1 // Copyright 2018 Amazon.com, Inc. or its affiliates. All Rights Reserved. 2 // 3 // Portions Copyright 2017 The Chromium OS Authors. All rights reserved. 4 // Use of this source code is governed by a BSD-style license that can be 5 // found in the LICENSE-BSD-3-Clause file. 6 // 7 // Copyright © 2020 Intel Corporation 8 // 9 // SPDX-License-Identifier: Apache-2.0 AND BSD-3-Clause 10 11 use std::collections::{BTreeMap, HashMap, VecDeque}; 12 use std::num::Wrapping; 13 use std::ops::Deref; 14 use std::os::unix::io::AsRawFd; 15 use std::path::PathBuf; 16 use std::sync::atomic::{AtomicBool, AtomicU64, Ordering}; 17 use std::sync::{Arc, Barrier}; 18 use std::{io, result}; 19 20 use anyhow::anyhow; 21 use block::async_io::{AsyncIo, AsyncIoError, DiskFile}; 22 use block::{build_serial, Request, RequestType, VirtioBlockConfig}; 23 use rate_limiter::group::{RateLimiterGroup, RateLimiterGroupHandle}; 24 use rate_limiter::TokenType; 25 use seccompiler::SeccompAction; 26 use serde::{Deserialize, Serialize}; 27 use thiserror::Error; 28 use virtio_bindings::virtio_blk::*; 29 use virtio_bindings::virtio_config::*; 30 use virtio_bindings::virtio_ring::{VIRTIO_RING_F_EVENT_IDX, VIRTIO_RING_F_INDIRECT_DESC}; 31 use virtio_queue::{Queue, QueueOwnedT, QueueT}; 32 use vm_memory::{ByteValued, Bytes, GuestAddressSpace, GuestMemoryAtomic, GuestMemoryError}; 33 use vm_migration::{Migratable, MigratableError, Pausable, Snapshot, Snapshottable, Transportable}; 34 use vm_virtio::AccessPlatform; 35 use vmm_sys_util::eventfd::EventFd; 36 37 use super::{ 38 ActivateError, ActivateResult, EpollHelper, EpollHelperError, EpollHelperHandler, 39 Error as DeviceError, VirtioCommon, VirtioDevice, VirtioDeviceType, VirtioInterruptType, 40 EPOLL_HELPER_EVENT_LAST, 41 }; 42 use crate::seccomp_filters::Thread; 43 use crate::thread_helper::spawn_virtio_thread; 44 use crate::{GuestMemoryMmap, VirtioInterrupt}; 45 46 const SECTOR_SHIFT: u8 = 9; 47 pub const SECTOR_SIZE: u64 = 0x01 << SECTOR_SHIFT; 48 49 // New descriptors are pending on the virtio queue. 50 const QUEUE_AVAIL_EVENT: u16 = EPOLL_HELPER_EVENT_LAST + 1; 51 // New completed tasks are pending on the completion ring. 52 const COMPLETION_EVENT: u16 = EPOLL_HELPER_EVENT_LAST + 2; 53 // New 'wake up' event from the rate limiter 54 const RATE_LIMITER_EVENT: u16 = EPOLL_HELPER_EVENT_LAST + 3; 55 56 // latency scale, for reduce precision loss in calculate. 57 const LATENCY_SCALE: u64 = 10000; 58 59 pub const MINIMUM_BLOCK_QUEUE_SIZE: u16 = 2; 60 61 #[derive(Error, Debug)] 62 pub enum Error { 63 #[error("Failed to parse the request: {0}")] 64 RequestParsing(block::Error), 65 #[error("Failed to execute the request: {0}")] 66 RequestExecuting(block::ExecuteError), 67 #[error("Failed to complete the request: {0}")] 68 RequestCompleting(block::Error), 69 #[error("Missing the expected entry in the list of requests")] 70 MissingEntryRequestList, 71 #[error("The asynchronous request returned with failure")] 72 AsyncRequestFailure, 73 #[error("Failed synchronizing the file: {0}")] 74 Fsync(AsyncIoError), 75 #[error("Failed adding used index: {0}")] 76 QueueAddUsed(virtio_queue::Error), 77 #[error("Failed creating an iterator over the queue: {0}")] 78 QueueIterator(virtio_queue::Error), 79 #[error("Failed to update request status: {0}")] 80 RequestStatus(GuestMemoryError), 81 #[error("Failed to enable notification: {0}")] 82 QueueEnableNotification(virtio_queue::Error), 83 } 84 85 pub type Result<T> = result::Result<T, Error>; 86 87 // latency will be records as microseconds, average latency 88 // will be save as scaled value. 