Changes in libcfa/src/concurrency/io.cfa [df40a56:732b406]
- File:
-
- 1 edited
-
libcfa/src/concurrency/io.cfa (modified) (22 diffs)
Legend:
- Unmodified
- Added
- Removed
-
libcfa/src/concurrency/io.cfa
rdf40a56 r732b406 38 38 39 39 #else 40 #define _GNU_SOURCE /* See feature_test_macros(7) */ 41 #include <errno.h> 42 #include <stdint.h> 43 #include <string.h> 44 #include <unistd.h> 45 #include <sys/mman.h> 46 40 47 extern "C" { 41 #define _GNU_SOURCE /* See feature_test_macros(7) */42 #include <errno.h>43 #include <stdint.h>44 #include <string.h>45 #include <unistd.h>46 #include <sys/mman.h>47 48 #include <sys/syscall.h> 48 49 … … 108 109 volatile uint32_t * head; 109 110 volatile uint32_t * tail; 111 volatile uint32_t prev_head; 110 112 111 113 // The actual kernel ring which uses head/tail … … 128 130 129 131 __spinlock_t lock; 132 __spinlock_t release_lock; 130 133 131 134 // A buffer of sqes (not the actual ring) … … 181 184 //============================================================================================= 182 185 void __kernel_io_startup( cluster & this, unsigned io_flags, bool main_cluster ) { 186 if( (io_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS) && (io_flags & CFA_CLUSTER_IO_EAGER_SUBMITS) ) { 187 abort("CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS and CFA_CLUSTER_IO_EAGER_SUBMITS cannot be mixed\n"); 188 } 189 183 190 this.io = malloc(); 184 191 … … 186 193 struct io_uring_params params; 187 194 memset(¶ms, 0, sizeof(params)); 195 if( io_flags & CFA_CLUSTER_IO_KERNEL_POLL_SUBMITS ) params.flags |= IORING_SETUP_SQPOLL; 196 if( io_flags & CFA_CLUSTER_IO_KERNEL_POLL_COMPLETES ) params.flags |= IORING_SETUP_IOPOLL; 188 197 189 198 uint32_t nentries = entries_per_cluster(); … … 252 261 sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped); 253 262 sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array); 263 sq.prev_head = *sq.head; 254 264 255 265 { … … 260 270 } 261 271 262 if( io_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) { 272 (sq.lock){}; 273 (sq.release_lock){}; 274 275 if( io_flags & ( CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS | CFA_CLUSTER_IO_EAGER_SUBMITS ) ) { 263 276 /* paranoid */ verify( is_pow2( io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET ) || ((io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET) < 8) ); 264 277 sq.ready_cnt = max(io_flags >> CFA_CLUSTER_IO_BUFFLEN_OFFSET, 8); … … 420 433 // I/O Polling 421 434 //============================================================================================= 422 struct io_user_data { 423 int32_t result; 424 $thread * thrd; 425 }; 435 static unsigned __collect_submitions( struct __io_data & ring ); 436 static uint32_t __release_consumed_submission( struct __io_data & ring ); 426 437 427 438 // Process a single completion message from the io_uring 428 439 // This is NOT thread-safe 429 440 static [int, bool] __drain_io( & struct __io_data ring, * sigset_t mask, int waitcnt, bool in_kernel ) { 441 /* paranoid */ verify( !kernelTLS.preemption_state.enabled ); 442 const uint32_t smask = *ring.submit_q.mask; 443 430 444 unsigned to_submit = 0; 431 445 if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) { 432 433 446 // If the poller thread also submits, then we need to aggregate the submissions which are ready 434 uint32_t tail = *ring.submit_q.tail; 435 const uint32_t mask = *ring.submit_q.mask; 436 437 // Go through the list of ready submissions 