } 468 469 let req: VirtioIommuReqMap = desc_chain 470 .memory() 471 .read_obj(req_addr as GuestAddress) 472 .map_err(Error::GuestMemory)?; 473 debug!("Map request {:?}", req); 474 475 // Copy the value to use it as a proper reference. 476 let domain_id = req.domain; 477 478 if let Some(domain) = mapping.domains.read().unwrap().get(&domain_id) { 479 if domain.bypass { 480 status = VIRTIO_IOMMU_S_INVAL; 481 return Err(Error::InvalidMapRequestBypassDomain); 482 } 483 } else { 484 status = VIRTIO_IOMMU_S_INVAL; 485 return Err(Error::InvalidMapRequestMissingDomain); 486 } 487 488 // Find the list of endpoints attached to the given domain. 489 let endpoints: Vec<u32> = mapping 490 .endpoints 491 .write() 492 .unwrap() 493 .iter() 494 .filter(|(_, &d)| d == domain_id) 495 .map(|(&e, _)| e) 496 .collect(); 497 498 // For viommu all endpoints receive their own VFIO container, as a result 499 // Each endpoint within the domain needs to be separately mapped, as the 500 // mapping is done on a per-container level, not a per-domain level 501 for endpoint in endpoints { 502 if let Some(ext_map) = ext_mapping.get(&endpoint) { 503 let size = req.virt_end - req.virt_start + 1; 504 ext_map 505 .map(req.virt_start, req.phys_start, size) 506 .map_err(Error::ExternalMapping)?; 507 } 508 } 509 510 // Add new mapping associated with the domain 511 mapping 512 .domains 513 .write() 514 .unwrap() 515 .get_mut(&domain_id) 516 .unwrap() 517 .mappings 518 .insert( 519 req.virt_start, 520 Mapping { 521 gpa: req.phys_start, 522 size: req.virt_end - req.virt_start + 1, 523 }, 524 ); 525 } 526 VIRTIO_IOMMU_T_UNMAP => { 527 if desc_size_left != size_of::<VirtioIommuReqUnmap>() { 528 status = VIRTIO_IOMMU_S_INVAL; 529 return Err(Error::InvalidUnmapRequest); 530 } 531 532 let req: VirtioIommuReqUnmap = desc_chain 533 .memory() 534 .read_obj(req_addr as GuestAddress) 535 .map_err(Error::GuestMemory)?; 536 debug!("Unmap request {:?}", req); 537 538 // Copy the value to use it as a proper reference. 539 let domain_id = req.domain; 540 let virt_start = req.virt_start; 541 542 if let Some(domain) = mapping.domains.read().unwrap().get(&domain_id) { 543 if domain.bypass { 544 status = VIRTIO_IOMMU_S_INVAL; 545 return Err(Error::InvalidUnmapRequestBypassDomain); 546 } 547 } else { 548 status = VIRTIO_IOMMU_S_INVAL; 549 return Err(Error::InvalidUnmapRequestMissingDomain); 550 } 551 552 // Find the list of endpoints attached to the given domain. 553 let endpoints: Vec<u32> = mapping 554 .endpoints 555 .write() 556 .unwrap() 557 .iter() 558 .filter(|(_, &d)| d == domain_id) 559 .map(|(&e, _)| e) 560 .collect(); 561 562 // Trigger external unmapping if necessary. 563 for endpoint in endpoints { 564 if let Some(ext_map) = ext_mapping.get(&endpoint) { 565 let size = req.virt_end - virt_start + 1; 566 ext_map 567 .unmap(virt_start, size) 568 .map_err(Error::ExternalUnmapping)?; 569 } 570 } 571 572 // Remove all mappings associated with the domain within the requested range 573 mapping 574 .domains 575 .write() 576 .unwrap() 577 .get_mut(&domain_id) 578 .unwrap() 579 .mappings 580 .retain(|&x, _| (x < req.virt_start || x > req.virt_end)); 581 } 582 VIRTIO_IOMMU_T_PROBE => { 583 if desc_size_left != size_of::<VirtioIommuReqProbe>() { 584 status = VIRTIO_IOMMU_S_INVAL; 585 return Err(Error::InvalidProbeRequest); 586 } 587 588 let req: VirtioIommuReqProbe = desc_chain 589 .memory() 590 .read_obj(req_addr as GuestAddress) 591 .map_err(Error::GuestMemory)?; 592 debug!