1 ============================== 2 RT-mutex implementation design 3 ============================== 4 5 Copyright (c) 2006 Steven Rostedt 6 7 Licensed under the GNU Free Documentation Lice 8 9 10 This document tries to describe the design of 11 It doesn't describe the reasons why rtmutex.c 12 Documentation/locking/rt-mutex.rst. Although 13 that happen without this code, but that is in 14 what the code actually is doing. 15 16 The goal of this document is to help others un 17 inheritance (PI) algorithm that is used, as we 18 decisions that were made to implement PI in th 19 20 21 Unbounded Priority Inversion 22 ---------------------------- 23 24 Priority inversion is when a lower priority pr 25 priority process wants to run. This happens f 26 most of the time it can't be helped. Anytime 27 to use a resource that a lower priority proces 28 the high priority process must wait until the 29 with the resource. This is a priority inversi 30 is something called unbounded priority inversi 31 priority process is prevented from running by 32 an undetermined amount of time. 33 34 The classic example of unbounded priority inve 35 processes, let's call them processes A, B, and 36 priority process, C is the lowest, and B is in 37 that C owns and must wait and lets C run to re 38 meantime, B executes, and since B is of a high 39 but by doing so, it is in fact preempting A wh 40 Now there's no way of knowing how long A will 41 to release the lock, because for all we know, 42 never give C a chance to release the lock. Th 43 inversion. 44 45 Here's a little ASCII art to show the problem: 46 47 grab lock L1 (owned by C) 48 | 49 A ---+ 50 C preempted by B 51 | 52 C +----+ 53 54 B +--------> 55 B now keeps A from running. 56 57 58 Priority Inheritance (PI) 59 ------------------------- 60 61 There are several ways to solve this issue, bu 62 for this document. Here we only discuss PI. 63 64 PI is where a process inherits the priority of 65 process blocks on a lock owned by the current 66 to understand, let's use the previous example, 67 68 This time, when A blocks on the lock owned by 69 of A. So now if B becomes runnable, it would 70 the high priority of A. As soon as C releases 71 inherited priority, and A then can continue wi 72 73 Terminology 74 ----------- 75 76 Here I explain some terminology that is used i 77 the design that is used to implement PI. 78 79 PI chain 80 - The PI chain is an ordered series o 81 processes to inherit priorities fro 82 blocked on one of its locks. This 83 later in this document. 84 85 mutex 86 - In this document, to differentiate 87 PI and spin locks that are used in 88 the PI locks will be called a mutex 89 90 lock 91 - In this document from now on, I wil 92 referring to spin locks that are us 93 algorithm. These locks disable pre 94 CONFIG_PREEMPT is enabled) and on S 95 entering critical sections simultan 96 97 spin lock 98 - Same as lock above. 99 100 waiter 101 - A waiter is a struct that is stored 102 process. Since the scope of the wa 103 a process being blocked on the mute 104 the waiter on the process's stack ( 105 structure holds a pointer to the ta 106 the task is blocked on. It also ha 107 place the task in the waiters rbtre 108 pi_waiters rbtree of a mutex owner 109 110 waiter is sometimes used in referen 111 on a mutex. This is the same as wai 112 113 waiters 114 - A list of processes that are blocke 115 116 top waiter 117 - The highest priority process waitin 118 119 top pi waiter 120 - The highest priority process w 121 that a specific process owns. 122 123 Note: 124 task and process are used interchangeab 125 differentiate between two processes tha 126 127 128 PI chain 129 -------- 130 131 The PI chain is a list of processes and mutexe 132 inheritance to take place. Multiple chains ma 133 would never diverge, since a process can't be 134 mutex at a time. 135 136 Example:: 137 138 Process: A, B, C, D, E 139 Mutexes: L1, L2, L3, L4 140 141 A owns: L1 142 B blocked on L1 143 B owns L2 144 C blocked on L2 145 C owns L3 146 D blocked on L3 147 D owns L4 148 E blocked on L 149 150 The chain would be:: 151 152 E->L4->D->L3->C->L2->B->L1->A 153 154 To show where two chains merge, we could add a 155 another mutex L5 where B owns L5 and F is bloc 156 157 The chain for F would be:: 158 159 F->L5->B->L1->A 160 161 Since a process may own more than one mutex, b 162 one, the chains merge. 