Changeset 8e4aa05 for libcfa/src/concurrency

libcfa/src/concurrency/coroutine.cfa

-              r342af53
+              r8e4aa05
 //-----------------------------------------------------------------------------
 FORALL_DATA_INSTANCE(CoroutineCancelled, (dtype coroutine_t), (coroutine_t))
 forall(dtype T)
+FORALL_DATA_INSTANCE(CoroutineCancelled, (coroutine_t &), (coroutine_t))
+forall(T &)
 void mark_exception(CoroutineCancelled(T) *) {}
 forall(dtype T)
+forall(T &)
 void copy(CoroutineCancelled(T) * dst, CoroutineCancelled(T) * src) {
         dst->virtual_table = src->virtual_table;
 …
+}
 forall(dtype T)
+forall(T &)
 const char * msg(CoroutineCancelled(T) *) {
         return "CoroutineCancelled(...)";
 …
 // This code should not be inlined. It is the error path on resume.
 forall(dtype T | is_coroutine(T))
+forall(T & | is_coroutine(T))
 void __cfaehm_cancelled_coroutine( T & cor, $coroutine * desc ) {
         verify( desc->cancellation );
 …
 // Part of the Public API
 // Not inline since only ever called once per coroutine
 forall(dtype T | is_coroutine(T))
+forall(T & | is_coroutine(T))
 void prime(T& cor) {
         $coroutine* this = get_coroutine(cor);
 …
 void __stack_clean  ( __stack_info_t * this ) {
-        size_t size = ((intptr_t)this->storage->base) - ((intptr_t)this->storage->limit) + sizeof(__stack_t);
         void * storage = this->storage->limit;
         #if CFA_COROUTINE_USE_MMAP
+                size_t size = ((intptr_t)this->storage->base) - ((intptr_t)this->storage->limit) + sizeof(__stack_t);
                 storage = (void *)(((intptr_t)storage) - __page_size);
                 if(munmap(storage, size + __page_size) == -1) {

libcfa/src/concurrency/coroutine.hfa

-              r342af53
+              r8e4aa05
 //-----------------------------------------------------------------------------
 // Exception thrown from resume when a coroutine stack is cancelled.
 FORALL_DATA_EXCEPTION(CoroutineCancelled, (dtype coroutine_t), (coroutine_t)) (
+FORALL_DATA_EXCEPTION(CoroutineCancelled, (coroutine_t &), (coroutine_t)) (
         coroutine_t * the_coroutine;
         exception_t * the_exception;
 );
 forall(dtype T)
+forall(T &)
 void copy(CoroutineCancelled(T) * dst, CoroutineCancelled(T) * src);
 forall(dtype T)
+forall(T &)
 const char * msg(CoroutineCancelled(T) *);
 …
 // Anything that implements this trait can be resumed.
 // Anything that is resumed is a coroutine.
 trait is_coroutine(dtype T | IS_RESUMPTION_EXCEPTION(CoroutineCancelled, (T))) {
+trait is_coroutine(T & | IS_RESUMPTION_EXCEPTION(CoroutineCancelled, (T))) {
         void main(T & this);
         $coroutine * get_coroutine(T & this);
 …
 //-----------------------------------------------------------------------------
 // Public coroutine API
 forall(dtype T | is_coroutine(T))
+forall(T & | is_coroutine(T))
 void prime(T & cor);
 …
         void __cfactx_invoke_coroutine(void (*main)(void *), void * this);
         forall(dtype T)
+        forall(T &)
         void __cfactx_start(void (*main)(T &), struct $coroutine * cor, T & this, void (*invoke)(void (*main)(void *), void *));
 …
+}
 forall(dtype T | is_coroutine(T))
+forall(T & | is_coroutine(T))
 void __cfaehm_cancelled_coroutine( T & cor, $coroutine * desc );
 // Resume implementation inlined for performance
 forall(dtype T | is_coroutine(T))
+forall(T & | is_coroutine(T))
 static inline T & resume(T & cor) {
         // optimization : read TLS once and reuse it

libcfa/src/concurrency/future.hfa

-              r342af53
+              r8e4aa05
 #include "monitor.hfa"
 forall( otype T ) {
+forall( T ) {
         struct future {
                 inline future_t;
 …
+}
 forall( otype T ) {
+forall( T ) {
         monitor multi_future {
                 inline future_t;