("Probe request {:?}", req); 593 594 let probe_prop = VirtioIommuProbeProperty { 595 type_: VIRTIO_IOMMU_PROBE_T_RESV_MEM, 596 length: size_of::<VirtioIommuProbeResvMem>() as u16, 597 }; 598 reply.extend_from_slice(probe_prop.as_slice()); 599 600 let resv_mem = VirtioIommuProbeResvMem { 601 subtype: VIRTIO_IOMMU_RESV_MEM_T_MSI, 602 start: msi_iova_start, 603 end: msi_iova_end, 604 ..Default::default() 605 }; 606 reply.extend_from_slice(resv_mem.as_slice()); 607 608 hdr_len = PROBE_PROP_SIZE; 609 } 610 _ => { 611 status = VIRTIO_IOMMU_S_INVAL; 612 return Err(Error::InvalidRequest); 613 } 614 } 615 Ok(()) 616 })(); 617 618 let status_desc = desc_chain.next().ok_or(Error::DescriptorChainTooShort)?; 619 620 // The status MUST always be writable 621 if !status_desc.is_write_only() { 622 return Err(Error::UnexpectedReadOnlyDescriptor); 623 } 624 625 if status_desc.len() < hdr_len + size_of::<VirtioIommuReqTail>() as u32 { 626 return Err(Error::BufferLengthTooSmall); 627 } 628 629 let tail = VirtioIommuReqTail { 630 status, 631 ..Default::default() 632 }; 633 reply.extend_from_slice(tail.as_slice()); 634 635 // Make sure we return the result of the request to the guest before 636 // we return a potential error internally. 637 desc_chain 638 .memory() 639 .write_slice(reply.as_slice(), status_desc.addr()) 640 .map_err(Error::GuestMemory)?; 641 642 // Return the error if the result was not Ok(). 643 result?; 644 645 Ok((hdr_len as usize) + size_of::<VirtioIommuReqTail>()) 646 } 647 } 648 649 fn detach_endpoint_from_domain( 650 endpoint: u32, 651 domain_id: u32, 652 mapping: &Arc<IommuMapping>, 653 ext_mapping: &BTreeMap<u32, Arc<dyn ExternalDmaMapping>>, 654 ) -> result::Result<(), Error> { 655 // Remove endpoint associated with specific domain 656 mapping.endpoints.write().unwrap().remove(&endpoint); 657 658 // Trigger external unmapping for the endpoint if necessary. 659 if let Some(domain_mappings) = &mapping.domains.read().unwrap().get(&domain_id) { 660 if let Some(ext_map) = ext_mapping.get(&endpoint) { 661 for (virt_start, addr_map) in &domain_mappings.mappings { 662 ext_map 663 .unmap(*virt_start, addr_map.size) 664 .map_err(Error::ExternalUnmapping)?; 665 } 666 } 667 } 668 669 if mapping 670 .endpoints 671 .write() 672 .unwrap() 673 .iter() 674 .filter(|(_, &d)| d == domain_id) 675 .count() 676 == 0 677 { 678 mapping.domains.write().unwrap().remove(&domain_id); 679 } 680 681 Ok(()) 682 } 683 684 struct IommuEpollHandler { 685 mem: GuestMemoryAtomic<GuestMemoryMmap>, 686 request_queue: Queue, 687 _event_queue: Queue, 688 interrupt_cb: Arc<dyn VirtioInterrupt>, 689 request_queue_evt: EventFd, 690 _event_queue_evt: EventFd, 691 kill_evt: EventFd, 692 pause_evt: EventFd, 693 mapping: Arc<IommuMapping>, 694 ext_mapping: Arc<Mutex<BTreeMap<u32, Arc<dyn ExternalDmaMapping>>>>, 695 msi_iova_space: (u64, u64), 696 } 697 698 impl IommuEpollHandler { 699 fn request_queue(&mut self) -> Result<bool, Error> { 700 let mut used_descs = false; 701 while let Some(mut desc_chain) = self.request_queue.pop_descriptor_chain(self.mem.memory()) 702 { 703 let len = Request::parse( 704 &mut desc_chain, 705 &self.mapping, 706 &self.ext_mapping.lock().unwrap(), 707 self.msi_iova_space, 708 )?; 709 710 self.request_queue 711 .add_used(desc_chain.memory(), desc_chain.head_index(), len as u32) 712 .map_err(Error::QueueAddUsed)?; 713 714 used_descs = true; 715 } 716 717 Ok(used_descs) 718 } 719 720 fn signal_used_queue(&self, queue_index: u16) -> result::Result<(), DeviceError> { 721 self.interrupt_cb 722 .trigger(VirtioInterruptType::Queue(queue_index)) 723 .map_err(|e| { 724 error!("Failed to signal used queue: {:?}", e); 725 DeviceError::FailedSignalingUsedQueue(e) 726 }) 727 } 728 729 fn run( 730 &mut self, 731 paused: Arc<AtomicBool>, 732 paused_sync: Arc<Barrier>, 733 ) -> result::Result<(), EpollHelperError> { 734 let mut helper = EpollHelper::new(&self.kill_evt, &self.pause_evt)?; 735 helper.add_event(self.request_queue_evt.as_raw_fd(), REQUEST_Q_EVENT)?; 736 helper.run(paused, paused_sync, self)?; 737 738 Ok(()) 739 } 740 } 741 742 impl EpollHelperHandler for IommuEpollHandler { 743 fn handle_event( 744 &mut self, 745 _helper: &mut EpollHelper, 746 event: &epoll::Event, 747 ) -> result::Result<(), EpollHelperError> { 748 let ev_type = event.data as u16; 749 match ev_type { 750 REQUEST_Q_EVENT => { 751 self.request_queue_evt.read().map_err(|e| { 752 EpollHelperError::HandleEvent(anyhow!("Failed to get queue event: {:?}", e)) 753 })?; 754 755 let needs_notification = self.request_queue().map_err(|e| { 756 EpollHelperError::HandleEvent(anyhow!( 757 "Failed to process request queue : {:?}", 758 e 759 )) 760 })?; 761 if needs_notification { 762 self.signal_used_queue(0).map_err(|e| { 763 EpollHelperError::HandleEvent(anyhow!( 764 "Failed to signal used queue: {:?}", 765 e 766 )) 767 })?; 768 } 769 } 770 _ => { 771 return Err(EpollHelperError::HandleEvent(anyhow!( 772 "Unexpected event: {}", 773 ev_type 774 ))); 775 } 776 } 777 Ok(()) 778 } 779 } 780 781 #[derive(Clone, Copy, Debug, Serialize, Deserialize)] 782 struct Mapping { 783 gpa: u64, 784 size: u64, 785 } 786 787 #[derive(Clone, Debug)] 788 struct Domain { 789 mappings: BTreeMap<u64, Mapping>, 790 bypass: bool, 791 } 792 793 #[derive(Debug)] 794 pub struct IommuMapping { 795 // Domain related to an endpoint. 796 endpoints: Arc<RwLock<BTreeMap<u32, u32>>>, 797 // Information related to each domain. 798 domains: Arc<RwLock<BTreeMap<u32, Domain>>>, 799 // Global flag indicating if endpoints that are not attached to any domain 800 // are in bypass mode. 801 bypass: AtomicBool, 802 } 803 804 impl DmaRemapping for IommuMapping { 805 fn translate_gva(&self, id: u32, addr: u64) -> std::result::Result<u64, std::io::Error> { 806 debug!