1 // Copyright © 2019 Intel Corporation 2 // 3 // SPDX-License-Identifier: Apache-2.0 AND BSD-3-Clause 4 5 use std::collections::BTreeMap; 6 use std::io; 7 use std::mem::size_of; 8 use std::ops::Bound::Included; 9 use std::os::unix::io::AsRawFd; 10 use std::result; 11 use std::sync::atomic::{AtomicBool, Ordering}; 12 use std::sync::{Arc, Barrier, Mutex, RwLock}; 13 14 use anyhow::anyhow; 15 use seccompiler::SeccompAction; 16 use serde::{Deserialize, Serialize}; 17 use thiserror::Error; 18 use virtio_queue::{DescriptorChain, Queue, QueueT}; 19 use vm_device::dma_mapping::ExternalDmaMapping; 20 use vm_memory::{ 21 Address, ByteValued, Bytes, GuestAddress, GuestAddressSpace, GuestMemoryAtomic, 22 GuestMemoryError, GuestMemoryLoadGuard, 23 }; 24 use vm_migration::{Migratable, MigratableError, Pausable, Snapshot, Snapshottable, Transportable}; 25 use vm_virtio::AccessPlatform; 26 use vmm_sys_util::eventfd::EventFd; 27 28 use super::Error as DeviceError; 29 use super::{ 30 ActivateResult, EpollHelper, EpollHelperError, EpollHelperHandler, VirtioCommon, VirtioDevice, 31 VirtioDeviceType, EPOLL_HELPER_EVENT_LAST, VIRTIO_F_VERSION_1, 32 }; 33 use crate::seccomp_filters::Thread; 34 use crate::thread_helper::spawn_virtio_thread; 35 use crate::GuestMemoryMmap; 36 use crate::{DmaRemapping, VirtioInterrupt, VirtioInterruptType}; 37 38 /// Queues sizes 39 const QUEUE_SIZE: u16 = 256; 40 const NUM_QUEUES: usize = 2; 41 const QUEUE_SIZES: &[u16] = &[QUEUE_SIZE; NUM_QUEUES]; 42 43 /// New descriptors are pending on the request queue. 44 /// "requestq" is meant to be used anytime an action is required to be 45 /// performed on behalf of the guest driver. 46 const REQUEST_Q_EVENT: u16 = EPOLL_HELPER_EVENT_LAST + 1; 47 /// New descriptors are pending on the event queue. 48 /// "eventq" lets the device report any fault or other asynchronous event to 49 /// the guest driver. 50 const _EVENT_Q_EVENT: u16 = EPOLL_HELPER_EVENT_LAST + 2; 51 52 /// PROBE properties size. 53 /// This is the minimal size to provide at least one RESV_MEM property. 54 /// Because virtio-iommu expects one MSI reserved region, we must provide it, 55 /// otherwise the driver in the guest will define a predefined one between 56 /// 0x8000000 and 0x80FFFFF, which is only relevant for ARM architecture, but 57 /// will conflict with x86. 58 const PROBE_PROP_SIZE: u32 = 59 (size_of::<VirtioIommuProbeProperty>() + size_of::<VirtioIommuProbeResvMem>()) as u32; 60 61 /// Virtio IOMMU features 62 #[allow(unused)] 63 const VIRTIO_IOMMU_F_INPUT_RANGE: u32 = 0; 64 #[allow(unused)] 65 const VIRTIO_IOMMU_F_DOMAIN_RANGE: u32 = 1; 66 #[allow(unused)] 67 const VIRTIO_IOMMU_F_MAP_UNMAP: u32 = 2; 68 #[allow(unused)] 69 const VIRTIO_IOMMU_F_BYPASS: u32 = 3; 70 const VIRTIO_IOMMU_F_PROBE: u32 = 4; 71 #[allow(unused)] 72 const VIRTIO_IOMMU_F_MMIO: u32 = 5; 73 const VIRTIO_IOMMU_F_BYPASS_CONFIG: u32 = 6; 74 75 // Support 2MiB and 4KiB page sizes. 76 const VIRTIO_IOMMU_PAGE_SIZE_MASK: u64 = (2 << 20) | (4 << 10); 77 78 #[derive(Copy, Clone, Debug, Default)] 79 #[repr(packed)] 80 #[allow(dead_code)] 81 struct VirtioIommuRange32 { 82 start: u32, 83 end: u32, 84 } 85 86 #[derive(Copy, Clone, Debug, Default)] 87 #[repr(packed)] 88 #[allow(dead_code)] 89 struct VirtioIommuRange64 { 90 start: u64, 91 end: u64, 92 } 93 94 #[derive(Copy, Clone, Debug, Default)] 95 #[repr(packed)] 96 #[allow(dead_code)] 97 struct VirtioIommuConfig { 98 page_size_mask: u64, 99 input_range: VirtioIommuRange64, 100 domain_range: VirtioIommuRange32, 101 probe_size: u32, 102 bypass: u8, 103 _reserved: [u8; 7], 104 } 105 106 /// Virtio IOMMU request type 107 const VIRTIO_IOMMU_T_ATTACH: u8 = 1; 108 const VIRTIO_IOMMU_T_DETACH: u8 = 2; 109 const VIRTIO_IOMMU_T_MAP: u8 = 3; 110 const VIRTIO_IOMMU_T_UNMAP: u8 = 4; 111 const VIRTIO_IOMMU_T_PROBE: u8 = 5; 112 113 #[derive(Copy, Clone, Debug, Default)] 114 #[repr(packed)] 115 struct VirtioIommuReqHead { 116 type_: u8, 117 _reserved: [u8; 3], 118 } 119 120 /// Virtio IOMMU request status 121 const VIRTIO_IOMMU_S_OK: u8 = 0; 122 #[allow(unused)] 123 const VIRTIO_IOMMU_S_IOERR: u8 = 1; 124 #[allow(unused)] 125 const VIRTIO_IOMMU_S_UNSUPP: u8 = 2; 126 #[allow(unused)] 127 const VIRTIO_IOMMU_S_DEVERR: u8 = 3; 128 #[allow(unused)] 129 const VIRTIO_IOMMU_S_INVAL: u8 = 4; 130 #[allow(unused)] 