Index: benchmark/io/readv.cfa =================================================================== --- benchmark/io/readv.cfa (revision c680a4ba5e93338ec036f1c5135c217e61c20ac1) +++ benchmark/io/readv.cfa (revision 44aad8f4ae26b1fe3eef138bd44cc12610ad6e9a) @@ -18,5 +18,5 @@ extern bool traceHeapOn(); -extern ssize_t async_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); +extern ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); int fd; @@ -34,5 +34,5 @@ while(__atomic_load_n(&run, __ATOMIC_RELAXED)) { - async_preadv2(fd, &iov, 1, 0, 0); + cfa_preadv2(fd, &iov, 1, 0, 0); __atomic_fetch_add( &count, 1, __ATOMIC_SEQ_CST ); } @@ -40,4 +40,8 @@ int main(int argc, char * argv[]) { + #if !defined(__CFA_NO_STATISTICS__) + print_stats_at_exit( *active_cluster() ); + #endif + double duration = 5.0; unsigned long int nthreads = 2; @@ -46,5 +50,4 @@ printf("Setting local\n"); setlocale(LC_NUMERIC, ""); - arg_loop: Index: examples/io/simple/server.cfa =================================================================== --- examples/io/simple/server.cfa (revision c680a4ba5e93338ec036f1c5135c217e61c20ac1) +++ examples/io/simple/server.cfa (revision 44aad8f4ae26b1fe3eef138bd44cc12610ad6e9a) @@ -51,7 +51,7 @@ //---------- -extern ssize_t async_recvmsg(int sockfd, struct msghdr *msg, int flags); -extern int async_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags); -extern int async_close(int fd); +extern ssize_t cfa_recvmsg(int sockfd, struct msghdr *msg, int flags); +extern int cfa_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags); +extern int cfa_close(int fd); //---------- @@ -88,5 +88,5 @@ struct sockaddr_in cli_addr; __socklen_t clilen = sizeof(cli_addr); - int newsock = async_accept4(sock, (struct sockaddr *) &cli_addr, &clilen, 0); + int newsock = cfa_accept4(sock, (struct sockaddr *) &cli_addr, &clilen, 0); if (newsock < 0) { error( printer, "accept", -newsock); @@ -97,5 +97,5 @@ while(1) { - int res = async_recvmsg(newsock, &msg, 0); + int res = cfa_recvmsg(newsock, &msg, 0); if(res == 0) break; if(res < 0) { @@ -107,5 +107,5 @@ } - ret = async_close(newsock); + ret = cfa_close(newsock); if(ret < 0) { error( printer, "close new", -ret); @@ -113,5 +113,5 @@ } - ret = async_close(sock); + ret = cfa_close(sock); if(ret < 0) { error( printer, "close old", -ret); Index: libcfa/src/concurrency/io.cfa =================================================================== --- libcfa/src/concurrency/io.cfa (revision c680a4ba5e93338ec036f1c5135c217e61c20ac1) +++ libcfa/src/concurrency/io.cfa (revision 44aad8f4ae26b1fe3eef138bd44cc12610ad6e9a) @@ -1,2 +1,17 @@ +// +// Cforall Version 1.0.0 Copyright (C) 2020 University of Waterloo +// +// The contents of this file are covered under the licence agreement in the +// file "LICENCE" distributed with Cforall. +// +// io.cfa -- +// +// Author : Thierry Delisle +// Created On : Thu Apr 23 17:31:00 2020 +// Last Modified By : +// Last Modified On : +// Update Count : +// + #include "kernel.hfa" @@ -8,8 +23,4 @@ void __kernel_io_shutdown( cluster & this ) { // Nothing to do without io_uring - } - - bool is_async( void (*)() ) { - return false; } @@ -90,8 +101,10 @@ // Requires features - // // adjust the size according to the parameters - // if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { - // cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz); - // } + #if defined(IORING_FEAT_SINGLE_MMAP) + // adjust the size according to the parameters + if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { + cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz); + } + #endif // mmap the Submit Queue into existence @@ -101,10 +114,13 @@ } - // mmap the Completion Queue into existence (may or may not be needed) // Requires features - // if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { - // cq->ring_ptr = sq->ring_ptr; - // } - // else { + #if defined(IORING_FEAT_SINGLE_MMAP) + // mmap the Completion Queue into existence (may or may not be needed) + if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { + cq->ring_ptr = sq->ring_ptr; + } + else + #endif + { // We need multiple call to MMAP cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING); @@ -113,5 +129,5 @@ abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno)); } - // } + } // mmap the submit queue entries @@ -160,4 +176,12 @@ (this.io.submit){ min(*sq.num, *cq.num) }; + // Initialize statistics + #if !defined(__CFA_NO_STATISTICS__) + this.io.submit_q.stats.submit_avg.val = 0; + this.io.submit_q.stats.submit_avg.cnt = 0; + this.io.completion_q.stats.completed_avg.val = 0; + this.io.completion_q.stats.completed_avg.cnt = 0; + #endif + // Create the poller thread this.io.stack = __create_pthread( &this.io.poller, __io_poller, &this ); @@ -174,4 +198,21 @@ pthread_join( this.io.poller, 0p ); free( this.io.stack ); + + // print statistics + #if !defined(__CFA_NO_STATISTICS__) + if(this.print_stats) { + __cfaabi_bits_print_safe( STDERR_FILENO, + "----- I/O uRing Stats -----\n" + "- total submit calls : %llu\n" + "- avg submit : %lf\n" + "- total wait calls : %llu\n" + "- avg completion/wait : %lf\n", + this.io.submit_q.stats.submit_avg.cnt, + ((double)this.io.submit_q.stats.submit_avg.val) / this.io.submit_q.stats.submit_avg.cnt, + this.io.completion_q.stats.completed_avg.cnt, + ((double)this.io.completion_q.stats.completed_avg.val) / this.io.completion_q.stats.completed_avg.cnt + ); + } + #endif // Shutdown the io rings @@ -204,29 +245,56 @@ // Process a single completion message from the io_uring // This is NOT thread-safe - static bool __io_process(struct io_ring & ring) { + int __drain_io( struct io_ring & ring, sigset_t & mask, int waitcnt ) { + int ret = syscall( __NR_io_uring_enter, ring.fd, 0, waitcnt, IORING_ENTER_GETEVENTS, &mask, _NSIG / 8); + if( ret < 0 ) { + switch((int)errno) { + case EAGAIN: + case EINTR: + return -EAGAIN; + default: + abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) ); + } + } + + // Drain the queue unsigned head = *ring.completion_q.head; unsigned tail = __atomic_load_n(ring.completion_q.tail, __ATOMIC_ACQUIRE); - if (head == tail) return false; - - unsigned idx = head & (*ring.completion_q.mask); - struct io_uring_cqe & cqe = ring.completion_q.cqes[idx]; - - /* paranoid */ verify(&cqe); - - struct io_user_data * data = (struct io_user_data *)cqe.user_data; - // __cfaabi_bits_print_safe( STDERR_FILENO, "Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd ); - - data->result = cqe.res; - __unpark( data->thrd __cfaabi_dbg_ctx2 ); + // Nothing was new return 0 + if (head == tail) { + #if !defined(__CFA_NO_STATISTICS__) + ring.completion_q.stats.completed_avg.cnt += 1; + #endif + return 0; + } + + uint32_t count = tail - head; + for(i; count) { + unsigned idx = (head + i) & (*ring.completion_q.mask); + struct io_uring_cqe & cqe = ring.completion_q.cqes[idx]; + + /* paranoid */ verify(&cqe); + + struct io_user_data * data = (struct io_user_data *)cqe.user_data; + // __cfaabi_bits_print_safe( STDERR_FILENO, "Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd ); + + data->result = cqe.res; + __unpark( data->thrd __cfaabi_dbg_ctx2 ); + } // Allow new submissions to happen - V(ring.submit); + V(ring.submit, count); // Mark to the kernel that the cqe has been seen // Ensure that the kernel only sees the new value of the head index after the CQEs have been read. - __atomic_fetch_add( ring.completion_q.head, 1, __ATOMIC_RELAXED ); - - return true; + __atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED ); + + // Update statistics + #if !defined(__CFA_NO_STATISTICS__) + ring.completion_q.stats.completed_avg.val += count; + ring.completion_q.stats.completed_avg.cnt += 1; + #endif + + return count; } @@ -246,18 +314,6 @@ verify( (*ring.completion_q.head) == (*ring.completion_q.tail) ); - LOOP: while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) { - int ret = syscall( __NR_io_uring_enter, ring.fd, 0, 1, IORING_ENTER_GETEVENTS, &mask, _NSIG / 8); - if( ret < 0 ) { - switch((int)errno) { - case EAGAIN: - case EINTR: - continue LOOP; - default: - abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) ); - } - } - - // Drain the queue - while(__io_process(ring)) {} + while(!