("Translate GVA addr 0x{:x}", addr); 807 if let Some(domain_id) = self.endpoints.read().unwrap().get(&id) { 808 if let Some(domain) = self.domains.read().unwrap().get(domain_id) { 809 // Directly return identity mapping in case the domain is in 810 // bypass mode. 811 if domain.bypass { 812 return Ok(addr); 813 } 814 815 let range_start = if VIRTIO_IOMMU_PAGE_SIZE_MASK > addr { 816 0 817 } else { 818 addr - VIRTIO_IOMMU_PAGE_SIZE_MASK 819 }; 820 for (&key, &value) in domain 821 .mappings 822 .range((Included(&range_start), Included(&addr))) 823 { 824 if addr >= key && addr < key + value.size { 825 let new_addr = addr - key + value.gpa; 826 debug!("Into GPA addr 0x{:x}", new_addr); 827 return Ok(new_addr); 828 } 829 } 830 } 831 } else if self.bypass.load(Ordering::Acquire) { 832 return Ok(addr); 833 } 834 835 Err(io::Error::new( 836 io::ErrorKind::Other, 837 format!("failed to translate GVA addr 0x{addr:x}"), 838 )) 839 } 840 841 fn translate_gpa(&self, id: u32, addr: u64) -> std::result::Result<u64, std::io::Error> { 842 debug!("Translate GPA addr 0x{:x}", addr); 843 if let Some(domain_id) = self.endpoints.read().unwrap().get(&id) { 844 if let Some(domain) = self.domains.read().unwrap().get(domain_id) { 845 // Directly return identity mapping in case the domain is in 846 // bypass mode. 847 if domain.bypass { 848 return Ok(addr); 849 } 850 851 for (&key, &value) in domain.mappings.iter() { 852 if addr >= value.gpa && addr < value.gpa + value.size { 853 let new_addr = addr - value.gpa + key; 854 debug!("Into GVA addr 0x{:x}", new_addr); 855 return Ok(new_addr); 856 } 857 } 858 } 859 } else if self.bypass.load(Ordering::Acquire) { 860 return Ok(addr); 861 } 862 863 Err(io::Error::new( 864 io::ErrorKind::Other, 865 format!("failed to translate GPA addr 0x{addr:x}"), 866 )) 867 } 868 } 869 870 #[derive(Debug)] 871 pub struct AccessPlatformMapping { 872 id: u32, 873 mapping: Arc<IommuMapping>, 874 } 875 876 impl AccessPlatformMapping { 877 pub fn new(id: u32, mapping: Arc<IommuMapping>) -> Self { 878 AccessPlatformMapping { id, mapping } 879 } 880 } 881 882 impl AccessPlatform for AccessPlatformMapping { 883 fn translate_gva(&self, base: u64, _size: u64) -> std::result::Result<u64, std::io::Error> { 884 self.mapping.translate_gva(self.id, base) 885 } 886 fn translate_gpa(&self, base: u64, _size: u64) -> std::result::Result<u64, std::io::Error> { 887 self.mapping.translate_gpa(self.id, base) 888 } 889 } 890 891 pub struct Iommu { 892 common: VirtioCommon, 893 id: String, 894 config: VirtioIommuConfig, 895 mapping: Arc<IommuMapping>, 896 ext_mapping: Arc<Mutex<BTreeMap<u32, Arc<dyn ExternalDmaMapping>>>>, 897 seccomp_action: SeccompAction, 898 exit_evt: EventFd, 899 msi_iova_space: (u64, u64), 900 } 901 902 type EndpointsState = Vec<(u32, u32)>; 903 type DomainsState = Vec<(u32, (Vec<(u64, Mapping)>, bool))>; 904 905 #[derive(Serialize, Deserialize)] 906 pub struct IommuState { 907 avail_features: u64, 908 acked_features: u64, 909 endpoints: EndpointsState, 910 domains: DomainsState, 911 } 912 913 impl Iommu { 914 pub fn new( 915 id: String, 916 seccomp_action: SeccompAction, 917 exit_evt: EventFd, 918 msi_iova_space: (u64, u64), 919 state: Option<IommuState>, 920 ) -> io::Result<(Self, Arc<IommuMapping>)> { 921 let (avail_features, acked_features, endpoints, domains, paused) = 922 if let Some(state) = state { 923 info!