131 const VIRTIO_IOMMU_S_RANGE: u8 = 5; 132 #[allow(unused)] 133 const VIRTIO_IOMMU_S_NOENT: u8 = 6; 134 #[allow(unused)] 135 const VIRTIO_IOMMU_S_FAULT: u8 = 7; 136 #[allow(unused)] 137 const VIRTIO_IOMMU_S_NOMEM: u8 = 8; 138 139 #[derive(Copy, Clone, Debug, Default)] 140 #[repr(packed)] 141 #[allow(dead_code)] 142 struct VirtioIommuReqTail { 143 status: u8, 144 _reserved: [u8; 3], 145 } 146 147 /// ATTACH request 148 #[derive(Copy, Clone, Debug, Default)] 149 #[repr(packed)] 150 struct VirtioIommuReqAttach { 151 domain: u32, 152 endpoint: u32, 153 flags: u32, 154 _reserved: [u8; 4], 155 } 156 157 const VIRTIO_IOMMU_ATTACH_F_BYPASS: u32 = 1; 158 159 /// DETACH request 160 #[derive(Copy, Clone, Debug, Default)] 161 #[repr(packed)] 162 struct VirtioIommuReqDetach { 163 domain: u32, 164 endpoint: u32, 165 _reserved: [u8; 8], 166 } 167 168 /// Virtio IOMMU request MAP flags 169 #[allow(unused)] 170 const VIRTIO_IOMMU_MAP_F_READ: u32 = 1; 171 #[allow(unused)] 172 const VIRTIO_IOMMU_MAP_F_WRITE: u32 = 1 << 1; 173 #[allow(unused)] 174 const VIRTIO_IOMMU_MAP_F_MMIO: u32 = 1 << 2; 175 #[allow(unused)] 176 const VIRTIO_IOMMU_MAP_F_MASK: u32 = 177 VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE | VIRTIO_IOMMU_MAP_F_MMIO; 178 179 /// MAP request 180 #[derive(Copy, Clone, Debug, Default)] 181 #[repr(packed)] 182 struct VirtioIommuReqMap { 183 domain: u32, 184 virt_start: u64, 185 virt_end: u64, 186 phys_start: u64, 187 _flags: u32, 188 } 189 190 /// UNMAP request 191 #[derive(Copy, Clone, Debug, Default)] 192 #[repr(packed)] 193 struct VirtioIommuReqUnmap { 194 domain: u32, 195 virt_start: u64, 196 virt_end: u64, 197 _reserved: [u8; 4], 198 } 199 200 /// Virtio IOMMU request PROBE types 201 #[allow(unused)] 202 const VIRTIO_IOMMU_PROBE_T_NONE: u16 = 0; 203 const VIRTIO_IOMMU_PROBE_T_RESV_MEM: u16 = 1; 204 #[allow(unused)] 205 const VIRTIO_IOMMU_PROBE_T_MASK: u16 = 0xfff; 206 207 /// PROBE request 208 #[derive(Copy, Clone, Debug, Default)] 209 #[repr(packed)] 210 #[allow(dead_code)] 211 struct VirtioIommuReqProbe { 212 endpoint: u32, 213 _reserved: [u64; 8], 214 } 215 216 #[derive(Copy, Clone, Debug, Default)] 217 #[repr(packed)] 218 #[allow(dead_code)] 219 struct VirtioIommuProbeProperty { 220 type_: u16, 221 length: u16, 222 } 223 224 /// Virtio IOMMU request PROBE property RESV_MEM subtypes 225 #[allow(unused)] 226 const VIRTIO_IOMMU_RESV_MEM_T_RESERVED: u8 = 0; 227 const VIRTIO_IOMMU_RESV_MEM_T_MSI: u8 = 1; 228 229 #[derive(Copy, Clone, Debug, Default)] 230 #[repr(packed)] 231 #[allow(dead_code)] 232 struct VirtioIommuProbeResvMem { 233 subtype: u8, 234 _reserved: [u8; 3], 235 start: u64, 236 end: u64, 237 } 238 239 /// Virtio IOMMU fault flags 240 #[allow(unused)] 241 const VIRTIO_IOMMU_FAULT_F_READ: u32 = 1; 242 #[allow(unused)] 243 const VIRTIO_IOMMU_FAULT_F_WRITE: u32 = 1 << 1; 244 #[allow(unused)] 245 const VIRTIO_IOMMU_FAULT_F_EXEC: u32 = 1 << 2; 246 #[allow(unused)] 247 const VIRTIO_IOMMU_FAULT_F_ADDRESS: u32 = 1 << 8; 248 249 /// Virtio IOMMU fault reasons 250 #[allow(unused)] 251 const VIRTIO_IOMMU_FAULT_R_UNKNOWN: u32 = 0; 252 #[allow(unused)] 253 const VIRTIO_IOMMU_FAULT_R_DOMAIN: u32 = 1; 254 #[allow(unused)] 255 const VIRTIO_IOMMU_FAULT_R_MAPPING: u32 = 2; 256 257 /// Fault reporting through eventq 258 #[allow(unused)] 259 #[derive(Copy, Clone, Debug, Default)] 260 #[repr(packed)] 261 struct VirtioIommuFault { 262 reason: u8, 263 reserved: [u8; 3], 264 flags: u32, 265 endpoint: u32, 266 reserved2: [u8; 4], 267 address: u64, 268 } 269 270 // SAFETY: data structure only contain integers and have no implicit padding 271 unsafe impl ByteValued for VirtioIommuRange32 {} 272 // SAFETY: data structure only contain integers and have no implicit padding 273 unsafe impl ByteValued for VirtioIommuRange64 {} 274 // SAFETY: data structure only contain integers and have no implicit padding 275 unsafe impl ByteValued for VirtioIommuConfig {} 276 // SAFETY: data structure only contain integers and have no implicit padding 277 unsafe impl ByteValued for VirtioIommuReqHead {} 278 // SAFETY: data structure only contain integers and have no implicit padding 279 unsafe impl ByteValued for VirtioIommuReqTail {} 280 // SAFETY: data structure only contain integers and have no implicit padding 281 unsafe impl ByteValued