diff --git a/src/util/meson.build b/src/util/meson.build index 0b5e2305204..7fc9ea85f2d 100644 --- a/src/util/meson.build +++ b/src/util/meson.build @@ -138,6 +138,7 @@ files_mesa_util = files( 'u_queue.c', 'u_queue.h', 'u_string.h', + 'u_thread.c', 'u_thread.h', 'u_vector.c', 'u_vector.h', diff --git a/src/util/u_thread.c b/src/util/u_thread.c new file mode 100644 index 00000000000..55b6b68589c --- /dev/null +++ b/src/util/u_thread.c @@ -0,0 +1,236 @@ +/* + * Copyright 1999-2006 Brian Paul + * Copyright 2008 VMware, Inc. + * Copyright 2022 Yonggang Luo + * SPDX-License-Identifier: MIT + */ + +#include "util/u_thread.h" + +#include "macros.h" + +#ifdef HAVE_PTHREAD +#include +#ifdef HAVE_PTHREAD_NP_H +#include +#endif +#endif + +#ifdef __HAIKU__ +#include +#endif + +#if DETECT_OS_LINUX && !defined(ANDROID) +#include +#elif defined(_WIN32) && !defined(HAVE_PTHREAD) +#include +#endif + +#ifdef __FreeBSD__ +/* pthread_np.h -> sys/param.h -> machine/param.h + * - defines ALIGN which clashes with our ALIGN + */ +#undef ALIGN +#define cpu_set_t cpuset_t +#endif + +int +util_get_current_cpu(void) +{ +#if DETECT_OS_LINUX && !defined(ANDROID) + return sched_getcpu(); + +#elif defined(_WIN32) && !defined(HAVE_PTHREAD) + return GetCurrentProcessorNumber(); + +#else + return -1; +#endif +} + +int u_thread_create(thrd_t *thrd, int (*routine)(void *), void *param) +{ + int ret = thrd_error; +#ifdef HAVE_PTHREAD + sigset_t saved_set, new_set; + + sigfillset(&new_set); + sigdelset(&new_set, SIGSYS); + + /* SIGSEGV is commonly used by Vulkan API tracing layers in order to track + * accesses in device memory mapped to user space. Blocking the signal hinders + * that tracking mechanism. + */ + sigdelset(&new_set, SIGSEGV); + pthread_sigmask(SIG_BLOCK, &new_set, &saved_set); + ret = thrd_create(thrd, routine, param); + pthread_sigmask(SIG_SETMASK, &saved_set, NULL); +#else + ret = thrd_create(thrd, routine, param); +#endif + + return ret; +} + +void u_thread_setname( const char *name ) +{ +#if defined(HAVE_PTHREAD) +#if DETECT_OS_LINUX || DETECT_OS_CYGWIN || DETECT_OS_SOLARIS || defined(__GLIBC__) + int ret = pthread_setname_np(pthread_self(), name); + if (ret == ERANGE) { + char buf[16]; + const size_t len = MIN2(strlen(name), ARRAY_SIZE(buf) - 1); + memcpy(buf, name, len); + buf[len] = '\0'; + pthread_setname_np(pthread_self(), buf); + } +#elif DETECT_OS_FREEBSD || DETECT_OS_OPENBSD + pthread_set_name_np(pthread_self(), name); +#elif DETECT_OS_NETBSD + pthread_setname_np(pthread_self(), "%s", (void *)name); +#elif DETECT_OS_APPLE + pthread_setname_np(name); +#elif DETECT_OS_HAIKU + rename_thread(find_thread(NULL), name); +#else +#warning Not sure how to call pthread_setname_np +#endif +#endif + (void)name; +} + +bool +util_set_thread_affinity(thrd_t thread, + const uint32_t *mask, + uint32_t *old_mask, + unsigned num_mask_bits) +{ +#if defined(HAVE_PTHREAD_SETAFFINITY) + cpu_set_t cpuset; + + if (old_mask) { + if (pthread_getaffinity_np(thread, sizeof(cpuset), &cpuset) != 0) + return false; + + memset(old_mask, 0, num_mask_bits / 8); + for (unsigned i = 0; i < num_mask_bits && i < CPU_SETSIZE; i++) { + if (CPU_ISSET(i, &cpuset)) + old_mask[i / 32] |= 