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 std::collections::BTreeMap; 17 use std::io; 18 use std::mem::size_of; 19 use std::ops::Bound::Included; 20 use std::os::unix::io::AsRawFd; 21 use std::result; 22 use std::sync::atomic::{AtomicBool, Ordering}; 23 use std::sync::{Arc, Barrier, Mutex, RwLock}; 24 use thiserror::Error; 25 use versionize::{VersionMap, Versionize, VersionizeResult}; 26 use versionize_derive::Versionize; 27 use virtio_queue::{DescriptorChain, Queue, QueueT}; 28 use vm_device::dma_mapping::ExternalDmaMapping; 29 use vm_memory::{ 30 Address, ByteValued, Bytes, GuestAddress, GuestAddressSpace, GuestMemoryAtomic, 31 GuestMemoryError, GuestMemoryLoadGuard, 32 }; 33 use vm_migration::VersionMapped; 34 use vm_migration::{Migratable, MigratableError, Pausable, Snapshot, Snapshottable, Transportable}; 35 use vm_virtio::AccessPlatform; 36 use vmm_sys_util::eventfd::EventFd; 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 .map_err(|e| { 358 error!("Missing head descriptor"); 359 e 360 })?; 361 362 // The descriptor contains the request type which MUST be readable. 363 if desc.is_write_only() { 364 return Err(Error::UnexpectedWriteOnlyDescriptor); 365 } 366 367 if (desc.len() as usize) < size_of::<VirtioIommuReqHead>() { 368 return Err(Error::InvalidRequest); 369 } 370 371 let req_head: VirtioIommuReqHead = desc_chain 372 .memory() 373 .read_obj(desc.addr()) 374 .map_err(Error::GuestMemory)?; 375 let req_offset = size_of::<VirtioIommuReqHead>(); 376 let desc_size_left = (desc.len() as usize) - req_offset; 377 let req_addr = if let Some(addr) = desc.addr().checked_add(req_offset as u64) { 378 addr 379 } else { 380 return Err(Error::InvalidRequest); 381 }; 382 383 let (msi_iova_start, msi_iova_end) = msi_iova_space; 384 385 // Create the reply 386 let mut reply: Vec<u8> = Vec::new(); 387 let mut status = VIRTIO_IOMMU_S_OK; 388 let mut hdr_len = 0; 389 390 let result = (|| { 391 match req_head.type_ { 392 VIRTIO_IOMMU_T_ATTACH => { 393 if desc_size_left != size_of::<VirtioIommuReqAttach>() { 394 status = VIRTIO_IOMMU_S_INVAL; 395 return Err(Error::InvalidAttachRequest); 396 } 397 398 let req: VirtioIommuReqAttach = desc_chain 399 .memory() 400 .read_obj(req_addr as GuestAddress) 401 .map_err(Error::GuestMemory)?; 402 debug!("Attach request {:?}", req); 403 404 // Copy the value to use it as a proper reference. 405 let domain_id = req.domain; 406 let endpoint = req.endpoint; 407 let bypass = 408 (req.flags & VIRTIO_IOMMU_ATTACH_F_BYPASS) == VIRTIO_IOMMU_ATTACH_F_BYPASS; 409 410 // Add endpoint associated with specific domain 411 mapping 412 .endpoints 413 .write() 414 .unwrap() 415 .insert(endpoint, domain_id); 416 417 // Add new domain with no mapping if the entry didn't exist yet 418 let mut domains = mapping.domains.write().unwrap(); 419 let domain = Domain { 420 mappings: BTreeMap::new(), 421 bypass, 422 }; 423 domains.entry(domain_id).or_insert_with(|| domain); 424 } 425 VIRTIO_IOMMU_T_DETACH => { 426 if desc_size_left != size_of::<VirtioIommuReqDetach>() { 427 status = VIRTIO_IOMMU_S_INVAL; 428 return Err(Error::InvalidDetachRequest); 429 } 430 431 let req: VirtioIommuReqDetach = desc_chain 432 .memory() 433 .read_obj(req_addr as GuestAddress) 434 .map_err(Error::GuestMemory)?; 435 debug!