for VirtioIommuReqAttach {} 282 // SAFETY: data structure only contain integers and have no implicit padding 283 unsafe impl ByteValued for VirtioIommuReqDetach {} 284 // SAFETY: data structure only contain integers and have no implicit padding 285 unsafe impl ByteValued for VirtioIommuReqMap {} 286 // SAFETY: data structure only contain integers and have no implicit padding 287 unsafe impl ByteValued for VirtioIommuReqUnmap {} 288 // SAFETY: data structure only contain integers and have no implicit padding 289 unsafe impl ByteValued for VirtioIommuReqProbe {} 290 // SAFETY: data structure only contain integers and have no implicit padding 291 unsafe impl ByteValued for VirtioIommuProbeProperty {} 292 // SAFETY: data structure only contain integers and have no implicit padding 293 unsafe impl ByteValued for VirtioIommuProbeResvMem {} 294 // SAFETY: data structure only contain integers and have no implicit padding 295 unsafe impl ByteValued for VirtioIommuFault {} 296 297 #[derive(Error, Debug)] 298 enum Error { 299 #[error("Guest gave us bad memory addresses: {0}")] 300 GuestMemory(GuestMemoryError), 301 #[error("Guest gave us a write only descriptor that protocol says to read from")] 302 UnexpectedWriteOnlyDescriptor, 303 #[error("Guest gave us a read only descriptor that protocol says to write to")] 304 UnexpectedReadOnlyDescriptor, 305 #[error("Guest gave us too few descriptors in a descriptor chain")] 306 DescriptorChainTooShort, 307 #[error("Guest gave us a buffer that was too short to use")] 308 BufferLengthTooSmall, 309 #[error("Guest sent us invalid request")] 310 InvalidRequest, 311 #[error("Guest sent us invalid ATTACH request")] 312 InvalidAttachRequest, 313 #[error("Guest sent us invalid DETACH request")] 314 InvalidDetachRequest, 315 #[error("Guest sent us invalid MAP request")] 316 InvalidMapRequest, 317 #[error("Invalid to map because the domain is in bypass mode")] 318 InvalidMapRequestBypassDomain, 319 #[error("Invalid to map because the domain is missing")] 320 InvalidMapRequestMissingDomain, 321 #[error("Guest sent us invalid UNMAP request")] 322 InvalidUnmapRequest, 323 #[error("Invalid to unmap because the domain is in bypass mode")] 324 InvalidUnmapRequestBypassDomain, 325 #[error("Invalid to unmap because the domain is missing")] 326 InvalidUnmapRequestMissingDomain, 327 #[error("Guest sent us invalid PROBE request")] 328 InvalidProbeRequest, 329 #[error("Failed to performing external mapping: {0}")] 330 ExternalMapping(io::Error), 331 #[error("Failed to performing external unmapping: {0}")] 332 ExternalUnmapping(io::Error), 333 #[error("Failed adding used index: {0}")] 334 QueueAddUsed(virtio_queue::Error), 335 } 336 337 struct Request {} 338 339 impl Request { 340 // Parse the available vring buffer. Based on the hashmap table of external 341 // mappings required from various devices such as VFIO or vhost-user ones, 342 // this function might update the hashmap table of external mappings per 343 // domain. 344 // Basically, the VMM knows about the device_id <=> mapping relationship 345 // before running the VM, but at runtime, a new domain <=> mapping hashmap 346 // is created based on the information provided from the guest driver for 347 // virtio-iommu (giving the link device_id <=> domain). 348 fn parse( 349 desc_chain: &mut DescriptorChain<GuestMemoryLoadGuard<GuestMemoryMmap>>, 350 mapping: &Arc<IommuMapping>, 351 ext_mapping: &BTreeMap<u32, Arc<dyn ExternalDmaMapping>>, 352 msi_iova_space: (u64, u64), 353 ) -> result::Result<usize, Error> { 354 let desc = desc_chain 355 .next() 356 .ok_or(Error::DescriptorChainTooShort) 357 .inspect_err(|_| { 358 error!("Missing head descriptor"); 359 })?; 360 361 // The descriptor contains the request type which MUST be readable. 