libcfa/src/concurrency/io.cfa

-              r342af53
+              r8e4aa05
         extern "C" {
                 #include <sys/syscall.h>
+                #include <sys/eventfd.h>
                 #include <linux/io_uring.h>
 …
         #include "io/types.hfa"
         static const char * opcodes[] = {
+        __attribute__((unused)) static const char * opcodes[] = {
                 "OP_NOP",
                 "OP_READV",
 …
         };
+        // returns true of acquired as leader or second leader
+        static inline bool try_lock( __leaderlock_t & this ) {
+                const uintptr_t thrd = 1z | (uintptr_t)active_thread();
+                bool block;
+                disable_interrupts();
+                for() {
+                        struct $thread * expected = this.value;
+                        if( 1p != expected && 0p != expected ) {
+                                /* paranoid */ verify( thrd != (uintptr_t)expected ); // We better not already be the next leader
+                                enable_interrupts( __cfaabi_dbg_ctx );
+                                return false;
+                        }
+                        struct $thread * desired;
+                        if( 0p == expected ) {
+                                // If the lock isn't locked acquire it, no need to block
+                                desired = 1p;
+                                block = false;
+                        }
+                        else {
+                                // If the lock is already locked try becomming the next leader
+                                desired = (struct $thread *)thrd;
+                                block = true;
+                        }
+                        if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
+                }
+                if( block ) {
+                        enable_interrupts( __cfaabi_dbg_ctx );
+                        park();
+                        disable_interrupts();
+                }
+                return true;
+        }
+        static inline bool next( __leaderlock_t & this ) {
+        static $io_context * __ioarbiter_allocate( $io_arbiter & mutex this, processor *, __u32 idxs[], __u32 want );
+        static void __ioarbiter_submit( $io_arbiter & mutex this, $io_context * , __u32 idxs[], __u32 have, bool lazy );
+        static void __ioarbiter_flush ( $io_arbiter & mutex this, $io_context * );
+        static inline void __ioarbiter_notify( $io_context & ctx );
+//=============================================================================================
+// I/O Polling
+//=============================================================================================
+        static inline unsigned __flush( struct $io_context & );
+        static inline __u32 __release_sqes( struct $io_context & );
+        void __cfa_io_drain( processor * proc ) {
                 /* paranoid */ verify( ! __preemption_enabled() );
+                struct $thread * nextt;
+                for() {
+                        struct $thread * expected = this.value;
+                        /* paranoid */ verify( (1 & (uintptr_t)expected) == 1 ); // The lock better be locked
+                        struct $thread * desired;
+                        if( 1p == expected ) {
+                                // No next leader, just unlock
+                                desired = 0p;
+                                nextt   = 0p;
+                        }
+                        else {
+                                // There is a next leader, remove but keep locked
+                                desired = 1p;
+                                nextt   = (struct $thread *)(~1z & (uintptr_t)expected);
+                        }
+                        if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break;
+                }
+                if(nextt) {
+                        unpark( nextt );
+                        enable_interrupts( __cfaabi_dbg_ctx );
+                        return true;
+                }
+                enable_interrupts( __cfaabi_dbg_ctx );
+                return false;
+        }
+//=============================================================================================
+// I/O Syscall
+//=============================================================================================
+        static int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) {
+                bool need_sys_to_submit = false;
+                bool need_sys_to_complete = false;
+                unsigned flags = 0;
+                TO_SUBMIT:
+                if( to_submit > 0 ) {
+                        if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
+                                need_sys_to_submit = true;
+                                break TO_SUBMIT;
+                        }
+                        if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) {
+                                need_sys_to_submit = true;
+                                flags |= IORING_ENTER_SQ_WAKEUP;
+                        }
+                }
+                if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) {
+                        flags |= IORING_ENTER_GETEVENTS;
+                        if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) {
+                                need_sys_to_complete = true;
+                        }
+                }
+                int ret = 0;
+                if( need_sys_to_submit || need_sys_to_complete ) {
+                        __cfadbg_print_safe(io_core, "Kernel I/O : IO_URING enter %d %u %u\n", ring.fd, to_submit, flags);
+                        ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, (sigset_t *)0p, _NSIG / 8);
+                        if( ret < 0 ) {
+                                switch((int)errno) {
+                                case EAGAIN:
+                                case EINTR:
+                                        ret = -1;
+                                        break;
+                                default:
+                                        abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
+                                }
+                        }
+                }
+                // Memory barrier
+                __atomic_thread_fence( __ATOMIC_SEQ_CST );
+                return ret;
+        }
+//=============================================================================================
+// I/O Polling
+//=============================================================================================
+        static unsigned __collect_submitions( struct __io_data & ring );
+        static __u32 __release_consumed_submission( struct __io_data & ring );
+        static inline void __clean( volatile struct io_uring_sqe * sqe );
+        // Process a single completion message from the io_uring
+        // This is NOT thread-safe
+        static inline void process( volatile struct io_uring_cqe & cqe ) {
+                struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
+                __cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
+                fulfil( *future, cqe.res );
+        }
+        static [int, bool] __drain_io( & struct __io_data ring ) {
+                /* paranoid */ verify( ! __preemption_enabled() );
+                unsigned to_submit = 0;
+                if( ring.poller_submits ) {
+                        // If the poller thread also submits, then we need to aggregate the submissions which are ready
+                        to_submit = __collect_submitions( ring );
+                }
+                int ret = __io_uring_enter(ring, to_submit, true);
+                if( ret < 0 ) {
+                        return [0, true];
+                }
+                // update statistics
+                if (to_submit > 0) {
+                        __STATS__( true,
+                                if( to_submit > 0 ) {
+                                        io.submit_q.submit_avg.rdy += to_submit;
+                                        io.submit_q.submit_avg.csm += ret;
+                                        io.submit_q.submit_avg.cnt += 1;
+                                }
+                        )
+                }
+                __atomic_thread_fence( __ATOMIC_SEQ_CST );
+                // Release the consumed SQEs
+                __release_consumed_submission( ring );
+                /* paranoid */ verify( proc );
+                /* paranoid */ verify( proc->io.ctx );
                 // Drain the queue
+                unsigned head = *ring.completion_q.head;
+                unsigned tail = *ring.completion_q.tail;
+                const __u32 mask = *ring.completion_q.mask;
+                // Nothing was new return 0
+                if (head == tail) {
+                        return [0, to_submit > 0];
+                }
+                $io_context * ctx = proc->io.ctx;
+                unsigned head = *ctx->cq.head;
+                unsigned tail = *ctx->cq.tail;
+                const __u32 mask = *ctx->cq.mask;
                 __u32 count = tail - head;
+                /* paranoid */ verify( count != 0 );
+                __STATS__( false, io.calls.drain++; io.calls.completed += count; )
                 for(i; count) {
                         unsigned idx = (head + i) & mask;
                         volatile struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
+                        volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
                         /* paranoid */ verify(&cqe);
+                        process( cqe );
+                }
+                        struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
+                        __cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
+                        fulfil( *future, cqe.res );
+                }
+                __cfadbg_print_safe(io, "Kernel I/O : %u completed\n", 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, count, __ATOMIC_SEQ_CST );
+                return [count, count > 0 || to_submit > 0];
+        }
+        void main( $io_ctx_thread & this ) {
+                __ioctx_register( this );
+                __cfadbg_print_safe(io_core, "Kernel I/O : IO poller %d (%p) ready\n", this.ring->fd, &this);
+                const int reset_cnt = 5;
+                int reset = reset_cnt;
+                // Then loop until we need to start
+                LOOP:
+                while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) {
+                        // Drain the io
+                        int count;
+                        bool again;
+                        disable_interrupts();
+                                [count, again] = __drain_io( *this.ring );
+                                if(!again) reset--;
+                __atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
+                /* paranoid */ verify( ! __preemption_enabled() );
+                return;
+        }
+        void __cfa_io_flush( processor * proc ) {
+                /* paranoid */ verify( ! __preemption_enabled() );
+                /* paranoid */ verify( proc );
+                /* paranoid */ verify( proc->io.ctx );
+                $io_context & ctx = *proc->io.ctx;
+                if(!ctx.ext_sq.empty) {
+                        __ioarbiter_flush( *ctx.arbiter, &ctx );
+                }
+                __STATS__( true, io.calls.flush++; )
+                int ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, 0, 0, (sigset_t *)0p, _NSIG / 8);
+                if( ret < 0 ) {
+                        switch((int)errno) {
+                        case EAGAIN:
+                        case EINTR:
+                        case EBUSY:
                                 // Update statistics
+                                __STATS__( true,
+                                        io.complete_q.completed_avg.val += count;
+                                        io.complete_q.completed_avg.cnt += 1;
+                                )
+                        enable_interrupts( __cfaabi_dbg_ctx );
+                        // If we got something, just yield and check again
+                        if(reset > 1) {
+                                yield();
+                                continue LOOP;
+                                __STATS__( false, io.calls.errors.busy ++; )
+                                return;
+                        default:
+                                abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
+                        }
+                        // We alread failed to find completed entries a few time.
+                        if(reset == 1) {
+                                // Rearm the context so it can block
+                                // but don't block right away
+                                // we need to retry one last time in case
+                                // something completed *just now*
+                                __ioctx_prepare_block( this );
+                                continue LOOP;
+                        }
+                                __STATS__( false,
+                                        io.complete_q.blocks += 1;
+                                )
+                                __cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %d (%p)\n", this.ring->fd, &this);
+                                // block this thread
+                                wait( this.sem );
+                        // restore counter
+                        reset = reset_cnt;
+                }
+                __cfadbg_print_safe(io_core, "Kernel I/O : Fast poller %d (%p) stopping\n", this.ring->fd, &this);
+                }
+                __cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
+                __STATS__( true, io.calls.submitted += ret; )
+                /* paranoid */ verify( ctx.sq.to_submit <= *ctx.sq.num );
+                /* paranoid */ verify( ctx.sq.to_submit >= ret );
+                ctx.sq.to_submit -= ret;
+                /* paranoid */ verify( ctx.sq.to_submit <= *ctx.sq.num );
+                // Release the consumed SQEs
+                __release_sqes( ctx );
+                /* paranoid */ verify( ! __preemption_enabled() );
+                ctx.proc->io.pending = false;
+        }
 …
 //         head and tail must be fully filled and shouldn't ever be touched again.
 //
+        //=============================================================================================
+        // Allocation
+        // for user's convenience fill the sqes from the indexes
+        static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx)  {
+                struct io_uring_sqe * sqes = ctx->sq.sqes;
+                for(i; want) {
+                        __cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
+                        out_sqes[i] = &sqes[idxs[i]];
+                }
+        }
+        // Try to directly allocate from the a given context
+        // Not thread-safe
+        static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
+                __sub_ring_t & sq = ctx->sq;
+                const __u32 mask  = *sq.mask;
+                __u32 fhead = sq.free_ring.head;    // get the current head of the queue
+                __u32 ftail = sq.free_ring.tail;    // get the current tail of the queue
+                // If we don't have enough sqes, fail
+                if((ftail - fhead) < want) { return false; }
+                // copy all the indexes we want from the available list
+                for(i; want) {
+                        __cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
+                        idxs[i] = sq.free_ring.array[(fhead + i) & mask];
+                }
+                // Advance the head to mark the indexes as consumed
+                __atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
+                // return success
+                return true;
+        }
         // Allocate an submit queue entry.
 …
         // for convenience, return both the index and the pointer to the sqe
         // sqe == &sqes[idx]
+        [* volatile struct io_uring_sqe, __u32] __submit_alloc( struct __io_data & ring, __u64 data ) {
+                /* paranoid */ verify( data != 0 );
+                // Prepare the data we need
+                __attribute((unused)) int len   = 0;
+                __attribute((unused)) int block = 0;
+                __u32 cnt = *ring.submit_q.num;
+                __u32 mask = *ring.submit_q.mask;
+                __u32 off = thread_rand();
+                // Loop around looking for an available spot
+                for() {
+                        // Look through the list starting at some offset
+                        for(i; cnt) {
+                                __u64 expected = 3;
+                                __u32 idx = (i + off) & mask; // Get an index from a random
+                                volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
+                                volatile __u64 * udata = &sqe->user_data;
+                                // Allocate the entry by CASing the user_data field from 0 to the future address
+                                if( *udata == expected &&
+                                        __atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) )
+                                {
+                                        // update statistics
+                                        __STATS__( false,
+                                                io.submit_q.alloc_avg.val   += len;
+                                                io.submit_q.alloc_avg.block += block;
+                                                io.submit_q.alloc_avg.cnt   += 1;
+                                        )
+                                        // debug log
+                                        __cfadbg_print_safe( io, "Kernel I/O : allocated [%p, %u] for %p (%p)\n", sqe, idx, active_thread(), (void*)data );
+                                        // Success return the data
+                                        return [sqe, idx];
+                                }
+                                verify(expected != data);
+                                // This one was used
+                                len ++;
+                        }
+                        block++;
+                        abort( "Kernel I/O : all submit queue entries used, yielding\n" );
+                        yield();
+                }
+        }
+        static inline __u32 __submit_to_ready_array( struct __io_data & ring, __u32 idx, const __u32 mask ) {
+                /* paranoid */ verify( idx <= mask   );
+                /* paranoid */ verify( idx != -1ul32 );
+                // We need to find a spot in the ready array
+                __attribute((unused)) int len   = 0;
+                __attribute((unused)) int block = 0;
+                __u32 ready_mask = ring.submit_q.ready_cnt - 1;
+                __u32 off = thread_rand();
+                __u32 picked;
+                LOOKING: for() {
+                        for(i; ring.submit_q.ready_cnt) {
+                                picked = (i + off) & ready_mask;
+                                __u32 expected = -1ul32;
+                                if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) {
+                                        break LOOKING;
+                                }
+                                verify(expected != idx);
+                                len ++;
+                        }
+                        block++;
+                        __u32 released = __release_consumed_submission( ring );
+                        if( released == 0 ) {
+                                yield();
+                        }
+                }
+                // update statistics
+                __STATS__( false,
+                        io.submit_q.look_avg.val   += len;
+                        io.submit_q.look_avg.block += block;
+                        io.submit_q.look_avg.cnt   += 1;
+                )
+                return picked;
+        }
+        void __submit( struct io_context * ctx, __u32 idx ) __attribute__((nonnull (1))) {
+                __io_data & ring = *ctx->thrd.ring;
+        struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
+                __cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
+                disable_interrupts();
+                processor * proc = __cfaabi_tls.this_processor;
+                $io_context * ctx = proc->io.ctx;
+                /* paranoid */ verify( __cfaabi_tls.this_processor );
+                /* paranoid */ verify( ctx );
+                __cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
+                // We can proceed to the fast path
+                if( __alloc(ctx, idxs, want) ) {
+                        // Allocation was successful
+                        __STATS__( true, io.alloc.fast += 1; )
+                        enable_interrupts( __cfaabi_dbg_ctx );
+                        __cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
+                        __fill( sqes, want, idxs, ctx );
+                        return ctx;
+                }
+                // The fast path failed, fallback
+                __STATS__( true, io.alloc.fail += 1; )
+                // Fast path failed, fallback on arbitration
+                __STATS__( true, io.alloc.slow += 1; )
+                enable_interrupts( __cfaabi_dbg_ctx );
+                $io_arbiter * ioarb = proc->cltr->io.arbiter;
+                /* paranoid */ verify( ioarb );
+                __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
+                struct $io_context * ret = __ioarbiter_allocate(*ioarb, proc, idxs, want);
+                __cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
+                __fill( sqes, want, idxs,ret );
+                return ret;
+        }
+        //=============================================================================================
+        // submission
+        static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
+                // We can proceed to the fast path
+                // Get the right objects
+                __sub_ring_t & sq = ctx->sq;
+                const __u32 mask  = *sq.mask;
+                __u32 tail = *sq.kring.tail;
+                // Add the sqes to the array
+                for( i; have ) {
+                        __cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
+                        sq.kring.array[ (tail + i) & mask ] = idxs[i];
+                }
+                // Make the sqes visible to the submitter
+                __atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
+                sq.to_submit++;
+                ctx->proc->io.pending = true;
+                ctx->proc->io.dirty   = true;
+                if(sq.to_submit > 30 || !lazy) {
+                        __cfa_io_flush( ctx->proc );
+                }
+        }
+        void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
+                __cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
+                disable_interrupts();
+                processor * proc = __cfaabi_tls.this_processor;
+                $io_context * ctx = proc->io.ctx;
+                /* paranoid */ verify( __cfaabi_tls.this_processor );
+                /* paranoid */ verify( ctx );
+                // Can we proceed to the fast path
+                if( ctx == inctx )              // We have the right instance?
+                {
+                        __attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx];
+                        __cfadbg_print_safe( io,
+                                "Kernel I/O : submitting %u (%p) for %p\n"
+                                "    data: %p\n"
+                                "    opcode: %s\n"
+                                "    fd: %d\n"
+                                "    flags: %d\n"
+                                "    prio: %d\n"
+                                "    off: %p\n"
+                                "    addr: %p\n"
+                                "    len: %d\n"
+                                "    other flags: %d\n"
+                                "    splice fd: %d\n"
+                                "    pad[0]: %llu\n"
+                                "    pad[1]: %llu\n"
+                                "    pad[2]: %llu\n",
+                                idx, sqe,
+                                active_thread(),
+                                (void*)sqe->user_data,
+                                opcodes[sqe->opcode],
+                                sqe->fd,
+                                sqe->flags,
+                                sqe->ioprio,
+                                sqe->off,
+                                sqe->addr,
+                                sqe->len,
+                                sqe->accept_flags,
+                                sqe->splice_fd_in,
+                                sqe->__pad2[0],
+                                sqe->__pad2[1],
+                                sqe->__pad2[2]
+                        );
+                }
+                // Get now the data we definetely need
+                volatile __u32 * const tail = ring.submit_q.tail;
+                const __u32 mask  = *ring.submit_q.mask;
+                // There are 2 submission schemes, check which one we are using
+                if( ring.poller_submits ) {
+                        // If the poller thread submits, then we just need to add this to the ready array
+                        __submit_to_ready_array( ring, idx, mask );
+                        post( ctx->thrd.sem );
+                        __cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() );
+                }
+                else if( ring.eager_submits ) {
+                        __u32 picked = __submit_to_ready_array( ring, idx, mask );
+                        #if defined(LEADER_LOCK)
+                                if( !try_lock(ring.submit_q.submit_lock) ) {
+                                        __STATS__( false,
+                                                io.submit_q.helped += 1;
+                                        )
+                                        return;
+                                }
+                                /* paranoid */ verify( ! __preemption_enabled() );
+                                __STATS__( true,
+                                        io.submit_q.leader += 1;
+                                )
+                        #else
+                                for() {
+                                        yield();
+                                        if( try_lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2) ) {
+                                                __STATS__( false,
+                                                        io.submit_q.leader += 1;
+                                                )
+                                                break;
+                                        }
+                                        // If some one else collected our index, we are done
+                                        #warning ABA problem
+                                        if( ring.submit_q.ready[picked] != idx ) {
+                                                __STATS__( false,
+                                                        io.submit_q.helped += 1;
+                                                )
+                                                return;
+                                        }
+                                        __STATS__( false,
+                                                io.submit_q.busy += 1;
+                                        )
+                                }
+                        #endif
+                        // We got the lock
+                        // Collect the submissions
+                        unsigned to_submit = __collect_submitions( ring );
+                        // Actually submit
+                        int ret = __io_uring_enter( ring, to_submit, false );
+                        #if defined(LEADER_LOCK)
+                                /* paranoid */ verify( ! __preemption_enabled() );
+                                next(ring.submit_q.submit_lock);
+                        #else
+                                unlock(ring.submit_q.submit_lock);
+                        #endif
+                        if( ret < 0 ) {
+                                return;
+                        }
+                        // Release the consumed SQEs
+                        __release_consumed_submission( ring );
+                        // update statistics
+                        __STATS__( false,
+                                io.submit_q.submit_avg.rdy += to_submit;
+                                io.submit_q.submit_avg.csm += ret;
+                                io.submit_q.submit_avg.cnt += 1;
+                        )
+                        __cfadbg_print_safe( io, "Kernel I/O : submitted %u (among %u) for %p\n", idx, ret, active_thread() );
+                }
+                else
+                {
+                        // get mutual exclusion
+                        #if defined(LEADER_LOCK)
+                                while(!try_lock(ring.submit_q.submit_lock));
+                        #else
+                                lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2);
+                        #endif
+                        /* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 3ul64,
+                        /* paranoid */  "index %u already reclaimed\n"
+                        /* paranoid */  "head %u, prev %u, tail %u\n"
+                        /* paranoid */  "[-0: %u,-1: %u,-2: %u,-3: %u]\n",
+                        /* paranoid */  idx,
+                        /* paranoid */  *ring.submit_q.head, ring.submit_q.prev_head, *tail
+                        /* paranoid */  ,ring.submit_q.array[ ((*ring.submit_q.head) - 0) & (*ring.submit_q.mask) ]
+                        /* paranoid */  ,ring.submit_q.array[ ((*ring.submit_q.head) - 1) & (*ring.submit_q.mask) ]
+                        /* paranoid */  ,ring.submit_q.array[ ((*ring.submit_q.head) - 2) & (*ring.submit_q.mask) ]
+                        /* paranoid */  ,ring.submit_q.array[ ((*ring.submit_q.head) - 3) & (*ring.submit_q.mask) ]
+                        /* paranoid */ );
+                        // Append to the list of ready entries
+                        /* paranoid */ verify( idx <= mask );
+                        ring.submit_q.array[ (*tail) & mask ] = idx;
+                        __atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST);
+                        // Submit however, many entries need to be submitted
+                        int ret = __io_uring_enter( ring, 1, false );
+                        if( ret < 0 ) {
+                                switch((int)errno) {
+                                default:
+                                        abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) );
+                                }
+                        }
+                        /* paranoid */ verify(ret == 1);
+                        // update statistics
+                        __STATS__( false,
+                                io.submit_q.submit_avg.csm += 1;
+                                io.submit_q.submit_avg.cnt += 1;
+                        )
+                        {
+                                __attribute__((unused)) volatile __u32 * const head = ring.submit_q.head;
+                                __attribute__((unused)) __u32 last_idx = ring.submit_q.array[ ((*head) - 1) & mask ];
+                                __attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[last_idx];
+                                __cfadbg_print_safe( io,
+                                        "Kernel I/O : last submitted is %u (%p)\n"
+                                        "    data: %p\n"
+                                        "    opcode: %s\n"
+                                        "    fd: %d\n"
+                                        "    flags: %d\n"
+                                        "    prio: %d\n"
+                                        "    off: %p\n"
+                                        "    addr: %p\n"
+                                        "    len: %d\n"
+                                        "    other flags: %d\n"
+                                        "    splice fd: %d\n"
+                                        "    pad[0]: %llu\n"
+                                        "    pad[1]: %llu\n"
+                                        "    pad[2]: %llu\n",
+                                        last_idx, sqe,
+                                        (void*)sqe->user_data,
+                                        opcodes[sqe->opcode],
+                                        sqe->fd,
+                                        sqe->flags,
+                                        sqe->ioprio,
+                                        sqe->off,
+                                        sqe->addr,
+                                        sqe->len,
+                                        sqe->accept_flags,
+                                        sqe->splice_fd_in,
+                                        sqe->__pad2[0],
+                                        sqe->__pad2[1],
+                                        sqe->__pad2[2]
+                                );
+                        }
+                        __atomic_thread_fence( __ATOMIC_SEQ_CST );
+                        // Release the consumed SQEs
+                        __release_consumed_submission( ring );
+                        // ring.submit_q.sqes[idx].user_data = 3ul64;
+                        #if defined(LEADER_LOCK)
+                                next(ring.submit_q.submit_lock);
+                        #else
+                                unlock(ring.submit_q.submit_lock);
+                        #endif
+                        __cfadbg_print_safe( io, "Kernel I/O : submitted %u for %p\n", idx, active_thread() );
+                }
+        }
+        // #define PARTIAL_SUBMIT 32
+        // go through the list of submissions in the ready array and moved them into
+        // the ring's submit queue
+        static unsigned __collect_submitions( struct __io_data & ring ) {
+                /* paranoid */ verify( ring.submit_q.ready != 0p );
+                /* paranoid */ verify( ring.submit_q.ready_cnt > 0 );
+                unsigned to_submit = 0;
+                __u32 tail = *ring.submit_q.tail;
+                const __u32 mask = *ring.submit_q.mask;
+                #if defined(PARTIAL_SUBMIT)
+                        #if defined(LEADER_LOCK)
+                                #error PARTIAL_SUBMIT and LEADER_LOCK cannot co-exist
+                        #endif
+                        const __u32 cnt = ring.submit_q.ready_cnt > PARTIAL_SUBMIT ? PARTIAL_SUBMIT : ring.submit_q.ready_cnt;
+                        const __u32 offset = ring.submit_q.prev_ready;
+                        ring.submit_q.prev_ready += cnt;
+                #else
+                        const __u32 cnt = ring.submit_q.ready_cnt;
+                        const __u32 offset = 0;
+                #endif
+                // Go through the list of ready submissions
+                for( c; cnt ) {
+                        __u32 i = (offset + c) % ring.submit_q.ready_cnt;
+                        // replace any submission with the sentinel, to consume it.
+                        __u32 idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED);
+                        // If it was already the sentinel, then we are done
+                        if( idx == -1ul32 ) continue;
+                        // If we got a real submission, append it to the list
+                        ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask;
+                        to_submit++;
+                }
+                // Increment the tail based on how many we are ready to submit
+                __atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST);
+                return to_submit;
+        }
+                        __submit(ctx, idxs, have, lazy);
+                        // Mark the instance as no longer in-use, re-enable interrupts and return
+                        __STATS__( true, io.submit.fast += 1; )
+                        enable_interrupts( __cfaabi_dbg_ctx );
+                        __cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
+                        return;
+                }
+                // Fast path failed, fallback on arbitration
+                __STATS__( true, io.submit.slow += 1; )
+                enable_interrupts( __cfaabi_dbg_ctx );
+                __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
+                __ioarbiter_submit(*inctx->arbiter, inctx, idxs, have, lazy);
+        }
+        //=============================================================================================
+        // Flushing
         // Go through the ring's submit queue and release everything that has already been consumed
         // by io_uring
+        static __u32 __release_consumed_submission( struct __io_data & ring ) {
+                const __u32 smask = *ring.submit_q.mask;
+                // We need to get the lock to copy the old head and new head
+                if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0;
+        // This cannot be done by multiple threads
+        static __u32 __release_sqes( struct $io_context & ctx ) {
+                const __u32 mask = *ctx.sq.mask;
                 __attribute__((unused))
                 __u32 ctail = *ring.submit_q.tail;        // get the current tail of the queue
                 __u32 chead = *ring.submit_q.head;              // get the current head of the queue
                 __u32 phead = ring.submit_q.prev_head;  // get the head the last time we were here
+                ring.submit_q.prev_head = chead;                // note up to were we processed
                 unlock(ring.submit_q.release_lock);
+                __u32 ctail = *ctx.sq.kring.tail;    // get the current tail of the queue
+                __u32 chead = *ctx.sq.kring.head;        // get the current head of the queue
+                __u32 phead = ctx.sq.kring.released; // get the head the last time we were here
+                __u32 ftail = ctx.sq.free_ring.tail;  // get the current tail of the queue
                 // the 3 fields are organized like this diagram
 …
                 __u32 count = chead - phead;
+                if(count == 0) {
+                        return 0;
+                }
                 // We acquired an previous-head/current-head range
                 // go through the range and release the sqes
                 for( i; count ) {
+                        __u32 idx = ring.submit_q.array[ (phead + i) & smask ];
+                        /* paranoid */ verify( 0 != ring.submit_q.sqes[ idx ].user_data );
+                        __clean( &ring.submit_q.sqes[ idx ] );
+                }
+                        __cfadbg_print_safe(io, "Kernel I/O : release loop\n");
+                        __u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
+                        ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
+                }
+                ctx.sq.kring.released = chead;          // note up to were we processed
+                __atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
+                __ioarbiter_notify(ctx);
                 return count;
+        }
+        void __sqe_clean( volatile struct io_uring_sqe * sqe ) {
+                __clean( sqe );
+        }
+        static inline void __clean( volatile struct io_uring_sqe * sqe ) {
+                // If we are in debug mode, thrash the fields to make sure we catch reclamation errors
+                __cfaabi_dbg_debug_do(
+                        memset(sqe, 0xde, sizeof(*sqe));
+                        sqe->opcode = (sizeof(opcodes) / sizeof(const char *)) - 1;
+                );
+                // Mark the entry as unused
+                __atomic_store_n(&sqe->user_data, 3ul64, __ATOMIC_SEQ_CST);
+//=============================================================================================
+// I/O Arbiter
+//=============================================================================================
+        static $io_context * __ioarbiter_allocate( $io_arbiter & mutex this, processor * proc, __u32 idxs[], __u32 want ) {
+                __cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
+                __STATS__( false, io.alloc.block += 1; )
+                // No one has any resources left, wait for something to finish
+                // Mark as pending
+                __atomic_store_n( &this.pending.flag, true, __ATOMIC_SEQ_CST );
+                // Wait for our turn to submit
+                wait( this.pending.blocked, want );
+                __attribute((unused)) bool ret =
+                __alloc( this.pending.ctx, idxs, want);
+                /* paranoid */ verify( ret );
+                return this.pending.ctx;
+        }
+        static void __ioarbiter_notify( $io_arbiter & mutex this, $io_context * ctx ) {
+                /* paranoid */ verify( !is_empty(this.pending.blocked) );
+                this.pending.ctx = ctx;
+                while( !is_empty(this.pending.blocked) ) {
+                        __cfadbg_print_safe(io, "Kernel I/O : notifying\n");
+                        __u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
+                        __u32 want = front( this.pending.blocked );
+                        if( have > want ) return;
+                        signal_block( this.pending.blocked );
+                }
+                this.pending.flag = false;
+        }
+        static void __ioarbiter_notify( $io_context & ctx ) {
+                if(__atomic_load_n( &ctx.arbiter->pending.flag, __ATOMIC_SEQ_CST)) {
+                        __ioarbiter_notify( *ctx.arbiter, &ctx );
+                }
+        }
+        // Simply append to the pending
+        static void __ioarbiter_submit( $io_arbiter & mutex this, $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
+                __cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
+                /* paranoid */ verify( &this == ctx->arbiter );
+                // Mark as pending
+                __atomic_store_n( &ctx->ext_sq.empty, false, __ATOMIC_SEQ_CST );
+                __cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
+                // Wait for our turn to submit
+                wait( ctx->ext_sq.blocked );
+                // Submit our indexes
+                __submit(ctx, idxs, have, lazy);
+                __cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
+        }
+        static void __ioarbiter_flush( $io_arbiter & mutex this, $io_context * ctx ) {
+                /* paranoid */ verify( &this == ctx->arbiter );
+                __STATS__( false, io.flush.external += 1; )
+                __cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
+                condition & blcked = ctx->ext_sq.blocked;
+                /* paranoid */ verify( ctx->ext_sq.empty == is_empty( blcked ) );
+                while(!is_empty( blcked )) {
+                        signal_block( blcked );
+                }
+                ctx->ext_sq.empty = true;
+        }
 #endif