("Detach request {:?}", req); 436 437 // Copy the value to use it as a proper reference. 438 let domain_id = req.domain; 439 let endpoint = req.endpoint; 440 441 // Remove endpoint associated with specific domain 442 mapping.endpoints.write().unwrap().remove(&endpoint); 443 444 // After all endpoints have been successfully detached from a 445 // domain, the domain can be removed. This means we must remove 446 // the mappings associated with this domain. 447 if mapping 448 .endpoints 449 .write() 450 .unwrap() 451 .iter() 452 .filter(|(_, &d)| d == domain_id) 453 .count() 454 == 0 455 { 456 mapping.domains.write().unwrap().remove(&domain_id); 457 } 458 } 459 VIRTIO_IOMMU_T_MAP => { 460 if desc_size_left != size_of::<VirtioIommuReqMap>() { 461 status = VIRTIO_IOMMU_S_INVAL; 462 return Err(Error::InvalidMapRequest); 463 } 464 465 let req: VirtioIommuReqMap = desc_chain 466 .memory() 467 .read_obj(req_addr as GuestAddress) 468 .map_err(Error::GuestMemory)?; 469 debug!("Map request {:?}", req); 470 471 // Copy the value to use it as a proper reference. 472 let domain_id = req.domain; 473 474 if let Some(domain) = mapping.domains.read().unwrap().get(&domain_id) { 475 if domain.bypass { 476 status = VIRTIO_IOMMU_S_INVAL; 477 return Err(Error::InvalidMapRequestBypassDomain); 478 } 479 } else { 480 status = VIRTIO_IOMMU_S_INVAL; 481 return Err(Error::InvalidMapRequestMissingDomain); 482 } 483 484 // Find the list of endpoints attached to the given domain. 485 let endpoints: Vec<u32> = mapping 486 .endpoints 487 .write() 488 .unwrap() 489 .iter() 490 .filter(|(_, &d)| d == domain_id) 491 .map(|(&e, _)| e) 492 .collect(); 493 494 // Trigger external mapping if necessary. 495 for endpoint in endpoints { 496 if let Some(ext_map) = ext_mapping.get(&endpoint) { 497 let size = req.virt_end - req.virt_start + 1; 498 ext_map 499 .map(req.virt_start, req.phys_start, size) 500 .map_err(Error::ExternalMapping)?; 501 } 502 } 503 504 // Add new mapping associated with the domain 505 mapping 506 .domains 507 .write() 508 .unwrap() 509 .get_mut(&domain_id) 510 .unwrap() 511 .mappings 512 .insert( 513 req.virt_start, 514 Mapping { 515 gpa: req.phys_start, 516 size: req.virt_end - req.virt_start + 1, 517 }, 518 ); 519 } 520 VIRTIO_IOMMU_T_UNMAP => { 521 if desc_size_left != size_of::<VirtioIommuReqUnmap>() { 522 status = VIRTIO_IOMMU_S_INVAL; 523 return Err(Error::InvalidUnmapRequest); 524 } 525 526 let req: VirtioIommuReqUnmap = desc_chain 527 .memory() 528 .read_obj(req_addr as GuestAddress) 529 .map_err(Error::GuestMemory)?; 530 debug!("Unmap request {:?}", req); 531 532 // Copy the value to use it as a proper reference. 