1 // Copyright 2020, The Android Open Source Project 2 // 3 // Licensed under the Apache License, Version 2.0 (the "License"); 4 // you may not use this file except in compliance with the License. 5 // You may obtain a copy of the License at 6 // 7 // http://www.apache.org/licenses/LICENSE-2.0 8 // 9 // Unless required by applicable law or agreed to in writing, software 10 // distributed under the License is distributed on an "AS IS" BASIS, 11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 // See the License for the specific language governing permissions and 13 // limitations under the License. 14 15 //! This crate implements the `IKeystoreOperation` AIDL interface, which represents 16 //! an ongoing key operation, as well as the operation database, which is mainly 17 //! required for tracking operations for the purpose of pruning. 18 //! This crate also implements an operation pruning strategy. 19 //! 20 //! Operations implement the API calls update, finish, and abort. 21 //! Additionally, an operation can be dropped and pruned. The former 22 //! happens if the client deletes a binder to the operation object. 23 //! An existing operation may get pruned when running out of operation 24 //! slots and a new operation takes precedence. 25 //! 26 //! ## Operation Lifecycle 27 //! An operation gets created when the client calls `IKeystoreSecurityLevel::create`. 28 //! It may receive zero or more update request. The lifecycle ends when: 29 //! * `update` yields an error. 30 //! * `finish` is called. 31 //! * `abort` is called. 32 //! * The operation gets dropped. 33 //! * The operation gets pruned. 34 //! `Operation` has an `Outcome` member. While the outcome is `Outcome::Unknown`, 35 //! the operation is active and in a good state. Any of the above conditions may 36 //! change the outcome to one of the defined outcomes Success, Abort, Dropped, 37 //! Pruned, or ErrorCode. The latter is chosen in the case of an unexpected error, during 38 //! `update` or `finish`. `Success` is chosen iff `finish` completes without error. 39 //! Note that all operations get dropped eventually in the sense that they lose 40 //! their last reference and get destroyed. At that point, the fate of the operation 41 //! gets logged. However, an operation will transition to `Outcome::Dropped` iff 42 //! the operation was still active (`Outcome::Unknown`) at that time. 43 //! 44 //! ## Operation Dropping 45 //! To observe the dropping of an operation, we have to make sure that there 46 //! are no strong references to the IBinder representing this operation. 47 //! This would be simple enough if the operation object would need to be accessed 48 //! only by transactions. But to perform pruning, we have to retain a reference to the 49 //! original operation object. 50 //! 51 //! ## Operation Pruning 52 //! Pruning an operation happens during the creation of a new operation. 53 //! We have to iterate through the operation database to find a suitable 54 //! candidate. Then we abort and finalize this operation setting its outcome to 55 //! `Outcome::Pruned`. The corresponding KeyMint operation slot will have been freed 56 //! up at this point, but the `Operation` object lingers. When the client 57 //! attempts to use the operation again they will receive 58 //! ErrorCode::INVALID_OPERATION_HANDLE indicating that the operation no longer 59 //! exits. This should be the cue for the client to destroy its binder. 60 //! At that point the operation gets dropped. 61 //! 62 //! ## Architecture 63 //! The `IKeystoreOperation` trait is implemented by `KeystoreOperation`. 64 //! This acts as a proxy object holding a strong reference to actual operation 65 //! implementation `Operation`. 66 //! 67 //! ``` 68 //! struct KeystoreOperation { 69 //! operation: Mutex<Option<Arc<Operation>>>, 70 //! } 71 //! ``` 72 //! 73 //! The `Mutex` serves two purposes. It provides interior mutability allowing 74 //! us to set the Option to None. We do this when the life cycle ends during 75 //! a call to `update`, `finish`, or `abort`. As a result most of the Operation 76 //! related resources are freed. The `KeystoreOperation` proxy object still 77 //! lingers until dropped by the client. 78 //! The second purpose is to protect operations against concurrent usage. 79 //! Failing to lock this mutex yields `ResponseCode::OPERATION_BUSY` and indicates 80 //! a programming error in the client. 