Synchronization: semaphores and some more stuff Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 1 What's wrong with busy waiting? The mutual exclusion algorithms we saw used busy-waiting. What’s wrong with that? Doesn't make sense for uni-processor o May cause priority inversion and deadlock Wastes CPU time o But is efficient if waiting-time is short Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 2 What's wrong with busy waiting? Busy waiting may cause priority-inversion and deadlock Process A's priority is higher than process B's Process B enters the CS Process A needs to enter the CS, busy-waits for B to exit the CS Process B cannot execute as long as the higher-priority process A is executing/ready Priority inversion and deadlock result Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 3 Outline Semaphores and the producer/consumer problem Counting semaphores from binary semaphores Event counters and message passing synchronization Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 4 Semaphores Two atomic operations are supported by a semaphore S: down(S) [the ‘p’ operation] If S≤0 the process is blocked. It will resume execution only after it is woken-up Else S-- up(S) [the `v’ operation] If there are blocked processes, wake-up one of them Else S++ S is non-negative Supported by Windows, Unix, … Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 5 Semaphores: is the following correct? Two atomic operations are supported by a semaphore S: down(S) [the ‘p’ operation] If S≤0 the process is blocked. It will resume execution only after it is woken-up S-- up(S) [the `v’ operation] S++ If there are blocked processes, wake-up one of them Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 6 Pseudo-code in previous slide is wrong Consider the following bad scneario: S=0 and process A performs down(S) – A is blocked Process B performs up(S) – S=1 A is ready Process C performs down(S) – S=0 & C proceeds Process A gets a time-slice and proceeds – S=0 A single up() freed 2 down()s Operating Systems, 2011, Danny Hendler & Amnon Meisels 7 Implementing mutex with semaphores Shared data: semaphore lock; /* initially lock = 1 */ down(lock) Critical section up(lock) Yes Does it satisfy deadlock-freedom? Yes Does it satisfy starvation-freedom? Depends… Does the algorithm satisfy mutex? Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 8 Semaphore as a General Synchronization Tool Execute B in Pjj only after A executed in Pii Use semaphore flag initialized to 0 Code: Pii … time A up(flag) 0 Pjj … down(flag) B Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 9 More on synchronization using semaphores Three processes p1; p2; p3 semaphores p1 down(s1); A up(s2); s1 = 1, s2 = 0; p2 down(s2); B up(s2); p3 down(s2); C up(s1); Which execution orders of A, B, C, are possible? (A B* C)* Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 10 No guarantee for correct synchronization P0 P1 down(S); down(Q); move1 up(S); up(Q) down(Q); down(S); move2 up(Q); up(S); 1 1 Example: move money between two accounts which are protected by semaphores S and Q Does this work? Deadlock! Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 11 Negative-valued semaphores Two atomic operations are supported by a semaphore S: down(S) S-If S<0 the process is blocked. It will resume execution only when S is non-negative up(S) S++ If there are blocked processes (i.e. S≤0), wake-up one of them -3 If S is negative, then there are –S blocked processes Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 12 Negative semaphore Implementation type semaphore = record value: integer; L: list of process; end; -3 L atomic down(S): S.value--; if (S.value < 0) { add this process to S.L; sleep; } atomic up(S): S.value++; if (S.value <= 0) { remove a process P from S.L; wakeup(P); } Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 13 Implementing a spin-lock with TSL In user space, one can use TSL (test-set-lock) mutex_lock: TSL CMP JZE CALL JMP ok: RET REG, mutex REG, #0 ok thread_yield mutex_lock mutex_unlock: MOV mutex, #0 RET Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 14 Implementing a negative semaphore with TSL type semaphore = record value, flag: integer; L: list of process; end; down(S): repeat until test-and-set(S.flag) S.value--; if (S.value < 0) { add this process to S.L; S.flag=0 sleep; } else S.flag=0 -3 L up(S): repeat until test-and-set(S.flag) S.value++; if (S.value <= 0) { remove a process P from S.L; wakeup(P); } S.flag=0 Any problem with this code? In down(), resetting flag and sleeping should be atomic. Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 15 More on semaphore implementation On a uni-processor, disabling interrupts may be used TSL implementation works for multi-processors On a multi-processor, we can use spin-lock mutual exclusion to protect semaphore access Why is this better than busy-waiting in the 1st place? Busy-waiting is now guaranteed to be very short Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 16 Producer-Consumer Problem buffer in out Paradigm for cooperating processes, • producer process produces information that is consumed by a consumer process Two versions • unbounded-buffer places no practical limit on the size of the buffer • bounded-buffer assumes that there is a fixed buffer size Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 17 Bounded Buffer buffer 0 1 2 item1 3 item2 2 Out 4 item3 producer consumer 5 item4 6 In 7 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 18 Implementation using semaphores Two processes or more use a shared buffer in memory The buffer has finite size(i.e., it is bounded) The producer writes to the buffer and the consumer reads from it A full buffer stops the producer An empty buffer stops the consumer Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 19 Producer-Consumer implementation with semaphores #define N typedef int semaphore semaphore semaphore 100 semaphore; mutex = 1; empty = N; full = 0; /* Buffer size */ /* access control to critical section */ /* counts empty buffer slots */ /* counts full slots */ void producer(void) { int item; while(TRUE) { produce_item(&item); /* generate something... */ down(&empty); /* decrement count of empty */ down(&mutex); /* enter critical section */ enter_item(item); /* insert into buffer */ up(&mutex); /* leave critical section */ up(&full); /* increment count of full slots */ } } Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 20 Producer-Consumer implementation with semaphores void consumer(void) { int item; while(TRUE) { down(&full); down(&mutex); remove_item(&item); up(&mutex); up(&empty); consume_item(item); } /* decrement count of full */ /* enter critical section */ /* take item from buffer) */ /* leave critical section */ /* update count of empty */ /* do something... */ } Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 21 Outline Semaphores and the producer/consumer problem Counting semaphores from binary semaphores Event counters and message passing synchronization Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 22 Binary Semaphore Assumes only values 0 or 1 Wait blocks if semaphore=0 Signal (up operation) either wakes up a waiting process, if there is one, or sets value to 1 (if value is already 1, signal is “wasted”) How can we implement a counting semaphore by using binary semaphores? Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 23 Implementing a counting semaphore with binary semaphores (user space): take 1 binary-semaphore S1 initially 1, S2 initially 0, S.value initially 1 down(S): down(S1); S.value--; if(S.value < 0){ L1: up(S1); L2: down(S2); } else up(S1); up(S): down(S1); S.value++; if(S.value ≤ 0) up(S2); up(S1) This code does not work. Why? Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 24 Race condition for counting semaphore take 1 1. Processes Q1 – Q4 perform down(S), Q2 – Q4 are preempted between lines L1 and L2: the value of the counting semaphore is now -3 2. Processes Q5-Q7 now perform up(S): the value of the counting semaphore is now 0 3. Now, Q2-Q4 wake-up in turn and perform line L2 (down S2) 4. Q2 runs but Q3-Q4 block. There is a discrepancy between the value of S and the number of processes waiting on it Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 25 Implementing a counting semaphore with binary semaphores (user space): take 2 binary-semaphore S1 initially 1, S2 initially 0, S.value initially 1 down(S): down(S1); S.value--; if(S.value < 0){ up(S1); //L1 down(S2); } //L2 up(S1); up(S): down(S1); S.value++; if(S.value ≤ 0) up(S2); else up(S1) Does this code work? Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 26 The effect of the added ‘else’ up(S1) is performed by up(S) only if no process waits on S2 Q5 leaves up(S) without releasing S1 Q6 cannot enter the critical section that protects the counter It can only do so after one of Q2-Q4 releases S1 This generates a “lock-step” situation: an up(), a down(), an up()… The critical section that protects the counter is entered alternately by a producer or a consumer Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 27 Recall the bounded-buffer algorithm #define N 100 typedef int semaphore; semaphore mutex = 1; semaphore empty = N; semaphore full = 0; void producer(void) { int item; while(TRUE) { produce_item(&item); down(&empty); down(&mutex); enter_item(item); up(&mutex); up(&full); } } void consumer(void) { int item; while(TRUE) { down(&full); down(&mutex); remove_item(&item); up(&mutex); up(&empty); consume_item(item); } } Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 28 A Problematic Scheduling Scenario Consider a Bounded buffer of 5 slots. Assume there are 6 processes each filling five slots in turn. 1 2 3 4 5 6 Empty.Value = 5 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 29 A Problematic Scheduling Scenario 1. five slots are filled by the first producer 1 2 3 4 5 6 Empty.Value = 0 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 30 A Problematic Scheduling Scenario 1. The second producer is blocked 1 2 3 4 5 6 Empty.Value = -1 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 31 A Problematic Scheduling Scenario 1. The third producer is blocked 1 2 3 4 5 6 Empty.Value = -2 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 32 A Problematic Scheduling Scenario 1. The fourth producer is blocked 1 2 3 4 5 6 Empty.Value = -3 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 33 A Problematic Scheduling Scenario 1. The fifth producer is blocked 1 2 3 4 5 6 Empty.Value = -4 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 34 A Problematic Scheduling Scenario 2. All blocked producers are waiting on S2 1 2 3 4 5 6 Empty.Value = -5 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 35 A Problematic Scheduling Scenario 3. The consumer consumes an item and is blocked on Empty.S1 until a producer adds an item. 1 2 3 4 5 6 Empty.Value = -5 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 36 A Problematic Scheduling Scenario 3. The consumer consumes an item and is blocked on S1 , one producer adds an item. 1 2 3 4 5 6 Empty.Value = -4 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 37 A Problematic Scheduling Scenario 4. Consumer must consume, only then another producer wakes up and produces an item 1 2 3 4 5 6 Empty.Value = -3 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 38 A Problematic Scheduling Scenario 4. Same as in step 3. 1 2 3 4 5 6 Empty.Value = -2 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 39 A Problematic Scheduling Scenario 5. And again… 1 2 3 4 5 6 Empty.Value = -1 Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 40 Implementing a counting semaphore with binary semaphores (user space): take 3 (P.A. Kearns, 1988) binary-semaphore S1=1, S2=0, value initially 1, integer wake=0 down(S) down(S1); S.value--; if(S.value < 0){ up(S1); //L1 down(S2); //L2 down(S1); S.wake--; if(S.wake > 0) then up(S2);} //L3 up(S1); up(S): down(S1); S.value++; if(S.value <= 0) { S.wake++; up(S2); } up(S1); Does THIS work? Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 41 Correctness arguments (Kearns)… The counter S.wake is used when processes performing down(S) are preempted between lines L1 and L2 In such a case, up(S2) performed by processes during up(S) has no effect However, these processes accumulate their waking signals on the (protected) counter S.wake After preemption is over, any single process that wakes up from its block on down(S2) checks the value of S.wake The check is again protected For each count of the wake-up signals, the awakened process performs the up(S2) (in line L3) Each re-scheduled process wakes up the next one Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 42 Kearns' algorithm is wrong Processes P0..P7 perform down(S), P0 goes through, P1..P7 are preempted just after line L2 of the operation Processes P8..P11 perform up(S) and their up(S2) operations release, say, P1..P4 Processes P5, P6, P7 are still waiting on S2 and S.wake = 4 Processes P1..P4 are ready, just before line L3 Each of P1..P3 will decrement S.wake in its turn, check that it's positive and signal one of P5..P7 Four up operations have released 7 down operations Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 43 Implementing a counting semaphore with binary semaphores (user space): take 4 (Hemmendinger, 1989) binary-semaphore S1=1, S2=0, integer wake=0 down(S) down(S1); S.value--; if(S.value < 0){ up(S1); down(S2); down(S1); S.wake--; if(S.wake > 0) then up(S2);} // L3 up(S1); up(S): down(S1); S.value++; if(S.value <= 0) { S.wake++; if (S.wake == 1) up(S2); } up(S1); This works Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 44 Implementing a counting semaphore with binary semaphores (user space): take 5 (Barz, 1983) binary-semaphore S1=1, S2=min(1, init_value), value=init_value down(S) down(S2); down(S1); S.value--; if (S.value>0) then up(S2); up(S1); up(S): down(S1); S.value++; if(S.value == 1) { up(S2); } up(S1); This works, is simpler, and was published earlier(!)