__atomic_load_n(&ring.done, __ATOMIC_SEQ_CST)) { + __drain_io( ring, mask, 1 ); } @@ -293,83 +349,82 @@ // -static inline [* struct io_uring_sqe, uint32_t] __submit_alloc( struct io_ring & ring ) { - // Wait for a spot to be available - P(ring.submit); - - // Allocate the sqe - uint32_t idx = __atomic_fetch_add(&ring.submit_q.alloc, 1ul32, __ATOMIC_SEQ_CST); - - // Validate that we didn't overflow anything - // Check that nothing overflowed - /* paranoid */ verify( true ); - - // Check that it goes head -> tail -> alloc and never head -> alloc -> tail - /* paranoid */ verify( true ); - - // Return the sqe - return [&ring.submit_q.sqes[ idx & (*ring.submit_q.mask)], idx]; -} - -static inline void __submit( struct io_ring & ring, uint32_t idx ) { - // get mutual exclusion - lock(ring.submit_q.lock __cfaabi_dbg_ctx2); - - // Append to the list of ready entries - uint32_t * tail = ring.submit_q.tail; - const uint32_t mask = *ring.submit_q.mask; - - ring.submit_q.array[ (*tail) & mask ] = idx & mask; - __atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST); - - // Submit however, many entries need to be submitted - int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0); - // __cfaabi_bits_print_safe( STDERR_FILENO, "Performed io_submit, returned %d\n", ret ); - if( ret < 0 ) { - switch((int)errno) { - default: - abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) ); - } - } - - unlock(ring.submit_q.lock); - // Make sure that idx was submitted - // Be careful to not get false positive if we cycled the entire list or that someone else submitted for us -} - -static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd) { - this.opcode = opcode; - #if !defined(IOSQE_ASYNC) - this.flags = 0; - #else - this.flags = IOSQE_ASYNC; - #endif - this.ioprio = 0; - this.fd = fd; - this.off = 0; - this.addr = 0; - this.len = 0; - this.rw_flags = 0; - this.__pad2[0] = this.__pad2[1] = this.__pad2[2] = 0; -} - -static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd, void * addr, uint32_t len, uint64_t off ) { - (this){ opcode, fd }; - this.off = off; - this.addr = (uint64_t)addr; - this.len = len; -} + static inline [* struct io_uring_sqe, uint32_t] __submit_alloc( struct io_ring & ring ) { + // Wait for a spot to be available + P(ring.submit); + + // Allocate the sqe + uint32_t idx = __atomic_fetch_add(&ring.submit_q.alloc, 1ul32, __ATOMIC_SEQ_CST); + + // Validate that we didn't overflow anything + // Check that nothing overflowed + /* paranoid */ verify( true ); + + // Check that it goes head -> tail -> alloc and never head -> alloc -> tail + /* paranoid */ verify( true ); + + // Return the sqe + return [&ring.submit_q.sqes[ idx & (*ring.submit_q.mask)], idx]; + } + + static inline void __submit( struct io_ring & ring, uint32_t idx ) { + // get mutual exclusion + lock(ring.submit_q.lock __cfaabi_dbg_ctx2); + + // Append to the list of ready entries + uint32_t * tail = ring.submit_q.tail; + const uint32_t mask = *ring.submit_q.mask; + + ring.submit_q.array[ (*tail) & mask ] = idx & mask; + __atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST); + + // Submit however, many entries need to be submitted + int ret = syscall( __NR_io_uring_enter, ring.fd, 1, 0, 0, 0p, 0); + // __cfaabi_bits_print_safe( STDERR_FILENO, "Performed io_submit, returned %d\n", ret ); + if( ret < 0 ) { + switch((int)errno) { + default: + abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) ); + } + } + + // update statistics + #if !defined(__CFA_NO_STATISTICS__) + ring.submit_q.stats.submit_avg.val += 1; + ring.submit_q.stats.submit_avg.cnt += 1; + #endif + + unlock(ring.submit_q.lock); + // Make sure that idx was submitted + // Be careful to not get false positive if we cycled the entire list or that someone else submitted for us + } + + static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd) { + this.opcode = opcode; + #if !defined(IOSQE_ASYNC) + this.flags = 0; + #else + this.flags = IOSQE_ASYNC; + #endif + this.ioprio = 0; + this.fd = fd; + this.off = 0; + this.addr = 0; + this.len = 0; + this.rw_flags = 0; + this.__pad2[0] = this.__pad2[1] = this.