Changeset 2d8f7b0

Timestamp:

Apr 14, 2020, 11:53:25 AM (4 years ago)

Author:

Thierry Delisle <tdelisle@…>

Branches:

ADT, arm-eh, ast-experimental, enum, forall-pointer-decay, jacob/cs343-translation, master, new-ast, new-ast-unique-expr, pthread-emulation, qualifiedEnum

Children:

34d0a28

Parents:

7df014f

Message:

Implemented basic non-blocking io

Location:

libcfa/src/concurrency

Files:

• alarm.cfa (modified) (1 diff)
• asyncio.hfa (added)
• io.cfa (modified) (13 diffs)
• kernel.cfa (modified) (2 diffs)
• kernel.hfa (modified) (2 diffs)
• kernel_private.hfa (modified) (1 diff)
• preemption.cfa (modified) (1 diff)
• thread.hfa (modified) (1 diff)

libcfa/src/concurrency/alarm.cfa

@@ -176 +176 @@
 }
 
+//=============================================================================================
+// Utilities
+//=============================================================================================
+
+void sleep( Duration duration ) {
+        alarm_node_t node = { active_thread(), __kernel_get_time() + duration, 0`s };
+
+        register_self( &node );
+        park( __cfaabi_dbg_ctx );
+
+        /* paranoid */ verify( !node.set );
+        /* paranoid */ verify( node.next == 0p );
+}
+
 // Local Variables: //
 // mode: c //

libcfa/src/concurrency/io.cfa

@@ -45 +45 @@
                 memset(&params, 0, sizeof(params));
 
-                int fd = syscall(__NR_io_uring_setup, entries_per_cluster(), &params );
+                uint32_t nentries = entries_per_cluster();
+
+                int fd = syscall(__NR_io_uring_setup, nentries, &params );
                 if(fd < 0) {
                         abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
@@ -52 +54 @@
                 // Step 2 : mmap result
                 memset(&this.io, 0, sizeof(struct io_ring));
-                struct io_uring_sq * sq = &this.io.submit_q;
-                struct io_uring_cq * cq = &this.io.completion_q;
+                struct io_uring_sq & sq = this.io.submit_q;
+                struct io_uring_cq & cq = this.io.completion_q;
 
                 // calculate the right ring size
-                sq->ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned)           );
-                cq->ring_sz = params.cq_off.cqes  + (params.cq_entries * sizeof(struct io_uring_cqe));
+                sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned)           );
+                cq.ring_sz = params.cq_off.cqes  + (params.cq_entries * sizeof(struct io_uring_cqe));
 
                 // Requires features
@@ -66 +68 @@
 
                 // mmap the Submit Queue into existence
-                sq->ring_ptr = mmap(0, sq->ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING);
-                if (sq->ring_ptr == (void*)MAP_FAILED) {
+                sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING);
+                if (sq.ring_ptr == (void*)MAP_FAILED) {
                         abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
                 }
@@ -78 +80 @@
                 // else {
                         // 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);
-                        if (cq->ring_ptr == (void*)MAP_FAILED) {
-                                munmap(sq->ring_ptr, sq->ring_sz);
+                        cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING);
+                        if (cq.ring_ptr == (void*)MAP_FAILED) {
+                                munmap(sq.ring_ptr, sq.ring_sz);
                                 abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
                         }
@@ -87 +89 @@
                 // mmap the submit queue entries
                 size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
-                sq->sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES);
-                if (sq->sqes == (struct io_uring_sqe *)MAP_FAILED) {
-                        munmap(sq->ring_ptr, sq->ring_sz);
-                        if (cq->ring_ptr != sq->ring_ptr) munmap(cq->ring_ptr, cq->ring_sz);
+                sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES);
+                if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
+                        munmap(sq.ring_ptr, sq.ring_sz);
+                        if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
                         abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
                 }
@@ -96 +98 @@
                 // Get the pointers from the kernel to fill the structure
                 // submit queue
-                sq->head    = (uint32_t *)((intptr_t)sq->ring_ptr) + params.sq_off.head;
-                sq->tail    = (uint32_t *)((intptr_t)sq->ring_ptr) + params.sq_off.tail;
-                sq->mask    = (uint32_t *)((intptr_t)sq->ring_ptr) + params.sq_off.ring_mask;
-                sq->entries = (uint32_t *)((intptr_t)sq->ring_ptr) + params.sq_off.ring_entries;
-                sq->flags   = (uint32_t *)((intptr_t)sq->ring_ptr) + params.sq_off.flags;
-                sq->dropped = (uint32_t *)((intptr_t)sq->ring_ptr) + params.sq_off.dropped;
-                sq->array   = (uint32_t *)((intptr_t)sq->ring_ptr) + params.sq_off.array;
+                sq.head    = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
+                sq.tail    = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
+                sq.mask    = (   const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
+                sq.num     = (   const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
+                sq.flags   = (         uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
+                sq.dropped = (         uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
+                sq.array   = (         uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
+                sq.alloc = *sq.tail;
 