1u << (i % 32); + } + } + + CPU_ZERO(&cpuset); + for (unsigned i = 0; i < num_mask_bits && i < CPU_SETSIZE; i++) { + if (mask[i / 32] & (1u << (i % 32))) + CPU_SET(i, &cpuset); + } + return pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset) == 0; + +#elif defined(_WIN32) && !defined(HAVE_PTHREAD) + DWORD_PTR m = mask[0]; + + if (sizeof(m) > 4 && num_mask_bits > 32) + m |= (uint64_t)mask[1] << 32; + + m = SetThreadAffinityMask(thread.handle, m); + if (!m) + return false; + + if (old_mask) { + memset(old_mask, 0, num_mask_bits / 8); + + old_mask[0] = m; +#ifdef _WIN64 + old_mask[1] = m >> 32; +#endif + } + + return true; +#else + return false; +#endif +} + +int64_t +util_thread_get_time_nano(thrd_t thread) +{ +#if defined(HAVE_PTHREAD) && !defined(__APPLE__) && !defined(__HAIKU__) + struct timespec ts; + clockid_t cid; + + pthread_getcpuclockid(thread, &cid); + clock_gettime(cid, &ts); + return (int64_t)ts.tv_sec * 1000000000 + ts.tv_nsec; +#elif defined(_WIN32) + union { + FILETIME time; + ULONGLONG value; + } kernel_time, user_time; + GetThreadTimes((HANDLE)thread.handle, NULL, NULL, &kernel_time.time, &user_time.time); + return (kernel_time.value + user_time.value) * 100; +#else + (void)thread; + return 0; +#endif +} + +#if defined(HAVE_PTHREAD) && !defined(__APPLE__) && !defined(__HAIKU__) + +void util_barrier_init(util_barrier *barrier, unsigned count) +{ + pthread_barrier_init(barrier, NULL, count); +} + +void util_barrier_destroy(util_barrier *barrier) +{ + pthread_barrier_destroy(barrier); +} + +bool util_barrier_wait(util_barrier *barrier) +{ + return pthread_barrier_wait(barrier) == PTHREAD_BARRIER_SERIAL_THREAD; +} + +#else /* If the OS doesn't have its own, implement barriers using a mutex and a condvar */ + +void util_barrier_init(util_barrier *barrier, unsigned count) +{ + barrier->count = count; + barrier->waiters = 0; + barrier->sequence = 0; + (void) mtx_init(&barrier->mutex, mtx_plain); + cnd_init(&barrier->condvar); +} + +void util_barrier_destroy(util_barrier *barrier) +{ + assert(barrier->waiters == 0); + mtx_destroy(&barrier->mutex); + cnd_destroy(&barrier->condvar); +} + +bool util_barrier_wait(util_barrier *barrier) +{ + mtx_lock(&barrier->mutex); + + assert(barrier->waiters < barrier->count); + barrier->waiters++; + + if (barrier->waiters < barrier->count) { + uint64_t sequence = barrier->sequence; + + do { + cnd_wait(&barrier->condvar, &barrier->mutex); + } while (sequence == barrier->sequence); + } else { + barrier->waiters = 0; + barrier->sequence++; + cnd_broadcast(&barrier->condvar); + } + + mtx_unlock(&barrier->mutex); + + return true; +} + +#endif diff --git a/src/util/u_thread.h b/src/util/u_thread.h index 544eea55829..7c532bca22c 100644 --- a/src/util/u_thread.h +++ b/src/util/u_thread.h @@ -34,32 +34,6 @@ #include "c11/threads.h" #include "detect_os.h" -#include "macros.h" - -#ifdef HAVE_PTHREAD -#include -#ifdef HAVE_PTHREAD_NP_H -#include -#endif -#endif - -#ifdef __HAIKU__ -#include -#endif - -#if DETECT_OS_LINUX && !defined(ANDROID) -#include -#elif defined(_WIN32) && !defined(HAVE_PTHREAD) -#include -#endif - -#ifdef __FreeBSD__ -/* pthread_np.h -> sys/param.h -> machine/param.h - * - defines