81 //! 82 //! Note that the Mutex only protects the operation against concurrent client calls. 83 //! We still retain weak references to the operation in the operation database: 84 //! 85 //! ``` 86 //! struct OperationDb { 87 //! operations: Mutex<Vec<Weak<Operation>>> 88 //! } 89 //! ``` 90 //! 91 //! This allows us to access the operations for the purpose of pruning. 92 //! We do this in three phases. 93 //! 1. We gather the pruning information. Besides non mutable information, 94 //! we access `last_usage` which is protected by a mutex. 95 //! We only lock this mutex for single statements at a time. During 96 //! this phase we hold the operation db lock. 97 //! 2. We choose a pruning candidate by computing the pruning resistance 98 //! of each operation. We do this entirely with information we now 99 //! have on the stack without holding any locks. 100 //! (See `OperationDb::prune` for more details on the pruning strategy.) 101 //! 3. During pruning we briefly lock the operation database again to get the 102 //! the pruning candidate by index. We then attempt to abort the candidate. 103 //! If the candidate was touched in the meantime or is currently fulfilling 104 //! a request (i.e., the client calls update, finish, or abort), 105 //! we go back to 1 and try again. 106 //! 107 //! So the outer Mutex in `KeystoreOperation::operation` only protects 108 //! operations against concurrent client calls but not against concurrent 109 //! pruning attempts. This is what the `Operation::outcome` mutex is used for. 110 //! 111 //! ``` 112 //! struct Operation { 113 //! ... 114 //! outcome: Mutex<Outcome>, 115 //! ... 116 //! } 117 //! ``` 118 //! 119 //! Any request that can change the outcome, i.e., `update`, `finish`, `abort`, 120 //! `drop`, and `prune` has to take the outcome lock and check if the outcome 121 //! is still `Outcome::Unknown` before entering. `prune` is special in that 122 //! it will `try_lock`, because we don't want to be blocked on a potentially 123 //! long running request at another operation. If it fails to get the lock 124 //! the operation is either being touched, which changes its pruning resistance, 125 //! or it transitions to its end-of-life, which means we may get a free slot. 126 //! Either way, we have to revaluate the pruning scores. 127 128 use crate::enforcements::AuthInfo; 129 use crate::error::{ 130 error_to_serialized_error, into_binder, into_logged_binder, map_km_error, Error, ErrorCode, 131 ResponseCode, SerializedError, 132 }; 133 use crate::ks_err; 134 use crate::metrics_store::log_key_operation_event_stats; 135 use crate::utils::watchdog as wd; 136 use android_hardware_security_keymint::aidl::android::hardware::security::keymint::{ 137 IKeyMintOperation::IKeyMintOperation, KeyParameter::KeyParameter, KeyPurpose::KeyPurpose, 138 SecurityLevel::SecurityLevel, 139 }; 140 use android_hardware_security_keymint::binder::{BinderFeatures, Strong}; 141 use android_system_keystore2::aidl::android::system::keystore2::{ 142 IKeystoreOperation::BnKeystoreOperation, IKeystoreOperation::IKeystoreOperation, 143 }; 144 use anyhow::{anyhow, Context, Result}; 145 use std::{ 146 collections::HashMap, 147 sync::{Arc, Mutex, MutexGuard, Weak}, 148 time::Duration, 149 time::Instant, 150 }; 151 152 /// Operations have `Outcome::Unknown` as long as they are active. They transition 153 /// to one of the other variants exactly once. The distinction in outcome is mainly 154 /// for the statistic. 155 #[derive(Debug, Copy, Clone, Eq, PartialEq, Ord, PartialOrd)] 156 pub enum Outcome { 157 /// Operations have `Outcome::Unknown` as long as they are active. 158 Unknown, 159 /// Operation is successful. 160 Success, 161 /// Operation is aborted. 162 Abort, 163 /// Operation is dropped. 164 Dropped, 165 /// Operation is pruned. 166 Pruned, 167 /// Operation is failed with the error code. 168 ErrorCode(SerializedError), 169 } 170 171 /// Operation bundles all of the operation related resources and tracks the operation's 172 /// outcome. 173 #[derive(Debug)] 174 pub struct Operation { 175 // The index of this operation in the OperationDb. 176 index: usize, 177 km_op: Strong<dyn IKeyMintOperation>, 178 last_usage: Mutex<Instant>, 179 outcome: Mutex<Outcome>, 180 owner: u32, // Uid of the operation's owner. 181 auth_info: Mutex<AuthInfo>, 182 forced: bool, 183 logging_info: LoggingInfo, 184 } 185 186 /// Keeps track of the information required for logging operations. 