163 164 Here we show both chains:: 165 166 E->L4->D->L3->C->L2-+ 167 | 168 +->B->L1->A 169 | 170 F->L5-+ 171 172 For PI to work, the processes at the right end 173 also call it the Top of the chain) must be equ 174 than the processes to the left or below in the 175 176 Also since a mutex may have more than one proc 177 have multiple chains merge at mutexes. If we 178 blocked on mutex L2:: 179 180 G->L2->B->L1->A 181 182 And once again, to show how this can grow I wi 183 again:: 184 185 E->L4->D->L3->C-+ 186 +->L2-+ 187 | | 188 G-+ +->B->L1->A 189 | 190 F->L5-+ 191 192 If process G has the highest priority in the c 193 the chain (A and B in this example), must have 194 to that of G. 195 196 Mutex Waiters Tree 197 ------------------ 198 199 Every mutex keeps track of all the waiters tha 200 mutex has a rbtree to store these waiters by p 201 by a spin lock that is located in the struct o 202 wait_lock. 203 204 205 Task PI Tree 206 ------------ 207 208 To keep track of the PI chains, each process h 209 a tree of all top waiters of the mutexes that 210 Note that this tree only holds the top waiters 211 blocked on mutexes owned by the process. 212 213 The top of the task's PI tree is always the hi 214 is waiting on a mutex that is owned by the tas 215 inherited a priority, it will always be the pr 216 at the top of this tree. 217 218 This tree is stored in the task structure of a 219 pi_waiters. It is protected by a spin lock al 220 called pi_lock. This lock may also be taken i 221 locking the pi_lock, interrupts must be disabl 222 223 224 Depth of the PI Chain 225 --------------------- 226 227 The maximum depth of the PI chain is not dynam 228 defined. But is very complex to figure it out 229 the nesting of mutexes. Let's look at the exa 230 L1, L2, and L3, and four separate functions fu 231 The following shows a locking order of L1->L2- 232 be directly nested that way:: 233 234 void func1(void) 235 { 236 mutex_lock(L1); 237 238 /* do anything */ 239 240 mutex_unlock(L1); 241 } 242 243 void func2(void) 244 { 245 mutex_lock(L1); 246 mutex_lock(L2); 247 248 /* do something */ 249 250 mutex_unlock(L2); 251 mutex_unlock(L1); 252 } 253 254 void func3(void) 255 { 256 mutex_lock(L2); 257 mutex_lock(L3); 258 259 /* do something else */ 260 261 mutex_unlock(L3); 262 mutex_unlock(L2); 263 } 264 265 void func4(void) 266 { 267 mutex_lock(L3); 268 269 /* do something again */ 270 271 mutex_unlock(L3); 272 } 273 274 Now we add 4 processes that run each of these 275 Processes A, B, C, and D which run functions f 276 respectively, and such that D runs first and A 277 in func4 in the "do something again" area, we 278 279 D owns L3 280 C blocked on L3 281 C owns L2 282 B blocked on L2 283 B owns L1 284 A blocked on L1 285 286 And thus we have the chain A->L1->B->L2->C-> 287 288 This gives us a PI depth of 4 (four processes) 289 functions individually, it seems as though the 290 depth of two. So, although the locking depth 291 it still is very difficult to find the possibi 292 293 Now since mutexes can be defined by user-land 294 type of application that nests large amounts o 295 PI chain, and have the code holding spin locks 296 amount of data. So to prevent this, the imple 297 a maximum lock depth, but also only holds at m 298 time, as it walks the PI chain. More about th 299 300 301 Mutex owner and flags 302 --------------------- 303 304 The mutex structure contains a pointer to the 305 mutex is not owned, this owner is set to NULL. 306 have the task structure on at least a two byte 307 not true, the rtmutex.c code will be broken!), 308 significant bit to be used as a flag. Bit 0 i 309 flag. It's set whenever there are waiters on a 310 311 See Documentation/locking/rt-mutex.rst for fur 312 313 cmpxchg Tricks 314 -------------- 315 316 Some architectures implement an atomic cmpxchg 317 is used (when applicable) to keep the fast pat 318 mutexes short. 