438 for( i; ring.submit_q.ready_cnt ) { 439 // replace any submission with the sentinel, to consume it. 440 uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED); 441 442 // If it was already the sentinel, then we are done 443 if( idx == -1ul32 ) continue; 444 445 // If we got a real submission, append it to the list 446 ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask; 447 to_submit++; 448 } 449 450 // Increment the tail based on how many we are ready to submit 451 __atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST); 452 } 453 454 const uint32_t smask = *ring.submit_q.mask; 455 uint32_t shead = *ring.submit_q.head; 456 int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, waitcnt, IORING_ENTER_GETEVENTS, mask, _NSIG / 8); 457 if( ret < 0 ) { 458 switch((int)errno) { 459 case EAGAIN: 460 case EINTR: 461 return -EAGAIN; 462 default: 463 abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) ); 464 } 465 } 466 467 // Release the consumed SQEs 468 for( i; ret ) { 469 uint32_t idx = ring.submit_q.array[ (i + shead) & smask ]; 470 ring.submit_q.sqes[ idx ].user_data = 0; 471 } 472 473 uint32_t avail = 0; 474 uint32_t sqe_num = *ring.submit_q.num; 475 for(i; sqe_num) { 476 if( ring.submit_q.sqes[ i ].user_data == 0 ) avail++; 477 } 478 479 // update statistics 480 #if !defined(__CFA_NO_STATISTICS__) 481 __tls_stats()->io.submit_q.submit_avg.rdy += to_submit; 482 __tls_stats()->io.submit_q.submit_avg.csm += ret; 483 __tls_stats()->io.submit_q.submit_avg.avl += avail; 484 __tls_stats()->io.submit_q.submit_avg.cnt += 1; 485 #endif 447 to_submit = __collect_submitions( ring ); 448 } 449 450 if (to_submit > 0 || waitcnt > 0) { 451 int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, waitcnt, IORING_ENTER_GETEVENTS, mask, _NSIG / 8); 452 if( ret < 0 ) { 453 switch((int)errno) { 454 case EAGAIN: 455 case EINTR: 456 return [0, true]; 457 default: 458 abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) ); 459 } 460 } 461 462 // Release the consumed SQEs 463 __release_consumed_submission( ring ); 464 465 // update statistics 466 __STATS__( true, 467 if( to_submit > 0 ) { 468 io.submit_q.submit_avg.rdy += to_submit; 469 io.submit_q.submit_avg.csm += ret; 470 io.submit_q.submit_avg.cnt += 1; 471 } 472 ) 473 } 474 475 // Memory barrier 476 __atomic_thread_fence( __ATOMIC_SEQ_CST ); 486 477 487 478 // Drain the queue … … 490 481 const uint32_t mask = *ring.completion_q.mask; 491 482 492 // Memory barrier493 __atomic_thread_fence( __ATOMIC_SEQ_CST );494 495 483 // Nothing was new return 0 496 484 if (head == tail) { 497 return 0;485 return [0, to_submit > 0]; 498 486 } 499 487 500 488 uint32_t count = tail - head; 489 /* paranoid */ verify( count != 0 ); 501 490 for(i; count) { 502 491 unsigned idx = (head + i) & mask; … … 505 494 /* paranoid */ verify(&cqe); 506 495 507 struct io_user_data * data = (struct io_user_data*)(uintptr_t)cqe.user_data;496 struct __io_user_data_t * data = (struct __io_user_data_t *)(uintptr_t)cqe.user_data; 508 497 __cfadbg_print_safe( io, "Kernel I/O : Performed reading io cqe %p, result %d for %p\n", data, cqe.res, data->thrd ); 509 498 … … 563 552 564 553 // Update statistics 565 #if !defined(__CFA_NO_STATISTICS__)566 __tls_stats()->io.complete_q.completed_avg.val += count;567 __tls_stats()->io.complete_q.completed_avg.slow_cnt += 1;568 #endif554 __STATS__( true, 555 io.complete_q.completed_avg.val += count; 556 io.complete_q.completed_avg.slow_cnt += 1; 557 ) 