187 #[derive(Debug)] 188 pub struct LoggingInfo { 189 sec_level: SecurityLevel, 190 purpose: KeyPurpose, 191 op_params: Vec<KeyParameter>, 192 key_upgraded: bool, 193 } 194 195 impl LoggingInfo { 196 /// Constructor new( sec_level: SecurityLevel, purpose: KeyPurpose, op_params: Vec<KeyParameter>, key_upgraded: bool, ) -> LoggingInfo197 pub fn new( 198 sec_level: SecurityLevel, 199 purpose: KeyPurpose, 200 op_params: Vec<KeyParameter>, 201 key_upgraded: bool, 202 ) -> LoggingInfo { 203 Self { sec_level, purpose, op_params, key_upgraded } 204 } 205 } 206 207 struct PruningInfo { 208 last_usage: Instant, 209 owner: u32, 210 index: usize, 211 forced: bool, 212 } 213 214 // We don't except more than 32KiB of data in `update`, `updateAad`, and `finish`. 215 const MAX_RECEIVE_DATA: usize = 0x8000; 216 217 impl Operation { 218 /// Constructor new( index: usize, km_op: binder::Strong<dyn IKeyMintOperation>, owner: u32, auth_info: AuthInfo, forced: bool, logging_info: LoggingInfo, ) -> Self219 pub fn new( 220 index: usize, 221 km_op: binder::Strong<dyn IKeyMintOperation>, 222 owner: u32, 223 auth_info: AuthInfo, 224 forced: bool, 225 logging_info: LoggingInfo, 226 ) -> Self { 227 Self { 228 index, 229 km_op, 230 last_usage: Mutex::new(Instant::now()), 231 outcome: Mutex::new(Outcome::Unknown), 232 owner, 233 auth_info: Mutex::new(auth_info), 234 forced, 235 logging_info, 236 } 237 } 238 get_pruning_info(&self) -> Option<PruningInfo>239 fn get_pruning_info(&self) -> Option<PruningInfo> { 240 // An operation may be finalized. 241 if let Ok(guard) = self.outcome.try_lock() { 242 match *guard { 243 Outcome::Unknown => {} 244 // If the outcome is any other than unknown, it has been finalized, 245 // and we can no longer consider it for pruning. 246 _ => return None, 247 } 248 } 249 // Else: If we could not grab the lock, this means that the operation is currently 250 // being used and it may be transitioning to finalized or it was simply updated. 251 // In any case it is fair game to consider it for pruning. If the operation 252 // transitioned to a final state, we will notice when we attempt to prune, and 253 // a subsequent attempt to create a new operation will succeed. 254 Some(PruningInfo { 255 // Expect safety: 256 // `last_usage` is locked only for primitive single line statements. 257 // There is no chance to panic and poison the mutex. 258 last_usage: *self.last_usage.lock().expect("In get_pruning_info."), 259 owner: self.owner, 260 index: self.index, 261 forced: self.forced, 262 }) 263 } 264 prune(&self, last_usage: Instant) -> Result<(), Error>265 fn prune(&self, last_usage: Instant) -> Result<(), Error> { 266 let mut locked_outcome = match self.outcome.try_lock() { 267 Ok(guard) => match *guard { 268 Outcome::Unknown => guard, 269 _ => return Err(Error::Km(ErrorCode::INVALID_OPERATION_HANDLE)), 270 }, 271 Err(_) => return Err(Error::Rc(ResponseCode::OPERATION_BUSY)), 272 }; 273 274 // In `OperationDb::prune`, which is our caller, we first gather the pruning 275 // information including the last usage. When we select a candidate 276 // we call `prune` on that candidate passing the last_usage 277 // that we gathered earlier. If the actual last usage 278 // has changed since than, it means the operation was busy in the 279 // meantime, which means that we have to reevaluate the pruning score. 280 // 281 // Expect safety: 282 // `last_usage` is locked only for primitive single line statements. 283 // There is no chance to panic and poison the mutex. 284 if *self.last_usage.lock().expect("In Operation::prune()") != last_usage { 285 return Err(Error::Rc(ResponseCode::OPERATION_BUSY)); 286 } 287 *locked_outcome = Outcome::Pruned; 288 289 let _wp = wd::watch("In Operation::prune: calling abort()"); 290 291 // We abort the operation. If there was an error we log it but ignore it. 292 if let Err(e) = map_km_error(self.km_op.abort()) { 293 log::warn!("In prune: KeyMint::abort failed with {:?}.", e); 294 } 295 296 Ok(()) 297 } 298 299 // This function takes a Result from a KeyMint call and inspects it for errors. 300 // If an error was found it updates the given `locked_outcome` accordingly. 301 // It forwards the Result unmodified. 302 // The precondition to this call must be *locked_outcome == Outcome::Unknown. 