319 320 cmpxchg is basically the following function pe 321 322 unsigned long _cmpxchg(unsigned long *A, uns 323 { 324 unsigned long T = *A; 325 if (*A == *B) { 326 *A = *C; 327 } 328 return T; 329 } 330 #define cmpxchg(a,b,c) _cmpxchg(&a,&b,&c) 331 332 This is really nice to have, since it allows y 333 if the variable is what you expect it to be. 334 the return value (the old value of A) is equal 335 336 The macro rt_mutex_cmpxchg is used to try to l 337 the architecture does not support CMPXCHG, the 338 to fail every time. But if CMPXCHG is support 339 help out extremely to keep the fast path short 340 341 The use of rt_mutex_cmpxchg with the flags in 342 the system for architectures that support it. 343 later in this document. 344 345 346 Priority adjustments 347 -------------------- 348 349 The implementation of the PI code in rtmutex.c 350 process must adjust its priority. With the he 351 process this is rather easy to know what needs 352 353 The functions implementing the task adjustment 354 and rt_mutex_setprio. rt_mutex_setprio is only 355 356 rt_mutex_adjust_prio examines the priority of 357 priority process that is waiting any of mutexe 358 the pi_waiters of a task holds an order by pri 359 of all the mutexes that the task owns, we simp 360 pi waiter to its own normal/deadline priority 361 Then rt_mutex_setprio is called to adjust the 362 new priority. Note that rt_mutex_setprio is de 363 to implement the actual change in priority. 364 365 Note: 366 For the "prio" field in task_struct, t 367 higher the priority. A "prio" of 5 is 368 "prio" of 10. 369 370 It is interesting to note that rt_mutex_adjust 371 or decrease the priority of the task. In the 372 process has just blocked on a mutex owned by t 373 would increase/boost the task's priority. But 374 were for some reason to leave the mutex (timeo 375 would decrease/unboost the priority of the tas 376 always contains the highest priority task that 377 by the task, so we only need to compare the pr 378 to the normal priority of the given task. 379 380 381 High level overview of the PI chain walk 382 ---------------------------------------- 383 384 The PI chain walk is implemented by the functi 385 386 The implementation has gone through several it 387 with what we believe is the best. It walks th 388 at most two locks at a time, and is very effic 389 390 The rt_mutex_adjust_prio_chain can be used eit 391 priorities. 392 393 rt_mutex_adjust_prio_chain is called with a ta 394 (de)boosting (the owner of a mutex that a proc 395 check for deadlocking, the mutex that the task 396 that is the process's waiter struct that is bl 397 parameter may be NULL for deboosting), a point 398 is blocked, and a top_task as the top waiter o 399 400 For this explanation, I will not mention deadl 401 will try to stay at a high level. 402 403 When this function is called, there are no loc 404 that the state of the owner and lock can chang 405 406 Before this function is called, the task has a 407 performed on it. This means that the task is 408 should be at, but the rbtree nodes of the task 409 with the new priorities, and this task may not 410 in the pi_waiters and waiters trees that the t 411 solves all that. 412 413 The main operation of this function is summari 414 rtmutex.c. See the 'Chain walk basics and prot 415 details. 416 417 Taking of a mutex (The walk through) 418 ------------------------------------ 419 420 OK, now let's take a look at the detailed walk 421 taking a mutex. 422 423 The first thing that is tried is the fast taki 424 done when we have CMPXCHG enabled (otherwise t 425 fails). Only when the owner field of the mute 426 taken with the CMPXCHG and nothing else needs 427 428 If there is contention on the lock, we go abou 429 (rt_mutex_slowlock). 430 431 The slow path function is where the task's wai 432 the stack. This is because the waiter structu 433 scope of this function. The waiter structure 434 the task on the waiters tree of the mutex, and 435 tree of the owner. 436 437 The wait_lock of the mutex is taken since the 438 mutex also takes this lock. 439 440 We then call try_to_take_rt_mutex. This is wh 441 does not implement CMPXCHG would always grab t 442 contention). 