569 558 570 559 if(again) { … … 580 569 int count; 581 570 bool again; 582 [count, again] = __drain_io( ring, &mask, 0, true );571 [count, again] = __drain_io( ring, &mask, 1, true ); 583 572 584 573 // Update statistics 585 #if !defined(__CFA_NO_STATISTICS__)586 __tls_stats()->io.complete_q.completed_avg.val += count;587 __tls_stats()->io.complete_q.completed_avg.slow_cnt += 1;588 #endif574 __STATS__( true, 575 io.complete_q.completed_avg.val += count; 576 io.complete_q.completed_avg.slow_cnt += 1; 577 ) 589 578 } 590 579 } … … 623 612 624 613 // Update statistics 625 #if !defined(__CFA_NO_STATISTICS__)626 __tls_stats()->io.complete_q.completed_avg.val += count;627 __tls_stats()->io.complete_q.completed_avg.fast_cnt += 1;628 #endif614 __STATS__( true, 615 io.complete_q.completed_avg.val += count; 616 io.complete_q.completed_avg.fast_cnt += 1; 617 ) 629 618 enable_interrupts( __cfaabi_dbg_ctx ); 630 619 … … 662 651 663 652 // Submition steps : 664 // 1 - We need to make sure we don't overflow any of the buffer, P(ring.submit) to make sure 665 // entries are available. The semaphore make sure that there is no more operations in 666 // progress then the number of entries in the buffer. This probably limits concurrency 667 // more than necessary since submitted but not completed operations don't need any 668 // entries in user space. However, I don't know what happens if we overflow the buffers 669 // because too many requests completed at once. This is a safe approach in all cases. 670 // Furthermore, with hundreds of entries, this may be okay. 671 // 672 // 2 - Allocate a queue entry. The ring already has memory for all entries but only the ones 653 // 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones 673 654 // listed in sq.array are visible by the kernel. For those not listed, the kernel does not 674 655 // offer any assurance that an entry is not being filled by multiple flags. Therefore, we 675 656 // need to write an allocator that allows allocating concurrently. 676 657 // 677 // 3- Actually fill the submit entry, this is the only simple and straightforward step.658 // 2 - Actually fill the submit entry, this is the only simple and straightforward step. 678 659 // 679 // 4- Append the entry index to the array and adjust the tail accordingly. This operation660 // 3 - Append the entry index to the array and adjust the tail accordingly. This operation 680 661 // needs to arrive to two concensus at the same time: 681 662 // A - The order in which entries are listed in the array: no two threads must pick the … … 685 666 // 686 667 687 // [* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) __attribute__((noinline)); 688 static inline [* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) { 689 verify( data != 0 ); 690 668 [* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) { 669 /* paranoid */ verify( data != 0 ); 691 670 692 671 // Prepare the data we need … … 713 692 { 714 693 // update statistics 715 #if !defined(__CFA_NO_STATISTICS__) 716 disable_interrupts(); 717 __tls_stats()->io.submit_q.alloc_avg.val += len; 718 __tls_stats()->io.submit_q.alloc_avg.block += block; 719 __tls_stats()->io.submit_q.alloc_avg.cnt += 1; 720 enable_interrupts( __cfaabi_dbg_ctx ); 721 #endif 694 __STATS__( false, 695 io.submit_q.alloc_avg.val += len; 696 io.submit_q.alloc_avg.block += block; 697 io.submit_q.alloc_avg.cnt += 1; 698 ) 722 699 723 700 … … 762 739 763 740 block++; 764 yield(); 741 