89 #[derive(Clone)] 90 pub struct BlockCounters { 91 read_bytes: Arc<AtomicU64>, 92 read_ops: Arc<AtomicU64>, 93 read_latency_min: Arc<AtomicU64>, 94 read_latency_max: Arc<AtomicU64>, 95 read_latency_avg: Arc<AtomicU64>, 96 write_bytes: Arc<AtomicU64>, 97 write_ops: Arc<AtomicU64>, 98 write_latency_min: Arc<AtomicU64>, 99 write_latency_max: Arc<AtomicU64>, 100 write_latency_avg: Arc<AtomicU64>, 101 } 102 103 impl Default for BlockCounters { 104 fn default() -> Self { 105 BlockCounters { 106 read_bytes: Arc::new(AtomicU64::new(0)), 107 read_ops: Arc::new(AtomicU64::new(0)), 108 read_latency_min: Arc::new(AtomicU64::new(u64::MAX)), 109 read_latency_max: Arc::new(AtomicU64::new(u64::MAX)), 110 read_latency_avg: Arc::new(AtomicU64::new(u64::MAX)), 111 write_bytes: Arc::new(AtomicU64::new(0)), 112 write_ops: Arc::new(AtomicU64::new(0)), 113 write_latency_min: Arc::new(AtomicU64::new(u64::MAX)), 114 write_latency_max: Arc::new(AtomicU64::new(u64::MAX)), 115 write_latency_avg: Arc::new(AtomicU64::new(u64::MAX)), 116 } 117 } 118 } 119 120 struct BlockEpollHandler { 121 queue_index: u16, 122 queue: Queue, 123 mem: GuestMemoryAtomic<GuestMemoryMmap>, 124 disk_image: Box<dyn AsyncIo>, 125 disk_nsectors: u64, 126 interrupt_cb: Arc<dyn VirtioInterrupt>, 127 serial: Vec<u8>, 128 kill_evt: EventFd, 129 pause_evt: EventFd, 130 writeback: Arc<AtomicBool>, 131 counters: BlockCounters, 132 queue_evt: EventFd, 133 inflight_requests: VecDeque<(u16, Request)>, 134 rate_limiter: Option<RateLimiterGroupHandle>, 135 access_platform: Option<Arc<dyn AccessPlatform>>, 136 read_only: bool, 137 host_cpus: Option<Vec<usize>>, 138 } 139 140 impl BlockEpollHandler { 141 fn process_queue_submit(&mut self) -> Result<()> { 142 let queue = &mut self.queue; 143 144 while let Some(mut desc_chain) = queue.pop_descriptor_chain(self.mem.memory()) { 145 let mut request = Request::parse(&mut desc_chain, self.access_platform.as_ref()) 146 .map_err(Error::RequestParsing)?; 147 148 // For virtio spec compliance 149 // "A device MUST set the status byte to VIRTIO_BLK_S_IOERR for a write request 150 // if the VIRTIO_BLK_F_RO feature if offered, and MUST NOT write any data." 151 if self.read_only 152 && (request.request_type == RequestType::Out 153 || request.request_type == RequestType::Flush) 154 { 155 desc_chain 156 .memory() 157 .write_obj(VIRTIO_BLK_S_IOERR, request.status_addr) 158 .map_err(Error::RequestStatus)?; 159 160 // If no asynchronous operation has been submitted, we can 161 // simply return the used descriptor. 162 queue 163 .add_used(desc_chain.memory(), desc_chain.head_index(), 0) 164 .map_err(Error::QueueAddUsed)?; 165 queue 166 .enable_notification(self.mem.memory().deref()) 167 .map_err(Error::QueueEnableNotification)?; 168 continue; 169 } 170 171 if let Some(rate_limiter) = &mut self.rate_limiter { 172 // If limiter.consume() fails it means there is no more TokenType::Ops 173 // budget and rate limiting is in effect. 174 if !rate_limiter.consume(1, TokenType::Ops) { 175 // Stop processing the queue and return this descriptor chain to the 176 // avail ring, for later processing. 177 queue.go_to_previous_position(); 178 break; 179 } 180 // Exercise the rate limiter only if this request is of data transfer type. 181 if request.request_type == RequestType::In 182 || request.request_type == RequestType::Out 183 { 184 let mut bytes = Wrapping(0); 185 for (_, data_len) in &request.data_descriptors { 186 bytes += Wrapping(*data_len as u64); 187 } 188 189 // If limiter.consume() fails it means there is no more TokenType::Bytes 190 // budget and rate limiting is in effect. 191 if !rate_limiter.consume(bytes.0, TokenType::Bytes) { 192 // Revert the OPS consume(). 193 rate_limiter.manual_replenish(1, TokenType::Ops); 194 // Stop processing the queue and return this descriptor chain to the 195 // avail ring, for later processing. 196 queue.go_to_previous_position(); 197 break; 198 } 199 }; 200 } 201 202 request.set_writeback(self.writeback.load(Ordering::Acquire)); 203 204 if request 205 .execute_async( 206 desc_chain.memory(), 207 self.disk_nsectors, 208 self.disk_image.as_mut(), 209 &self.serial, 210 desc_chain.head_index() as u64, 211 ) 212 .map_err(Error::RequestExecuting)? 213 { 214 self.inflight_requests 215 .push_back((desc_chain.head_index(), request)); 216 } else { 217 desc_chain 218 .memory() 219 .write_obj(VIRTIO_BLK_S_OK as u8, request.status_addr) 220 .map_err(Error::RequestStatus)?; 221 222 // If no asynchronous operation has been submitted, we can 223 // simply return the used descriptor. 224 queue 225 .add_used(desc_chain.memory(), desc_chain.head_index(), 0) 226 .map_err(Error::QueueAddUsed)?; 227 queue 228 .enable_notification(self.mem.memory().deref()) 229 .map_err(Error::QueueEnableNotification)?; 230 } 231 } 232 233 Ok(()) 234 } 235 236 fn try_signal_used_queue(&mut self) -> result::Result<(), EpollHelperError> { 237 if self 238 .queue 239 .needs_notification(self.mem.memory().deref()) 240 .map_err(|e| { 241 EpollHelperError::HandleEvent(anyhow!( 242 "Failed to check needs_notification: {:?}", 243 e 244 )) 245 })? 246 { 247 self.signal_used_queue().map_err(|e| { 248 EpollHelperError::HandleEvent(anyhow!("Failed to signal used queue: {:?}", e)) 249 })?; 250 } 251 252 Ok(()) 253 } 254 255 fn process_queue_submit_and_signal(&mut self) -> result::Result<(), EpollHelperError> { 256 self.process_queue_submit().map_err(|e| { 257 EpollHelperError::HandleEvent(anyhow!("Failed to process queue (submit): {:?}", e)) 258 })?; 259 260 self.try_signal_used_queue() 261 } 262 263 #[inline] 264 fn find_inflight_request(&mut self, completed_head: u16) -> Result<Request> { 265 // This loop neatly handles the fast path where the completions are 266 // in order (it turns into just a pop_front()) and the 1% of the time 267 // (analysis during boot) where slight out of ordering has been 268 // observed e.g. 269 // Submissions: 1 2 3 4 5 6 7 270 // Completions: 2 1 3 5 4 7 6 271 // In this case find the corresponding item and swap it with the front 272 // This is a O(1) operation and is prepared for the future as it it likely 273 // the next completion would be for the one that was skipped which will 274 // now be the new front. 275 for (i, (head, _)) in self.inflight_requests.iter().enumerate() { 276 if head == &completed_head { 277 return Ok(self.inflight_requests.swap_remove_front(i).unwrap().1); 278 } 279 } 280 281 Err(Error::MissingEntryRequestList) 282 } 283 284 fn process_queue_complete(&mut self) -> Result<()> { 285 let mem = self.mem.memory(); 286 let mut read_bytes = Wrapping(0); 287 let mut write_bytes = Wrapping(0); 288 let mut read_ops = Wrapping(0); 289 let mut write_ops = Wrapping(0); 290 291 while let Some((user_data, result)) = self.disk_image.next_completed_request() { 292 let desc_index = user_data as u16; 293 294 let mut request = self.find_inflight_request(desc_index)?; 295 296 request.complete_async().map_err(Error::RequestCompleting)?; 297 298 let latency = request.start.elapsed().as_micros() as u64; 299 let read_ops_last = self.counters.read_ops.load(Ordering::Relaxed); 300 let write_ops_last = self.counters.write_ops.load(Ordering::Relaxed); 301 let read_max = self.counters.read_latency_max.load(Ordering::Relaxed); 302 let write_max = self.counters.write_latency_max.load(Ordering::Relaxed); 303 let mut read_avg = self.counters.read_latency_avg.load(Ordering::Relaxed); 304 let mut write_avg = self.counters.write_latency_avg.load(Ordering::Relaxed); 305 let (status, len) = if result >= 0 { 306 match request.request_type { 307 RequestType::In => { 308 for (_, data_len) in &request.data_descriptors { 309 read_bytes += Wrapping(*data_len as u64); 310 } 311 read_ops += Wrapping(1); 312 if latency < self.counters.read_latency_min.load(Ordering::Relaxed) { 313 self.counters 314 .read_latency_min 315 .store(latency, Ordering::Relaxed); 316 } 317 if latency > read_max || read_max == u64::MAX { 318 self.counters 319 .read_latency_max 320 .store(latency, Ordering::Relaxed); 321 } 322 323 // Special case the first real latency report 324 read_avg = if read_avg == u64::MAX { 325 latency * LATENCY_SCALE 326 } else { 327 // Cumulative average is guaranteed to be 328 // positive if being calculated properly 329 (read_avg as i64 330 + ((latency * LATENCY_SCALE) as i64 - read_avg as i64) 331 / (read_ops_last + read_ops.0) as i64) 332 .try_into() 333 .unwrap() 334 }; 335 } 336 RequestType::Out => { 337 if !request.writeback { 338 self.disk_image.fsync(None).map_err(Error::Fsync)?; 339 } 340 for (_, data_len) in &request.data_descriptors { 341 write_bytes += Wrapping(*data_len as u64); 342 } 343 write_ops += Wrapping(1); 344 if latency < self.counters.write_latency_min.load(Ordering::Relaxed) { 345 self.counters 346 .write_latency_min 347 .store(latency, Ordering::Relaxed); 348 } 349 if latency > write_max || write_max == u64::MAX { 350 self.counters 351 .write_latency_max 352 .store(latency, Ordering::Relaxed); 353 } 354 355 // Special case the first real latency report 356 write_avg = if write_avg == u64::MAX { 357 latency * LATENCY_SCALE 358 } else { 359 // Cumulative average is guaranteed to be 360 // positive if being calculated properly 361 (write_avg as i64 362 + ((latency * LATENCY_SCALE) as i64 - write_avg as i64) 363 / (write_ops_last + write_ops.0) as i64) 364 .try_into() 365 .unwrap() 366 } 367 } 368 _ => {} 369 } 370 371 self.counters 372 .read_latency_avg 373 .store(read_avg, Ordering::Relaxed); 374 375 self.counters 376 .write_latency_avg 377 .store(write_avg, Ordering::Relaxed); 378 379 (VIRTIO_BLK_S_OK as u8, result as u32) 380 } else { 381 error!( 382 "Request failed: {:x?} {:?}", 383 request, 384 io::Error::from_raw_os_error(-result) 385 ); 386 return Err(Error::AsyncRequestFailure); 387 }; 388 389 mem.write_obj(status, request.status_addr) 390 .map_err(Error::RequestStatus)?; 391 392 let queue = &mut self.queue; 393 394 queue 395 .add_used(mem.deref(), desc_index, len) 396 .map_err(Error::QueueAddUsed)?; 397 queue 398 .enable_notification(mem.deref()) 399 .map_err(Error::QueueEnableNotification)?; 400 } 401 402 self.counters 403 .write_bytes 404 .fetch_add(write_bytes.0, Ordering::AcqRel); 405 self.counters 406 .write_ops 407 .fetch_add(write_ops.0, Ordering::AcqRel); 408 409 self.counters 410 .read_bytes 411 .fetch_add(read_bytes.0, Ordering::AcqRel); 412 self.counters 413 .read_ops 414 .fetch_add(read_ops.0, Ordering::AcqRel); 415 416 Ok(()) 417 } 418 419 fn signal_used_queue(&self) -> result::Result<(), DeviceError> { 420 self.interrupt_cb 421 .trigger(VirtioInterruptType::Queue(self.queue_index)) 422 .map_err(|e| { 423 error!("Failed to signal used queue: {:?}", e); 424 DeviceError::FailedSignalingUsedQueue(e) 425 }) 426 } 427 428 fn set_queue_thread_affinity(&self) { 429 // Prepare the CPU set the current queue thread is expected to run onto. 430 let cpuset = self.host_cpus.as_ref().map(|host_cpus| { 431 // SAFETY: all zeros is a valid pattern 432 let mut cpuset: libc::cpu_set_t = unsafe { std::mem::zeroed() }; 433 // SAFETY: FFI call, trivially safe 434 unsafe { libc::CPU_ZERO(&mut cpuset) }; 435 for host_cpu in host_cpus { 436 // SAFETY: FFI call, trivially safe 437 unsafe { libc::CPU_SET(*host_cpu, &mut cpuset) }; 438 } 439 cpuset 440 }); 441 442 // Schedule the thread to run on the expected CPU set 443 if let Some(cpuset) = cpuset.as_ref() { 444 // SAFETY: FFI call with correct arguments 445 let ret = unsafe { 446 libc::sched_setaffinity( 447 0, 448 std::mem::size_of::<libc::cpu_set_t>(), 449 cpuset as *const libc::cpu_set_t, 450 ) 451 }; 452 453 if ret != 0 { 454 error!( 455 "Failed scheduling the virtqueue thread {} on the expected CPU set: {}", 456 self.queue_index, 457 io::Error::last_os_error() 458 ) 459 } 460 } 461 } 462 463 fn run( 464 &mut self, 465 paused: Arc<AtomicBool>, 466 paused_sync: Arc<Barrier>, 467 ) -> result::Result<(), EpollHelperError> { 468 let mut helper = EpollHelper::new(&self.kill_evt, &self.pause_evt)?; 469 helper.add_event(self.queue_evt.as_raw_fd(), QUEUE_AVAIL_EVENT)?; 470 helper.add_event(self.disk_image.notifier().as_raw_fd(), COMPLETION_EVENT)?; 471 if let Some(rate_limiter) = &self.rate_limiter { 472 helper.add_event(rate_limiter.as_raw_fd(), RATE_LIMITER_EVENT)?; 473 } 474 self.set_queue_thread_affinity(); 475 helper.run(paused, paused_sync, self)?; 476 477 Ok(()) 478 } 479 } 480 481 impl EpollHelperHandler for BlockEpollHandler { 482 fn handle_event( 483 &mut self, 484 _helper: &mut EpollHelper, 485 event: &epoll::Event, 486 ) -> result::Result<(), EpollHelperError> { 487 let ev_type = event.data as u16; 488 match ev_type { 489 QUEUE_AVAIL_EVENT => { 490 self.queue_evt.read().map_err(|e| { 491 EpollHelperError::HandleEvent(anyhow!("Failed to get queue event: {:?}", e)) 492 })?; 493 494 let rate_limit_reached = self.rate_limiter.as_ref().is_some_and(|r| r.is_blocked()); 495 496 // Process the queue only when the rate limit is not reached 497 if !rate_limit_reached { 498 self.process_queue_submit_and_signal()? 499 } 500 } 501 COMPLETION_EVENT => { 502 self.disk_image.notifier().read().map_err(|e| { 503 EpollHelperError::HandleEvent(anyhow!("Failed to get queue event: {:?}", e)) 504 })?; 505 506 self.process_queue_complete().map_err(|e| { 507 EpollHelperError::HandleEvent(anyhow!( 508 "Failed to process queue (complete): {:?}", 509 e 510 )) 511 })?; 512 513 let rate_limit_reached = self.rate_limiter.as_ref().is_some_and(|r| r.is_blocked()); 514 515 // Process the queue only when the rate limit is not reached 516 if !rate_limit_reached { 517 self.process_queue_submit().map_err(|e| { 518 EpollHelperError::HandleEvent(anyhow!( 519 "Failed to process queue (submit): {:?}", 520 e 521 )) 522 })?; 523 } 524 self.try_signal_used_queue()?; 525 } 526 RATE_LIMITER_EVENT => { 527 if let Some(rate_limiter) = &mut self.rate_limiter { 528 // Upon rate limiter event, call the rate limiter handler 529 // and restart processing the queue. 530 rate_limiter.event_handler().map_err(|e| { 531 EpollHelperError::HandleEvent(anyhow!( 532 "Failed to process rate limiter event: {:?}", 533 e 534 )) 535 })?; 536 537 self.process_queue_submit_and_signal()? 538 } else { 539 return Err(EpollHelperError::HandleEvent(anyhow!( 540 "Unexpected 'RATE_LIMITER_EVENT' when rate_limiter is not enabled." 541 ))); 542 } 543 } 544 _ => { 545 return Err(EpollHelperError::HandleEvent(anyhow!( 546 "Unexpected event: {}", 547 ev_type 548 ))); 549 } 550 } 551 Ok(()) 552 } 553 } 554 555 /// Virtio device for exposing block level read/write operations on a host file. 556 pub struct Block { 557 common: VirtioCommon, 558 id: String, 559 disk_image: Box<dyn DiskFile>, 560 disk_path: PathBuf, 561 disk_nsectors: u64, 562 config: VirtioBlockConfig, 563 writeback: Arc<AtomicBool>, 564 counters: BlockCounters, 565 seccomp_action: SeccompAction, 566 rate_limiter: Option<Arc<RateLimiterGroup>>, 567 exit_evt: EventFd, 568 read_only: bool, 569 serial: Vec<u8>, 570 queue_affinity: BTreeMap<u16, Vec<usize>>, 571 } 572 573 #[derive(Serialize, Deserialize)] 574 pub struct BlockState { 575 pub disk_path: String, 576 pub disk_nsectors: u64, 577 pub avail_features: u64, 578 pub acked_features: u64, 579 pub config: VirtioBlockConfig, 580 } 581 582 impl Block { 583 /// Create a new virtio block device that operates on the given file. 584 #[allow(clippy::too_many_arguments)] 585 pub fn new( 586 id: String, 587 mut disk_image: Box<dyn DiskFile>, 588 disk_path: PathBuf, 589 read_only: bool, 590 iommu: bool, 591 num_queues: usize, 592 queue_size: u16, 593 serial: Option<String>, 594 seccomp_action: SeccompAction, 595 rate_limiter: Option<Arc<RateLimiterGroup>>, 596 exit_evt: EventFd, 597 state: Option<BlockState>, 598 queue_affinity: BTreeMap<u16, Vec<usize>>, 599 ) -> io::Result<Self> { 600 let (disk_nsectors, avail_features, acked_features, config, paused) = 601 if let Some(state) = state { 602 info!("Restoring virtio-block {}", id); 603 ( 604 state.disk_nsectors, 605 state.avail_features, 606 state.acked_features, 607 state.config, 608 true, 609 ) 610 } else { 611 let disk_size = disk_image 612 .size() 613 .map_err(|e| io::Error::other(format!("Failed getting disk size: {e}")))?; 614 if disk_size % SECTOR_SIZE != 0 { 615 warn!( 616 "Disk size {} is not a multiple of sector size {}; \ 617 the remainder will not be visible to the guest.", 618 disk_size, SECTOR_SIZE 619 ); 620 } 621 622 let mut avail_features = (1u64 << VIRTIO_F_VERSION_1) 623 | (1u64 << VIRTIO_BLK_F_FLUSH) 624 | (1u64 << VIRTIO_BLK_F_CONFIG_WCE) 625 | (1u64 << VIRTIO_BLK_F_BLK_SIZE) 626 | (1u64 << VIRTIO_BLK_F_TOPOLOGY) 627 | (1u64 << VIRTIO_BLK_F_SEG_MAX) 628 | (1u64 << VIRTIO_RING_F_EVENT_IDX) 629 | (1u64 << VIRTIO_RING_F_INDIRECT_DESC); 630 if iommu { 631 avail_features |= 1u64 << VIRTIO_F_IOMMU_PLATFORM; 632 } 633 634 if read_only { 635 avail_features |= 1u64 << VIRTIO_BLK_F_RO; 636 } 637 638 let topology = disk_image.topology(); 639 info!("Disk topology: {:?}", topology); 640 641 let logical_block_size = if topology.logical_block_size > 512 { 642 topology.logical_block_size 643 } else { 644 512 645 }; 646 647 // Calculate the exponent that maps physical block to logical block 648 let mut physical_block_exp = 0; 649 let mut size = logical_block_size; 650 while size < topology.physical_block_size { 651 physical_block_exp += 1; 652 size <<= 1; 653 } 654 655 let disk_nsectors = disk_size / SECTOR_SIZE; 656 let mut config = VirtioBlockConfig { 657 capacity: disk_nsectors, 658 writeback: 1, 659 blk_size: topology.logical_block_size as u32, 660 physical_block_exp, 661 min_io_size: (topology.minimum_io_size / logical_block_size) as u16, 662 opt_io_size: (topology.optimal_io_size / logical_block_size) as u32, 663 seg_max: (queue_size - MINIMUM_BLOCK_QUEUE_SIZE) as u32, 664 ..Default::default() 665 }; 666 667 if num_queues > 1 { 668 avail_features |= 1u64 << VIRTIO_BLK_F_MQ; 669 config.num_queues = num_queues as u16; 670 } 671 672 (disk_nsectors, avail_features, 0, config, false) 673 }; 674 675 let serial = serial 676 .map(Vec::from) 677 .unwrap_or_else(|| build_serial(&disk_path)); 678 679 Ok(Block { 680 common: VirtioCommon { 681 device_type: VirtioDeviceType::Block as u32, 682 avail_features, 683 acked_features, 684 paused_sync: Some(Arc::new(Barrier::new(num_queues + 1))), 685 queue_sizes: vec![queue_size; num_queues], 686 min_queues: 1, 687 paused: Arc::new(AtomicBool::new(paused)), 688 ..Default::default() 689 }, 690 id, 691 disk_image, 692 disk_path, 693 disk_nsectors, 694 config, 695 writeback: Arc::new(AtomicBool::new(true)), 696 counters: BlockCounters::default(), 697 seccomp_action, 698 rate_limiter, 699 exit_evt, 700 read_only, 701 serial, 702 queue_affinity, 703 }) 704 } 705 706 fn state(&self) -> BlockState { 707 BlockState { 708 disk_path: self.disk_path.to_str().unwrap().to_owned(), 709 disk_nsectors: self.disk_nsectors, 710 avail_features: self.common.avail_features, 711 acked_features: self.common.acked_features, 712 config: self.config, 713 } 714 } 715 716 fn update_writeback(&mut self) { 717 // Use writeback from config if VIRTIO_BLK_F_CONFIG_WCE 718 let writeback = if self.common.feature_acked(VIRTIO_BLK_F_CONFIG_WCE.into()) { 719 self.config.writeback == 1 720 } else { 721 // Else check if VIRTIO_BLK_F_FLUSH negotiated 722 self.common.feature_acked(VIRTIO_BLK_F_FLUSH.into()) 723 }; 724 725 info!( 726 "Changing cache mode to {}", 727 if writeback { 728 "writeback" 729 } else { 730 "writethrough" 731 } 732 ); 733 self.writeback.store(writeback, Ordering::Release); 734 } 735 736 #[cfg(fuzzing)] 737 pub fn wait_for_epoll_threads(&mut self) { 738 self.common.wait_for_epoll_threads(); 739 } 740 } 741 742 impl Drop for Block { 743 fn drop(&mut self) { 744 if let Some(kill_evt) = self.common.kill_evt.take() { 745 // Ignore the result because there is nothing we can do about it. 746 let _ = kill_evt.write(1); 747 } 748 self.common.wait_for_epoll_threads(); 749 } 750 } 751 752 impl VirtioDevice for Block { 753 fn device_type(&self) -> u32 { 754 self.common.device_type 755 } 756 757 fn queue_max_sizes(&self) -> &[u16] { 758 &self.common.queue_sizes 759 } 760 761 fn features(&self) -> u64 { 762 self.common.avail_features 763 } 764 765 fn ack_features(&mut self, value: u64) { 766 self.common.ack_features(value) 767 } 768 769 fn read_config(&self, offset: u64, data: &mut [u8]) { 770 self.read_config_from_slice(self.config.as_slice(), offset, data); 771 } 772 773 fn write_config(&mut self, offset: u64, data: &[u8]) { 774 // The "writeback" field is the only mutable field 775 let writeback_offset = 776 (&self.config.writeback as *const _ as u64) - (&self.config as *const _ as u64); 777 if offset != writeback_offset || data.len() != std::mem::size_of_val(&self.config.writeback) 778 { 779 error!( 780 "Attempt to write to read-only field: offset {:x} length {}", 781 offset, 782 data.len() 783 ); 784 return; 785 } 786 787 self.config.writeback = data[0]; 788 self.update_writeback(); 789 } 790 791 fn activate( 792 &mut self, 793 mem: GuestMemoryAtomic<GuestMemoryMmap>, 794 interrupt_cb: Arc<dyn VirtioInterrupt>, 795 mut queues: Vec<(usize, Queue, EventFd)>, 796 ) -> ActivateResult { 797 self.common.activate(&queues, &interrupt_cb)?; 798 799 self.update_writeback(); 800 801 let mut epoll_threads = Vec::new(); 802 let event_idx = self.common.feature_acked(VIRTIO_RING_F_EVENT_IDX.into()); 803 804 for i in 0..queues.len() { 805 let (_, mut queue, queue_evt) = queues.remove(0); 806 queue.set_event_idx(event_idx); 807 808 let queue_size = queue.size(); 809 let (kill_evt, pause_evt) = self.common.dup_eventfds(); 810 let queue_idx = i as u16; 811 812 let mut handler = BlockEpollHandler { 813 queue_index: queue_idx, 814 queue, 815 mem: mem.clone(), 816 disk_image: self 817 .disk_image 818 .new_async_io(queue_size as u32) 819 .map_err(|e| { 820 error!