("Restoring virtio-iommu {}", id); 924 ( 925 state.avail_features, 926 state.acked_features, 927 state.endpoints.into_iter().collect(), 928 state 929 .domains 930 .into_iter() 931 .map(|(k, v)| { 932 ( 933 k, 934 Domain { 935 mappings: v.0.into_iter().collect(), 936 bypass: v.1, 937 }, 938 ) 939 }) 940 .collect(), 941 true, 942 ) 943 } else { 944 let avail_features = 1u64 << VIRTIO_F_VERSION_1 945 | 1u64 << VIRTIO_IOMMU_F_MAP_UNMAP 946 | 1u64 << VIRTIO_IOMMU_F_PROBE 947 | 1u64 << VIRTIO_IOMMU_F_BYPASS_CONFIG; 948 949 (avail_features, 0, BTreeMap::new(), BTreeMap::new(), false) 950 }; 951 952 let config = VirtioIommuConfig { 953 page_size_mask: VIRTIO_IOMMU_PAGE_SIZE_MASK, 954 probe_size: PROBE_PROP_SIZE, 955 ..Default::default() 956 }; 957 958 let mapping = Arc::new(IommuMapping { 959 endpoints: Arc::new(RwLock::new(endpoints)), 960 domains: Arc::new(RwLock::new(domains)), 961 bypass: AtomicBool::new(true), 962 }); 963 964 Ok(( 965 Iommu { 966 id, 967 common: VirtioCommon { 968 device_type: VirtioDeviceType::Iommu as u32, 969 queue_sizes: QUEUE_SIZES.to_vec(), 970 avail_features, 971 acked_features, 972 paused_sync: Some(Arc::new(Barrier::new(2))), 973 min_queues: NUM_QUEUES as u16, 974 paused: Arc::new(AtomicBool::new(paused)), 975 ..Default::default() 976 }, 977 config, 978 mapping: mapping.clone(), 979 ext_mapping: Arc::new(Mutex::new(BTreeMap::new())), 980 seccomp_action, 981 exit_evt, 982 msi_iova_space, 983 }, 984 mapping, 985 )) 986 } 987 988 fn state(&self) -> IommuState { 989 IommuState { 990 avail_features: self.common.avail_features, 991 acked_features: self.common.acked_features, 992 endpoints: self 993 .mapping 994 .endpoints 995 .read() 996 .unwrap() 997 .clone() 998 .into_iter() 999 .collect(), 1000 domains: self 1001 .mapping 1002 .domains 1003 .read() 1004 .unwrap() 1005 .clone() 1006 .into_iter() 1007 .map(|(k, v)| (k, (v.mappings.into_iter().collect(), v.bypass))) 1008 .collect(), 1009 } 1010 } 1011 1012 fn update_bypass(&mut self) { 1013 // Use bypass from config if VIRTIO_IOMMU_F_BYPASS_CONFIG has been negotiated 1014 if !self 1015 .common 1016 .feature_acked(VIRTIO_IOMMU_F_BYPASS_CONFIG.into()) 1017 { 1018 return; 1019 } 1020 1021 let bypass = self.config.bypass == 1; 1022 info!("Updating bypass mode to {}", bypass); 1023 self.mapping.bypass.store(bypass, Ordering::Release); 1024 } 1025 1026 pub fn add_external_mapping(&mut self, device_id: u32, mapping: Arc<dyn ExternalDmaMapping>) { 1027 self.ext_mapping.lock().unwrap().insert(device_id, mapping); 1028 } 1029 1030 #[cfg(fuzzing)] 1031 pub fn wait_for_epoll_threads(&mut self) { 1032 self.common.wait_for_epoll_threads(); 1033 } 1034 } 1035 1036 impl Drop for Iommu { 1037 fn drop(&mut self) { 1038 if let Some(kill_evt) = self.common.kill_evt.take() { 1039 // Ignore the result because there is nothing we can do about it. 1040 let _ = kill_evt.write(1); 1041 } 1042 self.common.wait_for_epoll_threads(); 