if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) { 742 __release_consumed_submission( ring ); 743 unlock( ring.submit_q.lock ); 744 } 745 else { 746 yield(); 747 } 765 748 } 766 749 767 750 // update statistics 768 #if !defined(__CFA_NO_STATISTICS__) 769 disable_interrupts(); 770 __tls_stats()->io.submit_q.look_avg.val += len; 771 __tls_stats()->io.submit_q.look_avg.block += block; 772 __tls_stats()->io.submit_q.look_avg.cnt += 1; 773 enable_interrupts( __cfaabi_dbg_ctx ); 774 #endif 751 __STATS__( false, 752 io.submit_q.look_avg.val += len; 753 io.submit_q.look_avg.block += block; 754 io.submit_q.look_avg.cnt += 1; 755 ) 775 756 776 757 return picked; 777 758 } 778 759 779 // void __submit( struct __io_data & ring, uint32_t idx ) __attribute__((noinline)); 780 static inline void __submit( struct __io_data & ring, uint32_t idx ) { 760 void __submit( struct __io_data & ring, uint32_t idx ) { 781 761 // Get now the data we definetely need 782 762 uint32_t * const tail = ring.submit_q.tail; … … 786 766 if( ring.cltr_flags & CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS ) { 787 767 // If the poller thread submits, then we just need to add this to the ready array 788 789 768 __submit_to_ready_array( ring, idx, mask ); 790 769 … … 792 771 793 772 __cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() ); 773 } 774 else if( ring.cltr_flags & CFA_CLUSTER_IO_EAGER_SUBMITS ) { 775 uint32_t picked = __submit_to_ready_array( ring, idx, mask ); 776 777 for() { 778 yield(); 779 780 // If some one else collected our index, we are done 781 #warning ABA problem 782 if( ring.submit_q.ready[picked] != idx ) { 783 __STATS__( false, 784 io.submit_q.helped += 1; 785 ) 786 return; 787 } 788 789 if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) { 790 __STATS__( false, 791 io.submit_q.leader += 1; 792 ) 793 break; 794 } 795 796 __STATS__( false, 797 io.submit_q.busy += 1; 798 ) 799 } 800 801 // We got the lock 802 unsigned to_submit = __collect_submitions( ring ); 803 int ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, 0, 0p, _NSIG / 8); 804 if( ret < 0 ) { 805 switch((int)errno) { 806 case EAGAIN: 807 case EINTR: 808 unlock(ring.submit_q.lock); 809 return; 810 default: 811 abort( "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) ); 812 } 813 } 814 815 /* paranoid */ verify( ret > 0 ); 816 817 // Release the consumed SQEs 818 __release_consumed_submission( ring ); 819 820 // update statistics 821 __STATS__( true, 822 io.submit_q.submit_avg.rdy += to_submit; 823 io.submit_q.submit_avg.csm += ret; 824 io.submit_q.submit_avg.cnt += 1; 825 ) 826 827 unlock(ring.submit_q.lock); 794 828 } 795 829 else { … … 803 837 ring.submit_q.array[ (*tail) & mask ] = idx & mask; 804 838 __atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST); 839 840 /* paranoid */ verify( ring.submit_q.sqes[ idx ].user_data != 0 ); 805 841 806 842 // Submit however, many entries need to be submitted … … 814 850 815 851 // update statistics 816 #if !defined(__CFA_NO_STATISTICS__) 817 __tls_stats()->io.submit_q.submit_avg.csm += 1; 818 __tls_stats()->io.submit_q.submit_avg.cnt += 1; 819 #endif 820 821 ring.submit_q.sqes[ idx & mask ].user_data = 0; 852 __STATS__( false, 853 io.submit_q.submit_avg.csm += 1; 854 io.submit_q.submit_avg.cnt += 1; 855 ) 856 857 // Release the consumed SQEs 858 __release_consumed_submission( ring ); 822 859 823 860 unlock(ring.submit_q.lock); … … 827 864 } 828 865 829 static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd) { 830 this.opcode = opcode; 831 #if !defined(IOSQE_ASYNC) 832 this.flags = 0; 833 #else 834 this.flags = IOSQE_ASYNC; 835 #endif 836 this.ioprio = 0; 837 this.fd = fd; 838 this.off = 0; 839 this.addr = 0; 840 this.len = 0; 841 this.rw_flags = 0; 842 this.