libcfa/src/concurrency/io/call.cfa.in

-              r342af53
+              r8e4aa05
                         | IOSQE_IO_DRAIN
                 #endif
+                #if defined(CFA_HAVE_IOSQE_IO_LINK)
+                        | IOSQE_IO_LINK
+                #endif
+                #if defined(CFA_HAVE_IOSQE_IO_HARDLINK)
+                        | IOSQE_IO_HARDLINK
+                #endif
                 #if defined(CFA_HAVE_IOSQE_ASYNC)
                         | IOSQE_ASYNC
                 #endif
+        ;
+        static const __u32 LINK_FLAGS = 0
+                #if defined(CFA_HAVE_IOSQE_IO_LINK)
+                        | IOSQE_IO_LINK
+                #endif
+                #if defined(CFA_HAVE_IOSQE_IO_HARDLINK)
+                        | IOSQE_IO_HARDLINK
+                #if defined(CFA_HAVE_IOSQE_BUFFER_SELECTED)
+                        | IOSQE_BUFFER_SELECTED
                 #endif
+        ;
 …
+        ;
+        extern [* volatile struct io_uring_sqe, __u32] __submit_alloc( struct __io_data & ring, __u64 data );
+        extern void __submit( struct io_context * ctx, __u32 idx ) __attribute__((nonnull (1)));
+        static inline io_context * __get_io_context( void ) {
+                cluster * cltr = active_cluster();
+                /* paranoid */ verifyf( cltr, "No active cluster for io operation\\n");
+                assertf( cltr->io.cnt > 0, "Cluster %p has no default io contexts and no context was specified\\n", cltr );
+                /* paranoid */ verifyf( cltr->io.ctxs, "default io contexts for cluster %p are missing\\n", cltr);
+                return &cltr->io.ctxs[ thread_rand() % cltr->io.cnt ];
+        }
+        extern struct $io_context * cfa_io_allocate(struct io_uring_sqe * out_sqes[], __u32 out_idxs[], __u32 want)  __attribute__((nonnull (1,2)));
+        extern void cfa_io_submit( struct $io_context * in_ctx, __u32 in_idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1,2)));
 #endif
 …
 extern "C" {
         #include <sys/types.h>
+        #include <asm/types.h>
         #include <sys/socket.h>
         #include <sys/syscall.h>
 …
         extern int epoll_ctl(int epfd, int op, int fd, struct epoll_event *event);
         extern ssize_t splice(int fd_in, loff_t *off_in, int fd_out, loff_t *off_out, size_t len, unsigned int flags);
+        extern ssize_t splice(int fd_in, __off64_t *off_in, int fd_out, __off64_t *off_out, size_t len, unsigned int flags);
         extern ssize_t tee(int fd_in, int fd_out, size_t len, unsigned int flags);
+}
 …
                 return ', '.join(args_a)
 AsyncTemplate = """inline void async_{name}(io_future_t & future, {params}, int submit_flags, io_cancellation * cancellation, io_context * context) {{
+AsyncTemplate = """inline void async_{name}(io_future_t & future, {params}, __u64 submit_flags) {{
         #if !defined(CFA_HAVE_LINUX_IO_URING_H) || !defined(CFA_HAVE_IORING_OP_{op})
                 ssize_t res = {name}({args});
 …
                 }}
         #else
-                // we don't support LINK yet
-                if( 0 != (submit_flags & LINK_FLAGS) ) {{
-                        errno = ENOTSUP; return -1;
-                }}
-                if( !context ) {{
-                        context = __get_io_context();
-                }}
-                if(cancellation) {{
-                        cancellation->target = (__u64)(uintptr_t)&future;
-                }}
                 __u8 sflags = REGULAR_FLAGS & submit_flags;
-                struct __io_data & ring = *context->thrd.ring;
                 __u32 idx;
                 struct io_uring_sqe * sqe;
                 [(volatile struct io_uring_sqe *) sqe, idx] = __submit_alloc( ring, (__u64)(uintptr_t)&future );
+                struct $io_context * ctx = cfa_io_allocate( &sqe, &idx, 1 );
                 sqe->opcode = IORING_OP_{op};
+                sqe->user_data = (__u64)(uintptr_t)&future;
                 sqe->flags = sflags;
                 sqe->ioprio = 0;
 …
                 verify( sqe->user_data == (__u64)(uintptr_t)&future );
                 __submit( context, idx );
+                cfa_io_submit( ctx, &idx, 1, 0 != (submit_flags & CFA_IO_LAZY) );
         #endif
 }}"""
+SyncTemplate = """{ret} cfa_{name}({params}, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context) {{
+        if( timeout >= 0 ) {{
+                errno = ENOTSUP;
+                return -1;
+        }}
+SyncTemplate = """{ret} cfa_{name}({params}, __u64 submit_flags) {{
         io_future_t future;
         async_{name}( future, {args}, submit_flags, cancellation, context );
+        async_{name}( future, {args}, submit_flags );
         wait( future );
 …
         }),
         # CFA_HAVE_IORING_OP_SPLICE
         Call('SPLICE', 'ssize_t splice(int fd_in, loff_t *off_in, int fd_out, loff_t *off_out, size_t len, unsigned int flags)', {
+        Call('SPLICE', 'ssize_t splice(int fd_in, __off64_t *off_in, int fd_out, __off64_t *off_out, size_t len, unsigned int flags)', {
                 'splice_fd_in': 'fd_in',
                 'splice_off_in': 'off_in ? (__u64)*off_in : (__u64)-1',
 …
         if c.define:
                 print("""#if defined({define})
         {ret} cfa_{name}({params}, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
+        {ret} cfa_{name}({params}, __u64 submit_flags);
 #endif""".format(define=c.define,ret=c.ret, name=c.name, params=c.params))
         else:
                 print("{ret} cfa_{name}({params}, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);"
+                print("{ret} cfa_{name}({params}, __u64 submit_flags);"
                 .format(ret=c.ret, name=c.name, params=c.params))
 …
         if c.define:
                 print("""#if defined({define})
         void async_{name}(io_future_t & future, {params}, int submit_flags, io_cancellation * cancellation, io_context * context);
+        void async_{name}(io_future_t & future, {params}, __u64 submit_flags);
 #endif""".format(define=c.define,name=c.name, params=c.params))
         else:
                 print("void async_{name}(io_future_t & future, {params}, int submit_flags, io_cancellation * cancellation, io_context * context);"
+                print("void async_{name}(io_future_t & future, {params}, __u64 submit_flags);"
                 .format(name=c.name, params=c.params))
 print("\n")
 …
 print("""
-//-----------------------------------------------------------------------------
-bool cancel(io_cancellation & this) {
-        #if !defined(CFA_HAVE_LINUX_IO_URING_H) || !defined(CFA_HAVE_IORING_OP_ASYNC_CANCEL)
-                return false;
-        #else
-                io_future_t future;
-                io_context * context = __get_io_context();
-                __u8 sflags = 0;
-                struct __io_data & ring = *context->thrd.ring;
-                __u32 idx;
-                volatile struct io_uring_sqe * sqe;
-                [sqe, idx] = __submit_alloc( ring, (__u64)(uintptr_t)&future );
-                sqe->__pad2[0] = sqe->__pad2[1] = sqe->__pad2[2] = 0;
-                sqe->opcode = IORING_OP_ASYNC_CANCEL;
-                sqe->flags = sflags;
-                sqe->addr = this.target;
-                verify( sqe->user_data == (__u64)(uintptr_t)&future );
-                __submit( context, idx );
-                wait(future);
-                if( future.result == 0 ) return true; // Entry found
-                if( future.result == -EALREADY) return true; // Entry found but in progress
-                if( future.result == -ENOENT ) return false; // Entry not found
-                return false;
-        #endif
+}
 //-----------------------------------------------------------------------------
 // Check if a function is has asynchronous

libcfa/src/concurrency/io/setup.cfa

-              r342af53
+              r8e4aa05
 #if !defined(CFA_HAVE_LINUX_IO_URING_H)
-        void __kernel_io_startup() {
-                // Nothing to do without io_uring
+        }
-        void __kernel_io_shutdown() {
-                // Nothing to do without io_uring
+        }
         void ?{}(io_context_params & this) {}
+        void ?{}(io_context & this, struct cluster & cl) {}
+        void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {}
+        void ^?{}(io_context & this) {}
+        void ^?{}(io_context & this, bool cluster_context) {}
+        void  ?{}($io_context & this, struct cluster & cl) {}
+        void ^?{}($io_context & this) {}
+        void __cfa_io_start( processor * proc ) {}
+        void __cfa_io_flush( processor * proc ) {}
+        void __cfa_io_stop ( processor * proc ) {}
+        $io_arbiter * create(void) { return 0p; }
+        void destroy($io_arbiter *) {}
 #else
 …
         void ?{}(io_context_params & this) {
                 this.num_entries = 256;
-                this.num_ready = 256;
-                this.submit_aff = -1;
-                this.eager_submits = false;
-                this.poller_submits = false;
-                this.poll_submit = false;
-                this.poll_complete = false;
+        }
 …
 //=============================================================================================
-// I/O Startup / Shutdown logic + Master Poller
-//=============================================================================================
-        // IO Master poller loop forward
-        static void * iopoll_loop( __attribute__((unused)) void * args );
-        static struct {
-                pthread_t     thrd;    // pthread handle to io poller thread
-                void *        stack;   // pthread stack for io poller thread
-                int           epollfd; // file descriptor to the epoll instance
-                volatile bool run;     // Whether or not to continue
-        } iopoll;
-        void __kernel_io_startup(void) {
-                __cfadbg_print_safe(io_core, "Kernel : Creating EPOLL instance\n" );
-                iopoll.epollfd = epoll_create1(0);
-                if (iopoll.epollfd == -1) {
-                        abort( "internal error, epoll_create1\n");
+                }
-                __cfadbg_print_safe(io_core, "Kernel : Starting io poller thread\n" );
-                iopoll.run = true;
-                iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
+        }
-        void __kernel_io_shutdown(void) {
-                // Notify the io poller thread of the shutdown
-                iopoll.run = false;
-                sigval val = { 1 };
-                pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
-                // Wait for the io poller thread to finish
-                __destroy_pthread( iopoll.thrd, iopoll.stack, 0p );
-                int ret = close(iopoll.epollfd);
-                if (ret == -1) {
-                        abort( "internal error, close epoll\n");
+                }
-                // Io polling is now fully stopped
-                __cfadbg_print_safe(io_core, "Kernel : IO poller stopped\n" );
+        }
-        static void * iopoll_loop( __attribute__((unused)) void * args ) {
-                __processor_id_t id;
-                id.full_proc = false;
-                id.id = doregister(&id);
-                __cfaabi_tls.this_proc_id = &id;
-                __cfadbg_print_safe(io_core, "Kernel : IO poller thread starting\n" );
-                // Block signals to control when they arrive
-                sigset_t mask;
-                sigfillset(&mask);
-                if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) {
-                abort( "internal error, pthread_sigmask" );
+                }
-                sigdelset( &mask, SIGUSR1 );
-                // Create sufficient events
-                struct epoll_event events[10];
-                // Main loop
-                while( iopoll.run ) {
-                        __cfadbg_print_safe(io_core, "Kernel I/O - epoll : waiting on io_uring contexts\n");
-                        // Wait for events
-                        int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask );
-                        __cfadbg_print_safe(io_core, "Kernel I/O - epoll : %d io contexts events, waking up\n", nfds);
-                        // Check if an error occured
-                        if (nfds == -1) {
-                                if( errno == EINTR ) continue;
-                                abort( "internal error, pthread_sigmask" );
+                        }
-                        for(i; nfds) {
-                                $io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64;
-                                /* paranoid */ verify( io_ctx );
-                                __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Unparking io poller %d (%p)\n", io_ctx->ring->fd, io_ctx);
-                                #if !defined( __CFA_NO_STATISTICS__ )
-                                        __cfaabi_tls.this_stats = io_ctx->self.curr_cluster->stats;
-                                #endif
-                                eventfd_t v;
-                                eventfd_read(io_ctx->ring->efd, &v);
-                                post( io_ctx->sem );
+                        }
+                }
-                __cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
-                unregister(&id);
-                return 0p;
+        }
-//=============================================================================================
 // I/O Context Constrution/Destruction
 //=============================================================================================
+        void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; }
+        void main( $io_ctx_thread & this );
+        static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; }
+        void ^?{}( $io_ctx_thread & mutex this ) {}
+        static void __io_create ( __io_data & this, const io_context_params & params_in );
+        static void __io_destroy( __io_data & this );
+        void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {
+                (this.thrd){ cl };
+                this.thrd.ring = malloc();
+                __cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this);
+                __io_create( *this.thrd.ring, params );
+                __cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this);
+                this.thrd.done = false;
+                __thrd_start( this.thrd, main );
+                __cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this);
+        }
+        void ?{}(io_context & this, struct cluster & cl) {
+                io_context_params params;
+                (this){ cl, params };
+        }
+        void ^?{}(io_context & this, bool cluster_context) {
+                __cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this);
+                // Notify the thread of the shutdown
+                __atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST);
+                // If this is an io_context within a cluster, things get trickier
+                $thread & thrd = this.thrd.self;
+                if( cluster_context ) {
+                        // We are about to do weird things with the threads
+                        // we don't need interrupts to complicate everything
+                        disable_interrupts();
+                        // Get cluster info
+                        cluster & cltr = *thrd.curr_cluster;
+                        /* paranoid */ verify( cltr.idles.total == 0 || &cltr == mainCluster );
+                        /* paranoid */ verify( !ready_mutate_islocked() );
+                        // We need to adjust the clean-up based on where the thread is
+                        if( thrd.state == Ready || thrd.preempted != __NO_PREEMPTION ) {
+                                // This is the tricky case
+                                // The thread was preempted or ready to run and now it is on the ready queue
+                                // but the cluster is shutting down, so there aren't any processors to run the ready queue
+                                // the solution is to steal the thread from the ready-queue and pretend it was blocked all along
+                                ready_schedule_lock();
+                                        // The thread should on the list
+                                        /* paranoid */ verify( thrd.link.next != 0p );
+                                        // Remove the thread from the ready queue of this cluster
+                                        // The thread should be the last on the list
+                                        __attribute__((unused)) bool removed = remove_head( &cltr, &thrd );
+                                        /* paranoid */ verify( removed );
+                                        thrd.link.next = 0p;
+                                        thrd.link.prev = 0p;
+                                        // Fixup the thread state
+                                        thrd.state = Blocked;
+                                        thrd.ticket = TICKET_BLOCKED;
+                                        thrd.preempted = __NO_PREEMPTION;
+                                ready_schedule_unlock();
+                                // Pretend like the thread was blocked all along
+                        }
+                        // !!! This is not an else if !!!
+                        // Ok, now the thread is blocked (whether we cheated to get here or not)
+                        if( thrd.state == Blocked ) {
+                                // This is the "easy case"
+                                // The thread is parked and can easily be moved to active cluster
+                                verify( thrd.curr_cluster != active_cluster() || thrd.curr_cluster == mainCluster );
+                                thrd.curr_cluster = active_cluster();
+                                // unpark the fast io_poller
+                                unpark( &thrd );
+                        }
+                        else {
+                                // The thread is in a weird state
+                                // I don't know what to do here
+                                abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n");
+                        }
+                        // The weird thread kidnapping stuff is over, restore interrupts.
+                        enable_interrupts( __cfaabi_dbg_ctx );
+                } else {
+                        post( this.thrd.sem );
+                }
+                ^(this.thrd){};
+                __cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this);
+                __io_destroy( *this.thrd.ring );
+                __cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this);
+                free(this.thrd.ring);
+        }
+        void ^?{}(io_context & this) {
+                ^(this){ false };
+        }
+        static void __io_create( __io_data & this, const io_context_params & params_in ) {
+        static void __io_uring_setup ( $io_context & this, const io_context_params & params_in, int procfd );
+        static void __io_uring_teardown( $io_context & this );
+        static void __epoll_register($io_context & ctx);
+        static void __epoll_unregister($io_context & ctx);
+        void __ioarbiter_register( $io_arbiter & mutex, $io_context & ctx );
+        void __ioarbiter_unregister( $io_arbiter & mutex, $io_context & ctx );
+        void ?{}($io_context & this, processor * proc, struct cluster & cl) {
+                /* paranoid */ verify( cl.io.arbiter );
+                this.proc = proc;
+                this.arbiter = cl.io.arbiter;
+                this.ext_sq.empty = true;
+                (this.ext_sq.blocked){};
+                __io_uring_setup( this, cl.io.params, proc->idle );
+                __cfadbg_print_safe(io_core, "Kernel I/O : Created ring for io_context %u (%p)\n", this.fd, &this);
+        }
+        void ^?{}($io_context & this) {
+                __cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %u\n", this.fd);
+                __io_uring_teardown( this );
+                __cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %u\n", this.fd);
+        }
+        extern void __disable_interrupts_hard();
+        extern void __enable_interrupts_hard();
+        static void __io_uring_setup( $io_context & this, const io_context_params & params_in, int procfd ) {
                 // Step 1 : call to setup
                 struct io_uring_params params;
                 memset(&params, 0, sizeof(params));
                 if( params_in.poll_submit   ) params.flags |= IORING_SETUP_SQPOLL;
                 if( params_in.poll_complete ) params.flags |= IORING_SETUP_IOPOLL;
+                // if( params_in.poll_submit   ) params.flags |= IORING_SETUP_SQPOLL;
+                // if( params_in.poll_complete ) params.flags |= IORING_SETUP_IOPOLL;
                 __u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
 …
                         abort("ERROR: I/O setup 'num_entries' must be a power of 2\n");
+                }
-                if( params_in.poller_submits && params_in.eager_submits ) {
-                        abort("ERROR: I/O setup 'poller_submits' and 'eager_submits' cannot be used together\n");
+                }
                 int fd = syscall(__NR_io_uring_setup, nentries, &params );
 …
                 // Step 2 : mmap result
+                memset( &this, 0, sizeof(struct __io_data) );
+                struct __submition_data  & sq = this.submit_q;
+                struct __completion_data & cq = this.completion_q;
+                struct __sub_ring_t & sq = this.sq;
+                struct __cmp_ring_t & cq = this.cq;
                 // calculate the right ring size
 …
                 // Get the pointers from the kernel to fill the structure
                 // submit queue
+                sq.head    = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
+                sq.tail    = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
+                sq.mask    = (   const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
+                sq.num     = (   const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
+                sq.flags   = (         __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
+                sq.dropped = (         __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
+                sq.array   = (         __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
+                sq.prev_head = *sq.head;
+                {
+                        const __u32 num = *sq.num;
+                        for( i; num ) {
+                                __sqe_clean( &sq.sqes[i] );
+                        }
+                }
+                (sq.submit_lock){};
+                (sq.release_lock){};
+                if( params_in.poller_submits || params_in.eager_submits ) {
+                        /* paranoid */ verify( is_pow2( params_in.num_ready ) || (params_in.num_ready < 8) );
+                        sq.ready_cnt = max( params_in.num_ready, 8 );
+                        sq.ready = alloc( sq.ready_cnt, 64`align );
+                        for(i; sq.ready_cnt) {
+                                sq.ready[i] = -1ul32;
+                        }
+                        sq.prev_ready = 0;
+                }
+                else {
+                        sq.ready_cnt = 0;
+                        sq.ready = 0p;
+                        sq.prev_ready = 0;
+                }
+                sq.kring.head  = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.head);
+                sq.kring.tail  = (volatile __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail);
+                sq.kring.array = (         __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.array);
+                sq.mask        = (   const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask);
+                sq.num         = (   const __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries);
+                sq.flags       = (         __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags);
+                sq.dropped     = (         __u32 *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped);
+                sq.kring.released = 0;
+                sq.free_ring.head = 0;
+                sq.free_ring.tail = *sq.num;
+                sq.free_ring.array = alloc( *sq.num, 128`align );
+                for(i; (__u32)*sq.num) {
+                        sq.free_ring.array[i] = i;
+                }
+                sq.to_submit = 0;
                 // completion queue
 …
                 // Step 4 : eventfd
+                int efd;
+                for() {
+                        efd = eventfd(0, 0);
+                        if (efd < 0) {
+                                if (errno == EINTR) continue;
+                                abort("KERNEL ERROR: IO_URING EVENTFD - %s\n", strerror(errno));
+                        }
+                        break;
+                }
+                int ret;
+                for() {
+                        ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &efd, 1);
+                        if (ret < 0) {
+                                if (errno == EINTR) continue;
+                                abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
+                        }
+                        break;
+                }
+                // io_uring_register is so f*cking slow on some machine that it
+                // will never succeed if preemption isn't hard blocked
+                __cfadbg_print_safe(io_core, "Kernel I/O : registering %d for completion with ring %d\n", procfd, fd);
+                __disable_interrupts_hard();
+                int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &procfd, 1);
+                if (ret < 0) {
+                        abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
+                }
+                __enable_interrupts_hard();
+                __cfadbg_print_safe(io_core, "Kernel I/O : registered %d for completion with ring %d\n", procfd, fd);
                 // some paranoid checks
 …
                 /* 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 );
+                /* paranoid */ verifyf( (*sq.kring.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.kring.head );
+                /* paranoid */ verifyf( (*sq.kring.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.kring.tail );
                 // Update the global ring info
                 this.ring_flags = params.flags;
+                this.ring_flags = 0;
                 this.fd         = fd;
+                this.efd        = efd;
+                this.eager_submits  = params_in.eager_submits;
+                this.poller_submits = params_in.poller_submits;
+        }
+        static void __io_destroy( __io_data & this ) {
+        }
+        static void __io_uring_teardown( $io_context & this ) {
                 // Shutdown the io rings
                 struct __submition_data  & sq = this.submit_q;
                 struct __completion_data & cq = this.completion_q;
+                struct __sub_ring_t & sq = this.sq;
+                struct __cmp_ring_t & cq = this.cq;
                 // unmap the submit queue entries
 …
                 // close the file descriptor
                 close(this.fd);
+                close(this.efd);
+                free( this.submit_q.ready ); // Maybe null, doesn't matter
+                free( this.sq.free_ring.array ); // Maybe null, doesn't matter
+        }
+        void __cfa_io_start( processor * proc ) {
+                proc->io.ctx = alloc();
+                (*proc->io.ctx){proc, *proc->cltr};
+        }
+        void __cfa_io_stop ( processor * proc ) {
+                ^(*proc->io.ctx){};
+                free(proc->io.ctx);
+        }
 …
 // I/O Context Sleep
 //=============================================================================================
+        #define IOEVENTS EPOLLIN | EPOLLONESHOT
+        static inline void __ioctx_epoll_ctl($io_ctx_thread & ctx, int op, const char * error) {
+                struct epoll_event ev;
+                ev.events = IOEVENTS;
+                ev.data.u64 = (__u64)&ctx;
+                int ret = epoll_ctl(iopoll.epollfd, op, ctx.ring->efd, &ev);
+                if (ret < 0) {
+                        abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
+                }
+        }
+        void __ioctx_register($io_ctx_thread & ctx) {
+                __ioctx_epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
+        }
+        void __ioctx_prepare_block($io_ctx_thread & ctx) {
+                __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.ring->fd, &ctx);
+                __ioctx_epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
+        }
+        // static inline void __epoll_ctl($io_context & ctx, int op, const char * error) {
+        //      struct epoll_event ev;
+        //      ev.events = EPOLLIN | EPOLLONESHOT;
+        //      ev.data.u64 = (__u64)&ctx;
+        //      int ret = epoll_ctl(iopoll.epollfd, op, ctx.efd, &ev);
+        //      if (ret < 0) {
+        //              abort( "KERNEL ERROR: EPOLL %s - (%d) %s\n", error, (int)errno, strerror(errno) );
+        //      }
+        // }
+        // static void __epoll_register($io_context & ctx) {
+        //      __epoll_ctl(ctx, EPOLL_CTL_ADD, "ADD");
+        // }
+        // static void __epoll_unregister($io_context & ctx) {
+        //      // Read the current epoch so we know when to stop
+        //      size_t curr = __atomic_load_n(&iopoll.epoch, __ATOMIC_SEQ_CST);
+        //      // Remove the fd from the iopoller
+        //      __epoll_ctl(ctx, EPOLL_CTL_DEL, "REMOVE");
+        //      // Notify the io poller thread of the shutdown
+        //      iopoll.run = false;
+        //      sigval val = { 1 };
+        //      pthread_sigqueue( iopoll.thrd, SIGUSR1, val );
+        //      // Make sure all this is done
+        //      __atomic_thread_fence(__ATOMIC_SEQ_CST);
+        //      // Wait for the next epoch
+        //      while(curr == iopoll.epoch && !iopoll.stopped) Pause();
+        // }
+        // void __ioctx_prepare_block($io_context & ctx) {
+        //      __cfadbg_print_safe(io_core, "Kernel I/O - epoll : Re-arming io poller %d (%p)\n", ctx.fd, &ctx);
+        //      __epoll_ctl(ctx, EPOLL_CTL_MOD, "REARM");
+        // }
 //=============================================================================================
 // I/O Context Misc Setup
 //=============================================================================================
+        void register_fixed_files( io_context & ctx, int * files, unsigned count ) {
+                int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count );
+                if( ret < 0 ) {
+                        abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
+                }
+                __cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret );
+        }
+        void register_fixed_files( cluster & cltr, int * files, unsigned count ) {
+                for(i; cltr.io.cnt) {
+                        register_fixed_files( cltr.io.ctxs[i], files, count );
+                }
+        }
+        void ?{}( $io_arbiter & this ) {
+                this.pending.flag = false;
+        }
+        void ^?{}( $io_arbiter & mutex this ) {
+                // /* paranoid */ verify( empty(this.assigned) );
+                // /* paranoid */ verify( empty(this.available) );
+                /* paranoid */ verify( is_empty(this.pending.blocked) );
+        }
+        $io_arbiter * create(void) {
+                return new();
+        }
+        void destroy($io_arbiter * arbiter) {
+                delete(arbiter);
+        }
+//=============================================================================================
+// I/O Context Misc Setup
+//=============================================================================================
 #endif