                 // completion queue
-                cq->head     = (uint32_t *)((intptr_t)cq->ring_ptr) + params.cq_off.head;
-                cq->tail     = (uint32_t *)((intptr_t)cq->ring_ptr) + params.cq_off.tail;
-                cq->mask     = (uint32_t *)((intptr_t)cq->ring_ptr) + params.cq_off.ring_mask;
-                cq->entries  = (struct io_uring_cqe *)((intptr_t)cq->ring_ptr) + params.cq_off.ring_entries;
-                cq->overflow = (uint32_t *)((intptr_t)cq->ring_ptr) + params.cq_off.overflow;
-                cq->cqes = (struct io_uring_cqe *)((intptr_t)cq->ring_ptr) + params.cq_off.cqes;
+                cq.head     = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
+                cq.tail     = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
+                cq.mask     = (   const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
+                cq.num      = (   const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
+                cq.overflow = (         uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
+                cq.cqes   = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);
+
+                // some paranoid checks
+                /* paranoid */ verifyf( (*cq.mask) == ((*cq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (*cq.num) - 1ul32, *cq.num, *cq.mask  );
+                /* paranoid */ verifyf( (*cq.num)  >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
+                /* paranoid */ verifyf( (*cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
+                /* paranoid */ verifyf( (*cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );
+
+                /* paranoid */ verifyf( (*sq.mask) == ((*sq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (*sq.num) - 1ul32, *sq.num, *sq.mask );
+                /* paranoid */ verifyf( (*sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
+                /* paranoid */ verifyf( (*sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head );
+                /* paranoid */ verifyf( (*sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail );
 
                 // Update the global ring info
@@ -116 +130 @@
                 this.io.fd    = fd;
                 this.io.done  = false;
+                (this.io.submit){ min(*sq.num, *cq.num) };
 
                 // Create the poller thread
@@ -137 +152 @@
 
                 // unmap the submit queue entries
-                munmap(sq.sqes, *sq.entries * sizeof(struct io_uring_sqe));
+                munmap(sq.sqes, (*sq.num) * sizeof(struct io_uring_sqe));
 
                 // unmap the Submit Queue ring
@@ -173 +188 @@
 
                 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 );
+                __unpark( data->thrd __cfaabi_dbg_ctx2 );
+
+                // Allow new submissions to happen
+                V(ring.submit);
 
                 // 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_RELEASE );
+                __atomic_fetch_add( ring.completion_q.head, 1, __ATOMIC_RELAXED );
 
                 return true;
@@ -200 +220 @@
                 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);
+                        __cfaabi_bits_print_safe( STDERR_FILENO, "Performed io_wait, returned %d errmsg %s\n", ret, strerror(errno) );
                         if( ret < 0 ) {
                                 switch((int)errno) {
@@ -221 +242 @@
 //=============================================================================================
 