… Can we switch the order of downs in down(S)? Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 45 Correctness arguments… The critical section is guarded by S1 and each of the operations down(S) and up(S) uses it to correctly update the value of S.value After updating (and inside the critical section) both operations release the S2 semaphore only if value is positive S.value is never negative, because any process performing down(S) is blocked at S2 Signals cannot be 'wasted' Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 46 Fairness of semaphores Order of releasing blocked processes: o Weak – up() performing process enters after (one of the) blocked processes o Strong – An upper bound on the number of entries of process that performed up() if others are waiting Unfair: o No guarantee about the number of times the up() performing process enters before the blocked o Open competition each time the lock is free o Imitating the Java 'wait' 'notify' mechanism o Or the spin-lock of XV6… Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 47 Outline Semaphores and the producer/consumer problem Counting semaphores from binary semaphores Event counters and message passing synchronization Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 48 Event Counters Integer counters with three operations: o Advance(E): increment E by 1, wake up relevant sleepers o Await(E,v): wait until E ≥ v. Sleep if E < v o Read(E): return the current value of E Counter value is ever increasing The Read() operation is not required for the bounded-buffer implementation in the next slide Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 49 producer-consumer with Event Counters (for a single producer and a single consumer) #define typedef event_counter event_counter N int in = 0; out = 0; 100 event_counter; /* counts inserted items */ /* items removed from buffer */ void producer(void) { int item, sequence = 0; while(TRUE) { produce_item(&item); sequence = sequence + 1; /* counts items produced */ await(out, sequence - N); /* wait for room in buffer */ enter_item(item); /* insert into buffer */ advance(&in); /* inform consumer */ } } Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 50 Event counters (producer-consumer) void consumer(void) { int item, sequence = 0; while(TRUE) { sequence = sequence + 1; /* count items consumed */ await(in, sequence); /* wait for item */ remove_item(&item); /* take item from buffer */ advance(&out); /* inform producer */ consume_item(item); } } Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 51 Message Passing – no shared memory In a multi-processor system without shared memory, synchronization can be implemented by message passing Implementation issues: o Acknowledgements may be required (messages may be lost) o Message sequence numbers required to avoid message duplication o Unique process addresses across CPUs (domains..) o Authentication (validate sender’s identity, a multi-machine environment…) Two main functions: o send(destination, &message); o receive(source, &message) block while waiting... Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 52 Producer-consumer with Message Passing #define N #define MSIZE 100 4 typedef int message(MSIZE); void producer(void) { int item; message m; while(TRUE) { produce_item(&item); receive(consumer, &m); construct_message(&m, item); send(consumer, &m); } /* message size */ /* message buffer */ /*wait for an empty */ /* send item */ } Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 53 Message passing (cont.) void consumer(void) { int item, i; message m; for(i = 0; i < N; i++) send(producer, &m); /* send N empties */ while(TRUE) { receive(producer, &m); extract_item(&m, &item); send(producer, &m); consume_item(item); } } /* get message with item */ /* send an empty reply */ Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 54 Message passing variations Messages can be addressed to a process address or to a mailbox o Mailboxes are generated with some capacity. When sending a message to a full mailbox, a process blocks o Buffer management done by mailbox Unix pipes - a generalization of messages … no fixed size message (blocking receive) If no buffer is maintained by the system, then send and receive must run in lock-step. Example: Unix rendezvous Operating Systems, 2014, Meni Adler, Danny Hendler & Amnon Meisels 55