1043 } 1044 } 1045 1046 impl VirtioDevice for Iommu { 1047 fn device_type(&self) -> u32 { 1048 self.common.device_type 1049 } 1050 1051 fn queue_max_sizes(&self) -> &[u16] { 1052 &self.common.queue_sizes 1053 } 1054 1055 fn features(&self) -> u64 { 1056 self.common.avail_features 1057 } 1058 1059 fn ack_features(&mut self, value: u64) { 1060 self.common.ack_features(value) 1061 } 1062 1063 fn read_config(&self, offset: u64, data: &mut [u8]) { 1064 self.read_config_from_slice(self.config.as_slice(), offset, data); 1065 } 1066 1067 fn write_config(&mut self, offset: u64, data: &[u8]) { 1068 // The "bypass" field is the only mutable field 1069 let bypass_offset = 1070 (&self.config.bypass as *const _ as u64) - (&self.config as *const _ as u64); 1071 if offset != bypass_offset || data.len() != std::mem::size_of_val(&self.config.bypass) { 1072 error!( 1073 "Attempt to write to read-only field: offset {:x} length {}", 1074 offset, 1075 data.len() 1076 ); 1077 return; 1078 } 1079 1080 self.config.bypass = data[0]; 1081 1082 self.update_bypass(); 1083 } 1084 1085 fn activate( 1086 &mut self, 1087 mem: GuestMemoryAtomic<GuestMemoryMmap>, 1088 interrupt_cb: Arc<dyn VirtioInterrupt>, 1089 mut queues: Vec<(usize, Queue, EventFd)>, 1090 ) -> ActivateResult { 1091 self.common.activate(&queues, &interrupt_cb)?; 1092 let (kill_evt, pause_evt) = self.common.dup_eventfds(); 1093 1094 let (_, request_queue, request_queue_evt) = queues.remove(0); 1095 let (_, _event_queue, _event_queue_evt) = queues.remove(0); 1096 1097 let mut handler = IommuEpollHandler { 1098 mem, 1099 request_queue, 1100 _event_queue, 1101 interrupt_cb, 1102 request_queue_evt, 1103 _event_queue_evt, 1104 kill_evt, 1105 pause_evt, 1106 mapping: self.mapping.clone(), 1107 ext_mapping: self.ext_mapping.clone(), 1108 msi_iova_space: self.msi_iova_space, 1109 }; 1110 1111 let paused = self.common.paused.clone(); 1112 let paused_sync = self.common.paused_sync.clone(); 1113 let mut epoll_threads = Vec::new(); 1114 spawn_virtio_thread( 1115 &self.id, 1116 &self.seccomp_action, 1117 Thread::VirtioIommu, 1118 &mut epoll_threads, 1119 &self.exit_evt, 1120 move || handler.run(paused, paused_sync.unwrap()), 1121 )?; 1122 1123 self.common.epoll_threads = Some(epoll_threads); 1124 1125 event!("virtio-device", "activated", "id", &self.id); 1126 Ok(()) 1127 } 1128 1129 fn reset(&mut self) -> Option<Arc<dyn VirtioInterrupt>> { 1130 let result = self.common.reset(); 1131 event!("virtio-device", "reset", "id", &self.id); 1132 result 1133 } 1134 } 1135 1136 impl Pausable for Iommu { 1137 fn pause(&mut self) -> result::Result<(), MigratableError> { 1138 self.common.pause() 1139 } 1140 1141 fn resume(&mut self) -> result::Result<(), MigratableError> { 1142 self.common.resume() 1143 } 1144 } 1145 1146 impl Snapshottable for Iommu { 1147 fn id(&self) -> String { 1148 self.id.clone() 1149 } 1150 1151 fn snapshot(&mut self) -> std::result::Result<Snapshot, MigratableError> { 1152 Snapshot::new_from_state(&self.state()) 1153 } 1154 } 1155 impl Transportable for Iommu {} 1156 impl Migratable for Iommu {} 1157