__pad2[2] = 0; + } + + static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd, void * addr, uint32_t len, uint64_t off ) { + (this){ opcode, fd }; + this.off = off; + this.addr = (uint64_t)addr; + this.len = len; + } +#endif //============================================================================================= // I/O Interface //============================================================================================= - extern "C" { - #define __USE_GNU - #define _GNU_SOURCE - #include - #include - #include - #include - } - +#if defined(HAVE_LINUX_IO_URING_H) #define __submit_prelude \ struct io_ring & ring = active_cluster()->io; \ @@ -385,261 +440,290 @@ park( __cfaabi_dbg_ctx ); \ return data.result; +#endif + +// Some forward declarations +extern "C" { + #include + struct iovec; + extern ssize_t preadv2 (int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); + extern ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); + + extern int fsync(int fd); + extern int sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags); + + struct msghdr; + struct sockaddr; + extern ssize_t sendmsg(int sockfd, const struct msghdr *msg, int flags); + extern ssize_t recvmsg(int sockfd, struct msghdr *msg, int flags); + extern ssize_t send(int sockfd, const void *buf, size_t len, int flags); + extern ssize_t recv(int sockfd, void *buf, size_t len, int flags); + extern int accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags); + extern int connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen); + + extern int fallocate(int fd, int mode, uint64_t offset, uint64_t len); + extern int posix_fadvise(int fd, uint64_t offset, uint64_t len, int advice); + extern int madvise(void *addr, size_t length, int advice); + + extern int openat(int dirfd, const char *pathname, int flags, mode_t mode); + extern int close(int fd); + + struct statx; + extern int statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf); + + extern ssize_t read (int fd, void *buf, size_t count); +} //----------------------------------------------------------------------------- // Asynchronous operations - ssize_t async_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) { - #if !defined(IORING_OP_READV) - return preadv2(fd, iov, iovcnt, offset, flags); - #else - __submit_prelude - - (*sqe){ IORING_OP_READV, fd, iov, iovcnt, offset }; - - __submit_wait - #endif - } - - ssize_t async_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) { - #if !defined(IORING_OP_WRITEV) - return pwritev2(fd, iov, iovcnt, offset, flags); - #else - __submit_prelude - - (*sqe){ IORING_OP_WRITEV, fd, iov, iovcnt, offset }; - - __submit_wait - #endif - } - - int async_fsync(int fd) { - #if !defined(IORING_OP_FSYNC) - return fsync(fd); - #else - __submit_prelude - - (*sqe){ IORING_OP_FSYNC, fd }; - - __submit_wait - #endif - } - - int async_sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags) { - #if !defined(IORING_OP_SYNC_FILE_RANGE) - return sync_file_range(fd, offset, nbytes, flags); - #else - __submit_prelude - - (*sqe){ IORING_OP_SYNC_FILE_RANGE, fd }; - sqe->off = offset; - sqe->len = nbytes; - sqe->sync_range_flags = flags; - - __submit_wait - #endif - } - - - ssize_t async_sendmsg(int sockfd, const struct msghdr *msg, int flags) { - #if !defined(IORING_OP_SENDMSG) - return recv(sockfd, msg, flags); - #else - __submit_prelude - - (*sqe){ IORING_OP_SENDMSG, sockfd, msg, 1, 0 }; - sqe->msg_flags = flags; - - __submit_wait - #endif - } - - ssize_t async_recvmsg(int sockfd, struct msghdr *msg, int flags) { - #if !defined(IORING_OP_RECVMSG) - return recv(sockfd, msg, flags); - #else - __submit_prelude - - (*sqe){ IORING_OP_RECVMSG, sockfd, msg, 1, 0 }; - sqe->msg_flags = flags; - - __submit_wait - #endif - } - - ssize_t async_send(int sockfd, const void *buf, size_t len, int flags) { - #if !defined(IORING_OP_SEND) - return send( sockfd, buf, len, flags ); - #else - __submit_prelude - - (*sqe){ IORING_OP_SEND, sockfd }; - sqe->addr = (uint64_t)buf; - sqe->len = len; - sqe->msg_flags = flags; - - __submit_wait - #endif - } - - ssize_t async_recv(int sockfd, void *buf, size_t len, int flags) { - #if !defined(IORING_OP_RECV) - return recv( sockfd, buf, len, flags ); - #else - __submit_prelude - - (*sqe){ IORING_OP_RECV, sockfd }; - sqe->addr = (uint64_t)buf; - sqe->len = len; - sqe->msg_flags = flags; - - __submit_wait - #endif - } - - int async_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags) { - #if !defined(IORING_OP_ACCEPT) - __SOCKADDR_ARG _addr; - _addr.