533 let domain_id = req.domain; 534 let virt_start = req.virt_start; 535 536 if let Some(domain) = mapping.domains.read().unwrap().get(&domain_id) { 537 if domain.bypass { 538 status = VIRTIO_IOMMU_S_INVAL; 539 return Err(Error::InvalidUnmapRequestBypassDomain); 540 } 541 } else { 542 status = VIRTIO_IOMMU_S_INVAL; 543 return Err(Error::InvalidUnmapRequestMissingDomain); 544 } 545 546 // Find the list of endpoints attached to the given domain. 547 let endpoints: Vec<u32> = mapping 548 .endpoints 549 .write() 550 .unwrap() 551 .iter() 552 .filter(|(_, &d)| d == domain_id) 553 .map(|(&e, _)| e) 554 .collect(); 555 556 // Trigger external unmapping if necessary. 557 for endpoint in endpoints { 558 if let Some(ext_map) = ext_mapping.get(&endpoint) { 559 let size = req.virt_end - virt_start + 1; 560 ext_map 561 .unmap(virt_start, size) 562 .map_err(Error::ExternalUnmapping)?; 563 } 564 } 565 566 // Remove mapping associated with the domain 567 mapping 568 .domains 569 .write() 570 .unwrap() 571 .get_mut(&domain_id) 572 .unwrap() 573 .mappings 574 .remove(&virt_start); 575 } 576 VIRTIO_IOMMU_T_PROBE => { 577 if desc_size_left != size_of::<VirtioIommuReqProbe>() { 578 status = VIRTIO_IOMMU_S_INVAL; 579 return Err(Error::InvalidProbeRequest); 580 } 581 582 let req: VirtioIommuReqProbe = desc_chain 583 .memory() 584 .read_obj(req_addr as GuestAddress) 585 .map_err(Error::GuestMemory)?; 586 debug!("Probe request {:?}", req); 587 588 let probe_prop = VirtioIommuProbeProperty { 589 type_: VIRTIO_IOMMU_PROBE_T_RESV_MEM, 590 length: size_of::<VirtioIommuProbeResvMem>() as u16, 591 }; 592 reply.extend_from_slice(probe_prop.as_slice()); 593 594 let resv_mem = VirtioIommuProbeResvMem { 595 subtype: VIRTIO_IOMMU_RESV_MEM_T_MSI, 596 start: msi_iova_start, 597 end: msi_iova_end, 598 ..Default::default() 599 }; 600 reply.extend_from_slice(resv_mem.as_slice()); 601 602 hdr_len = PROBE_PROP_SIZE; 603 } 604 _ => { 605 status = VIRTIO_IOMMU_S_INVAL; 606 return Err(Error::InvalidRequest); 607 } 608 } 609 Ok(()) 610 })(); 611 612 let status_desc = desc_chain.next().ok_or(Error::DescriptorChainTooShort)?; 613 614 // The status MUST always be writable 615 if !status_desc.is_write_only() { 616 return Err(Error::UnexpectedReadOnlyDescriptor); 617 } 618 619 if status_desc.len() < hdr_len + size_of::<VirtioIommuReqTail>() as u32 { 620 return Err(Error::BufferLengthTooSmall); 621 } 622 623 let tail = VirtioIommuReqTail { 624 status, 625 ..Default::default() 626 }; 627 reply.extend_from_slice(tail.as_slice()); 628 629 // Make sure we return the result of the request to the guest before 630 // we return a potential error internally. 631 desc_chain 632 .memory() 633 .write_slice(reply.as_slice(), status_desc.addr()) 634 .map_err(Error::GuestMemory)?; 635 636 // Return the error if the result was not Ok(). 