362 if desc.is_write_only() { 363 return Err(Error::UnexpectedWriteOnlyDescriptor); 364 } 365 366 if (desc.len() as usize) < size_of::<VirtioIommuReqHead>() { 367 return Err(Error::InvalidRequest); 368 } 369 370 let req_head: VirtioIommuReqHead = desc_chain 371 .memory() 372 .read_obj(desc.addr()) 373 .map_err(Error::GuestMemory)?; 374 let req_offset = size_of::<VirtioIommuReqHead>(); 375 let desc_size_left = (desc.len() as usize) - req_offset; 376 let req_addr = if let Some(addr) = desc.addr().checked_add(req_offset as u64) { 377 addr 378 } else { 379 return Err(Error::InvalidRequest); 380 }; 381 382 let (msi_iova_start, msi_iova_end) = msi_iova_space; 383 384 // Create the reply 385 let mut reply: Vec<u8> = Vec::new(); 386 let mut status = VIRTIO_IOMMU_S_OK; 387 let mut hdr_len = 0; 388 389 let result = (|| { 390 match req_head.type_ { 391 VIRTIO_IOMMU_T_ATTACH => { 392 if desc_size_left != size_of::<VirtioIommuReqAttach>() { 393 status = VIRTIO_IOMMU_S_INVAL; 394 return Err(Error::InvalidAttachRequest); 395 } 396 397 let req: VirtioIommuReqAttach = desc_chain 398 .memory() 399 .read_obj(req_addr as GuestAddress) 400 .map_err(Error::GuestMemory)?; 401 debug!("Attach request {:?}", req); 402 403 // Copy the value to use it as a proper reference. 404 let domain_id = req.domain; 405 let endpoint = req.endpoint; 406 let bypass = 407 (req.flags & VIRTIO_IOMMU_ATTACH_F_BYPASS) == VIRTIO_IOMMU_ATTACH_F_BYPASS; 408 409 let mut old_domain_id = domain_id; 410 if let Some(&id) = mapping.endpoints.read().unwrap().get(&endpoint) { 411 old_domain_id = id; 412 } 413 414 if old_domain_id != domain_id { 415 detach_endpoint_from_domain(endpoint, old_domain_id, mapping, ext_mapping)?; 416 } 417 418 // Add endpoint associated with specific domain 419 mapping 420 .endpoints 421 .write() 422 .unwrap() 423 .insert(endpoint, domain_id); 424 425 // If any other mappings exist in the domain for other containers, 426 // make sure to issue these mappings for the new endpoint/container 427 if let Some(domain_mappings) = &mapping.domains.read().unwrap().get(&domain_id) 428 { 429 if let Some(ext_map) = ext_mapping.get(&endpoint) { 430 for (virt_start, addr_map) in &domain_mappings.mappings { 431 ext_map 432 .map(*virt_start, addr_map.gpa, addr_map.size) 433 .map_err(Error::ExternalUnmapping)?; 434 } 435 } 436 } 437 438 // Add new domain with no mapping if the entry didn't exist yet 439 let mut domains = mapping.domains.write().unwrap(); 440 let domain = Domain { 441 mappings: BTreeMap::new(), 442 bypass, 443 }; 444 domains.entry(domain_id).or_insert_with(|| domain); 445 } 446 VIRTIO_IOMMU_T_DETACH => { 447 if desc_size_left != size_of::<VirtioIommuReqDetach>() { 448 status = VIRTIO_IOMMU_S_INVAL; 449 return Err(Error::InvalidDetachRequest); 450 } 451 452 let req: VirtioIommuReqDetach = desc_chain 453 .memory() 454 .read_obj(req_addr as GuestAddress) 455 .map_err(Error::GuestMemory)?; 456 debug!("Detach request {:?}", req); 457 458 // Copy the value to use it as a proper reference. 459 let domain_id = req.domain; 460 let endpoint = req.endpoint; 461 462 // Remove endpoint associated with specific domain 463 detach_endpoint_from_domain(endpoint, domain_id, mapping, ext_mapping)?; 464 } 465 VIRTIO_IOMMU_T_MAP => { 466 if desc_size_left != size_of::<VirtioIommuReqMap>() { 467 status = VIRTIO_IOMMU_S_INVAL; 468 return Err(Error::InvalidMapRequest); 469 } 470 471 let req: VirtioIommuReqMap = desc_chain 472 .memory() 473 .read_obj(req_addr as GuestAddress) 474 .map_err(Error::GuestMemory)?; 475 debug!("Map request {:?}", req); 476 477 // Copy the value to use it as a proper reference. 478 let domain_id = req.domain; 479 480 if let Some(domain) = mapping.domains.read().unwrap().get(&domain_id) { 481 if domain.bypass { 482 status = VIRTIO_IOMMU_S_INVAL; 483 return Err(Error::InvalidMapRequestBypassDomain); 484 } 485 } else { 486 status = VIRTIO_IOMMU_S_INVAL; 487 return Err(Error::InvalidMapRequestMissingDomain); 488 } 489 490 // Find the list of endpoints attached to the given domain. 491 let endpoints: Vec<u32> = mapping 492 .endpoints 493 .write() 494 .unwrap() 495 .iter() 496 .filter(|(_, &d)| d == domain_id) 497 .map(|(&e, _)| e) 498 .collect(); 499 500 // For viommu all endpoints receive their own VFIO container, as a result 501 // Each endpoint within the domain needs to be separately mapped, as the 502 // mapping is done on a per-container level, not a per-domain level 503 for endpoint in endpoints { 504 if let Some(ext_map) = ext_mapping.get(&endpoint) { 505 let size = req.virt_end - req.virt_start + 1; 506 ext_map 507 .map(req.virt_start, req.phys_start, size) 508 .map_err(Error::ExternalMapping)?; 509 } 510 } 511 512 // Add new mapping associated with the domain 513 mapping 514 .domains 515 .write() 516 .unwrap() 517 .get_mut(&domain_id) 518 .unwrap() 519 .mappings 520 .insert( 521 req.virt_start, 522 Mapping { 523 gpa: req.phys_start, 524 size: req.virt_end - req.virt_start + 1, 525 }, 526 ); 527 } 528 VIRTIO_IOMMU_T_UNMAP => { 529 if desc_size_left != size_of::<VirtioIommuReqUnmap>() { 530 status = VIRTIO_IOMMU_S_INVAL; 531 return Err(Error::InvalidUnmapRequest); 532 } 533 534 let req: VirtioIommuReqUnmap = desc_chain 535 .memory() 536 .read_obj(req_addr as GuestAddress) 537 .map_err(Error::GuestMemory)?; 538 debug!