X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fovs-thread.h;h=180b66f8926e0f308108481ec2e0552649860e2e;hb=05bf6d3c62e1d;hp=98428fd04048175706b830b1256e3617e713953a;hpb=5453ae2067671c0d40a5b3ac3cb3d4027bed6abb;p=sliver-openvswitch.git diff --git a/lib/ovs-thread.h b/lib/ovs-thread.h index 98428fd04..180b66f89 100644 --- a/lib/ovs-thread.h +++ b/lib/ovs-thread.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2013 Nicira, Inc. + * Copyright (c) 2013, 2014 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -23,91 +23,160 @@ #include "ovs-atomic.h" #include "util.h" -/* glibc has some non-portable mutex types and initializers: - * - * - PTHREAD_MUTEX_ADAPTIVE_NP is a mutex type that works as a spinlock that - * falls back to a mutex after spinning for some number of iterations. - * - * - PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP is a non-portable initializer - * for an error-checking mutex. - * - * We use these definitions to fall back to PTHREAD_MUTEX_NORMAL instead in - * these cases. - * - * (glibc has other non-portable initializers, but we can't reasonably - * substitute for them here.) */ -#ifdef PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP -#define PTHREAD_MUTEX_ADAPTIVE PTHREAD_MUTEX_ADAPTIVE_NP -#define PTHREAD_ADAPTIVE_MUTEX_INITIALIZER \ - PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP + +/* Mutex. */ +struct OVS_LOCKABLE ovs_mutex { + pthread_mutex_t lock; + const char *where; /* NULL if and only if uninitialized. */ +}; + +/* "struct ovs_mutex" initializer. */ +#ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP +#define OVS_MUTEX_INITIALIZER { PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, \ + "" } #else -#define PTHREAD_MUTEX_ADAPTIVE PTHREAD_MUTEX_NORMAL -#define PTHREAD_ADAPTIVE_MUTEX_INITIALIZER PTHREAD_MUTEX_INITIALIZER +#define OVS_MUTEX_INITIALIZER { PTHREAD_MUTEX_INITIALIZER, "" } #endif -#ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP -#define PTHREAD_ERRORCHECK_MUTEX_INITIALIZER \ - PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP +#ifdef PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP +#define OVS_ADAPTIVE_MUTEX_INITIALIZER \ + { PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP, NULL } #else -#define PTHREAD_ERRORCHECK_MUTEX_INITIALIZER PTHREAD_MUTEX_INITIALIZER +#define OVS_ADAPTIVE_MUTEX_INITIALIZER OVS_MUTEX_INITIALIZER #endif - -/* Simple wrappers for pthreads functions. Most of these functions abort the - * process with an error message on any error. The *_trylock() functions are - * exceptions: they pass through a 0 or EBUSY return value to the caller and - * abort on any other error. */ -void xpthread_mutex_init(pthread_mutex_t *, pthread_mutexattr_t *); -void xpthread_mutex_destroy(pthread_mutex_t *); -void xpthread_mutex_lock(pthread_mutex_t *mutex) OVS_ACQUIRES(mutex); -void xpthread_mutex_unlock(pthread_mutex_t *mutex) OVS_RELEASES(mutex); -int xpthread_mutex_trylock(pthread_mutex_t *); +/* ovs_mutex functions analogous to pthread_mutex_*() functions. + * + * Most of these functions abort the process with an error message on any + * error. ovs_mutex_trylock() is an exception: it passes through a 0 or EBUSY + * return value to the caller and aborts on any other error. */ +void ovs_mutex_init(const struct ovs_mutex *); +void ovs_mutex_init_recursive(const struct ovs_mutex *); +void ovs_mutex_init_adaptive(const struct ovs_mutex *); +void ovs_mutex_destroy(const struct ovs_mutex *); +void ovs_mutex_unlock(const struct ovs_mutex *mutex) OVS_RELEASES(mutex); +void ovs_mutex_lock_at(const struct ovs_mutex *mutex, const char *where) + OVS_ACQUIRES(mutex); +#define ovs_mutex_lock(mutex) \ + ovs_mutex_lock_at(mutex, SOURCE_LOCATOR) +int ovs_mutex_trylock_at(const struct ovs_mutex *mutex, const char *where) + OVS_TRY_LOCK(0, mutex); +#define ovs_mutex_trylock(mutex) \ + ovs_mutex_trylock_at(mutex, SOURCE_LOCATOR) + +void ovs_mutex_cond_wait(pthread_cond_t *, const struct ovs_mutex *); + +/* Wrappers for pthread_mutex_*() that abort the process on any error. + * This is still needed when ovs-atomic-pthreads.h is used. */ +void xpthread_mutex_lock(pthread_mutex_t *mutex); +void xpthread_mutex_unlock(pthread_mutex_t *mutex); + +/* Wrappers for pthread_mutexattr_*() that abort the process on any error. */ void xpthread_mutexattr_init(pthread_mutexattr_t *); void xpthread_mutexattr_destroy(pthread_mutexattr_t *); void xpthread_mutexattr_settype(pthread_mutexattr_t *, int type); void xpthread_mutexattr_gettype(pthread_mutexattr_t *, int *typep); -void xpthread_rwlock_init(pthread_rwlock_t *, pthread_rwlockattr_t *); -void xpthread_rwlock_destroy(pthread_rwlock_t *); -void xpthread_rwlock_rdlock(pthread_rwlock_t *rwlock) OVS_ACQUIRES(rwlock); -void xpthread_rwlock_wrlock(pthread_rwlock_t *rwlock) OVS_ACQUIRES(rwlock); -void xpthread_rwlock_unlock(pthread_rwlock_t *rwlock) OVS_RELEASES(rwlock); -int xpthread_rwlock_tryrdlock(pthread_rwlock_t *); -int xpthread_rwlock_trywrlock(pthread_rwlock_t *); +/* Read-write lock. + * + * An ovs_rwlock does not support recursive readers, because POSIX allows + * taking the reader lock recursively to deadlock when a thread is waiting on + * the write-lock. (NetBSD does deadlock.) glibc rwlocks in their default + * configuration do not deadlock, but ovs_rwlock_init() initializes rwlocks as + * non-recursive (which will deadlock) for two reasons: + * + * - glibc only provides fairness to writers in this mode. + * + * - It's better to find bugs in the primary Open vSwitch target rather + * than exposing them only to porters. */ +struct OVS_LOCKABLE ovs_rwlock { + pthread_rwlock_t lock; + const char *where; /* NULL if and only if uninitialized. */ +}; + +/* Initializer. */ +#ifdef PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP +#define OVS_RWLOCK_INITIALIZER \ + { PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP, "" } +#else +#define OVS_RWLOCK_INITIALIZER { PTHREAD_RWLOCK_INITIALIZER, "" } +#endif +/* ovs_rwlock functions analogous to pthread_rwlock_*() functions. + * + * Most of these functions abort the process with an error message on any + * error. The "trylock" functions are exception: they pass through a 0 or + * EBUSY return value to the caller and abort on any other error. */ +void ovs_rwlock_init(const struct ovs_rwlock *); +void ovs_rwlock_destroy(const struct ovs_rwlock *); +void ovs_rwlock_unlock(const struct ovs_rwlock *rwlock) OVS_RELEASES(rwlock); + +/* Wrappers for pthread_rwlockattr_*() that abort the process on any error. */ +void xpthread_rwlockattr_init(pthread_rwlockattr_t *); +void xpthread_rwlockattr_destroy(pthread_rwlockattr_t *); +#ifdef PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP +void xpthread_rwlockattr_setkind_np(pthread_rwlockattr_t *, int kind); +#endif + +void ovs_rwlock_wrlock_at(const struct ovs_rwlock *rwlock, const char *where) + OVS_ACQ_WRLOCK(rwlock); +#define ovs_rwlock_wrlock(rwlock) \ + ovs_rwlock_wrlock_at(rwlock, SOURCE_LOCATOR) + +int ovs_rwlock_trywrlock_at(const struct ovs_rwlock *rwlock, const char *where) + OVS_TRY_WRLOCK(0, rwlock); +#define ovs_rwlock_trywrlock(rwlock) \ + ovs_rwlock_trywrlock_at(rwlock, SOURCE_LOCATOR) + +void ovs_rwlock_rdlock_at(const struct ovs_rwlock *rwlock, const char *where) + OVS_ACQ_RDLOCK(rwlock); +#define ovs_rwlock_rdlock(rwlock) \ + ovs_rwlock_rdlock_at(rwlock, SOURCE_LOCATOR) + +int ovs_rwlock_tryrdlock_at(const struct ovs_rwlock *rwlock, const char *where) + OVS_TRY_RDLOCK(0, rwlock); +#define ovs_rwlock_tryrdlock(rwlock) \ + ovs_rwlock_tryrdlock_at(rwlock, SOURCE_LOCATOR) + +/* Wrappers for xpthread_cond_*() that abort the process on any error. + * + * Use ovs_mutex_cond_wait() to wait for a condition. */ void xpthread_cond_init(pthread_cond_t *, pthread_condattr_t *); void xpthread_cond_destroy(pthread_cond_t *); void xpthread_cond_signal(pthread_cond_t *); void xpthread_cond_broadcast(pthread_cond_t *); -void