303 // Ideally the `locked_outcome` came from a successful call to `check_active` 304 // see below. update_outcome<T>( &self, locked_outcome: &mut Outcome, err: Result<T, Error>, ) -> Result<T, Error>305 fn update_outcome<T>( 306 &self, 307 locked_outcome: &mut Outcome, 308 err: Result<T, Error>, 309 ) -> Result<T, Error> { 310 match &err { 311 Err(e) => *locked_outcome = Outcome::ErrorCode(error_to_serialized_error(e)), 312 Ok(_) => (), 313 } 314 err 315 } 316 317 // This function grabs the outcome lock and checks the current outcome state. 318 // If the outcome is still `Outcome::Unknown`, this function returns 319 // the locked outcome for further updates. In any other case it returns 320 // ErrorCode::INVALID_OPERATION_HANDLE indicating that this operation has 321 // been finalized and is no longer active. check_active(&self) -> Result<MutexGuard<Outcome>>322 fn check_active(&self) -> Result<MutexGuard<Outcome>> { 323 let guard = self.outcome.lock().expect("In check_active."); 324 match *guard { 325 Outcome::Unknown => Ok(guard), 326 _ => Err(Error::Km(ErrorCode::INVALID_OPERATION_HANDLE)) 327 .context(ks_err!("Call on finalized operation with outcome: {:?}.", *guard)), 328 } 329 } 330 331 // This function checks the amount of input data sent to us. We reject any buffer 332 // exceeding MAX_RECEIVE_DATA bytes as input to `update`, `update_aad`, and `finish` 333 // in order to force clients into using reasonable limits. check_input_length(data: &[u8]) -> Result<()>334 fn check_input_length(data: &[u8]) -> Result<()> { 335 if data.len() > MAX_RECEIVE_DATA { 336 // This error code is unique, no context required here. 337 return Err(anyhow!(Error::Rc(ResponseCode::TOO_MUCH_DATA))); 338 } 339 Ok(()) 340 } 341 342 // Update the last usage to now. touch(&self)343 fn touch(&self) { 344 // Expect safety: 345 // `last_usage` is locked only for primitive single line statements. 346 // There is no chance to panic and poison the mutex. 347 *self.last_usage.lock().expect("In touch.") = Instant::now(); 348 } 349 350 /// Implementation of `IKeystoreOperation::updateAad`. 351 /// Refer to the AIDL spec at system/hardware/interfaces/keystore2 for details. update_aad(&self, aad_input: &[u8]) -> Result<()>352 fn update_aad(&self, aad_input: &[u8]) -> Result<()> { 353 let mut outcome = self.check_active().context("In update_aad")?; 354 Self::check_input_length(aad_input).context("In update_aad")?; 355 self.touch(); 356 357 let (hat, tst) = self 358 .auth_info 359 .lock() 360 .unwrap() 361 .before_update() 362 .context(ks_err!("Trying to get auth tokens."))?; 363 364 self.update_outcome(&mut outcome, { 365 let _wp = wd::watch("Operation::update_aad: calling updateAad"); 366 map_km_error(self.km_op.updateAad(aad_input, hat.as_ref(), tst.as_ref())) 367 }) 368 .context(ks_err!("Update failed."))?; 369 370 Ok(()) 371 } 372 373 /// Implementation of `IKeystoreOperation::update`. 374 /// Refer to the AIDL spec at system/hardware/interfaces/keystore2 for details. update(&self, input: &[u8]) -> Result<Option<Vec<u8>>>375 fn update(&self, input: &[u8]) -> Result<Option<Vec<u8>>> { 376 let mut outcome = self.check_active().context("In update")?; 377 Self::check_input_length(input).context("In update")?; 378 self.touch(); 379 380 let (hat, tst) = self 381 .auth_info 382 .lock() 383 .unwrap() 384 .before_update() 385 .context(ks_err!("Trying to get auth tokens."))?; 386 387 let output = self 388 .update_outcome(&mut outcome, { 389 let _wp = wd::watch("Operation::update: calling update"); 390 map_km_error(self.km_op.update(input, hat.as_ref(), tst.as_ref())) 391 }) 392 .context(ks_err!("Update failed."))?; 393 394 if output.is_empty() { 395 Ok(None) 396 } else { 397 Ok(Some(output)) 398 } 399 } 400 401 /// Implementation of `IKeystoreOperation::finish`. 402 /// Refer to the AIDL spec at system/hardware/interfaces/keystore2 for details. finish(&self, input: Option<&[u8]>, signature: Option<&[u8]>) -> Result<Option<Vec<u8>>>403 fn finish(&self, input: Option<&[u8]>, signature: Option<&[u8]>) -> Result<Option<Vec<u8>>> { 404 let mut outcome = self.check_active().context("In finish")?; 405 if let Some(input) = input { 406 Self::check_input_length(input).context("In finish")?; 407 } 408 self.touch(); 409 410 let (hat, tst, confirmation_token) = self 411 .auth_info 412 .lock() 413 .unwrap() 414 .before_finish() 415 .context(ks_err!("Trying to get auth tokens."))?; 416 417 let output = self 418 .update_outcome(&mut outcome, { 419 let _wp = wd::watch("Operation::finish: calling finish"); 420 map_km_error(self.km_op.finish( 421 input, 422 signature, 423 hat.as_ref(), 424 tst.as_ref(), 425 confirmation_token.as_deref(), 426 )) 427 }) 428 .context(ks_err!("Finish failed."))?; 429 430 self.auth_info.lock().unwrap().after_finish().context("In finish.")?; 431 432 // At this point the operation concluded successfully. 433 *outcome = Outcome::Success; 434 435 if output.is_empty() { 436 Ok(None) 437 } else { 438 Ok(Some(output)) 439 } 440 } 441 442 /// Aborts the operation if it is active. IFF the operation is aborted the outcome is 443 /// set to `outcome`. `outcome` must reflect the reason for the abort. Since the operation 444 /// gets aborted `outcome` must not be `Operation::Success` or `Operation::Unknown`. abort(&self, outcome: Outcome) -> Result<()>445 fn abort(&self, outcome: Outcome) -> Result<()> { 446 let mut locked_outcome = self.check_active().context("In abort")?; 447 *locked_outcome = outcome; 448 449 { 450 let _wp = wd::watch("Operation::abort: calling abort"); 451 map_km_error(self.km_op.abort()).context(ks_err!("KeyMint::abort failed.")) 452 } 453 } 454 } 455 456 impl Drop for Operation { drop(&mut self)457 fn drop(&mut self) { 458 let guard = self.outcome.lock().expect("In drop."); 459 log_key_operation_event_stats( 460 self.logging_info.sec_level, 461 self.logging_info.purpose, 462 &(self.logging_info.op_params), 463 &guard, 464 self.logging_info.key_upgraded, 465 ); 466 if let Outcome::Unknown = *guard { 467 drop(guard); 468 // If the operation was still active we call abort, setting 469 // the outcome to `Outcome::Dropped` 470 if let Err(e) = self.abort(Outcome::Dropped) { 471 log::error!("While dropping Operation: abort failed:\n {:?}", e); 472 } 473 } 474 } 475 } 476 477 /// The OperationDb holds weak references to all ongoing operations. 478 /// Its main purpose is to facilitate operation pruning. 479 #[derive(Debug, Default)] 480 pub struct OperationDb { 481 // TODO replace Vec with WeakTable when the weak_table crate becomes 482 // available. 483 operations: Mutex<Vec<Weak<Operation>>>, 484 } 485 486 impl OperationDb { 487 /// Creates a new OperationDb. new() -> Self488 pub fn new() -> Self { 489 Self { operations: Mutex::new(Vec::new()) } 490 } 491 492 /// Creates a new operation. 493 /// This function takes a KeyMint operation and an associated 494 /// owner uid and returns a new Operation wrapped in a `std::sync::Arc`. create_operation( &self, km_op: binder::Strong<dyn IKeyMintOperation>, owner: u32, auth_info: AuthInfo, forced: bool, logging_info: LoggingInfo, ) -> Arc<Operation>495 pub fn create_operation( 496 &self, 497 km_op: binder::Strong<dyn IKeyMintOperation>, 498 owner: u32, 499 auth_info: AuthInfo, 500 forced: bool, 501 logging_info: LoggingInfo, 502 ) -> Arc<Operation> { 503 // We use unwrap because we don't allow code that can panic while locked. 504 let mut operations = self.operations.lock().expect("In create_operation."); 505 506 let mut index: usize = 0; 507 // First we iterate through the operation slots to try and find an unused 508 // slot. If we don't find one, we append the new entry instead. 509 match (*operations).iter_mut().find(|s| { 510 index += 1; 511 s.upgrade().is_none() 512 }) { 513 Some(free_slot) => { 514 let new_op = Arc::new(Operation::new( 515 index - 1, 516 km_op, 517 owner, 518 auth_info, 519 forced, 520 logging_info, 521 )); 522 *free_slot = Arc::downgrade(&new_op); 523 new_op 524 } 525 None => { 526 let new_op = Arc::new(Operation::new( 527 operations.len(), 528 km_op, 529 owner, 530 auth_info, 531 forced, 532 logging_info, 533 )); 534 operations.push(Arc::downgrade(&new_op)); 535 new_op 536 } 537 } 538 } 539 get(&self, index: usize) -> Option<Arc<Operation>>540 fn get(&self, index: usize) -> Option<Arc<Operation>> { 541 self.operations.lock().expect("In OperationDb::get.").get(index).and_then(|op| op.upgrade()) 542 } 543 544 /// Attempts to prune an operation. 545 /// 546 /// This function is used during operation creation, i.e., by 547 /// `KeystoreSecurityLevel::create_operation`, to try and free up an operation slot 548 /// if it got `ErrorCode::TOO_MANY_OPERATIONS` from the KeyMint backend. It is not 549 /// guaranteed that an operation slot is available after this call successfully 550 /// returned for various reasons. E.g., another thread may have snatched up the newly 551 /// available slot. Callers may have to call prune multiple times before they get a 552 /// free operation slot. Prune may also return `Err(Error::Rc(ResponseCode::BACKEND_BUSY))` 553 /// which indicates that no prunable operation was found. 