ALIGN which clashes with our ALIGN - */ -#undef ALIGN -#define cpu_set_t cpuset_t -#endif /* For util_set_thread_affinity to size the mask. */ #define UTIL_MAX_CPUS 1024 /* this should be enough */ @@ -87,70 +61,16 @@ #define __THREAD_INITIAL_EXEC thread_local #endif -static inline int -util_get_current_cpu(void) -{ -#if DETECT_OS_LINUX && !defined(ANDROID) - return sched_getcpu(); - -#elif defined(_WIN32) && !defined(HAVE_PTHREAD) - return GetCurrentProcessorNumber(); - -#else - return -1; -#endif -} - -static inline int u_thread_create(thrd_t *thrd, int (*routine)(void *), void *param) -{ - int ret = thrd_error; -#ifdef HAVE_PTHREAD - sigset_t saved_set, new_set; - - sigfillset(&new_set); - sigdelset(&new_set, SIGSYS); - - /* SIGSEGV is commonly used by Vulkan API tracing layers in order to track - * accesses in device memory mapped to user space. Blocking the signal hinders - * that tracking mechanism. - */ - sigdelset(&new_set, SIGSEGV); - pthread_sigmask(SIG_BLOCK, &new_set, &saved_set); - ret = thrd_create(thrd, routine, param); - pthread_sigmask(SIG_SETMASK, &saved_set, NULL); -#else - ret = thrd_create(thrd, routine, param); +#ifdef __cplusplus +extern "C" { #endif - return ret; -} +int +util_get_current_cpu(void); -static inline void u_thread_setname( const char *name ) -{ -#if defined(HAVE_PTHREAD) -#if DETECT_OS_LINUX || DETECT_OS_CYGWIN || DETECT_OS_SOLARIS || defined(__GLIBC__) - int ret = pthread_setname_np(pthread_self(), name); - if (ret == ERANGE) { - char buf[16]; - const size_t len = MIN2(strlen(name), ARRAY_SIZE(buf) - 1); - memcpy(buf, name, len); - buf[len] = '\0'; - pthread_setname_np(pthread_self(), buf); - } -#elif DETECT_OS_FREEBSD || DETECT_OS_OPENBSD - pthread_set_name_np(pthread_self(), name); -#elif DETECT_OS_NETBSD - pthread_setname_np(pthread_self(), "%s", (void *)name); -#elif DETECT_OS_APPLE - pthread_setname_np(name); -#elif DETECT_OS_HAIKU - rename_thread(find_thread(NULL), name); -#else -#warning Not sure how to call pthread_setname_np -#endif -#endif - (void)name; -} +int u_thread_create(thrd_t *thrd, int (*routine)(void *), void *param); + +void u_thread_setname( const char *name ); /** * Set thread affinity. @@ -161,57 +81,11 @@ static inline void u_thread_setname( const char *name ) * \param num_mask_bits Number of bits in both masks * \return true on success */ -static inline bool +bool util_set_thread_affinity(thrd_t thread, const uint32_t *mask, uint32_t *old_mask, - unsigned num_mask_bits) -{ -#if defined(HAVE_PTHREAD_SETAFFINITY) - cpu_set_t cpuset; - - if (old_mask) { - if (pthread_getaffinity_np(thread, sizeof(cpuset), &cpuset) != 0) - return false; - - memset(old_mask, 0, num_mask_bits / 8); - for (unsigned i = 0; i < num_mask_bits && i < CPU_SETSIZE; i++) { - if (CPU_ISSET(i, &cpuset)) - old_mask[i / 32] |= 1u << (i % 32); - } - } - - CPU_ZERO(&cpuset); - for (unsigned i = 0; i < num_mask_bits && i < CPU_SETSIZE; i++) { - if (mask[i / 32] & (1u << (i % 32))) - CPU_SET(i, &cpuset); - } - return pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset) == 0; - -#elif defined(_WIN32) && !defined(HAVE_PTHREAD) - DWORD_PTR m = mask[0]; - - if (sizeof(m) > 4 && num_mask_bits > 32) - m |= (uint64_t)mask[1] << 32; - - m = SetThreadAffinityMask(thread.handle, m); - if (!m) - return false; - - if (old_mask) { - memset(old_mask, 0, num_mask_bits / 8); - - old_mask[0] = m; -#ifdef _WIN64 - old_mask[1] = m >> 32; -#endif - } - - return true; -#else - return false; -#endif -} + unsigned num_mask_bits); static inline bool util_set_current_thread_affinity(const uint32_t *mask, @@ -227,28 +101,8 @@ util_set_current_thread_affinity(const uint32_t *mask, */ /* Return the time of a thread's CPU time clock. */ -static inline int64_t -util_thread_get_time_nano(thrd_t thread) -{ -#if defined(HAVE_PTHREAD) && !defined(__APPLE__) && !defined(__HAIKU__) - struct timespec ts; - clockid_t cid; - - pthread_getcpuclockid(thread, &cid); - clock_gettime(cid, &ts); - return (int64_t)ts.tv_sec * 1000000000 + ts.tv_nsec; -#elif defined(_WIN32) - union { - FILETIME time; - ULONGLONG value; - } kernel_time, user_time; - GetThreadTimes((HANDLE)thread.handle, NULL, NULL, &kernel_time.time, &user_time.time); - return (kernel_time.value + user_time.value) * 100; -#else - (void)thread; - return 0; -#endif -} +int64_t +util_thread_get_time_nano(thrd_t thread); /* Return the time of the current thread's CPU time clock. */ static inline int64_t @@ -270,21 +124,11 @@ static inline bool u_thread_is_self(thrd_t thread) typedef pthread_barrier_t util_barrier; -static inline void util_barrier_init(util_barrier *barrier, unsigned count) -{ - pthread_barrier_init(barrier, NULL, count); -} +void util_barrier_init(util_barrier *barrier, unsigned count); -static inline void util_barrier_destroy(util_barrier *barrier) -{ - pthread_barrier_destroy(barrier); -} - -static inline bool util_barrier_wait(util_barrier *barrier) -{ - return pthread_barrier_wait(barrier) == PTHREAD_BARRIER_SERIAL_THREAD; -} +void util_barrier_destroy(util_barrier *barrier); +bool util_barrier_wait(util_barrier *barrier); #else /* If the OS doesn't have its own, implement barriers using a mutex and a condvar */ @@ -296,46 +140,16 @@ typedef struct { cnd_t condvar; } util_barrier; -static inline void util_barrier_init(util_barrier *barrier, unsigned count) -{ - barrier->count = count; - barrier->waiters = 0; - barrier->sequence = 0; - (void) mtx_init(&barrier->mutex, mtx_plain); - cnd_init(&barrier->condvar); -} +void util_barrier_init(util_barrier *barrier, unsigned count); -static inline void util_barrier_destroy(util_barrier *barrier) -{ - assert(barrier->waiters == 0); - mtx_destroy(&barrier->mutex); - cnd_destroy(&barrier->condvar); -} +void util_barrier_destroy(util_barrier *barrier); -static inline bool util_barrier_wait(util_barrier *barrier) -{ - mtx_lock(&barrier->mutex); - - assert(barrier->waiters < barrier->count); - barrier->waiters++; - - if (barrier->waiters < barrier->count) { - uint64_t sequence = barrier->sequence; - - do { - cnd_wait(&barrier->condvar, &barrier->mutex); - } while (sequence == barrier->sequence); - } else { - barrier->waiters = 0; - barrier->sequence++; - cnd_broadcast(&barrier->condvar); - } - - mtx_unlock(&barrier->mutex); - - return true; -} +bool util_barrier_wait(util_barrier *barrier); #endif +#ifdef __cplusplus +} +#endif + #endif /* U_THREAD_H_ */