__pad2[0] = this.__pad2[1] = this.__pad2[2] = 0; 843 } 844 845 static inline void ?{}(struct io_uring_sqe & this, uint8_t opcode, int fd, void * addr, uint32_t len, uint64_t off ) { 846 (this){ opcode, fd }; 847 this.off = off; 848 this.addr = (uint64_t)(uintptr_t)addr; 849 this.len = len; 850 } 851 852 853 //============================================================================================= 854 // I/O Interface 855 //============================================================================================= 856 857 #define __submit_prelude \ 858 io_user_data data = { 0, active_thread() }; \ 859 struct __io_data & ring = *data.thrd->curr_cluster->io; \ 860 struct io_uring_sqe * sqe; \ 861 uint32_t idx; \ 862 [sqe, idx] = __submit_alloc( ring, (uint64_t)(uintptr_t)&data ); 863 864 #define __submit_wait \ 865 /*__cfaabi_bits_print_safe( STDERR_FILENO, "Preparing user data %p for %p\n", &data, data.thrd );*/ \ 866 verify( sqe->user_data == (uint64_t)(uintptr_t)&data ); \ 867 __submit( ring, idx ); \ 868 park( __cfaabi_dbg_ctx ); \ 869 return data.result; 866 static unsigned __collect_submitions( struct __io_data & ring ) { 867 /* paranoid */ verify( ring.submit_q.ready != 0p ); 868 /* paranoid */ verify( ring.submit_q.ready_cnt > 0 ); 869 870 unsigned to_submit = 0; 871 uint32_t tail = *ring.submit_q.tail; 872 const uint32_t mask = *ring.submit_q.mask; 873 874 // Go through the list of ready submissions 875 for( i; ring.submit_q.ready_cnt ) { 876 // replace any submission with the sentinel, to consume it. 877 uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED); 878 879 // If it was already the sentinel, then we are done 880 if( idx == -1ul32 ) continue; 881 882 // If we got a real submission, append it to the list 883 ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask; 884 to_submit++; 885 } 886 887 // Increment the tail based on how many we are ready to submit 888 __atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST); 889 890 return to_submit; 891 } 892 893 static uint32_t __release_consumed_submission( struct __io_data & ring ) { 894 const uint32_t smask = *ring.submit_q.mask; 895 896 if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0; 897 uint32_t chead = *ring.submit_q.head; 898 uint32_t phead = ring.submit_q.prev_head; 899 ring.submit_q.prev_head = chead; 900 unlock(ring.submit_q.release_lock); 901 902 uint32_t count = chead - phead; 903 for( i; count ) { 904 uint32_t idx = ring.submit_q.array[ (phead + i) & smask ]; 905 ring.submit_q.sqes[ idx ].user_data = 0; 906 } 907 return count; 908 } 870 909 #endif 871 872 // Some forward declarations873 extern "C" {874 #include <unistd.h>875 #include <sys/types.h>876 #include <sys/socket.h>877 #include <sys/syscall.h>878 879 #if defined(HAVE_PREADV2)880 struct iovec;881 extern ssize_t preadv2 (int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags);882 #endif883 #if defined(HAVE_PWRITEV2)884 struct iovec;885 extern ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags);886 #endif887 888 extern int fsync(int fd);889 extern int sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags);890 891 struct msghdr;892 struct sockaddr;893 extern ssize_t