("Unmap request {:?}", req); 539 540 // Copy the value to use it as a proper reference. 541 let domain_id = req.domain; 542 let virt_start = req.virt_start; 543 544 if let Some(domain) = mapping.domains.read().unwrap().get(&domain_id) { 545 if domain.bypass { 546 status = VIRTIO_IOMMU_S_INVAL; 547 return Err(Error::InvalidUnmapRequestBypassDomain); 548 } 549 } else { 550 status = VIRTIO_IOMMU_S_INVAL; 551 return Err(Error::InvalidUnmapRequestMissingDomain); 552 } 553 554 // Find the list of endpoints attached to the given domain. 555 let endpoints: Vec<u32> = mapping 556 .endpoints 557 .write() 558 .unwrap() 559 .iter() 560 .filter(|(_, &d)| d == domain_id) 561 .map(|(&e, _)| e) 562 .collect(); 563 564 // Trigger external unmapping if necessary. 565 for endpoint in endpoints { 566 if let Some(ext_map) = ext_mapping.get(&endpoint) { 567 let size = req.virt_end - virt_start + 1; 568 ext_map 569 .unmap(virt_start, size) 570 .map_err(Error::ExternalUnmapping)?; 571 } 572 } 573 574 // Remove all mappings associated with the domain within the requested range 575 mapping 576 .domains 577 .write() 578 .unwrap() 579 .get_mut(&domain_id) 580 .unwrap() 581 .mappings 582 .retain(|&x, _| (x < req.virt_start || x > req.virt_end)); 583 } 584 VIRTIO_IOMMU_T_PROBE => { 585 if desc_size_left != size_of::<VirtioIommuReqProbe>() { 586 status = VIRTIO_IOMMU_S_INVAL; 587 return Err(Error::InvalidProbeRequest); 588 } 589 590 let req: VirtioIommuReqProbe = desc_chain 591 .memory() 592 .read_obj(req_addr as GuestAddress) 593 .map_err(Error::GuestMemory)?; 594 debug!("Probe request {:?}", req); 595 596 let probe_prop = VirtioIommuProbeProperty { 597 type_: VIRTIO_IOMMU_PROBE_T_RESV_MEM, 598 length: size_of::<VirtioIommuProbeResvMem>() as u16, 599 }; 600 reply.extend_from_slice(probe_prop.as_slice()); 601 602 let resv_mem = VirtioIommuProbeResvMem { 603 subtype: VIRTIO_IOMMU_RESV_MEM_T_MSI, 604 start: msi_iova_start, 605 end: msi_iova_end, 606 ..Default::default() 607 }; 608 reply.extend_from_slice(resv_mem.as_slice()); 609 610 hdr_len = PROBE_PROP_SIZE; 611 } 612 _ => { 613 status = VIRTIO_IOMMU_S_INVAL; 614 return Err(Error::InvalidRequest); 615 } 616 } 617 Ok(()) 618 })(); 619 620 let status_desc = desc_chain.next().ok_or(Error::DescriptorChainTooShort)?; 621 622 // The status MUST always be writable 623 if !status_desc.is_write_only() { 624 return Err(Error::UnexpectedReadOnlyDescriptor); 625 } 626 627 if status_desc.len() < hdr_len + size_of::<VirtioIommuReqTail>() as u32 { 628 return Err(Error::BufferLengthTooSmall); 629 } 630 631 let tail = VirtioIommuReqTail { 632 status, 633 ..Default::default() 634 }; 635 reply.extend_from_slice(tail.as_slice()); 636 637 // Make sure we return the result of the request to the guest before 638 // we return a potential error internally. 639 desc_chain 640 .memory() 641 .write_slice(reply.as_slice(), status_desc.addr()) 642 .map_err(Error::GuestMemory)?; 643 644 // Return the error if the result was not Ok(). 645 result?; 646 647 Ok((hdr_len as usize) + size_of::<VirtioIommuReqTail>()) 648 } 649 } 650 651 fn detach_endpoint_from_domain( 652 endpoint: u32, 653 domain_id: u32, 654 mapping: &Arc<IommuMapping>, 655 ext_mapping: &BTreeMap<u32, Arc<dyn ExternalDmaMapping>>, 656 ) -> result::Result<(), Error> { 657 // Remove endpoint associated with specific domain 658 mapping.endpoints.write().unwrap().remove(&endpoint); 659 660 // Trigger external unmapping for the endpoint if necessary. 