xpthread_cond_wait(pthread_cond_t *, pthread_mutex_t *mutex) - OVS_MUST_HOLD(mutex); - -#ifdef __CHECKER__ -/* Replace these functions by the macros already defined in the - * annotations, because the macro definitions have correct semantics for the - * conditional acquisition that can't be captured in a function annotation. - * The difference in semantics from pthread_*() to xpthread_*() does not matter - * because sparse is not a compiler. */ -#define xpthread_mutex_trylock pthread_mutex_trylock -#define xpthread_rwlock_tryrdlock pthread_rwlock_tryrdlock -#define xpthread_rwlock_trywrlock pthread_rwlock_trywrlock -#endif + +/* Wrappers for pthread_barrier_*() that abort the process on any error. */ +void xpthread_barrier_init(pthread_barrier_t *, pthread_barrierattr_t *, + unsigned int count); +int xpthread_barrier_wait(pthread_barrier_t *); +void xpthread_barrier_destroy(pthread_barrier_t *); void xpthread_key_create(pthread_key_t *, void (*destructor)(void *)); +void xpthread_key_delete(pthread_key_t); +void xpthread_setspecific(pthread_key_t, const void *); -void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *); +pthread_t ovs_thread_create(const char *name, void *(*)(void *), void *); +void xpthread_join(pthread_t, void **); /* Per-thread data. * - * Multiple forms of per-thread data exist, each with its own pluses and - * minuses: + * + * Standard Forms + * ============== + * + * Multiple forms of standard per-thread data exist, each with its own pluses + * and minuses. In general, if one of these forms is appropriate, then it's a + * good idea to use it: * * - POSIX per-thread data via pthread_key_t is portable to any pthreads * implementation, and allows a destructor function to be defined. It * only (directly) supports per-thread pointers, which are always * initialized to NULL. It requires once-only allocation of a - * pthread_key_t value. It is relatively slow. + * pthread_key_t value. It is relatively slow. Typically few + * "pthread_key_t"s are available (POSIX requires only at least 128, + * glibc supplies only 1024). * * - The thread_local feature newly defined in C11 works with * any data type and initializer, and it is fast. thread_local does not @@ -115,7 +184,8 @@ void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *); * define what happens if one attempts to access a thread_local object * from a thread other than the one to which that object belongs. There * is no provision to call a user-specified destructor when a thread - * ends. + * ends. Typical implementations allow for an arbitrary amount of + * thread_local storage, but statically allocated only. * * - The __thread keyword is a GCC extension similar to thread_local but * with a longer history. __thread is not portable to every GCC version @@ -132,9 +202,36 @@ void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *); * needs key allocation? yes no no * arbitrary initializer? no yes yes * cross-thread access? yes no yes + * amount available? few arbitrary arbitrary + * dynamically allocated? yes no no + * + * + * Extensions + * ========== + * + * OVS provides some extensions and wrappers: + * + * - In a situation where the performance of thread_local or __thread is + * desirable, but portability is required, DEFINE_STATIC_PER_THREAD_DATA + * and DECLARE_EXTERN_PER_THREAD_DATA/DEFINE_EXTERN_PER_THREAD_DATA may + * be appropriate (see below). + * + * - DEFINE_PER_THREAD_MALLOCED_DATA can be convenient for simple + * per-thread malloc()'d buffers. + * + * - struct ovs_tsd provides an alternative to pthread_key_t that isn't + * limited to a small number of keys. */ -/* DEFINE_PER_THREAD_DATA(TYPE, NAME, INITIALIZER). +/* For static data, use this macro in a source file: + * + * DEFINE_STATIC_PER_THREAD_DATA(TYPE, NAME, INITIALIZER). + * + * For global data, "declare" the data in the header and "define" it in + * the source file, with: + * + * DECLARE_EXTERN_PER_THREAD_DATA(TYPE, NAME). + * DEFINE_EXTERN_PER_THREAD_DATA(NAME, INITIALIZER). * * One