libcfa/src/concurrency/io/types.hfa

-              r342af53
+              r8e4aa05
 // file "LICENCE" distributed with Cforall.
 //
+// io/types.hfa --
+// io/types.hfa -- PRIVATE
+// Types used by the I/O subsystem
 //
 // Author           : Thierry Delisle
 …
 #include "bits/locks.hfa"
+#include "kernel/fwd.hfa"
 #if defined(CFA_HAVE_LINUX_IO_URING_H)
+        #define LEADER_LOCK
+        struct __leaderlock_t {
+                struct $thread * volatile value;        // ($thread) next_leader | (bool:1) is_locked
+        };
+        #include "bits/sequence.hfa"
+        #include "monitor.hfa"
+        static inline void ?{}( __leaderlock_t & this ) { this.value = 0p; }
+        struct processor;
+        monitor $io_arbiter;
         //-----------------------------------------------------------------------
         // Ring Data structure
+      struct __submition_data {
+                // Head and tail of the ring (associated with array)
+                volatile __u32 * head;
+                volatile __u32 * tail;
+                volatile __u32 prev_head;
+      struct __sub_ring_t {
+                struct {
+                        // Head and tail of the ring (associated with array)
+                        volatile __u32 * head;   // one passed last index consumed by the kernel
+                        volatile __u32 * tail;   // one passed last index visible to the kernel
+                        volatile __u32 released; // one passed last index released back to the free list
+                // The actual kernel ring which uses head/tail
+                // indexes into the sqes arrays
+                __u32 * array;
+                        // The actual kernel ring which uses head/tail
+                        // indexes into the sqes arrays
+                        __u32 * array;
+                } kring;
+                struct {
+                        volatile __u32 head;
+                        volatile __u32 tail;
+                        // The ring which contains free allocations
+                        // indexes into the sqes arrays
+                        __u32 * array;
+                } free_ring;
+                // number of sqes to submit on next system call.
+                __u32 to_submit;
                 // number of entries and mask to go with it
 …
                 const __u32 * mask;
                 // Submission flags (Not sure what for)
+                // Submission flags, currently only IORING_SETUP_SQPOLL
                 __u32 * flags;
+                // number of sqes not submitted (whatever that means)
+                // number of sqes not submitted
+                // From documentation : [dropped] is incremented for each invalid submission queue entry encountered in the ring buffer.
                 __u32 * dropped;
-                // Like head/tail but not seen by the kernel
-                volatile __u32 * ready;
-                __u32 ready_cnt;
-                __u32 prev_ready;
-                #if defined(LEADER_LOCK)
-                        __leaderlock_t submit_lock;
-                #else
-                        __spinlock_t submit_lock;
-                #endif
-                __spinlock_t  release_lock;
                 // A buffer of sqes (not the actual ring)
                 volatile struct io_uring_sqe * sqes;
+                struct io_uring_sqe * sqes;
                 // The location and size of the mmaped area
 …
         };
         struct __completion_data {
+        struct __cmp_ring_t {
                 // Head and tail of the ring
                 volatile __u32 * head;
 …
                 const __u32 * num;
                 // number of cqes not submitted (whatever that means)
+                // I don't know what this value is for
                 __u32 * overflow;
 …
         };
+        struct __io_data {
+                struct __submition_data submit_q;
+                struct __completion_data completion_q;
+        struct __attribute__((aligned(128))) $io_context {
+                $io_arbiter * arbiter;
+                processor * proc;
+                struct {
+                        volatile bool empty;
+                        condition blocked;
+                } ext_sq;
+                struct __sub_ring_t sq;
+                struct __cmp_ring_t cq;
                 __u32 ring_flags;
                 int fd;
+                int efd;
+                bool eager_submits:1;
+                bool poller_submits:1;
+        };
+        monitor __attribute__((aligned(128))) $io_arbiter {
+                struct {
+                        condition blocked;
+                        $io_context * ctx;
+                        volatile bool flag;
+                } pending;
         };
 …
         #endif
+        struct $io_ctx_thread;
+        void __ioctx_register($io_ctx_thread & ctx);
+        void __ioctx_prepare_block($io_ctx_thread & ctx);
+        void __sqe_clean( volatile struct io_uring_sqe * sqe );
+        // void __ioctx_prepare_block($io_context & ctx);
 #endif

libcfa/src/concurrency/iofwd.hfa

-              r342af53
+              r8e4aa05
 #include <unistd.h>
 extern "C" {
         #include <sys/types.h>
+        #include <asm/types.h>
         #if CFA_HAVE_LINUX_IO_URING_H
                 #include <linux/io_uring.h>
 …
 struct cluster;
 struct io_future_t;
+struct io_context;
+struct io_cancellation;
+struct $io_context;
 struct iovec;
 …
 struct sockaddr;
 struct statx;
+struct epoll_event;
+//----------
+// underlying calls
+extern struct $io_context * cfa_io_allocate(struct io_uring_sqe * out_sqes[], __u32 out_idxs[], __u32 want)  __attribute__((nonnull (1,2)));
+extern void cfa_io_submit( struct $io_context * in_ctx, __u32 in_idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1,2)));
 //----------
 // synchronous calls
 #if defined(CFA_HAVE_PREADV2)
         extern ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
+        extern ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags, __u64 submit_flags);
 #endif
 #if defined(CFA_HAVE_PWRITEV2)
         extern ssize_t cfa_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
+        extern ssize_t cfa_pwritev2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags, __u64 submit_flags);
 #endif
 extern int cfa_fsync(int fd, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern int cfa_epoll_ctl(int epfd, int op, int fd, struct epoll_event *event, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern int cfa_sync_file_range(int fd, off64_t offset, off64_t nbytes, unsigned int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern  ssize_t cfa_sendmsg(int sockfd, const struct msghdr *msg, int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern ssize_t cfa_recvmsg(int sockfd, struct msghdr *msg, int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern ssize_t cfa_send(int sockfd, const void *buf, size_t len, int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern ssize_t cfa_recv(int sockfd, void *buf, size_t len, int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern int cfa_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern int cfa_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern int cfa_fallocate(int fd, int mode, off_t offset, off_t len, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern int cfa_posix_fadvise(int fd, off_t offset, off_t len, int advice, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern int cfa_madvise(void *addr, size_t length, int advice, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern int cfa_openat(int dirfd, const char *pathname, int flags, mode_t mode, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
+extern int cfa_fsync(int fd, __u64 submit_flags);
+extern int cfa_epoll_ctl(int epfd, int op, int fd, struct epoll_event *event, __u64 submit_flags);
+extern int cfa_sync_file_range(int fd, off64_t offset, off64_t nbytes, unsigned int flags, __u64 submit_flags);
+extern  ssize_t cfa_sendmsg(int sockfd, const struct msghdr *msg, int flags, __u64 submit_flags);
+extern ssize_t cfa_recvmsg(int sockfd, struct msghdr *msg, int flags, __u64 submit_flags);
+extern ssize_t cfa_send(int sockfd, const void *buf, size_t len, int flags, __u64 submit_flags);
+extern ssize_t cfa_recv(int sockfd, void *buf, size_t len, int flags, __u64 submit_flags);
+extern int cfa_accept4(int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags, __u64 submit_flags);
+extern int cfa_connect(int sockfd, const struct sockaddr *addr, socklen_t addrlen, __u64 submit_flags);
+extern int cfa_fallocate(int fd, int mode, off_t offset, off_t len, __u64 submit_flags);
+extern int cfa_posix_fadvise(int fd, off_t offset, off_t len, int advice, __u64 submit_flags);
+extern int cfa_madvise(void *addr, size_t length, int advice, __u64 submit_flags);
+extern int cfa_openat(int dirfd, const char *pathname, int flags, mode_t mode, __u64 submit_flags);
 #if defined(CFA_HAVE_OPENAT2)
         extern int cfa_openat2(int dirfd, const char *pathname, struct open_how * how, size_t size, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
+        extern int cfa_openat2(int dirfd, const char *pathname, struct open_how * how, size_t size, __u64 submit_flags);
 #endif
 extern int cfa_close(int fd, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
+extern int cfa_close(int fd, __u64 submit_flags);
 #if defined(CFA_HAVE_STATX)
         extern int cfa_statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
+        extern int cfa_statx(int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf, __u64 submit_flags);
 #endif
 extern ssize_t cfa_read(int fd, void * buf, size_t count, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern ssize_t cfa_write(int fd, void * buf, size_t count, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern ssize_t cfa_splice(int fd_in, loff_t *off_in, int fd_out, loff_t *off_out, size_t len, unsigned int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
 extern ssize_t cfa_tee(int fd_in, int fd_out, size_t len, unsigned int flags, int submit_flags, Duration timeout, io_cancellation * cancellation, io_context * context);
+extern ssize_t cfa_read(int fd, void * buf, size_t count, __u64 submit_flags);
+extern ssize_t cfa_write(int fd, void * buf, size_t count, __u64 submit_flags);
+extern ssize_t cfa_splice(int fd_in, __off64_t *off_in, int fd_out, __off64_t *off_out, size_t len, unsigned int flags, __u64 submit_flags);
+extern ssize_t cfa_tee(int fd_in, int fd_out, size_t len, unsigned int flags, __u64 submit_flags);
 //----------
 // asynchronous calls
 #if defined(CFA_HAVE_PREADV2)
         extern void async_preadv2(io_future_t & future, int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
+        extern void async_preadv2(io_future_t & future, int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags, __u64 submit_flags);
 #endif
 #if defined(CFA_HAVE_PWRITEV2)
         extern void async_pwritev2(io_future_t & future, int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
+        extern void async_pwritev2(io_future_t & future, int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags, __u64 submit_flags);
 #endif
 extern void async_fsync(io_future_t & future, int fd, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_epoll_ctl(io_future_t & future, int epfd, int op, int fd, struct epoll_event *event, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_sync_file_range(io_future_t & future, int fd, off64_t offset, off64_t nbytes, unsigned int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_sendmsg(io_future_t & future, int sockfd, const struct msghdr *msg, int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_recvmsg(io_future_t & future, int sockfd, struct msghdr *msg, int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_send(io_future_t & future, int sockfd, const void *buf, size_t len, int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_recv(io_future_t & future, int sockfd, void *buf, size_t len, int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_accept4(io_future_t & future, int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_connect(io_future_t & future, int sockfd, const struct sockaddr *addr, socklen_t addrlen, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_fallocate(io_future_t & future, int fd, int mode, off_t offset, off_t len, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_posix_fadvise(io_future_t & future, int fd, off_t offset, off_t len, int advice, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_madvise(io_future_t & future, void *addr, size_t length, int advice, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_openat(io_future_t & future, int dirfd, const char *pathname, int flags, mode_t mode, int submit_flags, io_cancellation * cancellation, io_context * context);
+extern void async_fsync(io_future_t & future, int fd, __u64 submit_flags);
+extern void async_epoll_ctl(io_future_t & future, int epfd, int op, int fd, struct epoll_event *event, __u64 submit_flags);
+extern void async_sync_file_range(io_future_t & future, int fd, off64_t offset, off64_t nbytes, unsigned int flags, __u64 submit_flags);
+extern void async_sendmsg(io_future_t & future, int sockfd, const struct msghdr *msg, int flags, __u64 submit_flags);
+extern void async_recvmsg(io_future_t & future, int sockfd, struct msghdr *msg, int flags, __u64 submit_flags);
+extern void async_send(io_future_t & future, int sockfd, const void *buf, size_t len, int flags, __u64 submit_flags);
+extern void async_recv(io_future_t & future, int sockfd, void *buf, size_t len, int flags, __u64 submit_flags);
+extern void async_accept4(io_future_t & future, int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags, __u64 submit_flags);
+extern void async_connect(io_future_t & future, int sockfd, const struct sockaddr *addr, socklen_t addrlen, __u64 submit_flags);
+extern void async_fallocate(io_future_t & future, int fd, int mode, off_t offset, off_t len, __u64 submit_flags);
+extern void async_posix_fadvise(io_future_t & future, int fd, off_t offset, off_t len, int advice, __u64 submit_flags);
+extern void async_madvise(io_future_t & future, void *addr, size_t length, int advice, __u64 submit_flags);
+extern void async_openat(io_future_t & future, int dirfd, const char *pathname, int flags, mode_t mode, __u64 submit_flags);
 #if defined(CFA_HAVE_OPENAT2)
         extern void async_openat2(io_future_t & future, int dirfd, const char *pathname, struct open_how * how, size_t size, int submit_flags, io_cancellation * cancellation, io_context * context);
+        extern void async_openat2(io_future_t & future, int dirfd, const char *pathname, struct open_how * how, size_t size, __u64 submit_flags);
 #endif
 extern void async_close(io_future_t & future, int fd, int submit_flags, io_cancellation * cancellation, io_context * context);
+extern void async_close(io_future_t & future, int fd, __u64 submit_flags);
 #if defined(CFA_HAVE_STATX)
         extern void async_statx(io_future_t & future, int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf, int submit_flags, io_cancellation * cancellation, io_context * context);
+        extern void async_statx(io_future_t & future, int dirfd, const char *pathname, int flags, unsigned int mask, struct statx *statxbuf, __u64 submit_flags);
 #endif
 void async_read(io_future_t & future, int fd, void * buf, size_t count, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_write(io_future_t & future, int fd, void * buf, size_t count, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_splice(io_future_t & future, int fd_in, loff_t *off_in, int fd_out, loff_t *off_out, size_t len, unsigned int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
 extern void async_tee(io_future_t & future, int fd_in, int fd_out, size_t len, unsigned int flags, int submit_flags, io_cancellation * cancellation, io_context * context);
+void async_read(io_future_t & future, int fd, void * buf, size_t count, __u64 submit_flags);
+extern void async_write(io_future_t & future, int fd, void * buf, size_t count, __u64 submit_flags);
+extern void async_splice(io_future_t & future, int fd_in, __off64_t *off_in, int fd_out, __off64_t *off_out, size_t len, unsigned int flags, __u64 submit_flags);
+extern void async_tee(io_future_t & future, int fd_in, int fd_out, size_t len, unsigned int flags, __u64 submit_flags);
 …
 // Check if a function is blocks a only the user thread
 bool has_user_level_blocking( fptr_t func );
-//-----------------------------------------------------------------------------
-void register_fixed_files( io_context & ctx , int * files, unsigned count );
-void register_fixed_files( cluster    & cltr, int * files, unsigned count );