sendmsg(int sockfd, const struct msghdr *msg, int flags);894 extern ssize_t recvmsg(int sockfd, struct msghdr *msg, int flags);895 extern ssize_t send(int sockfd, const void *buf, size_t len, int flags);896 extern ssize_t recv(int sockfd, void *buf, size_t len, int flags);897 extern int accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags);898 extern int connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen);899 900 extern int fallocate(int fd, int mode, uint64_t offset, uint64_t len);901 extern int posix_fadvise(int fd, uint64_t offset, uint64_t len, int advice);902 extern int madvise(void *addr, size_t length, int advice);903 904 extern int openat(int dirfd, const char *pathname, int flags, mode_t mode);905 extern int close(int fd);906 907 extern ssize_t read (int fd, void *buf, size_t count);908 }909 910 //-----------------------------------------------------------------------------911 // Asynchronous operations912 #if defined(HAVE_PREADV2)913 ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) {914 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_READV)915 return preadv2(fd, iov, iovcnt, offset, flags);916 #else917 __submit_prelude918 919 (*sqe){ IORING_OP_READV, fd, iov, iovcnt, offset };920 921 __submit_wait922 #endif923 }924 #endif925 926 #if defined(HAVE_PWRITEV2)927 ssize_t cfa_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags) {928 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_WRITEV)929 return pwritev2(fd, iov, iovcnt, offset, flags);930 #else931 __submit_prelude932 933 (*sqe){ IORING_OP_WRITEV, fd, iov, iovcnt, offset };934 935 __submit_wait936 #endif937 }938 #endif939 940 int cfa_fsync(int fd) {941 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FSYNC)942 return fsync(fd);943 #else944 __submit_prelude945 946 (*sqe){ IORING_OP_FSYNC, fd };947 948 __submit_wait949 #endif950 }951 952 int cfa_sync_file_range(int fd, int64_t offset, int64_t nbytes, unsigned int flags) {953 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SYNC_FILE_RANGE)954 return sync_file_range(fd, offset, nbytes, flags);955 #else956 __submit_prelude957 958 (*sqe){ IORING_OP_SYNC_FILE_RANGE, fd };959 sqe->off = offset;960 sqe->len = nbytes;961 sqe->sync_range_flags = flags;962 963 __submit_wait964 #endif965 }966 967 968 ssize_t cfa_sendmsg(int sockfd, const struct msghdr *msg, int flags) {969 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SENDMSG)970 return sendmsg(sockfd, msg, flags);971 #else972 __submit_prelude973 974 (*sqe){ IORING_OP_SENDMSG, sockfd, msg, 1, 0 };975 sqe->msg_flags = flags;976 977 __submit_wait978 #endif979 }980 981 ssize_t cfa_recvmsg(int sockfd, struct msghdr *msg, int flags) {982 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_RECVMSG)983 return recvmsg(sockfd, msg, flags);984 #else985 __submit_prelude986 987 (*sqe){ IORING_OP_RECVMSG, sockfd, msg, 1, 0 };988 sqe->msg_flags = flags;989 990 __submit_wait991 #endif992 }993 994 ssize_t cfa_send(int sockfd, const void *buf, size_t len, int flags) {995 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_SEND)996 return send( sockfd, buf, len, flags );997 #else998 __submit_prelude999 1000 (*sqe){ IORING_OP_SEND, sockfd };1001 sqe->addr = (uint64_t)buf;1002 sqe->len = len;1003 sqe->msg_flags = flags;1004 1005 __submit_wait1006 #endif1007 }1008 1009 ssize_t cfa_recv(int sockfd, void *buf, size_t len, int flags) {1010 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_RECV)1011 return recv( sockfd, buf, len, flags );1012 #else1013 __submit_prelude1014 1015 (*sqe){ IORING_OP_RECV, sockfd };1016 sqe->addr = (uint64_t)buf;1017 sqe->len = len;1018 sqe->msg_flags = flags;1019 1020 __submit_wait1021 #endif1022 }1023 1024 int cfa_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags) {1025 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_ACCEPT)1026 return accept4( sockfd, addr, addrlen, flags );1027 #else1028 __submit_prelude1029 1030 (*sqe){ IORING_OP_ACCEPT, sockfd };1031 sqe->addr = addr;1032 sqe->addr2 = addrlen;1033 sqe->accept_flags = flags;1034 1035 __submit_wait1036 #endif1037 }1038 1039 int cfa_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen) {1040 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_CONNECT)1041 return connect( sockfd, addr, addrlen );1042 #else1043 __submit_prelude1044 1045 (*sqe){ IORING_OP_CONNECT, sockfd };1046 sqe->addr = (uint64_t)addr;1047 sqe->off = addrlen;1048 1049 __submit_wait1050 #endif1051 }1052 1053 int cfa_fallocate(int fd, int mode, uint64_t offset, uint64_t len) {1054 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FALLOCATE)1055 return fallocate( fd, mode, offset, len );1056 #else1057 __submit_prelude1058 1059 (*sqe){ IORING_OP_FALLOCATE, fd };1060 sqe->off = offset;1061 sqe->len = length;1062 sqe->mode = mode;1063 1064 __submit_wait1065 #endif1066 }1067 1068 int cfa_fadvise(int fd, uint64_t offset, uint64_t len, int advice) {1069 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_FADVISE)1070 return posix_fadvise( fd, offset, len, advice );1071 #else1072 __submit_prelude1073 1074 (*sqe){ IORING_OP_FADVISE, fd };1075 sqe->off = (uint64_t)offset;1076 sqe->len = length;1077 sqe->fadvise_advice = advice;1078 1079 __submit_wait1080 #endif1081 }1082 1083 int cfa_madvise(void *addr, size_t length, int advice) {1084 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_MADVISE)1085 return madvise( addr, length, advice );1086 #else1087 __submit_prelude1088 1089 (*sqe){ IORING_OP_MADVISE, 0 };1090 sqe->addr = (uint64_t)addr;1091 sqe->len = length;1092 sqe->fadvise_advice = advice;1093 1094 __submit_wait1095 #endif1096 }1097 1098 int cfa_openat(int dirfd, const char *pathname, int flags, mode_t mode) {1099 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_OPENAT)1100 return openat( dirfd, pathname, flags, mode );1101 #else1102 __submit_prelude1103 1104 (*sqe){ IORING_OP_OPENAT, dirfd };1105 sqe->addr = (uint64_t)pathname;1106 sqe->open_flags = flags;1107 sqe->mode = mode;1108 1109 __submit_wait1110 #endif1111 }1112 1113 int cfa_close(int fd) {1114 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_CLOSE)1115 return close( fd );1116 #else1117 __submit_prelude1118 1119 (*sqe){ IORING_OP_CLOSE, fd };1120 1121 __submit_wait1122 #endif1123 }1124 1125 1126 ssize_t cfa_read(int fd, void *buf, size_t count) {1127 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_READ)1128 return read( fd, buf, count );1129 #else1130 __submit_prelude1131 1132 (*sqe){ IORING_OP_READ, fd, buf, count, 0 };1133 1134 __submit_wait1135 #endif1136 }1137 1138 ssize_t cfa_write(int fd, void *buf, size_t count) {1139 #if !defined(HAVE_LINUX_IO_URING_H) || !defined(IORING_OP_WRITE)1140 return read( fd, buf, count );1141 #else1142 __submit_prelude1143 1144 (*sqe){ IORING_OP_WRITE, fd, buf, count, 0 };1145 1146 __submit_wait1147 #endif1148 }1149 1150 //-----------------------------------------------------------------------------1151 // Check if a function is asynchronous1152 