should prefer to use POSIX per-thread data, via pthread_key_t, when its * performance is acceptable, because of its portability (see the table above). @@ -172,23 +269,40 @@ void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *); #error #endif -#define DEFINE_PER_THREAD_DATA(TYPE, NAME, ...) \ - typedef TYPE NAME##_type; \ - static thread_local NAME##_type NAME##_var = __VA_ARGS__; \ - \ - static NAME##_type * \ - NAME##_get_unsafe(void) \ - { \ - return &NAME##_var; \ - } \ - \ - static NAME##_type * \ - NAME##_get(void) \ - { \ - return NAME##_get_unsafe(); \ +#define DEFINE_STATIC_PER_THREAD_DATA(TYPE, NAME, ...) \ + typedef TYPE NAME##_type; \ + \ + static NAME##_type * \ + NAME##_get_unsafe(void) \ + { \ + static thread_local NAME##_type var = __VA_ARGS__; \ + return &var; \ + } \ + \ + static NAME##_type * \ + NAME##_get(void) \ + { \ + return NAME##_get_unsafe(); \ + } +#define DECLARE_EXTERN_PER_THREAD_DATA(TYPE, NAME) \ + typedef TYPE NAME##_type; \ + extern thread_local NAME##_type NAME##_var; \ + \ + static inline NAME##_type * \ + NAME##_get_unsafe(void) \ + { \ + return &NAME##_var; \ + } \ + \ + static inline NAME##_type * \ + NAME##_get(void) \ + { \ + return NAME##_get_unsafe(); \ } +#define DEFINE_EXTERN_PER_THREAD_DATA(NAME, ...) \ + thread_local NAME##_type NAME##_var = __VA_ARGS__; #else /* no C implementation support for thread-local storage */ -#define DEFINE_PER_THREAD_DATA(TYPE, NAME, ...) \ +#define DEFINE_STATIC_PER_THREAD_DATA(TYPE, NAME, ...) \ typedef TYPE NAME##_type; \ static pthread_key_t NAME##_key; \ \ @@ -217,9 +331,52 @@ void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *); if (!value) { \ static const NAME##_type initial_value = __VA_ARGS__; \ \ - value = xmalloc(sizeof *value); \ + value = malloc(sizeof *value); \ + if (value == NULL) { \ + out_of_memory(); \ + } \ + *value = initial_value; \ + xpthread_setspecific(NAME##_key, value); \ + } \ + return value; \ + } +#define DECLARE_EXTERN_PER_THREAD_DATA(TYPE, NAME) \ + typedef TYPE NAME##_type; \ + static pthread_key_t NAME##_key; \ + \ + static inline NAME##_type * \ + NAME##_get_unsafe(void) \ + { \ + return pthread_getspecific(NAME##_key); \ + } \ + \ + NAME##_type *NAME##_get(void); +#define DEFINE_EXTERN_PER_THREAD_DATA(NAME, ...) \ + static void \ + NAME##_once_init(void) \ + { \ + if (pthread_key_create(&NAME##_key, free)) { \ + abort(); \ + } \ + } \ + \ + NAME##_type * \ + NAME##_get(void) \ + { \ + static pthread_once_t once = PTHREAD_ONCE_INIT; \ + NAME##_type *value; \ + \ + pthread_once(&once, NAME##_once_init); \ + value = NAME##_get_unsafe(); \ + if (!value) { \ + static const NAME##_type initial_value = __VA_ARGS__; \ + \ + value = malloc(sizeof *value); \ + if (value == NULL) { \ + out_of_memory(); \ + } \ *value = initial_value; \ - pthread_setspecific(NAME##_key, value); \ + xpthread_setspecific(NAME##_key, value); \ } \ return value; \ } @@ -290,7 +447,7 @@ void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *); NAME##_set_unsafe(TYPE value) \ { \ TYPE old_value = NAME##_get_unsafe(); \ - pthread_setspecific(NAME##_key, value); \ + xpthread_setspecific(NAME##_key, value); \ return old_value; \ } \ \ @@ -300,6 +457,37 @@ void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *); NAME##_init(); \ return NAME##_set_unsafe(value); \ } + +/* Dynamically allocated thread-specific data with lots of slots. + * + * pthread_key_t can provide as few as 128 pieces of thread-specific data (even + * glibc is limited to 1,024). Thus, one must be careful to allocate only a + * few keys globally. One cannot, for example, allocate a key for every + * instance of a data structure if there might be an arbitrary number of those + * data structures. + * + * This API is similar to the pthread one (simply search and replace pthread_ + * by ovsthread_) but it a much larger limit that can be raised if necessary + * (by recompiling). Thus, one may more freely