libcfa/src/concurrency/kernel.cfa

-              r342af53
+              r8e4aa05
 #include <signal.h>
 #include <unistd.h>
+extern "C" {
+        #include <sys/eventfd.h>
+}
 //CFA Includes
 …
 static [unsigned idle, unsigned total, * processor] query( & __cluster_idles idles );
+extern void __cfa_io_start( processor * );
+extern void __cfa_io_drain( processor * );
+extern void __cfa_io_flush( processor * );
+extern void __cfa_io_stop ( processor * );
+static inline void __maybe_io_drain( processor * );
+extern void __disable_interrupts_hard();
+extern void __enable_interrupts_hard();
 //=============================================================================================
 …
         verify(this);
+        __cfa_io_start( this );
         __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
         #if !defined(__CFA_NO_STATISTICS__)
 …
                 preemption_scope scope = { this };
+                #if !defined(__CFA_NO_STATISTICS__)
+                        unsigned long long last_tally = rdtscl();
+                #endif
                 __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);
 …
                 MAIN_LOOP:
                 for() {
+                        // Check if there is pending io
+                        __maybe_io_drain( this );
                         // Try to get the next thread
                         readyThread = __next_thread( this->cltr );
                         if( !readyThread ) {
+                                __cfa_io_flush( this );
                                 readyThread = __next_thread_slow( this->cltr );
+                        }
 …
                                 #endif
+                                wait( this->idle );
+                                __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle);
+                                __disable_interrupts_hard();
+                                eventfd_t val;
+                                eventfd_read( this->idle, &val );
+                                __enable_interrupts_hard();
                                 #if !defined(__CFA_NO_STATISTICS__)
 …
                         /* paranoid */ verify( readyThread );
+                        // Reset io dirty bit
+                        this->io.dirty = false;
                         // We found a thread run it
                         __run_thread(this, readyThread);
 …
                         // Are we done?
                         if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
+                        #if !defined(__CFA_NO_STATISTICS__)
+                                unsigned long long curr = rdtscl();
+                                if(curr > (last_tally + 500000000)) {
+                                        __tally_stats(this->cltr->stats, __cfaabi_tls.this_stats);
+                                        last_tally = curr;
+                                }
+                        #endif
+                        if(this->io.pending && !this->io.dirty) {
+                                __cfa_io_flush( this );
+                        }
+                }
 …
+        }
+        V( this->terminated );
+        __cfa_io_stop( this );
+        post( this->terminated );
         if(this == mainProcessor) {
 …
         /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
         __builtin_prefetch( thrd_dst->context.SP );
+        __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
         $coroutine * proc_cor = get_coroutine(this->runner);
 …
         // Just before returning to the processor, set the processor coroutine to active
         proc_cor->state = Active;
+        __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
         /* paranoid */ verify( ! __preemption_enabled() );
 …
         // We found a processor, wake it up
+        post( p->idle );
+        eventfd_t val;
+        val = 1;
+        eventfd_write( p->idle, val );
         #if !defined(__CFA_NO_STATISTICS__)
 …
         disable_interrupts();
                 /* paranoid */ verify( ! __preemption_enabled() );
+                post( this->idle );
+                eventfd_t val;
+                val = 1;
+                eventfd_write( this->idle, val );
         enable_interrupts( __cfaabi_dbg_ctx );
+}
 …
 // Unexpected Terminating logic
 //=============================================================================================
+static __spinlock_t kernel_abort_lock;
+static bool kernel_abort_called = false;
+void * kernel_abort(void) __attribute__ ((__nothrow__)) {
+        // abort cannot be recursively entered by the same or different processors because all signal handlers return when
+        // the globalAbort flag is true.
+        lock( kernel_abort_lock __cfaabi_dbg_ctx2 );
+        // disable interrupts, it no longer makes sense to try to interrupt this processor
+        disable_interrupts();
+        // first task to abort ?
+        if ( kernel_abort_called ) {                    // not first task to abort ?
+                unlock( kernel_abort_lock );
+                sigset_t mask;
+                sigemptyset( &mask );
+                sigaddset( &mask, SIGALRM );            // block SIGALRM signals
+                sigaddset( &mask, SIGUSR1 );            // block SIGALRM signals
+                sigsuspend( &mask );                            // block the processor to prevent further damage during abort
+                _exit( EXIT_FAILURE );                          // if processor unblocks before it is killed, terminate it
+        }
+        else {
+                kernel_abort_called = true;
+                unlock( kernel_abort_lock );
+        }
+        return __cfaabi_tls.this_thread;
+}
+void kernel_abort_msg( void * kernel_data, char * abort_text, int abort_text_size ) {
+        $thread * thrd = ( $thread * ) kernel_data;
+void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
+        $thread * thrd = __cfaabi_tls.this_thread;
         if(thrd) {
 …
+}
 int kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
         return get_coroutine(kernelTLS().this_thread) == get_coroutine(mainThread) ? 4 : 2;
+int __kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
+        return get_coroutine(__cfaabi_tls.this_thread) == get_coroutine(mainThread) ? 4 : 2;
+}
 …
 // Kernel Utilities
 //=============================================================================================
+//-----------------------------------------------------------------------------
+// Locks
+void  ?{}( semaphore & this, int count = 1 ) {
+        (this.lock){};
+        this.count = count;
+        (this.waiting){};
+}
+void ^?{}(semaphore & this) {}
+bool P(semaphore & this) with( this ){
+        lock( lock __cfaabi_dbg_ctx2 );
+        count -= 1;
+        if ( count < 0 ) {
+                // queue current task
+                append( waiting, active_thread() );
+                // atomically release spin lock and block
+                unlock( lock );
+                park();
+                return true;
+        }
+        else {
+            unlock( lock );
+            return false;
+        }
+}
+bool V(semaphore & this) with( this ) {
+        $thread * thrd = 0p;
+        lock( lock __cfaabi_dbg_ctx2 );
+        count += 1;
+        if ( count <= 0 ) {
+                // remove task at head of waiting list
+                thrd = pop_head( waiting );
+        }
+        unlock( lock );
+        // make new owner
+        unpark( thrd );
+        return thrd != 0p;
+}
+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 ) );
+        }
+        unlock( lock );
+        return thrd != 0p;
+#if defined(CFA_HAVE_LINUX_IO_URING_H)
+#include "io/types.hfa"
+#endif
+static inline void __maybe_io_drain( processor * proc ) {
+        #if defined(CFA_HAVE_LINUX_IO_URING_H)
+                __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);
+                // Check if we should drain the queue
+                $io_context * ctx = proc->io.ctx;
+                unsigned head = *ctx->cq.head;
+                unsigned tail = *ctx->cq.tail;
+                if(head != tail) __cfa_io_drain( proc );
+        #endif
+}

libcfa/src/concurrency/kernel.hfa

-              r342af53
+              r8e4aa05
 // file "LICENCE" distributed with Cforall.
 //
 // kernel --
+// kernel -- Header containing the core of the kernel API
 //
 // Author           : Thierry Delisle
 …
 extern "C" {
         #include <bits/pthreadtypes.h>
+        #include <pthread.h>
         #include <linux/types.h>
+}
+//-----------------------------------------------------------------------------
+// Locks
+struct semaphore {
+        __spinlock_t lock;
+        int count;
+        __queue_t($thread) waiting;
+};
+void  ?{}(semaphore & this, int count = 1);
+void ^?{}(semaphore & this);
+bool   P (semaphore & this);
+bool   V (semaphore & this);
+bool   V (semaphore & this, unsigned count);
+#ifdef __CFA_WITH_VERIFY__
+        extern bool __cfaabi_dbg_in_kernel();
+#endif
+//-----------------------------------------------------------------------------
+// I/O
+struct cluster;
+struct $io_context;
+struct $io_arbiter;
+struct io_context_params {
+        int num_entries;
+};
+void  ?{}(io_context_params & this);
 //-----------------------------------------------------------------------------
 …
         pthread_t kernel_thread;
+        struct {
+                $io_context * ctx;
+                bool pending;
+                bool dirty;
+        } io;
         // Preemption data
         // Node which is added in the discrete event simulaiton
 …
         // Idle lock (kernel semaphore)
         __bin_sem_t idle;
+        int idle;
         // Termination synchronisation (user semaphore)
         semaphore terminated;
+        oneshot terminated;
         // pthread Stack
 …
 DLISTED_MGD_IMPL_OUT(processor)
-//-----------------------------------------------------------------------------
-// I/O
-struct __io_data;
-// IO poller user-thread
-// Not using the "thread" keyword because we want to control
-// more carefully when to start/stop it
-struct $io_ctx_thread {
-        struct __io_data * ring;
-        single_sem sem;
-        volatile bool done;
-        $thread self;
-};
-struct io_context {
-        $io_ctx_thread thrd;
-};
-struct io_context_params {
-        int num_entries;
-        int num_ready;
-        int submit_aff;
-        bool eager_submits:1;
-        bool poller_submits:1;
-        bool poll_submit:1;
-        bool poll_complete:1;
-};
-void  ?{}(io_context_params & this);
-void  ?{}(io_context & this, struct cluster & cl);
-void  ?{}(io_context & this, struct cluster & cl, const io_context_params & params);
-void ^?{}(io_context & this);
-struct io_cancellation {
-        __u64 target;
-};
-static inline void  ?{}(io_cancellation & this) { this.target = -1u; }
-static inline void ^?{}(io_cancellation &) {}
-bool cancel(io_cancellation & this);
 //-----------------------------------------------------------------------------
 …
         struct {
                 io_context * ctxs;
                 unsigned cnt;
+                $io_arbiter * arbiter;
+                io_context_params params;
         } io;