443 444 try_to_take_rt_mutex is used every time the ta 445 slow path. The first thing that is done here 446 the "Has Waiters" flag of the mutex's owner fi 447 now, the current owner of the mutex being cont 448 without going into the slow unlock path, and i 449 wait_lock, which this code currently holds. So 450 forces the current owner to synchronize with t 451 452 The lock is taken if the following are true: 453 454 1) The lock has no owner 455 2) The current task is the highest priority 456 waiters of the lock 457 458 If the task succeeds to acquire the lock, then 459 owner of the lock, and if the lock still has w 460 (highest priority task waiting on the lock) is 461 pi_waiters tree. 462 463 If the lock is not taken by try_to_take_rt_mut 464 task_blocks_on_rt_mutex() function is called. 465 the lock's waiter tree and propagate the pi ch 466 as the lock's owner's pi_waiters tree. This is 467 section. 468 469 Task blocks on mutex 470 -------------------- 471 472 The accounting of a mutex and process is done 473 the process. The "task" field is set to the p 474 to the mutex. The rbtree node of waiter are i 475 current priority. 476 477 Since the wait_lock was taken at the entry of 478 add the waiter to the task waiter tree. If th 479 highest priority process currently waiting on 480 previous top waiter process (if it exists) fro 481 and add the current process to that tree. Sin 482 has changed, we call rt_mutex_adjust_prio on t 483 should adjust its priority accordingly. 484 485 If the owner is also blocked on a lock, and ha 486 (or deadlock checking is on), we unlock the wa 487 and run rt_mutex_adjust_prio_chain on the owne 488 489 Now all locks are released, and if the current 490 mutex (waiter "task" field is not NULL), then 491 492 Waking up in the loop 493 --------------------- 494 495 The task can then wake up for a couple of reas 496 1) The previous lock owner released the lock 497 2) we received a signal or timeout 498 499 In both cases, the task will try again to acqu 500 does, then it will take itself off the waiters 501 to the TASK_RUNNING state. 502 503 In first case, if the lock was acquired by ano 504 could get the lock, then it will go back to sl 505 506 The second case is only applicable for tasks t 507 that can wake up before getting the lock, eith 508 a timeout (i.e. rt_mutex_timed_futex_lock()). 509 take the lock again, if it succeeds, then the 510 lock held, otherwise it will return with -EINT 511 by a signal, or -ETIMEDOUT if it timed out. 512 513 514 Unlocking the Mutex 515 ------------------- 516 517 The unlocking of a mutex also has a fast path 518 CMPXCHG. Since the taking of a mutex on conte 519 "Has Waiters" flag of the mutex's owner, we us 520 take the slow path when unlocking the mutex. 521 waiters, the owner field of the mutex would eq 522 the mutex can be unlocked by just replacing th 523 524 If the owner field has the "Has Waiters" bit s 525 the slow unlock path is taken. 526 527 The first thing done in the slow unlock path i 528 mutex. This synchronizes the locking and unlo 529 530 A check is made to see if the mutex has waiter 531 do not have CMPXCHG, this is the location that 532 determine if a waiter needs to be awoken or no 533 do have CMPXCHG, that check is done in the fas 534 in the slow path too. If a waiter of a mutex 535 or timeout between the time the owner failed t 536 the grabbing of the wait_lock, the mutex may n 537 owner still needs to make this check. If there 538 owner field is set to NULL, the wait_lock is r 539 needed. 540 541 If there are waiters, then we need to wake one 542 543 On the wake up code, the pi_lock of the curren 544 waiter of the lock is found and removed from t 545 as well as the pi_waiters tree of the current 546 marked to prevent lower priority tasks from st 547 548 Finally we unlock the pi_lock of the pending o 549 550 551 Contact 552 ------- 553 554 For updates on this document, please email Ste< 555 556 557 Credits 558 ------- 559 560 Author: Steven Rostedt <rostedt@goodmis.org> 561 562 Updated: Alex Shi <alex.shi@linaro.org> - 7/6/2 563 564 Original Reviewers: 565 Ingo Molnar, Thomas Gleix 566 Randy Dunlap 567 568 Update (7/6/2017) Reviewers: Steven Rostedt an 569 570 Updates 571 ------- 572 573 This document was originally written for 2.6.1 574 was updated on 4.12
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.