661 if let Some(domain_mappings) = &mapping.domains.read().unwrap().get(&domain_id) { 662 if let Some(ext_map) = ext_mapping.get(&endpoint) { 663 for (virt_start, addr_map) in &domain_mappings.mappings { 664 ext_map 665 .unmap(*virt_start, addr_map.size) 666 .map_err(Error::ExternalUnmapping)?; 667 } 668 } 669 } 670 671 if mapping 672 .endpoints 673 .write() 674 .unwrap() 675 .iter() 676 .filter(|(_, &d)| d == domain_id) 677 .count() 678 == 0 679 { 680 mapping.domains.write().unwrap().remove(&domain_id); 681 } 682 683 Ok(()) 684 } 685 686 struct IommuEpollHandler { 687 mem: GuestMemoryAtomic<GuestMemoryMmap>, 688 request_queue: Queue, 689 _event_queue: Queue, 690 interrupt_cb: Arc<dyn VirtioInterrupt>, 691 request_queue_evt: EventFd, 692 _event_queue_evt: EventFd, 693 kill_evt: EventFd, 694 pause_evt: EventFd, 695 mapping: Arc<IommuMapping>, 696 ext_mapping: Arc<Mutex<BTreeMap<u32, Arc<dyn ExternalDmaMapping>>>>, 697 msi_iova_space: (u64, u64), 698 } 699 700 impl IommuEpollHandler { 701 fn request_queue(&mut self) -> Result<bool, Error> { 702 let mut used_descs = false; 703 while let Some(mut desc_chain) = self.request_queue.pop_descriptor_chain(self.mem.memory()) 704 { 705 let len = Request::parse( 706 &mut desc_chain, 707 &self.mapping, 708 &self.ext_mapping.lock().unwrap(), 709 self.msi_iova_space, 710 )?; 711 712 self.request_queue 713 .add_used(desc_chain.memory(), desc_chain.head_index(), len as u32) 714 .map_err(Error::QueueAddUsed)?; 715 716 used_descs = true; 717 } 718 719 Ok(used_descs) 720 } 721 722 fn signal_used_queue(&self, queue_index: u16) -> result::Result<(), DeviceError> { 723 self.interrupt_cb 724 .trigger(VirtioInterruptType::Queue(queue_index)) 725 .map_err(|e| { 726 error!("Failed to signal used queue: {:?}", e); 727 DeviceError::FailedSignalingUsedQueue(e) 728 }) 729 } 730 731 fn run( 732 &mut self, 733 paused: Arc<AtomicBool>, 734 paused_sync: Arc<Barrier>, 735 ) -> result::Result<(), EpollHelperError> { 736 let mut helper = EpollHelper::new(&self.kill_evt, &self.pause_evt)?; 737 helper.add_event(self.request_queue_evt.as_raw_fd(), REQUEST_Q_EVENT)?; 738 helper.run(paused, paused_sync, self)?; 739 740 Ok(()) 741 } 742 } 743 744 impl EpollHelperHandler for IommuEpollHandler { 745 fn handle_event( 746 &mut self, 747 _helper: &mut EpollHelper, 748 event: &epoll::Event, 749 ) -> result::Result<(), EpollHelperError> { 750 let ev_type = event.data as u16; 751 match ev_type { 752 REQUEST_Q_EVENT => { 753 self.request_queue_evt.read().map_err(|e| { 754 EpollHelperError::HandleEvent(anyhow!("Failed to get queue event: {:?}", e)) 755 })?; 756 757 let needs_notification = self.request_queue().map_err(|e| { 758 EpollHelperError::HandleEvent(anyhow!( 759 "Failed to process request queue : {:?}", 760 e 761 )) 762 })?; 763 if needs_notification { 764 self.signal_used_queue(0).map_err(|e| { 765 EpollHelperError::HandleEvent(anyhow!( 766 "Failed to signal used queue: {:?}", 767 e 768 )) 769 })?; 770 } 771 } 772 _ => { 773 return Err(EpollHelperError::HandleEvent(anyhow!( 774 "Unexpected event: {}", 775 ev_type 776 ))); 777 } 778 } 779 Ok(()) 780 } 781 } 782 783 #[derive(Clone, Copy, Debug, Serialize, Deserialize)] 784 struct Mapping { 785 gpa: u64, 786 size: u64, 787 } 788 789 #[derive(Clone, Debug)] 790 struct Domain { 791 mappings: BTreeMap<u64, Mapping>, 792 bypass: bool, 793 } 794 795 #[derive(Debug)] 796 pub struct IommuMapping { 797 // Domain related to an endpoint. 798 endpoints: Arc<RwLock<BTreeMap<u32, u32>>>, 799 // Information related to each domain. 800 domains: Arc<RwLock<BTreeMap<u32, Domain>>>, 801 // Global flag indicating if endpoints that are not attached to any domain 802 // are in bypass mode. 803 bypass: AtomicBool, 804 } 805 806 impl DmaRemapping for IommuMapping { 807 fn translate_gva(&self, id: u32, addr: u64) -> std::result::Result<u64, std::io::Error> { 808 debug!("Translate GVA addr 0x{:x}", addr); 809 if let Some(domain_id) = self.endpoints.read().unwrap().get(&id) { 810 if let Some(domain) = self.domains.read().unwrap().get(domain_id) { 811 // Directly return identity mapping in case the domain is in 812 // bypass mode. 813 if domain.bypass { 814 return Ok(addr); 815 } 816 817 let range_start = if VIRTIO_IOMMU_PAGE_SIZE_MASK > addr { 818 0 819 } else { 820 addr - VIRTIO_IOMMU_PAGE_SIZE_MASK 821 }; 822 for (&key, &value) in domain 823 .mappings 824 .range((Included(&range_start), Included(&addr))) 825 { 826 if addr >= key && addr < key + value.size { 827 let new_addr = addr - key + value.gpa; 828 debug!("Into GPA addr 0x{:x}", new_addr); 829 return Ok(new_addr); 830 } 831 } 832 } 833 } else if self.bypass.load(Ordering::Acquire) { 834 return Ok(addr); 835 } 836 837 Err(io::Error::new( 838 io::ErrorKind::Other, 839 format!