-[* struct io_uring_sqe, uint32_t] io_submit_prelude( struct io_ring & ring ) {
-        return [0p, 0];
+// Submition steps :
+// 1 - We need to make sure we don't overflow any of the buffer, P(ring.submit) to make sure
+//     entries are available. The semaphore make sure that there is no more operations in
+//     progress then the number of entries in the buffer. This probably limits concurrency
+//     more than necessary since submitted but not completed operations don't need any
+//     entries in user space. However, I don't know what happens if we overflow the buffers
+//     because too many requests completed at once. This is a safe approach in all cases.
+//     Furthermore, with hundreds of entries, this may be okay.
+//
+// 2 - Allocate a queue entry. The ring already has memory for all entries but only the ones
+//     listed in sq.array are visible by the kernel. For those not listed, the kernel does not
+//     offer any assurance that an entry is not being filled by multiple flags. Therefore, we
+//     need to write an allocator that allows allocating concurrently.
+//
+// 3 - Actually fill the submit entry, this is the only simple and straightforward step.
+//
+// 4 - Append the entry index to the array and adjust the tail accordingly. This operation
+//     needs to arrive to two concensus at the same time:
+//     A - The order in which entries are listed in the array: no two threads must pick the
+//         same index for their entries
+//     B - When can the tail be update for the kernel. EVERY entries in the array between
+//         head and tail must be fully filled and shouldn't ever be touched again.
+//
+
+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;
 }
 