637 result?; 638 639 Ok((hdr_len as usize) + size_of::<VirtioIommuReqTail>()) 640 } 641 } 642 643 struct IommuEpollHandler { 644 mem: GuestMemoryAtomic<GuestMemoryMmap>, 645 request_queue: Queue, 646 _event_queue: Queue, 647 interrupt_cb: Arc<dyn VirtioInterrupt>, 648 request_queue_evt: EventFd, 649 _event_queue_evt: EventFd, 650 kill_evt: EventFd, 651 pause_evt: EventFd, 652 mapping: Arc<IommuMapping>, 653 ext_mapping: Arc<Mutex<BTreeMap<u32, Arc<dyn ExternalDmaMapping>>>>, 654 msi_iova_space: (u64, u64), 655 } 656 657 impl IommuEpollHandler { 658 fn request_queue(&mut self) -> Result<bool, Error> { 659 let mut used_descs = false; 660 while let Some(mut desc_chain) = self.request_queue.pop_descriptor_chain(self.mem.memory()) 661 { 662 let len = Request::parse( 663 &mut desc_chain, 664 &self.mapping, 665 &self.ext_mapping.lock().unwrap(), 666 self.msi_iova_space, 667 )?; 668 669 self.request_queue 670 .add_used(desc_chain.memory(), desc_chain.head_index(), len as u32) 671 .map_err(Error::QueueAddUsed)?; 672 673 used_descs = true; 674 } 675 676 Ok(used_descs) 677 } 678 679 fn signal_used_queue(&self, queue_index: u16) -> result::Result<(), DeviceError> { 680 self.interrupt_cb 681 .trigger(VirtioInterruptType::Queue(queue_index)) 682 .map_err(|e| { 683 error!("Failed to signal used queue: {:?}", e); 684 DeviceError::FailedSignalingUsedQueue(e) 685 }) 686 } 687 688 fn run( 689 &mut self, 690 paused: Arc<AtomicBool>, 691 paused_sync: Arc<Barrier>, 692 ) -> result::Result<(), EpollHelperError> { 693 let mut helper = EpollHelper::new(&self.kill_evt, &self.pause_evt)?; 694 helper.add_event(self.request_queue_evt.as_raw_fd(), REQUEST_Q_EVENT)?; 695 helper.run(paused, paused_sync, self)?; 696 697 Ok(()) 698 } 699 } 700 701 impl EpollHelperHandler for IommuEpollHandler { 702 fn handle_event( 703 &mut self, 704 _helper: &mut EpollHelper, 705 event: &epoll::Event, 706 ) -> result::Result<(), EpollHelperError> { 707 let ev_type = event.data as u16; 708 match ev_type { 709 REQUEST_Q_EVENT => { 710 self.request_queue_evt.read().map_err(|e| { 711 EpollHelperError::HandleEvent(anyhow!("Failed to get queue event: {:?}", e)) 712 })?; 713 714 let needs_notification = self.request_queue().map_err(|e| { 715 EpollHelperError::HandleEvent(anyhow!( 716 "Failed to process request queue : {:?}", 717 e 718 )) 719 })?; 720 if needs_notification { 721 self.signal_used_queue(0).map_err(|e| { 722 EpollHelperError::HandleEvent(anyhow!( 723 "Failed to signal used queue: {:?}", 724 e 725 )) 726 })?; 727 } 728 } 729 _ => { 730 return Err(EpollHelperError::HandleEvent(anyhow!( 731 "Unexpected event: {}", 732 ev_type 733 ))); 734 } 735 } 736 Ok(()) 737 } 738 } 739 740 #[derive(Clone, Copy, Debug, Versionize)] 741 struct Mapping { 742 gpa: u64, 743 size: u64, 744 } 745 746 #[derive(Clone, Debug)] 747 struct Domain { 748 mappings: BTreeMap<u64, Mapping>, 749 bypass: bool, 750 } 751 752 #[derive(Debug)] 753 pub struct IommuMapping { 754 // Domain related to an endpoint. 755 endpoints: Arc<RwLock<BTreeMap<u32, u32>>>, 756 // Information related to each domain. 757 domains: Arc<RwLock<BTreeMap<u32, Domain>>>, 758 // Global flag indicating if endpoints that are not attached to any domain 759 // are in bypass mode. 760 bypass: AtomicBool, 761 } 762 763 impl DmaRemapping for IommuMapping { 764 fn translate_gva(&self, id: u32, addr: u64) -> std::result::Result<u64, std::io::Error> { 765 debug!("Translate GVA addr 0x{:x}", addr); 766 if let Some(domain_id) = self.endpoints.read().unwrap().get(&id) { 767 if let Some(domain) = self.domains.read().unwrap().get(domain_id) { 768 // Directly return identity mapping in case the domain is in 769 // bypass mode. 770 if domain.bypass { 771 return Ok(addr); 772 } 773 774 let range_start = if VIRTIO_IOMMU_PAGE_SIZE_MASK > addr { 775 0 776 } else { 777 addr - VIRTIO_IOMMU_PAGE_SIZE_MASK 778 }; 779 for (&key, &value) in domain 780 .mappings 781 .range((Included(&range_start), Included(&addr))) 782 { 783 if addr >= key && addr < key + value.size { 784 let new_addr = addr - key + value.gpa; 785 debug!("Into GPA addr 0x{:x}", new_addr); 786 return Ok(new_addr); 787 } 788 } 789 } 790 } else if self.bypass.load(Ordering::Acquire) { 791 return Ok(addr); 792 } 793 794 Err(io::Error::new( 795 io::ErrorKind::Other, 796 format!("failed to translate GVA addr 0x{:x}", addr), 797 )) 798 } 799 800 fn translate_gpa(&self, id: u32, addr: u64) -> std::result::Result<u64, std::io::Error> { 801 debug!("Translate GPA addr 0x{:x}", addr); 802 if let Some(domain_id) = self.endpoints.read().unwrap().get(&id) { 803 if let Some(domain) = self.domains.read().unwrap().get(domain_id) { 804 // Directly return identity mapping in case the domain is in 805 // bypass mode. 806 if domain.bypass { 807 return Ok(addr); 808 } 809 810 for (&key, &value) in domain.mappings.iter() { 811 if addr >= value.gpa && addr < value.gpa + value.size { 812 let new_addr = addr - value.gpa + key; 813 debug!("Into GVA addr 0x{:x}", new_addr); 814 return Ok(new_addr); 815 } 816 } 817 } 818 } else if self.bypass.load(Ordering::Acquire) { 819 return Ok(addr); 820 } 821 822 Err(io::Error::new( 823 io::ErrorKind::Other, 824 format!("failed to translate GPA addr 0x{:x}", addr), 825 )) 826 } 827 } 828 829 #[derive(Debug)] 830 pub struct AccessPlatformMapping { 831 id: u32, 832 mapping: Arc<IommuMapping>, 833 } 834 835 impl AccessPlatformMapping { 836 pub fn new(id: u32, mapping: Arc<IommuMapping>) -> Self { 837 AccessPlatformMapping { id, mapping } 838 } 839 } 840 841 impl AccessPlatform for AccessPlatformMapping { 842 fn translate_gva(&self, base: u64, _size: u64) -> std::result::Result<u64, std::io::Error> { 843 self.mapping.translate_gva(self.id, base) 844 } 845 fn translate_gpa(&self, base: u64, _size: u64) -> std::result::Result<u64, std::io::Error> { 846 self.mapping.translate_gpa(self.id, base) 847 } 848 } 849 850 pub struct Iommu { 851 common: VirtioCommon, 852 id: String, 853 config: VirtioIommuConfig, 854 mapping: Arc<IommuMapping>, 855 ext_mapping: Arc<Mutex<BTreeMap<u32, Arc<dyn ExternalDmaMapping>>>>, 856 seccomp_action: SeccompAction, 857 exit_evt: EventFd, 858 msi_iova_space: (u64, u64), 859 } 860 861 type EndpointsState = Vec<(u32, u32)>; 862 type DomainsState = Vec<(u32, (Vec<(u64, Mapping)>, bool))>; 863 864 #[derive(Versionize)] 865 pub struct IommuState { 866 avail_features: u64, 867 acked_features: u64, 868 endpoints: EndpointsState, 869 domains: DomainsState, 870 } 871 872 impl VersionMapped for IommuState {} 873 874 impl Iommu { 875 pub fn new( 876 id: String, 877 seccomp_action: SeccompAction, 878 exit_evt: EventFd, 879 msi_iova_space: (u64, u64), 880 state: Option<IommuState>, 881 ) -> io::Result<(Self, Arc<IommuMapping>)> { 882 let (avail_features, acked_features, endpoints, domains) = if let Some(state) = state { 883 info!("Restoring virtio-iommu {}", id); 884 ( 885 state.avail_features, 886 state.acked_features, 887 state.endpoints.into_iter().collect(), 888 state 889 .domains 890 .into_iter() 891 .map(|(k, v)| { 892 ( 893 k, 894 Domain { 895 mappings: v.0.into_iter().collect(), 896 bypass: v.1, 897 }, 898 ) 899 }) 900 .collect(), 901 ) 902 } else { 903 let avail_features = 1u64 << VIRTIO_F_VERSION_1 904 | 1u64 << VIRTIO_IOMMU_F_MAP_UNMAP 905 | 1u64 << VIRTIO_IOMMU_F_PROBE 906 | 1u64 << VIRTIO_IOMMU_F_BYPASS_CONFIG; 907 908 (avail_features, 0, BTreeMap::new(), BTreeMap::new()) 909 }; 910 911 let config = VirtioIommuConfig { 912 page_size_mask: VIRTIO_IOMMU_PAGE_SIZE_MASK, 913 probe_size: PROBE_PROP_SIZE, 914 ..Default::default() 915 }; 916 917 let mapping = Arc::new(IommuMapping { 918 endpoints: Arc::new(RwLock::new(endpoints)), 919 domains: Arc::new(RwLock::new(domains)), 920 bypass: AtomicBool::new(true), 921 }); 922 923 Ok(( 924 Iommu { 925 id, 926 common: VirtioCommon { 927 device_type: VirtioDeviceType::Iommu as u32, 928 queue_sizes: QUEUE_SIZES.to_vec(), 929 avail_features, 930 acked_features, 931 paused_sync: Some(Arc::new(Barrier::new(2))), 932 min_queues: NUM_QUEUES as u16, 933 ..Default::default() 934 }, 935 config, 936 mapping: mapping.clone(), 937 ext_mapping: Arc::new(Mutex::new(BTreeMap::new())), 938 seccomp_action, 939 exit_evt, 940 msi_iova_space, 941 }, 942 mapping, 943 )) 944 } 945 946 fn state(&self) -> IommuState { 947 IommuState { 948 avail_features: self.common.avail_features, 949 acked_features: self.common.acked_features, 950 endpoints: self 951 .mapping 952 .endpoints 953 .read() 954 .unwrap() 955 .clone() 956 .into_iter() 957 .collect(), 958 domains: self 959 .mapping 960 .domains 961 .read() 962 .unwrap() 963 .clone() 964 .into_iter() 965 .map(|(k, v)| (k, (v.mappings.into_iter().collect(), v.bypass))) 966 .collect(), 967 } 968 } 969 970 fn update_bypass(&mut self) { 971 // Use bypass from config if VIRTIO_IOMMU_F_BYPASS_CONFIG has been negotiated 972 if !self 973 .common 974 .feature_acked(VIRTIO_IOMMU_F_BYPASS_CONFIG.into()) 975 { 976 return; 977 } 978 979 let bypass = self.config.bypass == 1; 980 info!("Updating bypass mode to {}", bypass); 981 self.mapping.bypass.store(bypass, Ordering::Release); 982 } 983 984 pub fn add_external_mapping(&mut self, device_id: u32, mapping: Arc<dyn ExternalDmaMapping>) { 985 self.ext_mapping.lock().unwrap().insert(device_id, mapping); 986 } 987 988 #[cfg(fuzzing)] 989 pub fn wait_for_epoll_threads(&mut self) { 990 self.common.wait_for_epoll_threads(); 991 } 992 } 993 994 impl Drop for Iommu { 995 fn drop(&mut self) { 996 if let Some(kill_evt) = self.common.kill_evt.take() { 997 // Ignore the result because there is nothing we can do about it. 998 let _ = kill_evt.write(1); 999 } 1000 } 1001 } 1002 1003 impl VirtioDevice for Iommu { 1004 fn device_type(&self) -> u32 { 1005 self.common.device_type 1006 } 1007 1008 fn queue_max_sizes(&self) -> &[u16] { 1009 &self.common.queue_sizes 1010 } 1011 1012 fn features(&self) -> u64 { 1013 self.common.avail_features 1014 } 1015 1016 fn ack_features(&mut self, value: u64) { 1017 self.common.ack_features(value) 1018 } 1019 1020 fn read_config(&self, offset: u64, data: &mut [u8]) { 1021 self.read_config_from_slice(self.config.as_slice(), offset, data); 1022 } 1023 1024 fn write_config(&mut self, offset: u64, data: &[u8]) { 1025 // The "bypass" field is the only mutable field 1026 let bypass_offset = 1027 (&self.config.bypass as *const _ as u64) - (&self.config as *const _ as u64); 1028 if offset != bypass_offset || data.len() != std::mem::size_of_val(&self.config.bypass) { 1029 error!( 1030 "Attempt to write to read-only field: offset {:x} length {}", 1031 offset, 1032 data.len() 1033 ); 1034 return; 1035 } 1036 1037 self.config.bypass = data[0]; 1038 1039 self.update_bypass(); 1040 } 1041 1042 fn activate( 1043 &mut self, 1044 mem: GuestMemoryAtomic<GuestMemoryMmap>, 1045 interrupt_cb: Arc<dyn VirtioInterrupt>, 1046 mut queues: Vec<(usize, Queue, EventFd)>, 1047 ) -> ActivateResult { 1048 self.common.activate(&queues, &interrupt_cb)?; 1049 let (kill_evt, pause_evt) = self.common.dup_eventfds(); 1050 1051 let (_, request_queue, request_queue_evt) = queues.remove(0); 1052 let (_, _event_queue, _event_queue_evt) = queues.remove(0); 1053 1054 let mut handler = IommuEpollHandler { 1055 mem, 1056 request_queue, 1057 _event_queue, 1058 interrupt_cb, 1059 request_queue_evt, 1060 _event_queue_evt, 1061 kill_evt, 1062 pause_evt, 1063 mapping: self.mapping.clone(), 1064 ext_mapping: self.ext_mapping.clone(), 1065 msi_iova_space: self.msi_iova_space, 1066 }; 1067 1068 let paused = self.common.paused.clone(); 1069 let paused_sync = self.common.paused_sync.clone(); 1070 let mut epoll_threads = Vec::new(); 1071 spawn_virtio_thread( 1072 &self.id, 1073 &self.seccomp_action, 1074 Thread::VirtioIommu, 1075 &mut epoll_threads, 1076 &self.exit_evt, 1077 move || handler.run(paused, paused_sync.unwrap()), 1078 )?; 1079 1080 self.common.epoll_threads = Some(epoll_threads); 1081 1082 event!("virtio-device", "activated", "id", &self.id); 1083 Ok(()) 1084 } 1085 1086 fn reset(&mut self) -> Option<Arc<dyn VirtioInterrupt>> { 1087 let result = self.common.reset(); 1088 event!("virtio-device", "reset", "id", &self.id); 1089 result 1090 } 1091 } 1092 1093 impl Pausable for Iommu { 1094 fn pause(&mut self) -> result::Result<(), MigratableError> { 1095 self.common.pause() 1096 } 1097 1098 fn resume(&mut self) -> result::Result<(), MigratableError> { 1099 self.common.resume() 1100 } 1101 } 1102 1103 impl Snapshottable for Iommu { 1104 fn id(&self) -> String { 1105 self.id.clone() 1106 } 1107 1108 fn snapshot(&mut self) -> std::result::Result<Snapshot, MigratableError> { 1109 Snapshot::new_from_versioned_state(&self.id, &self.state()) 1110 } 1111 } 1112 impl Transportable for Iommu {} 1113 impl Migratable for Iommu {} 1114