__sockaddr__ = addr; - return accept4( sockfd, _addr, addrlen, flags ); - #else - __submit_prelude - - (*sqe){ IORING_OP_ACCEPT, sockfd }; - sqe->addr = addr; - sqe->addr2 = addrlen; - sqe->accept_flags = flags; - - __submit_wait - #endif - } - - int async_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen) { - #if !defined(IORING_OP_CONNECT) - __CONST_SOCKADDR_ARG _addr; - _addr.__sockaddr__ = addr; - return connect( sockfd, _addr, addrlen ); - #else - __submit_prelude - - (*sqe){ IORING_OP_CONNECT, sockfd }; - sqe->addr = (uint64_t)addr; - sqe->off = addrlen; - - __submit_wait - #endif - } - - int async_fallocate(int fd, int mode, uint64_t offset, uint64_t len) { - #if !defined(IORING_OP_FALLOCATE) - return fallocate( fd, mode, offset, len ); - #else - __submit_prelude - - (*sqe){ IORING_OP_FALLOCATE, fd }; - sqe->off = offset; - sqe->len = length; - sqe->mode = mode; - - __submit_wait - #endif - } - - int async_fadvise(int fd, uint64_t offset, uint64_t len, int advice) { - #if !defined(IORING_OP_FADVISE) - return posix_fadvise( fd, offset, len, advice ); - #else - __submit_prelude - - (*sqe){ IORING_OP_FADVISE, fd }; - sqe->off = (uint64_t)offset; - sqe->len = length; - sqe->fadvise_advice = advice; - - __submit_wait - #endif - } - - int async_madvise(void *addr, size_t length, int advice) { - #if !defined(IORING_OP_MADVISE) - return madvise( addr, length, advice ); - #else - __submit_prelude - - (*sqe){ IORING_OP_MADVISE, 0 }; - sqe->addr = (uint64_t)addr; - sqe->len = length; - sqe->fadvise_advice = advice; - - __submit_wait - #endif - } - - int async_openat(int dirfd, const char *pathname, int flags, mode_t mode) { - #if !defined(IORING_OP_OPENAT) - return openat( dirfd, pathname, flags, mode ); - #else - __submit_prelude - - (*sqe){ IORING_OP_OPENAT, dirfd }; - sqe->addr = (uint64_t)pathname; - sqe->open_flags = flags; - sqe->mode = mode; - - __submit_wait - #endif - } - - int async_close(int fd) { - #if !defined(IORING_OP_CLOSE) - return close( fd ); - #else - __submit_prelude - - (*sqe){ IORING_OP_CLOSE, fd }; - - __submit_wait - #endif - } - - int async_statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf) { - #if !defined(IORING_OP_STATX) - //return statx( dirfd, pathname, flags, mask, statxbuf ); - return syscall( __NR_io_uring_setup, dirfd, pathname, flags, mask, statxbuf ); - #else - __submit_prelude - - (*sqe){ IORING_OP_STATX, dirfd }; - sqe->addr = (uint64_t)pathname; - sqe->statx_flags = flags; - sqe->len = mask; - sqe->off = (uint64_t)statxbuf; - - __submit_wait - #endif - } - - - ssize_t async_read(int fd, void *buf, size_t count) { - #if !defined(IORING_OP_READ) - return read( fd, buf, count ); - #else - __submit_prelude - - (*sqe){ IORING_OP_READ, fd, buf, count, 0 }; - - __submit_wait - #endif - } - - ssize_t async_write(int fd, void *buf, size_t count) { - #if !defined(IORING_OP_WRITE) - return read( fd, buf, count ); - #else - __submit_prelude - - (*sqe){ IORING_OP_WRITE, fd, buf, count, 0 }; - - __submit_wait - #endif - } +ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_READV) + return preadv2(fd, iov, iovcnt, offset, flags); + #else + __submit_prelude + + (*sqe){ IORING_OP_READV, fd, iov, iovcnt, offset }; + + __submit_wait + #endif +} + +ssize_t cfa_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_WRITEV) + return pwritev2(fd, iov, iovcnt, offset, flags); + #else + __submit_prelude + + (*sqe){ IORING_OP_WRITEV, fd, iov, iovcnt, offset }; + + __submit_wait + #endif +} + +int cfa_fsync(int fd) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FSYNC) + return fsync(fd); + #else + __submit_prelude + + (*sqe){ IORING_OP_FSYNC, fd }; + + __submit_wait + #endif +} + +int cfa_sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SYNC_FILE_RANGE) + return sync_file_range(fd, offset, nbytes, flags); + #else + __submit_prelude + + (*sqe){ IORING_OP_SYNC_FILE_RANGE, fd }; + sqe->off = offset; + sqe->len = nbytes; + sqe->sync_range_flags = flags; + + __submit_wait + #endif +} + + +ssize_t cfa_sendmsg(int sockfd, const struct msghdr *msg, int flags) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SENDMSG) + return recv(sockfd, msg, flags); + #else + __submit_prelude + + (*sqe){ IORING_OP_SENDMSG, sockfd, msg, 1, 0 }; + sqe->msg_flags = flags; + + __submit_wait + #endif +} + +ssize_t cfa_recvmsg(int sockfd, struct msghdr *msg, int flags) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_RECVMSG) + return recv(sockfd, msg, flags); + #else + __submit_prelude + + (*sqe){ IORING_OP_RECVMSG, sockfd, msg, 1, 0 }; + sqe->msg_flags = flags; + + __submit_wait + #endif +} + +ssize_t cfa_send(int sockfd, const void *buf, size_t len, int flags) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SEND) + return send( sockfd, buf, len, flags ); + #else + __submit_prelude + + (*sqe){ IORING_OP_SEND, sockfd }; + sqe->addr = (uint64_t)buf; + sqe->len = len; + sqe->msg_flags = flags; + + __submit_wait + #endif +} + +ssize_t cfa_recv(int sockfd, void *buf, size_t len, int flags) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_RECV) + return recv( sockfd, buf, len, flags ); + #else + __submit_prelude + + (*sqe){ IORING_OP_RECV, sockfd }; + sqe->addr = (uint64_t)buf; + sqe->len = len; + sqe->msg_flags = flags; + + __submit_wait + #endif +} + +int cfa_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_ACCEPT) + return accept4( sockfd, addr, addrlen, flags ); + #else + __submit_prelude + + (*sqe){ IORING_OP_ACCEPT, sockfd }; + sqe->addr = addr; + sqe->addr2 = addrlen; + sqe->accept_flags = flags; + + __submit_wait + #endif +} + +int cfa_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_CONNECT) + return connect( sockfd, addr, addrlen ); + #else + __submit_prelude + + (*sqe){ IORING_OP_CONNECT, sockfd }; + sqe->addr = (uint64_t)addr; + sqe->off = addrlen; + + __submit_wait + #endif +} + +int cfa_fallocate(int fd, int mode, uint64_t offset, uint64_t len) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FALLOCATE) + return fallocate( fd, mode, offset, len ); + #else + __submit_prelude + + (*sqe){ IORING_OP_FALLOCATE, fd }; + sqe->off = offset; + sqe->len = length; + sqe->mode = mode; + + __submit_wait + #endif +} + +int cfa_fadvise(int fd, uint64_t offset, uint64_t len, int advice) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FADVISE) + return posix_fadvise( fd, offset, len, advice ); + #else + __submit_prelude + + (*sqe){ IORING_OP_FADVISE, fd }; + sqe->off = (uint64_t)offset; + sqe->len = length; + sqe->fadvise_advice = advice; + + __submit_wait + #endif +} + +int cfa_madvise(void *addr, size_t length, int advice) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_MADVISE) + return madvise( addr, length, advice ); + #else + __submit_prelude + + (*sqe){ IORING_OP_MADVISE, 0 }; + sqe->addr = (uint64_t)addr; + sqe->len = length; + sqe->fadvise_advice = advice; + + __submit_wait + #endif +} + +int cfa_openat(int dirfd, const char *pathname, int flags, mode_t mode) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_OPENAT) + return openat( dirfd, pathname, flags, mode ); + #else + __submit_prelude + + (*sqe){ IORING_OP_OPENAT, dirfd }; + sqe->addr = (uint64_t)pathname; + sqe->open_flags = flags; + sqe->mode = mode; + + __submit_wait + #endif +} + +int cfa_close(int fd) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_CLOSE) + return close( fd ); + #else + __submit_prelude + + (*sqe){ IORING_OP_CLOSE, fd }; + + __submit_wait + #endif +} + +int cfa_statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_STATX) + //return statx( dirfd, pathname, flags, mask, statxbuf ); + return syscall( __NR_io_uring_setup, dirfd, pathname, flags, mask, statxbuf ); + #else + __submit_prelude + + (*sqe){ IORING_OP_STATX, dirfd }; + sqe->addr = (uint64_t)pathname; + sqe->statx_flags = flags; + sqe->len = mask; + sqe->off = (uint64_t)statxbuf; + + __submit_wait + #endif +} + + +ssize_t cfa_read(int fd, void *buf, size_t count) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_READ) + return read( fd, buf, count ); + #else + __submit_prelude + + (*sqe){ IORING_OP_READ, fd, buf, count, 0 }; + + __submit_wait + #endif +} + +ssize_t cfa_write(int fd, void *buf, size_t count) { + #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_WRITE) + return read( fd, buf, count ); + #else + __submit_prelude + + (*sqe){ IORING_OP_WRITE, fd, buf, count, 0 }; + + __submit_wait + #endif +} //----------------------------------------------------------------------------- @@ -647,103 +731,102 @@ // Macro magic to reduce the size of the following switch case - #define IS_DEFINED_APPLY(f, ...) f(__VA_ARGS__) - #define IS_DEFINED_SECOND(first, second, ...) second - #define IS_DEFINED_TEST(expansion) _CFA_IO_FEATURE_##expansion - #define IS_DEFINED(macro) IS_DEFINED_APPLY( IS_DEFINED_SECOND,IS_DEFINED_TEST(macro) false, true) - - bool is_async( fptr_t func ) { - +#define IS_DEFINED_APPLY(f, ...) f(__VA_ARGS__) +#define IS_DEFINED_SECOND(first, second, ...) second +#define IS_DEFINED_TEST(expansion) _CFA_IO_FEATURE_##expansion +#define IS_DEFINED(macro) IS_DEFINED_APPLY( IS_DEFINED_SECOND,IS_DEFINED_TEST(macro) false, true) + +bool has_user_level_blocking( fptr_t func ) { + #if defined(HAVE_LINUX_IO_URING_H) if( /*func == (fptr_t)preadv2 || */ - func == (fptr_t)async_preadv2 ) + func == (fptr_t)cfa_preadv2 ) #define _CFA_IO_FEATURE_IORING_OP_READV , return IS_DEFINED(IORING_OP_READV); if( /*func == (fptr_t)pwritev2 || */ - func == (fptr_t)async_pwritev2 ) + func == (fptr_t)cfa_pwritev2 ) #define _CFA_IO_FEATURE_IORING_OP_WRITEV , return IS_DEFINED(IORING_OP_WRITEV); if( /*func == (fptr_t)fsync || */ - func == (fptr_t)async_fsync ) + func == (fptr_t)cfa_fsync ) #define _CFA_IO_FEATURE_IORING_OP_FSYNC , return IS_DEFINED(IORING_OP_FSYNC); if( /*func == (fptr_t)ync_file_range || */ - func == (fptr_t)async_sync_file_range ) + func == (fptr_t)cfa_sync_file_range ) #define _CFA_IO_FEATURE_IORING_OP_SYNC_FILE_RANGE , return IS_DEFINED(IORING_OP_SYNC_FILE_RANGE); if( /*func == (fptr_t)sendmsg || */ - func == (fptr_t)async_sendmsg ) + func == (fptr_t)cfa_sendmsg ) #define _CFA_IO_FEATURE_IORING_OP_SENDMSG , return IS_DEFINED(IORING_OP_SENDMSG); if( /*func == (fptr_t)recvmsg || */ - func == (fptr_t)async_recvmsg ) + func == (fptr_t)cfa_recvmsg ) #define _CFA_IO_FEATURE_IORING_OP_RECVMSG , return IS_DEFINED(IORING_OP_RECVMSG); if( /*func == (fptr_t)send || */ - func == (fptr_t)async_send ) + func == (fptr_t)cfa_send ) #define _CFA_IO_FEATURE_IORING_OP_SEND , return IS_DEFINED(IORING_OP_SEND); if( /*func == (fptr_t)recv || */ - func == (fptr_t)async_recv ) + func == (fptr_t)cfa_recv ) #define _CFA_IO_FEATURE_IORING_OP_RECV , return IS_DEFINED(IORING_OP_RECV); if( /*func == (fptr_t)accept4 || */ - func == (fptr_t)async_accept4 ) + func == (fptr_t)cfa_accept4 ) #define _CFA_IO_FEATURE_IORING_OP_ACCEPT , return IS_DEFINED(IORING_OP_ACCEPT); if( /*func == (fptr_t)connect || */ - func == (fptr_t)async_connect ) + func == (fptr_t)cfa_connect ) #define _CFA_IO_FEATURE_IORING_OP_CONNECT , return IS_DEFINED(IORING_OP_CONNECT); if( /*func == (fptr_t)fallocate || */ - func == (fptr_t)async_fallocate ) + func == (fptr_t)cfa_fallocate ) #define _CFA_IO_FEATURE_IORING_OP_FALLOCATE , return IS_DEFINED(IORING_OP_FALLOCATE); - if( /*func == (fptr_t)fadvise || */ - func == (fptr_t)async_fadvise ) + if( /*func == (fptr_t)posix_fadvise || */ + func == (fptr_t)cfa_fadvise ) #define _CFA_IO_FEATURE_IORING_OP_FADVISE , return IS_DEFINED(IORING_OP_FADVISE); if( /*func == (fptr_t)madvise || */ - func == (fptr_t)async_madvise ) + func == (fptr_t)cfa_madvise ) #define _CFA_IO_FEATURE_IORING_OP_MADVISE , return IS_DEFINED(IORING_OP_MADVISE); if( /*func == (fptr_t)openat || */ - func == (fptr_t)async_openat ) + func == (fptr_t)cfa_openat ) #define _CFA_IO_FEATURE_IORING_OP_OPENAT , return IS_DEFINED(IORING_OP_OPENAT); if( /*func == (fptr_t)close || */ - func == (fptr_t)async_close ) + func == (fptr_t)cfa_close ) #define _CFA_IO_FEATURE_IORING_OP_CLOSE , return IS_DEFINED(IORING_OP_CLOSE); if( /*func == (fptr_t)statx || */ - func == (fptr_t)async_statx ) + func == (fptr_t)cfa_statx ) #define _CFA_IO_FEATURE_IORING_OP_STATX , return IS_DEFINED(IORING_OP_STATX); if( /*func == (fptr_t)read || */ - func == (fptr_t)async_read ) + func == (fptr_t)cfa_read ) #define _CFA_IO_FEATURE_IORING_OP_READ , return IS_DEFINED(IORING_OP_READ); if( /*func == (fptr_t)write || */ - func == (fptr_t)async_write ) + func == (fptr_t)cfa_write ) #define _CFA_IO_FEATURE_IORING_OP_WRITE , return IS_DEFINED(IORING_OP_WRITE); - - return false; - } - -#endif + #endif + + return false; +} Index: libcfa/src/concurrency/iofwd.hfa =================================================================== --- libcfa/src/concurrency/iofwd.hfa (revision 44aad8f4ae26b1fe3eef138bd44cc12610ad6e9a) +++ libcfa/src/concurrency/iofwd.hfa (revision 44aad8f4ae26b1fe3eef138bd44cc12610ad6e9a) @@ -0,0 +1,39 @@ +// +// Cforall Version 1.0.0 Copyright (C) 2020 University of Waterloo +// +// The contents of this file are covered under the licence agreement in the +// file "LICENCE" distributed with Cforall. +// +// iofwd.hfa -- +// +// Author : Thierry Delisle +// Created On : Thu Apr 23 17:31:00 2020 +// Last Modified By : +// Last Modified On : +// Update Count : +// + +#pragma once + +ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); +ssize_t cfa_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); +int cfa_fsync(int fd); +int cfa_sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags); +ssize_t cfa_sendmsg(int sockfd, const struct msghdr *msg, int flags); +ssize_t cfa_recvmsg(int sockfd, struct msghdr *msg, int flags); +ssize_t cfa_send(int sockfd, const void *buf, size_t len, int flags); +ssize_t cfa_recv(int sockfd, void *buf, size_t len, int flags); +int cfa_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags); +int cfa_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen); +int cfa_fallocate(int fd, int mode, uint64_t offset, uint64_t len); +int cfa_fadvise(int fd, uint64_t offset, uint64_t len, int advice); +int cfa_madvise(void *addr, size_t length, int advice); +int cfa_openat(int dirfd, const char *pathname, int flags, mode_t mode); +int cfa_close(int fd); +int cfa_statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf); +ssize_t cfa_read(int fd, void *buf, size_t count); +ssize_t cfa_write(int fd, void *buf, size_t count) + +//----------------------------------------------------------------------------- +// Check if a function is blocks a only the user thread +bool has_user_level_blocking( fptr_t func ); Index: libcfa/src/concurrency/kernel.cfa =================================================================== --- libcfa/src/concurrency/kernel.cfa (revision c680a4ba5e93338ec036f1c5135c217e61c20ac1) +++ libcfa/src/concurrency/kernel.cfa (revision 44aad8f4ae26b1fe3eef138bd44cc12610ad6e9a) @@ -257,4 +257,8 @@ ready_queue{}; ready_queue_lock{}; + + #if !defined(__CFA_NO_STATISTICS__) + print_stats = false; + #endif procs{ __get }; @@ -1004,4 +1008,18 @@ } +bool V(semaphore & this, unsigned diff) with( this ) { + $thread * thrd = 0p; + lock( lock __cfaabi_dbg_ctx2 ); + int release = max(-count, (int)diff); + count += diff; + for(release) { + unpark( pop_head( waiting ) __cfaabi_dbg_ctx2 ); + } + + unlock( lock ); + + return thrd != 0p; +} + //----------------------------------------------------------------------------- // Global Queues Index: libcfa/src/concurrency/kernel.hfa =================================================================== --- libcfa/src/concurrency/kernel.hfa (revision c680a4ba5e93338ec036f1c5135c217e61c20ac1) +++ libcfa/src/concurrency/kernel.hfa (revision 44aad8f4ae26b1fe3eef138bd44cc12610ad6e9a) @@ -40,4 +40,5 @@ void P (semaphore & this); bool V (semaphore & this); +bool V (semaphore & this, unsigned count); @@ -144,4 +145,14 @@ void * ring_ptr; size_t ring_sz; + + // Statistics + #if !defined(__CFA_NO_STATISTICS__) + struct { + struct { + unsigned long long int val; + unsigned long long int cnt; + } submit_avg; + } stats; + #endif }; @@ -164,4 +175,14 @@ void * ring_ptr; size_t ring_sz; + + // Statistics + #if !defined(__CFA_NO_STATISTICS__) + struct { + struct { + unsigned long long int val; + unsigned long long int cnt; + } completed_avg; + } stats; + #endif }; @@ -213,4 +234,8 @@ struct io_ring io; #endif + + #if !defined(__CFA_NO_STATISTICS__) + bool print_stats; + #endif }; extern Duration default_preemption(); @@ -227,4 +252,10 @@ static inline struct processor * active_processor() { return TL_GET( this_processor ); } // UNSAFE static inline struct cluster * active_cluster () { return TL_GET( this_processor )->cltr; } + +#if !defined(__CFA_NO_STATISTICS__) + static inline void print_stats_at_exit( cluster & this ) { + this.print_stats = true; + } +#endif // Local Variables: //