@@ -228 +336 @@
 // I/O Interface
 //=============================================================================================
-
-/*
         extern "C" {
                 #define __USE_GNU
+                #define _GNU_SOURCE
+                #include <fcntl.h>
                 #include <sys/uio.h>
-        }
-
+                #include <sys/socket.h>
+                #include <sys/stat.h>
+        }
+
+        #define __submit_prelude \
+                struct io_ring & ring = active_cluster()->io; \
+                struct io_uring_sqe * sqe; \
+                uint32_t idx; \
+                [sqe, idx] = __submit_alloc( ring );
+
+        #define __submit_wait \
+                io_user_data data = { 0, active_thread() }; \
+                __cfaabi_bits_print_safe( STDERR_FILENO, "Preparing user data %p for %p\n", &data, data.thrd ); \
+                sqe->user_data = (uint64_t)&data; \
+                __submit( ring, idx ); \
+                park( __cfaabi_dbg_ctx ); \
+                return data.result;
+
+//-----------------------------------------------------------------------------
+// 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
-                        sqe->opcode = IORING_OP_READV;
-                        sqe->flags = 0;
-                        sqe->ioprio = 0;
-                        sqe->fd = fd;
-                        sqe->off = offset;
-                        sqe->addr = (unsigned long) iov;
-                        sqe->len = iovcnt;
-                        sqe->rw_flags = 0;
-                        sqe->user_data = 0;
-                        sqe->__pad2[0] = sqe->__pad2[1] = sqe->__pad2[2] = 0;
+                        __submit_prelude
+
+                        (*sqe){ IORING_OP_READV, fd, iov, iovcnt, offset };
+
+                        __submit_wait
                 #endif
         }
@@ -256 +377 @@
                         return pwritev2(fd, iov, iovcnt, offset, flags);
                 #else
-                        #warning not implemented
-                #endif
-        }
-
-        bool is_async( void (*func)() ) {
-                switch((uintptr_t)func) {
-                case (uintptr_t)preadv2:
-                case (uintptr_t)async_preadv2:
-                        #if defined(IORING_OP_READV)
-                                return true;
-                        #else
-                                return false;
-                        #endif
-                default:
-                        return false;
-                }
-        }
-*/
+                        __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 );
+                #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
+        }
+
+//-----------------------------------------------------------------------------
+// Check if a function is asynchronous
+
+// 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 ) {
+
+                if( /*func == (fptr_t)preadv2 || */
+                        func == (fptr_t)async_preadv2 )
+                        #define _CFA_IO_FEATURE_IORING_OP_READV ,
+                        return IS_DEFINED(IORING_OP_READV);
+
+                if( /*func == (fptr_t)pwritev2 || */
+                      func == (fptr_t)async_pwritev2 )
+                        #define _CFA_IO_FEATURE_IORING_OP_WRITEV ,
+                        return IS_DEFINED(IORING_OP_WRITEV);
+
+                if( /*func == (fptr_t)fsync || */
+                      func == (fptr_t)async_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 )
+                        #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 )
+                        #define _CFA_IO_FEATURE_IORING_OP_SENDMSG ,
+                        return IS_DEFINED(IORING_OP_SENDMSG);
+
+                if( /*func == (fptr_t)recvmsg || */
+                      func == (fptr_t)async_recvmsg )
+                        #define _CFA_IO_FEATURE_IORING_OP_RECVMSG ,
+                        return IS_DEFINED(IORING_OP_RECVMSG);
+
+                if( /*func == (fptr_t)send || */
+                        func == (fptr_t)async_send )
+                        #define _CFA_IO_FEATURE_IORING_OP_SEND ,
+                        return IS_DEFINED(IORING_OP_SEND);
+
+                if( /*func == (fptr_t)recv || */
+                        func == (fptr_t)async_recv )
+                        #define _CFA_IO_FEATURE_IORING_OP_RECV ,
+                        return IS_DEFINED(IORING_OP_RECV);
+
+                if( /*func == (fptr_t)accept4 || */
+                        func == (fptr_t)async_accept4 )
+                        #define _CFA_IO_FEATURE_IORING_OP_ACCEPT ,
+                        return IS_DEFINED(IORING_OP_ACCEPT);
+
+                if( /*func == (fptr_t)connect || */
+                        func == (fptr_t)async_connect )
+                        #define _CFA_IO_FEATURE_IORING_OP_CONNECT ,
+                        return IS_DEFINED(IORING_OP_CONNECT);
+
+                if( /*func == (fptr_t)fallocate || */
+                        func == (fptr_t)async_fallocate )
+                        #define _CFA_IO_FEATURE_IORING_OP_FALLOCATE ,
+                        return IS_DEFINED(IORING_OP_FALLOCATE);
+
+                if( /*func == (fptr_t)fadvise || */
+                        func == (fptr_t)async_fadvise )
+                        #define _CFA_IO_FEATURE_IORING_OP_FADVISE ,
+                        return IS_DEFINED(IORING_OP_FADVISE);
+
+                if( /*func == (fptr_t)madvise || */
+                        func == (fptr_t)async_madvise )
+                        #define _CFA_IO_FEATURE_IORING_OP_MADVISE ,
+                        return IS_DEFINED(IORING_OP_MADVISE);
+
+                if( /*func == (fptr_t)openat || */
+                        func == (fptr_t)async_openat )
+                        #define _CFA_IO_FEATURE_IORING_OP_OPENAT ,
+                        return IS_DEFINED(IORING_OP_OPENAT);
+
+                if( /*func == (fptr_t)close || */
+                        func == (fptr_t)async_close )
+                        #define _CFA_IO_FEATURE_IORING_OP_CLOSE ,
+                        return IS_DEFINED(IORING_OP_CLOSE);
+
+                if( /*func == (fptr_t)statx || */
+                        func == (fptr_t)async_statx )
+                        #define _CFA_IO_FEATURE_IORING_OP_STATX ,
+                        return IS_DEFINED(IORING_OP_STATX);
+
+                if( /*func == (fptr_t)read || */
+                      func == (fptr_t)async_read )
+                        #define _CFA_IO_FEATURE_IORING_OP_READ ,
+                        return IS_DEFINED(IORING_OP_READ);
+
+                if( /*func == (fptr_t)write || */
+                      func == (fptr_t)async_write )
+                        #define _CFA_IO_FEATURE_IORING_OP_WRITE ,
+                        return IS_DEFINED(IORING_OP_WRITE);
+
+                return false;
+        }
 