use this form of + * thread-specific data. + * + * ovsthread_key_t also differs from pthread_key_t in the following ways: + * + * - Destructors must not access thread-specific data (via ovsthread_key). + * + * - The pthread_key_t API allows concurrently exiting threads to start + * executing the destructor after pthread_key_delete() returns. The + * ovsthread_key_t API guarantees that, when ovsthread_key_delete() + * returns, all destructors have returned and no new ones will start + * execution. + */ +typedef struct ovsthread_key *ovsthread_key_t; + +void ovsthread_key_create(ovsthread_key_t *, void (*destructor)(void *)); +void ovsthread_key_delete(ovsthread_key_t); + +void ovsthread_setspecific(ovsthread_key_t, const void *); +void *ovsthread_getspecific(ovsthread_key_t); /* Convenient once-only execution. * @@ -335,22 +523,25 @@ void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *); struct ovsthread_once { atomic_bool done; - pthread_mutex_t mutex; + struct ovs_mutex mutex; }; #define OVSTHREAD_ONCE_INITIALIZER \ { \ ATOMIC_VAR_INIT(false), \ - PTHREAD_ADAPTIVE_MUTEX_INITIALIZER, \ + OVS_MUTEX_INITIALIZER, \ } -static inline bool ovsthread_once_start(struct ovsthread_once *); -void ovsthread_once_done(struct ovsthread_once *once) OVS_RELEASES(once); +static inline bool ovsthread_once_start(struct ovsthread_once *once) + OVS_TRY_LOCK(true, once->mutex); +void ovsthread_once_done(struct ovsthread_once *once) + OVS_RELEASES(once->mutex); -bool ovsthread_once_start__(struct ovsthread_once *); +bool ovsthread_once_start__(struct ovsthread_once *once) + OVS_TRY_LOCK(false, once->mutex); static inline bool -ovsthread_once_is_done__(const struct ovsthread_once *once) +ovsthread_once_is_done__(struct ovsthread_once *once) { bool done; @@ -371,19 +562,71 @@ ovsthread_once_start(struct ovsthread_once *once) return OVS_UNLIKELY(!ovsthread_once_is_done__(once) && !ovsthread_once_start__(once)); } + +/* Thread ID. + * + * pthread_t isn't so nice for some purposes. Its size and representation are + * implementation dependent, which means that there is no way to hash it. + * This thread ID avoids the problem. + */ -#ifdef __CHECKER__ -#define ovsthread_once_start(ONCE) \ - ((ONCE)->done ? false : ({ OVS_ACQUIRE(ONCE); true; })) -#endif +DECLARE_EXTERN_PER_THREAD_DATA(unsigned int, ovsthread_id); + +/* Returns a per-thread identifier unique within the lifetime of the + * process. */ +static inline unsigned int +ovsthread_id_self(void) +{ + return *ovsthread_id_get(); +} +/* Simulated global counter. + * + * Incrementing such a counter is meant to be cheaper than incrementing a + * global counter protected by a lock. It is probably more expensive than + * incrementing a truly thread-local variable, but such a variable has no + * straightforward way to get the sum. + * + * + * Thread-safety + * ============= + * + * Fully thread-safe. */ + +struct ovsthread_stats { + struct ovs_mutex mutex; + void *volatile buckets[16]; +}; + +void ovsthread_stats_init(struct ovsthread_stats *); +void ovsthread_stats_destroy(struct ovsthread_stats *); + +void *ovsthread_stats_bucket_get(struct ovsthread_stats *, + void *(*new_bucket)(void)); + +#define OVSTHREAD_STATS_FOR_EACH_BUCKET(BUCKET, IDX, STATS) \ + for ((IDX) = ovs_thread_stats_next_bucket(STATS, 0); \ + ((IDX) < ARRAY_SIZE((STATS)->buckets) \ + ? ((BUCKET) = (STATS)->buckets[IDX], true) \ + : false); \ + (IDX) = ovs_thread_stats_next_bucket(STATS, (IDX) + 1)) +size_t ovs_thread_stats_next_bucket(const struct ovsthread_stats *, size_t); + +bool single_threaded(void); + void assert_single_threaded_at(const char *where); #define assert_single_threaded() assert_single_threaded_at(SOURCE_LOCATOR) +#ifndef _WIN32 pid_t xfork_at(const char *where); #define xfork() xfork_at(SOURCE_LOCATOR) +#endif void forbid_forking(const char *reason); bool may_fork(void); + +/* Useful functions related to threading. */ + +int count_cpu_cores(void); #endif /* ovs-thread.h */