1153 // Macro magic to reduce the size of the following switch case1154 #define IS_DEFINED_APPLY(f, ...) f(__VA_ARGS__)1155 #define IS_DEFINED_SECOND(first, second, ...) second1156 #define IS_DEFINED_TEST(expansion) _CFA_IO_FEATURE_##expansion1157 #define IS_DEFINED(macro) IS_DEFINED_APPLY( IS_DEFINED_SECOND,IS_DEFINED_TEST(macro) false, true)1158 1159 bool has_user_level_blocking( fptr_t func ) {1160 #if defined(HAVE_LINUX_IO_URING_H)1161 #if defined(HAVE_PREADV2)1162 if( /*func == (fptr_t)preadv2 || */1163 func == (fptr_t)cfa_preadv2 )1164 #define _CFA_IO_FEATURE_IORING_OP_READV ,1165 return IS_DEFINED(IORING_OP_READV);1166 #endif1167 1168 #if defined(HAVE_PWRITEV2)1169 if( /*func == (fptr_t)pwritev2 || */1170 func == (fptr_t)cfa_pwritev2 )1171 #define _CFA_IO_FEATURE_IORING_OP_WRITEV ,1172 return IS_DEFINED(IORING_OP_WRITEV);1173 #endif1174 1175 if( /*func == (fptr_t)fsync || */1176 func == (fptr_t)cfa_fsync )1177 #define _CFA_IO_FEATURE_IORING_OP_FSYNC ,1178 return IS_DEFINED(IORING_OP_FSYNC);1179 1180 if( /*func == (fptr_t)ync_file_range || */1181 func == (fptr_t)cfa_sync_file_range )1182 #define _CFA_IO_FEATURE_IORING_OP_SYNC_FILE_RANGE ,1183 return IS_DEFINED(IORING_OP_SYNC_FILE_RANGE);1184 1185 if( /*func == (fptr_t)sendmsg || */1186 func == (fptr_t)cfa_sendmsg )1187 #define _CFA_IO_FEATURE_IORING_OP_SENDMSG ,1188 return IS_DEFINED(IORING_OP_SENDMSG);1189 1190 if( /*func == (fptr_t)recvmsg || */1191 func == (fptr_t)cfa_recvmsg )1192 #define _CFA_IO_FEATURE_IORING_OP_RECVMSG ,1193 return IS_DEFINED(IORING_OP_RECVMSG);1194 1195 if( /*func == (fptr_t)send || */1196 func == (fptr_t)cfa_send )1197 #define _CFA_IO_FEATURE_IORING_OP_SEND ,1198 return IS_DEFINED(IORING_OP_SEND);1199 1200 if( /*func == (fptr_t)recv || */1201 func == (fptr_t)cfa_recv )1202 #define _CFA_IO_FEATURE_IORING_OP_RECV ,1203 return IS_DEFINED(IORING_OP_RECV);1204 1205 if( /*func == (fptr_t)accept4 || */1206 func == (fptr_t)cfa_accept4 )1207 #define _CFA_IO_FEATURE_IORING_OP_ACCEPT ,1208 return IS_DEFINED(IORING_OP_ACCEPT);1209 1210 if( /*func == (fptr_t)connect || */1211 func == (fptr_t)cfa_connect )1212 #define _CFA_IO_FEATURE_IORING_OP_CONNECT ,1213 return IS_DEFINED(IORING_OP_CONNECT);1214 1215 if( /*func == (fptr_t)fallocate || */1216 func == (fptr_t)cfa_fallocate )1217 #define _CFA_IO_FEATURE_IORING_OP_FALLOCATE ,1218 return IS_DEFINED(IORING_OP_FALLOCATE);1219 1220 if( /*func == (fptr_t)posix_fadvise || */1221 func == (fptr_t)cfa_fadvise )1222 #define _CFA_IO_FEATURE_IORING_OP_FADVISE ,1223 return IS_DEFINED(IORING_OP_FADVISE);1224 1225 if( /*func == (fptr_t)madvise || */1226 func == (fptr_t)cfa_madvise )1227 #define _CFA_IO_FEATURE_IORING_OP_MADVISE ,1228 return IS_DEFINED(IORING_OP_MADVISE);1229 1230 if( /*func == (fptr_t)openat || */1231 func == (fptr_t)cfa_openat )1232 #define _CFA_IO_FEATURE_IORING_OP_OPENAT ,1233 return IS_DEFINED(IORING_OP_OPENAT);1234 1235 if( /*func == (fptr_t)close || */1236 func == (fptr_t)cfa_close )1237 #define _CFA_IO_FEATURE_IORING_OP_CLOSE ,1238 return IS_DEFINED(IORING_OP_CLOSE);1239 1240 if( /*func == (fptr_t)read || */1241 func == (fptr_t)cfa_read )1242 #define _CFA_IO_FEATURE_IORING_OP_READ ,1243 return IS_DEFINED(IORING_OP_READ);1244 1245 if( /*func == (fptr_t)write || */1246 func == (fptr_t)cfa_write )1247 #define _CFA_IO_FEATURE_IORING_OP_WRITE ,1248 return IS_DEFINED(IORING_OP_WRITE);1249 #endif1250 1251 return false;1252 }
Note:
See TracChangeset
for help on using the changeset viewer.