libcfa/src/concurrency/kernel/fwd.hfa

-              r342af53
+              r8e4aa05
 // file "LICENCE" distributed with Cforall.
 //
+// kernel/fwd.hfa --
+// kernel/fwd.hfa -- PUBLIC
+// Fundamental code needed to implement threading M.E.S. algorithms.
 //
 // Author           : Thierry Delisle
 …
                 extern uint64_t thread_rand();
+                // Semaphore which only supports a single thread
+                struct single_sem {
+                        struct $thread * volatile ptr;
+                };
+                static inline {
+                        void  ?{}(single_sem & this) {
+                                this.ptr = 0p;
+                        }
+                        void ^?{}(single_sem &) {}
+                        bool wait(single_sem & this) {
+                                for() {
+                                        struct $thread * expected = this.ptr;
+                                        if(expected == 1p) {
+                                                if(__atomic_compare_exchange_n(&this.ptr, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
+                                                        return false;
+                                                }
+                                        }
+                                        else {
+                                                /* paranoid */ verify( expected == 0p );
+                                                if(__atomic_compare_exchange_n(&this.ptr, &expected, active_thread(), false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
+                                                        park();
+                                                        return true;
+                                                }
+                                        }
+                                }
+                        }
+                        bool post(single_sem & this) {
+                                for() {
+                                        struct $thread * expected = this.ptr;
+                                        if(expected == 1p) return false;
+                                        if(expected == 0p) {
+                                                if(__atomic_compare_exchange_n(&this.ptr, &expected, 1p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
+                                                        return false;
+                                                }
+                                        }
+                                        else {
+                                                if(__atomic_compare_exchange_n(&this.ptr, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
+                                                        unpark( expected );
+                                                        return true;
+                                                }
+                                        }
+                                }
+                        }
+                }
+                // Synchronozation primitive which only supports a single thread and one post
+                // Similar to a binary semaphore with a 'one shot' semantic
+                // is expected to be discarded after each party call their side
+                struct oneshot {
+                        // Internal state :
+                        //     0p     : is initial state (wait will block)
+                        //     1p     : fulfilled (wait won't block)
+                        // any thread : a thread is currently waiting
+                        struct $thread * volatile ptr;
+                };
+                static inline {
+                        void  ?{}(oneshot & this) {
+                                this.ptr = 0p;
+                        }
+                        void ^?{}(oneshot &) {}
+                        // Wait for the post, return immidiately if it already happened.
+                        // return true if the thread was parked
+                        bool wait(oneshot & this) {
+                                for() {
+                                        struct $thread * expected = this.ptr;
+                                        if(expected == 1p) return false;
+                                        /* paranoid */ verify( expected == 0p );
+                                        if(__atomic_compare_exchange_n(&this.ptr, &expected, active_thread(), false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
+                                                park();
+                                                /* paranoid */ verify( this.ptr == 1p );
+                                                return true;
+                                        }
+                                }
+                        }
+                        // Mark as fulfilled, wake thread if needed
+                        // return true if a thread was unparked
+                        bool post(oneshot & this) {
+                                struct $thread * got = __atomic_exchange_n( &this.ptr, 1p, __ATOMIC_SEQ_CST);
+                                if( got == 0p ) return false;
+                                unpark( got );
+                                return true;
+                        }
+                }
+                // base types for future to build upon
+                // It is based on the 'oneshot' type to allow multiple futures
+                // to block on the same instance, permitting users to block a single
+                // thread on "any of" [a given set of] futures.
+                // does not support multiple threads waiting on the same future
+                struct future_t {
+                        // Internal state :
+                        //     0p      : is initial state (wait will block)
+                        //     1p      : fulfilled (wait won't block)
+                        //     2p      : in progress ()
+                        //     3p      : abandoned, server should delete
+                        // any oneshot : a context has been setup to wait, a thread could wait on it
+                        struct oneshot * volatile ptr;
+                };
+                static inline {
+                        void  ?{}(future_t & this) {
+                                this.ptr = 0p;
+                        }
+                        void ^?{}(future_t &) {}
+                        void reset(future_t & this) {
+                                // needs to be in 0p or 1p
+                                __atomic_exchange_n( &this.ptr, 0p, __ATOMIC_SEQ_CST);
+                        }
+                        // check if the future is available
+                        bool available( future_t & this ) {
+                                return this.ptr == 1p;
+                        }
+                        // Prepare the future to be waited on
+                        // intented to be use by wait, wait_any, waitfor, etc. rather than used directly
+                        bool setup( future_t & this, oneshot & wait_ctx ) {
+                                /* paranoid */ verify( wait_ctx.ptr == 0p );
+                                // The future needs to set the wait context
+                                for() {
+                                        struct oneshot * expected = this.ptr;
+                                        // Is the future already fulfilled?
+                                        if(expected == 1p) return false; // Yes, just return false (didn't block)
+                                        // The future is not fulfilled, try to setup the wait context
+                                        /* paranoid */ verify( expected == 0p );
+                                        if(__atomic_compare_exchange_n(&this.ptr, &expected, &wait_ctx, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
+                                                return true;
+                                        }
+                                }
+                        }
+                        // Stop waiting on a future
+                        // When multiple futures are waited for together in "any of" pattern
+                        // futures that weren't fulfilled before the thread woke up
+                        // should retract the wait ctx
+                        // intented to be use by wait, wait_any, waitfor, etc. rather than used directly
+                        void retract( future_t & this, oneshot & wait_ctx ) {
+                                // Remove the wait context
+                                struct oneshot * got = __atomic_exchange_n( &this.ptr, 0p, __ATOMIC_SEQ_CST);
+                                // got == 0p: future was never actually setup, just return
+                                if( got == 0p ) return;
+                                // got == wait_ctx: since fulfil does an atomic_swap,
+                                // if we got back the original then no one else saw context
+                                // It is safe to delete (which could happen after the return)
+                                if( got == &wait_ctx ) return;
+                                // got == 1p: the future is ready and the context was fully consumed
+                                // the server won't use the pointer again
+                                // It is safe to delete (which could happen after the return)
+                                if( got == 1p ) return;
+                                // got == 2p: the future is ready but the context hasn't fully been consumed
+                                // spin until it is safe to move on
+                                if( got == 2p ) {
+                                        while( this.ptr != 1p ) Pause();
+                                        return;
+                                }
+                                // got == any thing else, something wen't wrong here, abort
+                                abort("Future in unexpected state");
+                        }
+                        // Mark the future as abandoned, meaning it will be deleted by the server
+                        bool abandon( future_t & this ) {
+                                /* paranoid */ verify( this.ptr != 3p );
+                                // Mark the future as abandonned
+                                struct oneshot * got = __atomic_exchange_n( &this.ptr, 3p, __ATOMIC_SEQ_CST);
+                                // If the future isn't already fulfilled, let the server delete it
+                                if( got == 0p ) return false;
+                                // got == 2p: the future is ready but the context hasn't fully been consumed
+                                // spin until it is safe to move on
+                                if( got == 2p ) {
+                                        while( this.ptr != 1p ) Pause();
+                                        got = 1p;
+                                }
+                                // The future is completed delete it now
+                                /* paranoid */ verify( this.ptr != 1p );
+                                free( &this );
+                                return true;
+                        }
+                        // from the server side, mark the future as fulfilled
+                        // delete it if needed
+                        bool fulfil( future_t & this ) {
+                                for() {
+                                        struct oneshot * expected = this.ptr;
+                                        // was this abandoned?
+                                        #if defined(__GNUC__) && __GNUC__ >= 7
+                                                #pragma GCC diagnostic push
+                                                #pragma GCC diagnostic ignored "-Wfree-nonheap-object"
+                                        #endif
+                                                if( expected == 3p ) { free( &this ); return false; }
+                                        #if defined(__GNUC__) && __GNUC__ >= 7
+                                                #pragma GCC diagnostic pop
+                                        #endif
+                                        /* paranoid */ verify( expected != 1p ); // Future is already fulfilled, should not happen
+                                        /* paranoid */ verify( expected != 2p ); // Future is bein fulfilled by someone else, this is even less supported then the previous case.
+                                        // If there is a wait context, we need to consume it and mark it as consumed after
+                                        // If there is no context then we can skip the in progress phase
+                                        struct oneshot * want = expected == 0p ? 1p : 2p;
+                                        if(__atomic_compare_exchange_n(&this.ptr, &expected, want, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
+                                                if( expected == 0p ) { /* paranoid */ verify( this.ptr == 1p); return false; }
+                                                bool ret = post( *expected );
+                                                __atomic_store_n( &this.ptr, 1p, __ATOMIC_SEQ_CST);
+                                                return ret;
+                                        }
+                                }
+                        }
+                        // Wait for the future to be fulfilled
+                        bool wait( future_t & this ) {
+                                oneshot temp;
+                                if( !setup(this, temp) ) return false;
+                                // Wait context is setup, just wait on it
+                                bool ret = wait( temp );
+                                // Wait for the future to tru
+                                while( this.ptr == 2p ) Pause();
+                                // Make sure the state makes sense
+                                // Should be fulfilled, could be in progress but it's out of date if so
+                                // since if that is the case, the oneshot was fulfilled (unparking this thread)
+                                // and the oneshot should not be needed any more
+                                __attribute__((unused)) struct oneshot * was = this.ptr;
+                                /* paranoid */ verifyf( was == 1p, "Expected this.ptr to be 1p, was %p\n", was );
+                                // Mark the future as fulfilled, to be consistent
+                                // with potential calls to avail
+                                // this.ptr = 1p;
+                                return ret;
+                        }
+                }
                 //-----------------------------------------------------------------------
                 // Statics call at the end of each thread to register statistics

libcfa/src/concurrency/kernel/startup.cfa

-              r342af53
+              r8e4aa05
 extern "C" {
       #include <limits.h>       // PTHREAD_STACK_MIN
+        #include <sys/eventfd.h>  // eventfd
       #include <sys/mman.h>     // mprotect
       #include <sys/resource.h> // getrlimit
 …
 extern void __kernel_alarm_startup(void);
 extern void __kernel_alarm_shutdown(void);
-extern void __kernel_io_startup (void);
-extern void __kernel_io_shutdown(void);
 //-----------------------------------------------------------------------------
 …
 KERNEL_STORAGE($thread,              mainThread);
 KERNEL_STORAGE(__stack_t,            mainThreadCtx);
-KERNEL_STORAGE(io_context,           mainPollerThread);
 KERNEL_STORAGE(__scheduler_RWLock_t, __scheduler_lock);
 #if !defined(__CFA_NO_STATISTICS__)
 …
         void ?{}(processor & this) with( this ) {
+                ( this.idle ){};
+                ( this.terminated ){ 0 };
+                ( this.terminated ){};
                 ( this.runner ){};
                 init( this, "Main Processor", *mainCluster );
 …
         __kernel_alarm_startup();
-        // Start IO
-        __kernel_io_startup();
         // Add the main thread to the ready queue
         // once resume is called on mainProcessor->runner the mainThread needs to be scheduled like any normal thread
 …
         // THE SYSTEM IS NOW COMPLETELY RUNNING
-        // SKULLDUGGERY: The constructor for the mainCluster will call alloc with a dimension of 0
-        // malloc *can* return a non-null value, we should free it if that is the case
-        free( mainCluster->io.ctxs );
-        // Now that the system is up, finish creating systems that need threading
-        mainCluster->io.ctxs = (io_context *)&storage_mainPollerThread;
-        mainCluster->io.cnt  = 1;
-        (*mainCluster->io.ctxs){ *mainCluster };
         __cfadbg_print_safe(runtime_core, "Kernel : Started\n--------------------------------------------------\n\n");
 …
 static void __kernel_shutdown(void) {
-        //Before we start shutting things down, wait for systems that need threading to shutdown
-        ^(*mainCluster->io.ctxs){};
-        mainCluster->io.cnt  = 0;
-        mainCluster->io.ctxs = 0p;
         /* paranoid */ verify( __preemption_enabled() );
         disable_interrupts();
 …
         // Disable preemption
         __kernel_alarm_shutdown();
-        // Stop IO
-        __kernel_io_shutdown();
         // Destroy the main processor and its context in reverse order of construction
 …
         pending_preemption = false;
+        this.io.ctx = 0p;
+        this.io.pending = false;
+        this.io.dirty   = false;
+        this.idle = eventfd(0, 0);
+        if (idle < 0) {
+                abort("KERNEL ERROR: PROCESSOR EVENTFD - %s\n", strerror(errno));
+        }
         #if !defined(__CFA_NO_STATISTICS__)
                 print_stats = 0;
 …
         // Finally we don't need the read_lock any more
         unregister((__processor_id_t*)&this);
+        close(this.idle);
+}
 void ?{}(processor & this, const char name[], cluster & _cltr) {
+        ( this.idle ){};
+        ( this.terminated ){ 0 };
+        ( this.terminated ){};
         ( this.runner ){};
 …
                 __wake_proc( &this );
                 P( terminated );
+                wait( terminated );
                 /* paranoid */ verify( active_processor() != &this);
+        }
 …
         threads{ __get };
+        io.arbiter = create();
+        io.params = io_params;
         doregister(this);
 …
         ready_mutate_unlock( last_size );
         enable_interrupts_noPoll(); // Don't poll, could be in main cluster
-        this.io.cnt  = num_io;
-        this.io.ctxs = aalloc(num_io);
-        for(i; this.io.cnt) {
-                (this.io.ctxs[i]){ this, io_params };
+        }
+}
 void ^?{}(cluster & this) {
+        for(i; this.io.cnt) {
+                ^(this.io.ctxs[i]){ true };
+        }
+        free(this.io.ctxs);
+        destroy(this.io.arbiter);
         // Lock the RWlock so no-one pushes/pops while we are changing the queue
 …
+}
 #if defined(__CFA_WITH_VERIFY__)
 static bool verify_fwd_bck_rng(void) {

libcfa/src/concurrency/kernel_private.hfa

-              r342af53
+              r8e4aa05
 //-----------------------------------------------------------------------------
 // I/O
+void ^?{}(io_context & this, bool );
+$io_arbiter * create(void);
+void destroy($io_arbiter *);
 //=======================================================================

libcfa/src/concurrency/locks.cfa

-              r342af53
+              r8e4aa05
+//
+// Cforall Version 1.0.0 Copyright (C) 2021 University of Waterloo
+//
+// The contents of this file are covered under the licence agreement in the
+// file "LICENCE" distributed with Cforall.
+//
+// locks.hfa -- LIBCFATHREAD
+// Runtime locks that used with the runtime thread system.
+//
+// Author           : Colby Alexander Parsons
+// Created On       : Thu Jan 21 19:46:50 2021
+// Last Modified By :
+// Last Modified On :
+// Update Count     :
+//
+#define __cforall_thread__
 #include "locks.hfa"
 #include "kernel_private.hfa"
 …
 //-----------------------------------------------------------------------------
 // info_thread
 forall(dtype L | is_blocking_lock(L)) {
+forall(L & | is_blocking_lock(L)) {
         struct info_thread {
                 // used to put info_thread on a dl queue (aka sequence)
 …
 void ^?{}( blocking_lock & this ) {}
+void  ?{}( single_acquisition_lock & this ) {((blocking_lock &)this){ false, false };}
+void ^?{}( single_acquisition_lock & this ) {}
+void  ?{}( owner_lock & this ) {((blocking_lock &)this){ true, true };}
+void ^?{}( owner_lock & this ) {}
+void  ?{}( multiple_acquisition_lock & this ) {((blocking_lock &)this){ true, false };}
+void ^?{}( multiple_acquisition_lock & this ) {}
 void lock( blocking_lock & this ) with( this ) {
 …
 //-----------------------------------------------------------------------------
-// Overloaded routines for traits
-// These routines are temporary until an inheritance bug is fixed
-void   lock      ( single_acquisition_lock & this ) { lock   ( (blocking_lock &)this ); }
-void   unlock    ( single_acquisition_lock & this ) { unlock ( (blocking_lock &)this ); }
-void   on_wait   ( single_acquisition_lock & this ) { on_wait( (blocking_lock &)this ); }
-void   on_notify ( single_acquisition_lock & this, struct $thread * t ) { on_notify( (blocking_lock &)this, t ); }
-void   set_recursion_count( single_acquisition_lock & this, size_t recursion ) { set_recursion_count( (blocking_lock &)this, recursion ); }
-size_t get_recursion_count( single_acquisition_lock & this ) { return get_recursion_count( (blocking_lock &)this ); }
-void   lock     ( owner_lock & this ) { lock   ( (blocking_lock &)this ); }
-void   unlock   ( owner_lock & this ) { unlock ( (blocking_lock &)this ); }
-void   on_wait  ( owner_lock & this ) { on_wait( (blocking_lock &)this ); }
-void   on_notify( owner_lock & this, struct $thread * t ) { on_notify( (blocking_lock &)this, t ); }
-void   set_recursion_count( owner_lock & this, size_t recursion ) { set_recursion_count( (blocking_lock &)this, recursion ); }
-size_t get_recursion_count( owner_lock & this ) { return get_recursion_count( (blocking_lock &)this ); }
-void   lock     ( multiple_acquisition_lock & this ) { lock   ( (blocking_lock &)this ); }
-void   unlock   ( multiple_acquisition_lock & this ) { unlock ( (blocking_lock &)this ); }
-void   on_wait  ( multiple_acquisition_lock & this ) { on_wait( (blocking_lock &)this ); }
-void   on_notify( multiple_acquisition_lock & this, struct $thread * t ){ on_notify( (blocking_lock &)this, t ); }
-void   set_recursion_count( multiple_acquisition_lock & this, size_t recursion ){ set_recursion_count( (blocking_lock &)this, recursion ); }
-size_t get_recursion_count( multiple_acquisition_lock & this ){ return get_recursion_count( (blocking_lock &)this ); }
-//-----------------------------------------------------------------------------
 // alarm node wrapper
 forall(dtype L | is_blocking_lock(L)) {
+forall(L & | is_blocking_lock(L)) {
         struct alarm_node_wrap {
                 alarm_node_t alarm_node;
 …
 //-----------------------------------------------------------------------------
 // condition variable
 forall(dtype L | is_blocking_lock(L)) {
+forall(L & | is_blocking_lock(L)) {
         void ?{}( condition_variable(L) & this ){
 …
         bool wait( condition_variable(L) & this, L & l, uintptr_t info, Time time         ) with(this) { WAIT_TIME( info, &l , time ) }
+}
+//-----------------------------------------------------------------------------
+// Semaphore
+void  ?{}( semaphore & this, int count = 1 ) {
+        (this.lock){};
+        this.count = count;
+        (this.waiting){};
+}
+void ^?{}(semaphore & this) {}
+bool P(semaphore & this) with( this ){
+        lock( lock __cfaabi_dbg_ctx2 );
+        count -= 1;
+        if ( count < 0 ) {
+                // queue current task
+                append( waiting, active_thread() );
+                // atomically release spin lock and block
+                unlock( lock );
+                park();
+                return true;
+        }
+        else {
+            unlock( lock );
+            return false;
+        }
+}
+bool V(semaphore & this) with( this ) {
+        $thread * thrd = 0p;
+        lock( lock __cfaabi_dbg_ctx2 );
+        count += 1;
+        if ( count <= 0 ) {
+                // remove task at head of waiting list
+                thrd = pop_head( waiting );
+        }
+        unlock( lock );
+        // make new owner
+        unpark( thrd );
+        return thrd != 0p;
+}
+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 ) );
+        }
+        unlock( lock );
+        return thrd != 0p;
+}