554 /// 555 /// To find a suitable candidate we compute the malus for the caller and each existing 556 /// operation. The malus is the inverse of the pruning power (caller) or pruning 557 /// resistance (existing operation). 558 /// 559 /// The malus is based on the number of sibling operations and age. Sibling 560 /// operations are operations that have the same owner (UID). 561 /// 562 /// Every operation, existing or new, starts with a malus of 1. Every sibling 563 /// increases the malus by one. The age is the time since an operation was last touched. 564 /// It increases the malus by log6(<age in seconds> + 1) rounded down to the next 565 /// integer. So the malus increases stepwise after 5s, 35s, 215s, ... 566 /// Of two operations with the same malus the least recently used one is considered 567 /// weaker. 568 /// 569 /// For the caller to be able to prune an operation it must find an operation 570 /// with a malus higher than its own. 571 /// 572 /// The malus can be expressed as 573 /// ``` 574 /// malus = 1 + no_of_siblings + floor(log6(age_in_seconds + 1)) 575 /// ``` 576 /// where the constant `1` accounts for the operation under consideration. 577 /// In reality we compute it as 578 /// ``` 579 /// caller_malus = 1 + running_siblings 580 /// ``` 581 /// because the new operation has no age and is not included in the `running_siblings`, 582 /// and 583 /// ``` 584 /// running_malus = running_siblings + floor(log6(age_in_seconds + 1)) 585 /// ``` 586 /// because a running operation is included in the `running_siblings` and it has 587 /// an age. 588 /// 589 /// ## Example 590 /// A caller with no running operations has a malus of 1. Young (age < 5s) operations 591 /// also with no siblings have a malus of one and cannot be pruned by the caller. 592 /// We have to find an operation that has at least one sibling or is older than 5s. 593 /// 594 /// A caller with one running operation has a malus of 2. Now even young siblings 595 /// or single child aging (5s <= age < 35s) operations are off limit. An aging 596 /// sibling of two, however, would have a malus of 3 and would be fair game. 597 /// 598 /// ## Rationale 599 /// Due to the limitation of KeyMint operation slots, we cannot get around pruning or 600 /// a single app could easily DoS KeyMint. 601 /// Keystore 1.0 used to always prune the least recently used operation. This at least 602 /// guaranteed that new operations can always be started. With the increased usage 603 /// of Keystore we saw increased pruning activity which can lead to a livelock 604 /// situation in the worst case. 605 /// 606 /// With the new pruning strategy we want to provide well behaved clients with 607 /// progress assurances while punishing DoS attempts. As a result of this 608 /// strategy we can be in the situation where no operation can be pruned and the 609 /// creation of a new operation fails. This allows single child operations which 610 /// are frequently updated to complete, thereby breaking up livelock situations 611 /// and facilitating system wide progress. 612 /// 613 /// ## Update 614 /// We also allow callers to cannibalize their own sibling operations if no other 615 /// slot can be found. In this case the least recently used sibling is pruned. prune(&self, caller: u32, forced: bool) -> Result<(), Error>616 pub fn prune(&self, caller: u32, forced: bool) -> Result<(), Error> { 617 loop { 618 // Maps the uid of the owner to the number of operations that owner has 619 // (running_siblings). More operations per owner lowers the pruning 620 // resistance of the operations of that owner. Whereas the number of 621 // ongoing operations of the caller lowers the pruning power of the caller. 622 let mut owners: HashMap<u32, u64> = HashMap::new(); 623 let mut pruning_info: Vec<PruningInfo> = Vec::new(); 624 625 let now = Instant::now(); 626 self.operations 627 .lock() 628 .expect("In OperationDb::prune: Trying to lock self.operations.") 629 .iter() 630 .for_each(|op| { 631 if let Some(op) = op.upgrade() { 632 if let Some(p_info) = op.get_pruning_info() { 633 let owner = p_info.owner; 634 pruning_info.push(p_info); 635 // Count operations per owner. 636 *owners.entry(owner).or_insert(0) += 1; 637 } 638 } 639 }); 640 641 // If the operation is forced, the caller has a malus of 0. 642 let caller_malus = if forced { 0 } else { 1u64 + *owners.entry(caller).or_default() }; 643 644 // We iterate through all operations computing the malus and finding 645 // the candidate with the highest malus which must also be higher 646 // than the caller_malus. 647 struct CandidateInfo { 648 index: usize, 649 malus: u64, 650 last_usage: Instant, 651 age: Duration, 652 } 653 let mut oldest_caller_op: Option<CandidateInfo> = None; 654 let candidate = pruning_info.iter().fold( 655 None, 656 |acc: Option<CandidateInfo>, &PruningInfo { last_usage, owner, index, forced }| { 657 // Compute the age of the current operation. 658 let age = now 659 .checked_duration_since(last_usage) 660 .unwrap_or_else(|| Duration::new(0, 0)); 661 662 // Find the least recently used sibling as an alternative pruning candidate. 663 if owner == caller { 664 if let Some(CandidateInfo { age: a, .. }) = oldest_caller_op { 665 if age > a { 666 oldest_caller_op = 667 Some(CandidateInfo { index, malus: 0, last_usage, age }); 668 } 669 } else { 670 oldest_caller_op = 671 Some(CandidateInfo { index, malus: 0, last_usage, age }); 672 } 673 } 674 675 // Compute the malus of the current operation. 676 let malus = if forced { 677 // Forced operations have a malus of 0. And cannot even be pruned 678 // by other forced operations. 679 0 680 } else { 681 // Expect safety: Every owner in pruning_info was counted in 682 // the owners map. So this unwrap cannot panic. 683 *owners.get(&owner).expect( 684 "This is odd. We should have counted every owner in pruning_info.", 685 ) + ((age.as_secs() + 1) as f64).log(6.0).floor() as u64 686 }; 687 688 // Now check if the current operation is a viable/better candidate 689 // the one currently stored in the accumulator. 690 match acc { 691 // First we have to find any operation that is prunable by the caller. 692 None => { 693 if caller_malus < malus { 694 Some(CandidateInfo { index, malus, last_usage, age }) 695 } else { 696 None 697 } 698 } 699 // If we have found one we look for the operation with the worst score. 700 // If there is a tie, the older operation is considered weaker. 701 Some(CandidateInfo { index: i, malus: m, last_usage: l, age: a }) => { 702 if malus > m || (malus == m && age > a) { 703 Some(CandidateInfo { index, malus, last_usage, age }) 704 } else { 705 Some(CandidateInfo { index: i, malus: m, last_usage: l, age: a }) 706 } 707 } 708 } 709 }, 710 ); 711 712 // If we did not find a suitable candidate we may cannibalize our oldest sibling. 713 let candidate = candidate.or(oldest_caller_op); 714 715 match candidate { 716 Some(CandidateInfo { index, malus: _, last_usage, age: _ }) => { 717 match self.get(index) { 718 Some(op) => { 719 match op.prune(last_usage) { 720 // We successfully freed up a slot. 721 Ok(()) => break Ok(()), 722 // This means the operation we tried to prune was on its way 723 // out. It also means that the slot it had occupied was freed up. 724 Err(Error::Km(ErrorCode::INVALID_OPERATION_HANDLE)) => break Ok(()), 725 // This means the operation we tried to prune was currently 726 // servicing a request. There are two options. 727 // * Assume that it was touched, which means that its 728 // pruning resistance increased. In that case we have 729 // to start over and find another candidate. 730 // * Assume that the operation is transitioning to end-of-life. 731 // which means that we got a free slot for free. 732 // If we assume the first but the second is true, we prune 733 // a good operation without need (aggressive approach). 734 // If we assume the second but the first is true, our 735 // caller will attempt to create a new KeyMint operation, 736 // fail with `ErrorCode::TOO_MANY_OPERATIONS`, and call 737 // us again (conservative approach). 738 Err(Error::Rc(ResponseCode::OPERATION_BUSY)) => { 739 // We choose the conservative approach, because 740 // every needlessly pruned operation can impact 741 // the user experience. 742 // To switch to the aggressive approach replace 743 // the following line with `continue`. 744 break Ok(()); 745 } 746 747 // The candidate may have been touched so the score 748 // has changed since our evaluation. 749 _ => continue, 750 } 751 } 752 // This index does not exist any more. The operation 753 // in this slot was dropped. Good news, a slot 754 // has freed up. 755 None => break Ok(()), 756 } 757 } 758 // We did not get a pruning candidate. 759 None => break Err(Error::Rc(ResponseCode::BACKEND_BUSY)), 760 } 761 } 762 } 763 } 764 765 /// Implementation of IKeystoreOperation. 766 pub struct KeystoreOperation { 767 operation: Mutex<Option<Arc<Operation>>>, 768 } 769 770 impl KeystoreOperation { 771 /// Creates a new operation instance wrapped in a 772 /// BnKeystoreOperation proxy object. It also enables 773 /// `BinderFeatures::set_requesting_sid` on the new interface, because 774 /// we need it for checking Keystore permissions. new_native_binder(operation: Arc<Operation>) -> binder::Strong<dyn IKeystoreOperation>775 pub fn new_native_binder(operation: Arc<Operation>) -> binder::Strong<dyn IKeystoreOperation> { 776 BnKeystoreOperation::new_binder( 777 Self { operation: Mutex::new(Some(operation)) }, 778 BinderFeatures { set_requesting_sid: true, ..BinderFeatures::default() }, 779 ) 780 } 781 782 /// Grabs the outer operation mutex and calls `f` on the locked operation. 783 /// The function also deletes the operation if it returns with an error or if 784 /// `delete_op` is true. with_locked_operation<T, F>(&self, f: F, delete_op: bool) -> Result<T> where for<'a> F: FnOnce(&'a Operation) -> Result<T>,785 fn with_locked_operation<T, F>(&self, f: F, delete_op: bool) -> Result<T> 786 where 787 for<'a> F: FnOnce(&'a Operation) -> Result<T>, 788 { 789 let mut delete_op: bool = delete_op; 790 match self.operation.try_lock() { 791 Ok(mut mutex_guard) => { 792 let result = match &*mutex_guard { 793 Some(op) => { 794 let result = f(op); 795 // Any error here means we can discard the operation. 796 if result.is_err() { 797 delete_op = true; 798 } 799 result 800 } 801 None => Err(Error::Km(ErrorCode::INVALID_OPERATION_HANDLE)) 802 .context(ks_err!("KeystoreOperation::with_locked_operation")), 803 }; 804 805 if delete_op { 806 // We give up our reference to the Operation, thereby freeing up our 807 // internal resources and ending the wrapped KeyMint operation. 808 // This KeystoreOperation object will still be owned by an SpIBinder 809 // until the client drops its remote reference. 810 *mutex_guard = None; 811 } 812 result 813 } 814 Err(_) => Err(Error::Rc(ResponseCode::OPERATION_BUSY)) 815 .context(ks_err!("KeystoreOperation::with_locked_operation")), 816 } 817 } 818 } 819 820 impl binder::Interface for KeystoreOperation {} 821 822 impl IKeystoreOperation for KeystoreOperation { updateAad(&self, aad_input: &[u8]) -> binder::Result<()>823 fn updateAad(&self, aad_input: &[u8]) -> binder::Result<()> { 824 let _wp = wd::watch("IKeystoreOperation::updateAad"); 825 self.with_locked_operation( 826 |op| op.update_aad(aad_input).context(ks_err!("KeystoreOperation::updateAad")), 827 false, 828 ) 829 .map_err(into_logged_binder) 830 } 831 update(&self, input: &[u8]) -> binder::Result<Option<Vec<u8>>>832 fn update(&self, input: &[u8]) -> binder::Result<Option<Vec<u8>>> { 833 let _wp = wd::watch("IKeystoreOperation::update"); 834 self.with_locked_operation( 835 |op| op.update(input).context(ks_err!("KeystoreOperation::update")), 836 false, 837 ) 838 .map_err(into_logged_binder) 839 } finish( &self, input: Option<&[u8]>, signature: Option<&[u8]>, ) -> binder::Result<Option<Vec<u8>>>840 fn finish( 841 &self, 842 input: Option<&[u8]>, 843 signature: Option<&[u8]>, 844 ) -> binder::Result<Option<Vec<u8>>> { 845 let _wp = wd::watch("IKeystoreOperation::finish"); 846 self.with_locked_operation( 847 |op| op.finish(input, signature).context(ks_err!("KeystoreOperation::finish")), 848 true, 849 ) 850 .map_err(into_logged_binder) 851 } 852 abort(&self) -> binder::Result<()>853 fn abort(&self) -> binder::Result<()> { 854 let _wp = wd::watch("IKeystoreOperation::abort"); 855 let result = self.with_locked_operation( 856 |op| op.abort(Outcome::Abort).context(ks_err!("KeystoreOperation::abort")), 857 true, 858 ); 859 result.map_err(|e| { 860 match e.root_cause().downcast_ref::<Error>() { 861 // Calling abort on expired operations is something very common. 862 // There is no reason to clutter the log with it. It is never the cause 863 // for a true problem. 864 Some(Error::Km(ErrorCode::INVALID_OPERATION_HANDLE)) => {} 865 _ => log::error!("{:?}", e), 866 }; 867 into_binder(e) 868 }) 869 } 870 } 871