 #endif

libcfa/src/concurrency/kernel.cfa

@@ -615 +615 @@
 }
 
-void unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
-        if( !thrd ) return;
-
-        disable_interrupts();
+// KERNEL ONLY unpark with out disabling interrupts
+void __unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
         static_assert(sizeof(thrd->state) == sizeof(int));
 
@@ -647 +645 @@
                         abort();
         }
+}
+
+void unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
+        if( !thrd ) return;
+
+        disable_interrupts();
+        __unpark( thrd __cfaabi_dbg_ctx_fwd2 );
         enable_interrupts( __cfaabi_dbg_ctx );
 }

libcfa/src/concurrency/kernel.hfa

@@ -115 +115 @@
 #if defined(HAVE_LINUX_IO_URING_H)
 struct io_uring_sq {
-        uint32_t * head;
-        uint32_t * tail;
-        uint32_t * mask;
-        uint32_t * entries;
-        uint32_t * flags;
-        uint32_t * dropped;
-        uint32_t * array;
-        struct io_uring_sqe * sqes;
-
-        uint32_t sqe_head;
-        uint32_t sqe_tail;
-
-        size_t ring_sz;
-        void * ring_ptr;
-};
-
-struct io_uring_cq {
+        // Head and tail of the ring (associated with array)
         volatile uint32_t * head;
         volatile uint32_t * tail;
-        uint32_t * mask;
-        struct io_uring_cqe * entries;
+
+        // The actual kernel ring which uses head/tail
+        // indexes into the sqes arrays
+        uint32_t * array;
+
+        // number of entries and mask to go with it
+        const uint32_t * num;
+        const uint32_t * mask;
+
+        // Submission flags (Not sure what for)
+        uint32_t * flags;
+
+        // number of sqes not submitted (whatever that means)
+        uint32_t * dropped;
+
+        // Like head/tail but not seen by the kernel
+        volatile uint32_t alloc;
+
+        __spinlock_t lock;
+
+        // A buffer of sqes (not the actual ring)
+        struct io_uring_sqe * sqes;
+
+        // The location and size of the mmaped area
+        void * ring_ptr;
+        size_t ring_sz;
+};
+
+struct io_uring_cq {
+        // Head and tail of the ring
+        volatile uint32_t * head;
+        volatile uint32_t * tail;
+
+        // number of entries and mask to go with it
+        const uint32_t * mask;
+        const uint32_t * num;
+
+        // number of cqes not submitted (whatever that means)
         uint32_t * overflow;
+
+        // the kernel ring
         struct io_uring_cqe * cqes;
 
+        // The location and size of the mmaped area
+        void * ring_ptr;
         size_t ring_sz;
-        void * ring_ptr;
 };
 
@@ -151 +174 @@
         void * stack;
         volatile bool done;
+        semaphore submit;
 };
 #endif

libcfa/src/concurrency/kernel_private.hfa

@@ -70 +70 @@
 )
 
+// KERNEL ONLY unpark with out disabling interrupts
+void __unpark( $thread * thrd __cfaabi_dbg_ctx_param2 );
+
 //-----------------------------------------------------------------------------
 // I/O

libcfa/src/concurrency/preemption.cfa

@@ -268 +268 @@
 // reserved for future use
 static void timeout( $thread * this ) {
-        //TODO : implement waking threads
+        __unpark( this __cfaabi_dbg_ctx2 );
 }
 

libcfa/src/concurrency/thread.hfa

@@ -117 +117 @@
 }
 
+//----------
+// sleep: force thread to block and be rescheduled after Duration duration
+void sleep( Duration duration );
+
 // Local Variables: //
 // mode: c //