("failed to create new AsyncIo: {}", e); 821 ActivateError::BadActivate 822 })?, 823 disk_nsectors: self.disk_nsectors, 824 interrupt_cb: interrupt_cb.clone(), 825 serial: self.serial.clone(), 826 kill_evt, 827 pause_evt, 828 writeback: self.writeback.clone(), 829 counters: self.counters.clone(), 830 queue_evt, 831 // Analysis during boot shows around ~40 maximum requests 832 // This gives head room for systems with slower I/O without 833 // compromising the cost of the reallocation or memory overhead 834 inflight_requests: VecDeque::with_capacity(64), 835 rate_limiter: self 836 .rate_limiter 837 .as_ref() 838 .map(|r| r.new_handle()) 839 .transpose() 840 .unwrap(), 841 access_platform: self.common.access_platform.clone(), 842 read_only: self.read_only, 843 host_cpus: self.queue_affinity.get(&queue_idx).cloned(), 844 }; 845 846 let paused = self.common.paused.clone(); 847 let paused_sync = self.common.paused_sync.clone(); 848 849 spawn_virtio_thread( 850 &format!("{}_q{}", self.id.clone(), i), 851 &self.seccomp_action, 852 Thread::VirtioBlock, 853 &mut epoll_threads, 854 &self.exit_evt, 855 move || handler.run(paused, paused_sync.unwrap()), 856 )?; 857 } 858 859 self.common.epoll_threads = Some(epoll_threads); 860 event!("virtio-device", "activated", "id", &self.id); 861 862 Ok(()) 863 } 864 865 fn reset(&mut self) -> Option<Arc<dyn VirtioInterrupt>> { 866 let result = self.common.reset(); 867 event!("virtio-device", "reset", "id", &self.id); 868 result 869 } 870 871 fn counters(&self) -> Option<HashMap<&'static str, Wrapping<u64>>> { 872 let mut counters = HashMap::new(); 873 874 counters.insert( 875 "read_bytes", 876 Wrapping(self.counters.read_bytes.load(Ordering::Acquire)), 877 ); 878 counters.insert( 879 "write_bytes", 880 Wrapping(self.counters.write_bytes.load(Ordering::Acquire)), 881 ); 882 counters.insert( 883 "read_ops", 884 Wrapping(self.counters.read_ops.load(Ordering::Acquire)), 885 ); 886 counters.insert( 887 "write_ops", 888 Wrapping(self.counters.write_ops.load(Ordering::Acquire)), 889 ); 890 counters.insert( 891 "write_latency_min", 892 Wrapping(self.counters.write_latency_min.load(Ordering::Acquire)), 893 ); 894 counters.insert( 895 "write_latency_max", 896 Wrapping(self.counters.write_latency_max.load(Ordering::Acquire)), 897 ); 898 counters.insert( 899 "write_latency_avg", 900 Wrapping(self.counters.write_latency_avg.load(Ordering::Acquire) / LATENCY_SCALE), 901 ); 902 counters.insert( 903 "read_latency_min", 904 Wrapping(self.counters.read_latency_min.load(Ordering::Acquire)), 905 ); 906 counters.insert( 907 "read_latency_max", 908 Wrapping(self.counters.read_latency_max.load(Ordering::Acquire)), 909 ); 910 counters.insert( 911 "read_latency_avg", 912 Wrapping(self.counters.read_latency_avg.load(Ordering::Acquire) / LATENCY_SCALE), 913 ); 914 915 Some(counters) 916 } 917 918 fn set_access_platform(&mut self, access_platform: Arc<dyn AccessPlatform>) { 919 self.common.set_access_platform(access_platform) 920 } 921 } 922 923 impl Pausable for Block { 924 fn pause(&mut self) -> result::Result<(), MigratableError> { 925 self.common.pause() 926 } 927 928 fn resume(&mut self) -> result::Result<(), MigratableError> { 929 self.common.resume() 930 } 931 } 932 933 impl Snapshottable for Block { 934 fn id(&self) -> String { 935 self.id.clone() 936 } 937 938 fn snapshot(&mut self) -> std::result::Result<Snapshot, MigratableError> { 939 Snapshot::new_from_state(&self.state()) 940 } 941 } 942 impl Transportable for Block {} 943 impl Migratable for Block {} 944