libcfa/src/concurrency/locks.hfa

-              r342af53
+              r8e4aa05
+//
+// Cforall Version 1.0.0 Copyright (C) 2021 University of Waterloo
+//
+// The contents of this file are covered under the licence agreement in the
+// file "LICENCE" distributed with Cforall.
+//
+// locks.hfa -- PUBLIC
+// Runtime locks that used with the runtime thread system.
+//
+// Author           : Colby Alexander Parsons
+// Created On       : Thu Jan 21 19:46:50 2021
+// Last Modified By :
+// Last Modified On :
+// Update Count     :
+//
 #pragma once
 #include <stdbool.h>
+#include "bits/locks.hfa"
+#include "bits/sequence.hfa"
+#include "invoke.h"
+#include "bits/weakso_locks.hfa"
 #include "time_t.hfa"
 #include "time.hfa"
+//----------
+struct single_acquisition_lock {
+        inline blocking_lock;
+};
+static inline void  ?{}( single_acquisition_lock & this ) {((blocking_lock &)this){ false, false };}
+static inline void ^?{}( single_acquisition_lock & this ) {}
+static inline void   lock      ( single_acquisition_lock & this ) { lock   ( (blocking_lock &)this ); }
+static inline void   unlock    ( single_acquisition_lock & this ) { unlock ( (blocking_lock &)this ); }
+static inline void   on_wait   ( single_acquisition_lock & this ) { on_wait( (blocking_lock &)this ); }
+static inline void   on_notify ( single_acquisition_lock & this, struct $thread * t ) { on_notify( (blocking_lock &)this, t ); }
+static inline void   set_recursion_count( single_acquisition_lock & this, size_t recursion ) { set_recursion_count( (blocking_lock &)this, recursion ); }
+static inline size_t get_recursion_count( single_acquisition_lock & this ) { return get_recursion_count( (blocking_lock &)this ); }
+//----------
+struct owner_lock {
+        inline blocking_lock;
+};
+static inline void  ?{}( owner_lock & this ) {((blocking_lock &)this){ true, true };}
+static inline void ^?{}( owner_lock & this ) {}
+static inline void   lock     ( owner_lock & this ) { lock   ( (blocking_lock &)this ); }
+static inline void   unlock   ( owner_lock & this ) { unlock ( (blocking_lock &)this ); }
+static inline void   on_wait  ( owner_lock & this ) { on_wait( (blocking_lock &)this ); }
+static inline void   on_notify( owner_lock & this, struct $thread * t ) { on_notify( (blocking_lock &)this, t ); }
+static inline void   set_recursion_count( owner_lock & this, size_t recursion ) { set_recursion_count( (blocking_lock &)this, recursion ); }
+static inline size_t get_recursion_count( owner_lock & this ) { return get_recursion_count( (blocking_lock &)this ); }
 //-----------------------------------------------------------------------------
 // is_blocking_lock
 trait is_blocking_lock(dtype L | sized(L)) {
+trait is_blocking_lock(L & | sized(L)) {
         // For synchronization locks to use when acquiring
         void on_notify( L &, struct $thread * );
 …
 // the info thread is a wrapper around a thread used
 // to store extra data for use in the condition variable
 forall(dtype L | is_blocking_lock(L)) {
+forall(L & | is_blocking_lock(L)) {
         struct info_thread;
 …
 //-----------------------------------------------------------------------------
-// Blocking Locks
-struct blocking_lock {
-        // Spin lock used for mutual exclusion
-        __spinlock_t lock;
-        // List of blocked threads
-        Sequence( $thread ) blocked_threads;
-        // Count of current blocked threads
-        size_t wait_count;
-        // Flag if the lock allows multiple acquisition
-        bool multi_acquisition;
-        // Flag if lock can be released by non owner
-        bool strict_owner;
-        // Current thread owning the lock
-        struct $thread * owner;
-        // Number of recursion level
-        size_t recursion_count;
-};
-struct single_acquisition_lock {
-        inline blocking_lock;
-};
-struct owner_lock {
-        inline blocking_lock;
-};
-struct multiple_acquisition_lock {
-        inline blocking_lock;
-};
-void  ?{}( blocking_lock & this, bool multi_acquisition, bool strict_owner );
-void ^?{}( blocking_lock & this );
-void  ?{}( single_acquisition_lock & this );
-void ^?{}( single_acquisition_lock & this );
-void  ?{}( owner_lock & this );
-void ^?{}( owner_lock & this );
-void  ?{}( multiple_acquisition_lock & this );
-void ^?{}( multiple_acquisition_lock & this );
-void lock( blocking_lock & this );
-bool try_lock( blocking_lock & this );
-void unlock( blocking_lock & this );
-void on_notify( blocking_lock & this, struct $thread * t );
-void on_wait( blocking_lock & this );
-size_t wait_count( blocking_lock & this );
-void set_recursion_count( blocking_lock & this, size_t recursion );
-size_t get_recursion_count( blocking_lock & this );
-void lock( single_acquisition_lock & this );
-void unlock( single_acquisition_lock & this );
-void on_notify( single_acquisition_lock & this, struct $thread * t );
-void on_wait( single_acquisition_lock & this );
-void set_recursion_count( single_acquisition_lock & this, size_t recursion );
-size_t get_recursion_count( single_acquisition_lock & this );
-void lock( owner_lock & this );
-void unlock( owner_lock & this );
-void on_notify( owner_lock & this, struct $thread * t );
-void on_wait( owner_lock & this );
-void set_recursion_count( owner_lock & this, size_t recursion );
-size_t get_recursion_count( owner_lock & this );
-void lock( multiple_acquisition_lock & this );
-void unlock( multiple_acquisition_lock & this );
-void on_notify( multiple_acquisition_lock & this, struct $thread * t );
-void on_wait( multiple_acquisition_lock & this );
-void set_recursion_count( multiple_acquisition_lock & this, size_t recursion );
-size_t get_recursion_count( multiple_acquisition_lock & this );
-//-----------------------------------------------------------------------------
 // Synchronization Locks
 forall(dtype L | is_blocking_lock(L)) {
+forall(L & | is_blocking_lock(L)) {
         struct condition_variable {
                 // Spin lock used for mutual exclusion
 …
         bool wait( condition_variable(L) & this, L & l, uintptr_t info, Time time );
+}
+//-----------------------------------------------------------------------------
+// Semaphore
+struct semaphore {
+        __spinlock_t lock;
+        int count;
+        __queue_t($thread) waiting;
+};
+void  ?{}(semaphore & this, int count = 1);
+void ^?{}(semaphore & this);
+bool   P (semaphore & this);
+bool   V (semaphore & this);
+bool   V (semaphore & this, unsigned count);

libcfa/src/concurrency/monitor.cfa

-              r342af53
+              r8e4aa05
 static inline [$thread *, int] search_entry_queue( const __waitfor_mask_t &, $monitor * monitors [], __lock_size_t count );
 forall(dtype T | sized( T ))
+forall(T & | sized( T ))
 static inline __lock_size_t insert_unique( T * array [], __lock_size_t & size, T * val );
 static inline __lock_size_t count_max    ( const __waitfor_mask_t & mask );
 …
+}
 forall(dtype T | sized( T ))
+forall(T & | sized( T ))
 static inline __lock_size_t insert_unique( T * array [], __lock_size_t & size, T * val ) {
         if( !val ) return size;

libcfa/src/concurrency/monitor.hfa

-              r342af53
+              r8e4aa05
 #include "stdlib.hfa"
 trait is_monitor(dtype T) {
+trait is_monitor(T &) {
         $monitor * get_monitor( T & );
         void ^?{}( T & mutex );
 …
 void ^?{}( monitor_dtor_guard_t & this );
 static inline forall( dtype T | sized(T) | { void ^?{}( T & mutex ); } )
+static inline forall( T & | sized(T) | { void ^?{}( T & mutex ); } )
 void delete( T * th ) {
         ^(*th){};
+        if(th) ^(*th){};
         free( th );
+}

libcfa/src/concurrency/mutex.cfa

-              r342af53
+              r8e4aa05
+}
 forall(dtype L | is_lock(L))
+forall(L & | is_lock(L))
 void wait(condition_variable & this, L & l) {
         lock( this.lock __cfaabi_dbg_ctx2 );
 …
 //-----------------------------------------------------------------------------
 // Scopes
 forall(dtype L | is_lock(L))
+forall(L & | is_lock(L))
 void lock_all  ( L * locks[], size_t count) {
         // Sort locks based on addresses
 …
+}
 forall(dtype L | is_lock(L))
+forall(L & | is_lock(L))
 void unlock_all( L * locks[], size_t count) {
         // Lock all

libcfa/src/concurrency/mutex.hfa

-              r342af53
+              r8e4aa05
 };
 void ?{}(mutex_lock & this);
 void ^?{}(mutex_lock & this);
 void lock(mutex_lock & this);
 bool try_lock(mutex_lock & this);
 void unlock(mutex_lock & this);
+void ?{}(mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+void ^?{}(mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+void lock(mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+bool try_lock(mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+void unlock(mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
 // Exclusive lock - recursive
 …
 };
 void ?{}(recursive_mutex_lock & this);
 void ^?{}(recursive_mutex_lock & this);
 void lock(recursive_mutex_lock & this);
 bool try_lock(recursive_mutex_lock & this);
 void unlock(recursive_mutex_lock & this);
+void ?{}(recursive_mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+void ^?{}(recursive_mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+void lock(recursive_mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+bool try_lock(recursive_mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+void unlock(recursive_mutex_lock & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
 trait is_lock(dtype L | sized(L)) {
+trait is_lock(L & | sized(L)) {
         void lock  (L &);
         void unlock(L &);
 …
 };
 void ?{}(condition_variable & this);
 void ^?{}(condition_variable & this);
+void ?{}(condition_variable & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+void ^?{}(condition_variable & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
 void notify_one(condition_variable & this);
 void notify_all(condition_variable & this);
+void notify_one(condition_variable & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
+void notify_all(condition_variable & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
 void wait(condition_variable & this);
+void wait(condition_variable & this) __attribute__((deprecated("use concurrency/locks.hfa instead")));
 forall(dtype L | is_lock(L))
 void wait(condition_variable & this, L & l);
+forall(L & | is_lock(L))
+void wait(condition_variable & this, L & l) __attribute__((deprecated("use concurrency/locks.hfa instead")));
 //-----------------------------------------------------------------------------
 // Scopes
 forall(dtype L | is_lock(L)) {
+forall(L & | is_lock(L)) {
         #if !defined( __TUPLE_ARRAYS_EXIST__ )
         void lock  ( L * locks [], size_t count);

libcfa/src/concurrency/preemption.cfa

-              r342af53
+              r8e4aa05
 static void timeout( $thread * this ) {
         unpark( this );
+}
+void __disable_interrupts_hard() {
+        sigset_t oldset;
+        int ret;
+        ret = pthread_sigmask(0, ( const sigset_t * ) 0p, &oldset);  // workaround trac#208: cast should be unnecessary
+        if(ret != 0) { abort("ERROR sigprocmask returned %d", ret); }
+        ret = sigismember(&oldset, SIGUSR1);
+        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
+        if(ret == 1) { abort("ERROR SIGUSR1 is disabled"); }
+        ret = sigismember(&oldset, SIGALRM);
+        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
+        if(ret == 0) { abort("ERROR SIGALRM is enabled"); }
+        signal_block( SIGUSR1 );
+}
+void __enable_interrupts_hard() {
+        signal_unblock( SIGUSR1 );
+        sigset_t oldset;
+        int ret;
+        ret = pthread_sigmask(0, ( const sigset_t * ) 0p, &oldset);  // workaround trac#208: cast should be unnecessary
+        if(ret != 0) { abort("ERROR sigprocmask returned %d", ret); }
+        ret = sigismember(&oldset, SIGUSR1);
+        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
+        if(ret == 1) { abort("ERROR SIGUSR1 is disabled"); }
+        ret = sigismember(&oldset, SIGALRM);
+        if(ret <  0) { abort("ERROR sigismember returned %d", ret); }
+        if(ret == 0) { abort("ERROR SIGALRM is enabled"); }
+}
 …
         // Setup proper signal handlers
         __cfaabi_sigaction( SIGUSR1, sigHandler_ctxSwitch, SA_SIGINFO | SA_RESTART ); // __cfactx_switch handler
         __cfaabi_sigaction( SIGALRM, sigHandler_alarm    , SA_SIGINFO | SA_RESTART ); // debug handler
+        __cfaabi_sigaction( SIGUSR1, sigHandler_ctxSwitch, SA_SIGINFO ); // __cfactx_switch handler
+        __cfaabi_sigaction( SIGALRM, sigHandler_alarm    , SA_SIGINFO ); // debug handler
         signal_block( SIGALRM );
 …
         __cfaabi_dbg_print_safe( "Kernel : Preemption stopped\n" );
+}
+// Prevent preemption since we are about to start terminating things
+void __kernel_abort_lock(void) {
+        signal_block( SIGUSR1 );
+}

libcfa/src/concurrency/ready_queue.cfa

r342af53	r8e4aa05
330	330	#if defined(BIAS)
331	331	// Don't bother trying locally too much
332		~~int local_tries = 8;~~
333	332	preferred = kernelTLS().this_processor->id * 4;
334	333	#endif