("failed to translate GVA addr 0x{addr:x}"), 840 )) 841 } 842 843 fn translate_gpa(&self, id: u32, addr: u64) -> std::result::Result<u64, std::io::Error> { 844 debug!("Translate GPA addr 0x{:x}", addr); 845 if let Some(domain_id) = self.endpoints.read().unwrap().get(&id) { 846 if let Some(domain) = self.domains.read().unwrap().get(domain_id) { 847 // Directly return identity mapping in case the domain is in 848 // bypass mode. 849 if domain.bypass { 850 return Ok(addr); 851 } 852 853 for (&key, &value) in domain.mappings.iter() { 854 if addr >= value.gpa && addr < value.gpa + value.size { 855 let new_addr = addr - value.gpa + key; 856 debug!("Into GVA addr 0x{:x}", new_addr); 857 return Ok(new_addr); 858 } 859 } 860 } 861 } else if self.bypass.load(Ordering::Acquire) { 862 return Ok(addr); 863 } 864 865 Err(io::Error::new( 866 io::ErrorKind::Other, 867 format!("failed to translate GPA addr 0x{addr:x}"), 868 )) 869 } 870 } 871 872 #[derive(Debug)] 873 pub struct AccessPlatformMapping { 874 id: u32, 875 mapping: Arc<IommuMapping>, 876 } 877 878 impl AccessPlatformMapping { 879 pub fn new(id: u32, mapping: Arc<IommuMapping>) -> Self { 880 AccessPlatformMapping { id, mapping } 881 } 882 } 883 884 impl AccessPlatform for AccessPlatformMapping { 885 fn translate_gva(&self, base: u64, _size: u64) -> std::result::Result<u64, std::io::Error> { 886 self.mapping.translate_gva(self.id, base) 887 } 888 fn translate_gpa(&self, base: u64, _size: u64) -> std::result::Result<u64, std::io::Error> { 889 self.mapping.translate_gpa(self.id, base) 890 } 891 } 892 893 pub struct Iommu { 894 common: VirtioCommon, 895 id: String, 896 config: VirtioIommuConfig, 897 mapping: Arc<IommuMapping>, 898 ext_mapping: Arc<Mutex<BTreeMap<u32, Arc<dyn ExternalDmaMapping>>>>, 899 seccomp_action: SeccompAction, 900 exit_evt: EventFd, 901 msi_iova_space: (u64, u64), 902 } 903 904 type EndpointsState = Vec<(u32, u32)>; 905 type DomainsState = Vec<(u32, (Vec<(u64, Mapping)>, bool))>; 906 907 #[derive(Serialize, Deserialize)] 908 pub struct IommuState { 909 avail_features: u64, 910 acked_features: u64, 911 endpoints: EndpointsState, 912 domains: DomainsState, 913 } 914 915 impl Iommu { 916 pub fn new( 917 id: String, 918 seccomp_action: SeccompAction, 919 exit_evt: EventFd, 920 msi_iova_space: (u64, u64), 921 state: Option<IommuState>, 922 ) -> io::Result<(Self, Arc<IommuMapping>)> { 923 let (avail_features, acked_features, endpoints, domains, paused) = 924 if let Some(state) = state { 925 info!("Restoring virtio-iommu {}", id); 926 ( 927 state.avail_features, 928 state.acked_features, 929 state.endpoints.into_iter().collect(), 930 state 931 .domains 932 .into_iter() 933 .map(|(k, v)| { 934 ( 935 k, 936 Domain { 937 mappings: v.0.into_iter().collect(), 938 bypass: v.1, 939 }, 940 ) 941 }) 942 .collect(), 943 true, 944 ) 945 } else { 946 let avail_features = 1u64 << VIRTIO_F_VERSION_1 947 | 1u64 << VIRTIO_IOMMU_F_MAP_UNMAP 948 | 1u64 << VIRTIO_IOMMU_F_PROBE 949 | 1u64 << VIRTIO_IOMMU_F_BYPASS_CONFIG; 950 951 (avail_features, 0, BTreeMap::new(), BTreeMap::new(), false) 952 }; 953 954 let config = VirtioIommuConfig { 955 page_size_mask: VIRTIO_IOMMU_PAGE_SIZE_MASK, 956 probe_size: PROBE_PROP_SIZE, 957 ..Default::default() 958 }; 959 960 let mapping = Arc::new(IommuMapping { 961 endpoints: Arc::new(RwLock::new(endpoints)), 962 domains: Arc::new(RwLock::new(domains)), 963 bypass: AtomicBool::new(true), 964 }); 965 966 Ok(( 967 Iommu { 968 id, 969 common: VirtioCommon { 970 device_type: VirtioDeviceType::Iommu as u32, 971 queue_sizes: QUEUE_SIZES.to_vec(), 972 avail_features, 973 acked_features, 974 paused_sync: Some(Arc::new(Barrier::new(2))), 975 min_queues: NUM_QUEUES as u16, 976 paused: Arc::new(AtomicBool::new(paused)), 977 ..Default::default() 978 }, 979 config, 980 mapping: mapping.clone(), 981 ext_mapping: Arc::new(Mutex::new(BTreeMap::new())), 982 seccomp_action, 983 exit_evt, 984 msi_iova_space, 985 }, 986 mapping, 987 )) 988 } 989 990 fn state(&self) -> IommuState { 991 IommuState { 992 avail_features: self.common.avail_features, 993 acked_features: self.common.acked_features, 994 endpoints: self 995 .mapping 996 .endpoints 997 .read() 998 .unwrap() 999 .clone() 1000 .into_iter() 1001 .collect(), 1002 domains: self 1003 .mapping 1004 .domains 1005 .read() 1006 .unwrap() 1007 .clone() 1008 .into_iter() 1009 .map(|(k, v)| (k, (v.mappings.into_iter().collect(), v.bypass))) 1010 .collect(), 1011 } 1012 } 1013 1014 fn update_bypass(&mut self) { 1015 // Use bypass from config if VIRTIO_IOMMU_F_BYPASS_CONFIG has been negotiated 1016 if !self 1017 .common 1018 .feature_acked(VIRTIO_IOMMU_F_BYPASS_CONFIG.into()) 1019 { 1020 return; 1021 } 1022 1023 let bypass = self.config.bypass == 1; 1024 info!("Updating bypass mode to {}", bypass); 1025 self.mapping.bypass.store(bypass, Ordering::Release); 1026 } 1027 1028 pub fn add_external_mapping(&mut self, device_id: u32, mapping: Arc<dyn ExternalDmaMapping>) { 1029 self.ext_mapping.lock().unwrap().insert(device_id, mapping); 1030 } 1031 1032 #[cfg(fuzzing)] 1033 pub fn wait_for_epoll_threads(&mut self) { 1034 self.common.wait_for_epoll_threads(); 1035 } 1036 } 1037 1038 impl Drop for Iommu { 1039 fn drop(&mut self) { 1040 if let Some(kill_evt) = self.common.kill_evt.take() { 1041 // Ignore the result because there is nothing we can do about it. 1042 let _ = kill_evt.write(1); 1043 } 1044 self.common.wait_for_epoll_threads(); 1045 } 1046 } 1047 1048 impl VirtioDevice for Iommu { 1049 fn device_type(&self) -> u32 { 1050 self.common.device_type 1051 } 1052 1053 fn queue_max_sizes(&self) -> &[u16] { 1054 &self.common.queue_sizes 1055 } 1056 1057 fn features(&self) -> u64 { 1058 self.common.avail_features 1059 } 1060 1061 fn ack_features(&mut self, value: u64) { 1062 self.common.ack_features(value) 1063 } 1064 1065 fn read_config(&self, offset: u64, data: &mut [u8]) { 1066 self.read_config_from_slice(self.config.as_slice(), offset, data); 1067 } 1068 1069 fn write_config(&mut self, offset: u64, data: &[u8]) { 1070 // The "bypass" field is the only mutable field 1071 let bypass_offset = 1072 (&self.config.bypass as *const _ as u64) - (&self.config as *const _ as u64); 1073 if offset != bypass_offset || data.len() != std::mem::size_of_val(&self.config.bypass) { 1074 error!( 1075 "Attempt to write to read-only field: offset {:x} length {}", 1076 offset, 1077 data.len() 1078 ); 1079 return; 1080 } 1081 1082 self.config.bypass = data[0]; 1083 1084 self.update_bypass(); 1085 } 1086 1087 fn activate( 1088 &mut self, 1089 mem: GuestMemoryAtomic<GuestMemoryMmap>, 1090 interrupt_cb: Arc<dyn VirtioInterrupt>, 1091 mut queues: Vec<(usize, Queue, EventFd)>, 1092 ) -> ActivateResult { 1093 self.common.activate(&queues, &interrupt_cb)?; 1094 let (kill_evt, pause_evt) = self.common.dup_eventfds(); 1095 1096 let (_, request_queue, request_queue_evt) = queues.remove(0); 1097 let (_, _event_queue, _event_queue_evt) = queues.remove(0); 1098 1099 let mut handler = IommuEpollHandler { 1100 mem, 1101 request_queue, 1102 _event_queue, 1103 interrupt_cb, 1104 request_queue_evt, 1105 _event_queue_evt, 1106 kill_evt, 1107 pause_evt, 1108 mapping: self.mapping.clone(), 1109 ext_mapping: self.ext_mapping.clone(), 1110 msi_iova_space: self.msi_iova_space, 1111 }; 1112 1113 let paused = self.common.paused.clone(); 1114 let paused_sync = self.common.paused_sync.clone(); 1115 let mut epoll_threads = Vec::new(); 1116 spawn_virtio_thread( 1117 &self.id, 1118 &self.seccomp_action, 1119 Thread::VirtioIommu, 1120 &mut epoll_threads, 1121 &self.exit_evt, 1122 move || handler.run(paused, paused_sync.unwrap()), 1123 )?; 1124 1125 self.common.epoll_threads = Some(epoll_threads); 1126 1127 event!("virtio-device", "activated", "id", &self.id); 1128 Ok(()) 1129 } 1130 1131 fn reset(&mut self) -> Option<Arc<dyn VirtioInterrupt>> { 1132 let result = self.common.reset(); 1133 event!("virtio-device", "reset", "id", &self.id); 1134 result 1135 } 1136 } 1137 1138 impl Pausable for Iommu { 1139 fn pause(&mut self) -> result::Result<(), MigratableError> { 1140 self.common.pause() 1141 } 1142 1143 fn resume(&mut self) -> result::Result<(), MigratableError> { 1144 self.common.resume() 1145 } 1146 } 1147 1148 impl Snapshottable for Iommu { 1149 fn id(&self) -> String { 1150 self.id.clone() 1151 } 1152 1153 fn snapshot(&mut self) -> std::result::Result<Snapshot, MigratableError> { 1154 Snapshot::new_from_state(&self.state()) 1155 } 1156 } 1157 impl Transportable for Iommu {} 1158 impl Migratable for Iommu {} 1159