libcfa/src/concurrency/stats.cfa

-              r342af53
+              r8e4aa05
                 #if defined(CFA_HAVE_LINUX_IO_URING_H)
+                        stats->io.submit_q.submit_avg.rdy = 0;
+                        stats->io.submit_q.submit_avg.csm = 0;
+                        stats->io.submit_q.submit_avg.cnt = 0;
+                        stats->io.submit_q.look_avg.val   = 0;
+                        stats->io.submit_q.look_avg.cnt   = 0;
+                        stats->io.submit_q.look_avg.block = 0;
+                        stats->io.submit_q.alloc_avg.val   = 0;
+                        stats->io.submit_q.alloc_avg.cnt   = 0;
+                        stats->io.submit_q.alloc_avg.block = 0;
+                        stats->io.submit_q.helped = 0;
+                        stats->io.submit_q.leader = 0;
+                        stats->io.submit_q.busy   = 0;
+                        stats->io.complete_q.completed_avg.val = 0;
+                        stats->io.complete_q.completed_avg.cnt = 0;
+                        stats->io.complete_q.blocks = 0;
+                        stats->io.alloc.fast        = 0;
+                        stats->io.alloc.slow        = 0;
+                        stats->io.alloc.fail        = 0;
+                        stats->io.alloc.revoke      = 0;
+                        stats->io.alloc.block       = 0;
+                        stats->io.submit.fast       = 0;
+                        stats->io.submit.slow       = 0;
+                        stats->io.flush.external    = 0;
+                        stats->io.calls.flush       = 0;
+                        stats->io.calls.submitted   = 0;
+                        stats->io.calls.drain       = 0;
+                        stats->io.calls.completed   = 0;
+                        stats->io.calls.errors.busy = 0;
+                        stats->io.poller.sleeps     = 0;
                 #endif
+        }
 …
                 #if defined(CFA_HAVE_LINUX_IO_URING_H)
+                        __atomic_fetch_add( &cltr->io.submit_q.submit_avg.rdy     , proc->io.submit_q.submit_avg.rdy     , __ATOMIC_SEQ_CST ); proc->io.submit_q.submit_avg.rdy      = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.submit_avg.csm     , proc->io.submit_q.submit_avg.csm     , __ATOMIC_SEQ_CST ); proc->io.submit_q.submit_avg.csm      = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.submit_avg.avl     , proc->io.submit_q.submit_avg.avl     , __ATOMIC_SEQ_CST ); proc->io.submit_q.submit_avg.avl      = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.submit_avg.cnt     , proc->io.submit_q.submit_avg.cnt     , __ATOMIC_SEQ_CST ); proc->io.submit_q.submit_avg.cnt      = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.look_avg.val       , proc->io.submit_q.look_avg.val       , __ATOMIC_SEQ_CST ); proc->io.submit_q.look_avg.val        = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.look_avg.cnt       , proc->io.submit_q.look_avg.cnt       , __ATOMIC_SEQ_CST ); proc->io.submit_q.look_avg.cnt        = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.look_avg.block     , proc->io.submit_q.look_avg.block     , __ATOMIC_SEQ_CST ); proc->io.submit_q.look_avg.block      = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.alloc_avg.val      , proc->io.submit_q.alloc_avg.val      , __ATOMIC_SEQ_CST ); proc->io.submit_q.alloc_avg.val       = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.alloc_avg.cnt      , proc->io.submit_q.alloc_avg.cnt      , __ATOMIC_SEQ_CST ); proc->io.submit_q.alloc_avg.cnt       = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.alloc_avg.block    , proc->io.submit_q.alloc_avg.block    , __ATOMIC_SEQ_CST ); proc->io.submit_q.alloc_avg.block     = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.helped             , proc->io.submit_q.helped             , __ATOMIC_SEQ_CST ); proc->io.submit_q.helped              = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.leader             , proc->io.submit_q.leader             , __ATOMIC_SEQ_CST ); proc->io.submit_q.leader              = 0;
+                        __atomic_fetch_add( &cltr->io.submit_q.busy               , proc->io.submit_q.busy               , __ATOMIC_SEQ_CST ); proc->io.submit_q.busy                = 0;
+                        __atomic_fetch_add( &cltr->io.complete_q.completed_avg.val, proc->io.complete_q.completed_avg.val, __ATOMIC_SEQ_CST ); proc->io.complete_q.completed_avg.val = 0;
+                        __atomic_fetch_add( &cltr->io.complete_q.completed_avg.cnt, proc->io.complete_q.completed_avg.cnt, __ATOMIC_SEQ_CST ); proc->io.complete_q.completed_avg.cnt = 0;
+                        __atomic_fetch_add( &cltr->io.complete_q.blocks           , proc->io.complete_q.blocks           , __ATOMIC_SEQ_CST ); proc->io.complete_q.blocks            = 0;
+                        __atomic_fetch_add( &cltr->io.alloc.fast       , proc->io.alloc.fast       , __ATOMIC_SEQ_CST ); proc->io.alloc.fast        = 0;
+                        __atomic_fetch_add( &cltr->io.alloc.slow       , proc->io.alloc.slow       , __ATOMIC_SEQ_CST ); proc->io.alloc.slow        = 0;
+                        __atomic_fetch_add( &cltr->io.alloc.fail       , proc->io.alloc.fail       , __ATOMIC_SEQ_CST ); proc->io.alloc.fail        = 0;
+                        __atomic_fetch_add( &cltr->io.alloc.revoke     , proc->io.alloc.revoke     , __ATOMIC_SEQ_CST ); proc->io.alloc.revoke      = 0;
+                        __atomic_fetch_add( &cltr->io.alloc.block      , proc->io.alloc.block      , __ATOMIC_SEQ_CST ); proc->io.alloc.block       = 0;
+                        __atomic_fetch_add( &cltr->io.submit.fast      , proc->io.submit.fast      , __ATOMIC_SEQ_CST ); proc->io.submit.fast       = 0;
+                        __atomic_fetch_add( &cltr->io.submit.slow      , proc->io.submit.slow      , __ATOMIC_SEQ_CST ); proc->io.submit.slow       = 0;
+                        __atomic_fetch_add( &cltr->io.flush.external   , proc->io.flush.external   , __ATOMIC_SEQ_CST ); proc->io.flush.external    = 0;
+                        __atomic_fetch_add( &cltr->io.calls.flush      , proc->io.calls.flush      , __ATOMIC_SEQ_CST ); proc->io.calls.flush       = 0;
+                        __atomic_fetch_add( &cltr->io.calls.submitted  , proc->io.calls.submitted  , __ATOMIC_SEQ_CST ); proc->io.calls.submitted   = 0;
+                        __atomic_fetch_add( &cltr->io.calls.drain      , proc->io.calls.drain      , __ATOMIC_SEQ_CST ); proc->io.calls.drain       = 0;
+                        __atomic_fetch_add( &cltr->io.calls.completed  , proc->io.calls.completed  , __ATOMIC_SEQ_CST ); proc->io.calls.completed   = 0;
+                        __atomic_fetch_add( &cltr->io.calls.errors.busy, proc->io.calls.errors.busy, __ATOMIC_SEQ_CST ); proc->io.calls.errors.busy = 0;
+                        __atomic_fetch_add( &cltr->io.poller.sleeps    , proc->io.poller.sleeps    , __ATOMIC_SEQ_CST ); proc->io.poller.sleeps     = 0;
                 #endif
+        }
 …
                 if( flags & CFA_STATS_READY_Q ) {
-                        double push_sur = (100.0 * ((double)ready.pick.push.success) / ready.pick.push.attempt);
-                        double pop_sur  = (100.0 * ((double)ready.pick.pop .success) / ready.pick.pop .attempt);
                         double push_len = ((double)ready.pick.push.attempt) / ready.pick.push.success;
                         double pop_len  = ((double)ready.pick.pop .attempt) / ready.pick.pop .success;
-                        double lpush_sur = (100.0 * ((double)ready.pick.push.lsuccess) / ready.pick.push.local);
-                        double lpop_sur  = (100.0 * ((double)ready.pick.pop .lsuccess) / ready.pick.pop .local);
                         double lpush_len = ((double)ready.pick.push.local) / ready.pick.push.lsuccess;
 …
                         __cfaabi_bits_print_safe( STDOUT_FILENO,
                                 "----- %s \"%s\" (%p) - Ready Q Stats -----\n"
+                                "- total threads run      : %'15" PRIu64 "\n"
+                                "- total threads scheduled: %'15" PRIu64 "\n"
+                                "- push average probe len : %'18.2lf, %'18.2lf%% (%'15" PRIu64 " attempts)\n"
+                                "- pop  average probe len : %'18.2lf, %'18.2lf%% (%'15" PRIu64 " attempts)\n"
+                                "- local push avg prb len : %'18.2lf, %'18.2lf%% (%'15" PRIu64 " attempts)\n"
+                                "- local pop  avg prb len : %'18.2lf, %'18.2lf%% (%'15" PRIu64 " attempts)\n"
+                                "- thread migrations      : %'15" PRIu64 "\n"
+                                "- Idle Sleep -\n"
+                                "-- halts                 : %'15" PRIu64 "\n"
+                                "-- cancelled halts       : %'15" PRIu64 "\n"
+                                "-- schedule wake         : %'15" PRIu64 "\n"
+                                "-- wake on exit          : %'15" PRIu64 "\n"
+                                "- total threads  : %'15" PRIu64 "run, %'15" PRIu64 "schd (%'" PRIu64 "mig )\n"
+                                "- push avg probe : %'3.2lf, %'3.2lfl (%'15" PRIu64 " attempts, %'15" PRIu64 " locals)\n"
+                                "- pop  avg probe : %'3.2lf, %'3.2lfl (%'15" PRIu64 " attempts, %'15" PRIu64 " locals)\n"
+                                "- Idle Sleep     : %'15" PRIu64 "h, %'15" PRIu64 "c, %'15" PRIu64 "w, %'15" PRIu64 "e\n"
                                 "\n"
                                 , type, name, id
                                 , ready.pick.pop.success
                                 , ready.pick.push.success
-                                , push_len, push_sur, ready.pick.push.attempt
-                                , pop_len , pop_sur , ready.pick.pop .attempt
-                                , lpush_len, lpush_sur, ready.pick.push.local
-                                , lpop_len , lpop_sur , ready.pick.pop .local
                                 , ready.threads.migration
+                                , push_len, lpush_len, ready.pick.push.attempt, ready.pick.push.local
+                                , pop_len , lpop_len , ready.pick.pop .attempt, ready.pick.pop .local
                                 , ready.sleep.halts, ready.sleep.cancels, ready.sleep.wakes, ready.sleep.exits
                         );
 …
                 #if defined(CFA_HAVE_LINUX_IO_URING_H)
                         if( flags & CFA_STATS_IO ) {
                                 double avgrdy = ((double)io.submit_q.submit_avg.rdy) / io.submit_q.submit_avg.cnt;
                                 double avgcsm = ((double)io.submit_q.submit_avg.csm) / io.submit_q.submit_avg.cnt;
+                                uint64_t total_allocs = io.alloc.fast + io.alloc.slow;
+                                double avgfasta = ((double)io.alloc.fast) / total_allocs;
+                                double lavgv = 0;
+                                double lavgb = 0;
+                                if(io.submit_q.look_avg.cnt != 0) {
+                                        lavgv = ((double)io.submit_q.look_avg.val  ) / io.submit_q.look_avg.cnt;
+                                        lavgb = ((double)io.submit_q.look_avg.block) / io.submit_q.look_avg.cnt;
+                                }
+                                uint64_t total_submits = io.submit.fast + io.submit.slow;
+                                double avgfasts = ((double)io.submit.fast) / total_submits;
+                                double aavgv = 0;
+                                double aavgb = 0;
+                                if(io.submit_q.alloc_avg.cnt != 0) {
+                                        aavgv = ((double)io.submit_q.alloc_avg.val  ) / io.submit_q.alloc_avg.cnt;
+                                        aavgb = ((double)io.submit_q.alloc_avg.block) / io.submit_q.alloc_avg.cnt;
+                                }
+                                double avgsubs = ((double)io.calls.submitted) / io.calls.flush;
+                                double avgcomp = ((double)io.calls.completed) / io.calls.drain;
                                 __cfaabi_bits_print_safe( STDOUT_FILENO,
                                         "----- %s \"%s\" (%p) - I/O Stats -----\n"
+                                        "- total submit calls     : %'15" PRIu64 "\n"
+                                        "- avg ready entries      : %'18.2lf\n"
+                                        "- avg submitted entries  : %'18.2lf\n"
+                                        "- total helped entries   : %'15" PRIu64 "\n"
+                                        "- total leader entries   : %'15" PRIu64 "\n"
+                                        "- total busy submit      : %'15" PRIu64 "\n"
+                                        "- total ready search     : %'15" PRIu64 "\n"
+                                        "- avg ready search len   : %'18.2lf\n"
+                                        "- avg ready search block : %'18.2lf\n"
+                                        "- total alloc search     : %'15" PRIu64 "\n"
+                                        "- avg alloc search len   : %'18.2lf\n"
+                                        "- avg alloc search block : %'18.2lf\n"
+                                        "- total wait calls       : %'15" PRIu64 "\n"
+                                        "- avg completion/wait    : %'18.2lf\n"
+                                        "- total completion blocks: %'15" PRIu64 "\n"
+                                        "- total allocations : %'" PRIu64 "f, %'" PRIu64 "s (%'2.2lff) \n"
+                                        "-     failures      : %'" PRIu64 "oom, %'" PRIu64 "rvk, %'" PRIu64 "blk\n"
+                                        "- total submits     : %'" PRIu64 "f, %'" PRIu64 "s (%'2.2lf) \n"
+                                        "- flush external    : %'" PRIu64 "\n"
+                                        "- io_uring_enter    : %'" PRIu64 " (%'" PRIu64 ", %'" PRIu64 " EBUSY)\n"
+                                        "-     submits       : %'" PRIu64 " (%'.2lf) \n"
+                                        "-     completes     : %'" PRIu64 " (%'.2lf) \n"
+                                        "- poller sleeping   : %'" PRIu64 "\n"
                                         "\n"
                                         , type,  name, id
+                                        , io.submit_q.submit_avg.cnt
+                                        , avgrdy, avgcsm
+                                        , io.submit_q.helped, io.submit_q.leader, io.submit_q.busy
+                                        , io.submit_q.look_avg.cnt
+                                        , lavgv, lavgb
+                                        , io.submit_q.alloc_avg.cnt
+                                        , aavgv, aavgb
+                                        , io.complete_q.completed_avg.cnt
+                                        , ((double)io.complete_q.completed_avg.val) / io.complete_q.completed_avg.cnt
+                                        , io.complete_q.blocks
+                                        , io.alloc.fast, io.alloc.slow, avgfasta
+                                        , io.alloc.fail, io.alloc.revoke, io.alloc.block
+                                        , io.submit.fast, io.submit.slow, avgfasts
+                                        , io.flush.external
+                                        , io.calls.flush, io.calls.drain, io.calls.errors.busy
+                                        , io.calls.submitted, avgsubs
+                                        , io.calls.completed, avgcomp
+                                        , io.poller.sleeps
                                 );
+                        }

libcfa/src/concurrency/stats.hfa

-              r342af53
+              r8e4aa05
 #include <stdint.h>
+enum {
+        CFA_STATS_READY_Q  = 0x01,
+        CFA_STATS_IO = 0x02,
+};
 #if defined(__CFA_NO_STATISTICS__)
 …
         static inline void __print_stats( struct __stats_t *, int, const char *, const char *, void * ) {}
 #else
-        enum {
-                CFA_STATS_READY_Q  = 0x01,
-                #if defined(CFA_HAVE_LINUX_IO_URING_H)
-                        CFA_STATS_IO = 0x02,
-                #endif
-        };
         struct __attribute__((aligned(64))) __stats_readQ_t {
 …
                 struct __attribute__((aligned(64))) __stats_io_t{
                         struct {
+                                volatile uint64_t fast;
+                                volatile uint64_t slow;
+                                volatile uint64_t fail;
+                                volatile uint64_t revoke;
+                                volatile uint64_t block;
+                        } alloc;
+                        struct {
+                                volatile uint64_t fast;
+                                volatile uint64_t slow;
+                        } submit;
+                        struct {
+                                volatile uint64_t external;
+                        } flush;
+                        struct {
+                                volatile uint64_t drain;
+                                volatile uint64_t completed;
+                                volatile uint64_t flush;
+                                volatile uint64_t submitted;
                                 struct {
+                                        volatile uint64_t rdy;
+                                        volatile uint64_t csm;
+                                        volatile uint64_t avl;
+                                        volatile uint64_t cnt;
+                                } submit_avg;
+                                struct {
+                                        volatile uint64_t val;
+                                        volatile uint64_t cnt;
+                                        volatile uint64_t block;
+                                } look_avg;
+                                struct {
+                                        volatile uint64_t val;
+                                        volatile uint64_t cnt;
+                                        volatile uint64_t block;
+                                } alloc_avg;
+                                volatile uint64_t helped;
+                                volatile uint64_t leader;
+                                volatile uint64_t busy;
+                        } submit_q;
+                                        volatile uint64_t busy;
+                                } errors;
+                        } calls;
                         struct {
+                                struct {
+                                        volatile uint64_t val;
+                                        volatile uint64_t cnt;
+                                } completed_avg;
+                                volatile uint64_t blocks;
+                        } complete_q;
+                                volatile uint64_t sleeps;
+                        } poller;
                 };
         #endif

libcfa/src/concurrency/thread.cfa

-              r342af53
+              r8e4aa05
+}
 FORALL_DATA_INSTANCE(ThreadCancelled, (dtype thread_t), (thread_t))
+FORALL_DATA_INSTANCE(ThreadCancelled, (thread_t &), (thread_t))
 forall(dtype T)
+forall(T &)
 void copy(ThreadCancelled(T) * dst, ThreadCancelled(T) * src) {
         dst->virtual_table = src->virtual_table;
 …
+}
 forall(dtype T)
+forall(T &)
 const char * msg(ThreadCancelled(T) *) {
         return "ThreadCancelled";
+}
 forall(dtype T)
+forall(T &)
 static void default_thread_cancel_handler(ThreadCancelled(T) & ) {
         abort( "Unhandled thread cancellation.\n" );
+}
 forall(dtype T | is_thread(T) | IS_EXCEPTION(ThreadCancelled, (T)))
+forall(T & | is_thread(T) | IS_EXCEPTION(ThreadCancelled, (T)))
 void ?{}( thread_dtor_guard_t & this,
                 T & thrd, void(*defaultResumptionHandler)(ThreadCancelled(T) &)) {
         $monitor * m = get_monitor(thrd);
+                T & thrd, void(*cancelHandler)(ThreadCancelled(T) &)) {
+        $monitor * m = get_monitor(thrd);
         $thread * desc = get_thread(thrd);
         // Setup the monitor guard
         void (*dtor)(T& mutex this) = ^?{};
         bool join = defaultResumptionHandler != (void(*)(ThreadCancelled(T)&))0;
+        bool join = cancelHandler != (void(*)(ThreadCancelled(T)&))0;
         (this.mg){&m, (void(*)())dtor, join};
 …
+        }
         desc->state = Cancelled;
+        if (!join) {
+                defaultResumptionHandler = default_thread_cancel_handler;
+        }
+        void(*defaultResumptionHandler)(ThreadCancelled(T) &) =
+                join ? cancelHandler : default_thread_cancel_handler;
         ThreadCancelled(T) except;
 …
 //-----------------------------------------------------------------------------
 // Starting and stopping threads
 forall( dtype T | is_thread(T) )
+forall( T & | is_thread(T) )
 void __thrd_start( T & this, void (*main_p)(T &) ) {
         $thread * this_thrd = get_thread(this);
 …
 //-----------------------------------------------------------------------------
 // Support for threads that don't ues the thread keyword
 forall( dtype T | sized(T) | is_thread(T) | { void ?{}(T&); } )
+forall( T & | sized(T) | is_thread(T) | { void ?{}(T&); } )
 void ?{}( scoped(T)& this ) with( this ) {
         handle{};
 …
+}
 forall( dtype T, ttype P | sized(T) | is_thread(T) | { void ?{}(T&, P); } )
+forall( T &, P... | sized(T) | is_thread(T) | { void ?{}(T&, P); } )
 void ?{}( scoped(T)& this, P params ) with( this ) {
         handle{ params };
 …
+}
 forall( dtype T | sized(T) | is_thread(T) )
+forall( T & | sized(T) | is_thread(T) )
 void ^?{}( scoped(T)& this ) with( this ) {
         ^handle{};
 …
 //-----------------------------------------------------------------------------
 forall(dtype T | is_thread(T) | IS_RESUMPTION_EXCEPTION(ThreadCancelled, (T)))
+forall(T & | is_thread(T) | IS_RESUMPTION_EXCEPTION(ThreadCancelled, (T)))
 T & join( T & this ) {
         thread_dtor_guard_t guard = { this, defaultResumptionHandler };

libcfa/src/concurrency/thread.hfa

-              r342af53
+              r8e4aa05
 //-----------------------------------------------------------------------------
 // thread trait
 trait is_thread(dtype T) {
+trait is_thread(T &) {
         void ^?{}(T& mutex this);
         void main(T& this);
 …
 };
 FORALL_DATA_EXCEPTION(ThreadCancelled, (dtype thread_t), (thread_t)) (
+FORALL_DATA_EXCEPTION(ThreadCancelled, (thread_t &), (thread_t)) (
         thread_t * the_thread;
         exception_t * the_exception;
 );
 forall(dtype T)
+forall(T &)
 void copy(ThreadCancelled(T) * dst, ThreadCancelled(T) * src);
 forall(dtype T)
+forall(T &)
 const char * msg(ThreadCancelled(T) *);
 …
 // Inline getters for threads/coroutines/monitors
 forall( dtype T | is_thread(T) )
+forall( T & | is_thread(T) )
 static inline $coroutine* get_coroutine(T & this) __attribute__((const)) { return &get_thread(this)->self_cor; }
 forall( dtype T | is_thread(T) )
+forall( T & | is_thread(T) )
 static inline $monitor  * get_monitor  (T & this) __attribute__((const)) { return &get_thread(this)->self_mon; }
 …
 extern struct cluster * mainCluster;
 forall( dtype T | is_thread(T) )
+forall( T & | is_thread(T) )
 void __thrd_start( T & this, void (*)(T &) );
 …
 };
 forall( dtype T | is_thread(T) | IS_EXCEPTION(ThreadCancelled, (T)) )
+forall( T & | is_thread(T) | IS_EXCEPTION(ThreadCancelled, (T)) )
 void ?{}( thread_dtor_guard_t & this, T & thrd, void(*)(ThreadCancelled(T) &) );
 void ^?{}( thread_dtor_guard_t & this );
 …
 // thread runner
 // Structure that actually start and stop threads
 forall( dtype T | sized(T) | is_thread(T) )
+forall( T & | sized(T) | is_thread(T) )
 struct scoped {
         T handle;
 };
 forall( dtype T | sized(T) | is_thread(T) | { void ?{}(T&); } )
+forall( T & | sized(T) | is_thread(T) | { void ?{}(T&); } )
 void ?{}( scoped(T)& this );
 forall( dtype T, ttype P | sized(T) | is_thread(T) | { void ?{}(T&, P); } )
+forall( T &, P... | sized(T) | is_thread(T) | { void ?{}(T&, P); } )
 void ?{}( scoped(T)& this, P params );
 forall( dtype T | sized(T) | is_thread(T) )
+forall( T & | sized(T) | is_thread(T) )
 void ^?{}( scoped(T)& this );
 …
 void unpark( $thread * this );
 forall( dtype T | is_thread(T) )
+forall( T & | is_thread(T) )
 static inline void unpark( T & this ) { if(!&this) return; unpark( get_thread( this ) );}
 …
 //----------
 // join
 forall( dtype T | is_thread(T) | IS_RESUMPTION_EXCEPTION(ThreadCancelled, (T)) )
+forall( T & | is_thread(T) | IS_RESUMPTION_EXCEPTION(ThreadCancelled, (T)) )
 T & join( T & this );

Context Navigation

Legend:

libcfa/src/concurrency/coroutine.cfa

libcfa/src/concurrency/coroutine.hfa

libcfa/src/concurrency/future.hfa

libcfa/src/concurrency/io.cfa

libcfa/src/concurrency/io/call.cfa.in

libcfa/src/concurrency/io/setup.cfa

libcfa/src/concurrency/io/types.hfa

libcfa/src/concurrency/iofwd.hfa

libcfa/src/concurrency/kernel.cfa

libcfa/src/concurrency/kernel.hfa

libcfa/src/concurrency/kernel/fwd.hfa

libcfa/src/concurrency/kernel/startup.cfa

libcfa/src/concurrency/kernel_private.hfa

libcfa/src/concurrency/locks.cfa

libcfa/src/concurrency/locks.hfa

libcfa/src/concurrency/monitor.cfa

libcfa/src/concurrency/monitor.hfa

libcfa/src/concurrency/mutex.cfa

libcfa/src/concurrency/mutex.hfa

libcfa/src/concurrency/preemption.cfa

libcfa/src/concurrency/ready_queue.cfa

libcfa/src/concurrency/stats.cfa

libcfa/src/concurrency/stats.hfa

libcfa/src/concurrency/thread.cfa

libcfa/src/concurrency/thread.hfa

Download in other formats: