Changes in libcfa/src/concurrency/io.cfa [e9c0b4c:2fab24e3]

File:

• libcfa/src/concurrency/io.cfa (modified) (6 diffs)

libcfa/src/concurrency/io.cfa

@@ -32 +32 @@
         extern "C" {
                 #include <sys/syscall.h>
-                #include <sys/eventfd.h>
 
                 #include <linux/io_uring.h>
@@ -40 +39 @@
         #include "kernel.hfa"
         #include "kernel/fwd.hfa"
-        #include "kernel_private.hfa"
         #include "io/types.hfa"
 
@@ -81 +79 @@
         };
 
-        static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want );
-        static void __ioarbiter_submit( $io_context * , __u32 idxs[], __u32 have, bool lazy );
-        static void __ioarbiter_flush ( $io_context & );
-        static inline void __ioarbiter_notify( $io_context & ctx );
+        // 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 ) {
+                /* 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);
+                        __cfadbg_print_safe(io_core, "Kernel I/O : IO_URING %d returned %d\n", ring.fd, ret);
+
+                        if( ret < 0 ) {
+                                switch((int)errno) {
+                                case EAGAIN:
+                                case EINTR:
+                                case EBUSY:
+                                        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 inline unsigned __flush( struct $io_context & );
-        static inline __u32 __release_sqes( struct $io_context & );
-        extern void __kernel_unpark( $thread * thrd );
-
-        bool __cfa_io_drain( processor * proc ) {
+        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() );
-                /* paranoid */ verify( ready_schedule_islocked() );
-                /* paranoid */ verify( proc );
-                /* paranoid */ verify( proc->io.ctx );
+
+                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 );
 
                 // Drain the queue
-                $io_context * ctx = proc->io.ctx;
-                unsigned head = *ctx->cq.head;
-                unsigned tail = *ctx->cq.tail;
-                const __u32 mask = *ctx->cq.mask;
+                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];
+                }
 
                 __u32 count = tail - head;
-                __STATS__( false, io.calls.drain++; io.calls.completed += count; )
-
-                if(count == 0) return false;
-
+                /* paranoid */ verify( count != 0 );
                 for(i; count) {
                         unsigned idx = (head + i) & mask;
-                        volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
+                        volatile struct io_uring_cqe & cqe = ring.completion_q.cqes[idx];
 
                         /* paranoid */ verify(&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 );
-
-                        __kernel_unpark( fulfil( *future, cqe.res, false ) );
-                }
-
-                __cfadbg_print_safe(io, "Kernel I/O : %u completed\n", count);
+                        process( cqe );
+                }
 
                 // 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_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
-
-                /* paranoid */ verify( ready_schedule_islocked() );
-                /* paranoid */ verify( ! __preemption_enabled() );
-
-                return true;
-        }
-
-        void __cfa_io_flush( processor * proc ) {
-                /* paranoid */ verify( ! __preemption_enabled() );
-                /* paranoid */ verify( proc );
-                /* paranoid */ verify( proc->io.ctx );
-
-                $io_context & ctx = *proc->io.ctx;
-
-                __ioarbiter_flush( 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:
+                __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--;
+
                                 // Update statistics
-                                __STATS__( false, io.calls.errors.busy ++; )
-                                return;
-                        default:
-                                abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
-                        }
-                }
-
-                __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;
+                                __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;
+                        }
+
+                        // 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);
+
+                __ioctx_unregister( this );
         }
 
@@ -193 +344 @@
 //         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.
@@ -233 +350 @@
         // for convenience, return both the index and the pointer to the sqe
         // sqe == &sqes[idx]
-        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();
-
-                        __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();
-
-                $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, 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?
+        [* 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++;
+
+                        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;
+
                 {
-                        __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();
-
-                        __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();
-
-                __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
-
-                __ioarbiter_submit(inctx, idxs, have, lazy);
-        }
-
-        //=============================================================================================
-        // Flushing
+                        __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,
+                                (void*)sqe->off,
+                                (void*)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 ) {
+                        __attribute__((unused)) __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,
+                                        (void*)sqe->off,
+                                        (void*)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;
+        }
+
         // Go through the ring's submit queue and release everything that has already been consumed
         // by io_uring
-        // This cannot be done by multiple threads
-        static __u32 __release_sqes( struct $io_context & ctx ) {
-                const __u32 mask = *ctx.sq.mask;
-
+        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;
                 __attribute__((unused))
-                __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
+                __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);
 
                 // the 3 fields are organized like this diagram
@@ -365 +734 @@
                 __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 ) {
-                        __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);
-
+                        __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 ] );
+                }
                 return count;
         }
 
-//=============================================================================================
-// I/O Arbiter
-//=============================================================================================
-        static inline void block(__outstanding_io_queue & queue, __outstanding_io & item) {
-                // Lock the list, it's not thread safe
-                lock( queue.lock __cfaabi_dbg_ctx2 );
-                {
-                        // Add our request to the list
-                        add( queue.queue, item );
-
-                        // Mark as pending
-                        __atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
-                }
-                unlock( queue.lock );
-
-                wait( item.sem );
-        }
-
-        static inline bool empty(__outstanding_io_queue & queue ) {
-                return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
-        }
-
-        static $io_context * __ioarbiter_allocate( $io_arbiter & this, __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
-                // We need to add ourself to a list of pending allocs and wait for an answer
-                __pending_alloc pa;
-                pa.idxs = idxs;
-                pa.want = want;
-
-                block(this.pending, (__outstanding_io&)pa);
-
-                return pa.ctx;
-
-        }
-
-        static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
-                /* paranoid */ verify( !empty(this.pending.queue) );
-
-                lock( this.pending.lock __cfaabi_dbg_ctx2 );
-                {
-                        while( !empty(this.pending.queue) ) {
-                                __cfadbg_print_safe(io, "Kernel I/O : notifying\n");
-                                __u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
-                                __pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
-
-                                if( have > pa.want ) goto DONE;
-                                drop( this.pending.queue );
-
-                                /* paranoid */__attribute__((unused)) bool ret =
-
-                                __alloc(ctx, pa.idxs, pa.want);
-
-                                /* paranoid */ verify( ret );
-
-                                pa.ctx = ctx;
-
-                                post( pa.sem );
-                        }
-
-                        this.pending.empty = true;
-                        DONE:;
-                }
-                unlock( this.pending.lock );
-        }
-
-        static void __ioarbiter_notify( $io_context & ctx ) {
-                if(!empty( ctx.arbiter->pending )) {
-                        __ioarbiter_notify( *ctx.arbiter, &ctx );
-                }
-        }
-
-        // Simply append to the pending
-        static void __ioarbiter_submit( $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);
-
-                __cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
-
-                __external_io ei;
-                ei.idxs = idxs;
-                ei.have = have;
-                ei.lazy = lazy;
-
-                block(ctx->ext_sq, (__outstanding_io&)ei);
-
-                __cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
-        }
-
-        static void __ioarbiter_flush( $io_context & ctx ) {
-                if(!empty( ctx.ext_sq )) {
-                        __STATS__( false, io.flush.external += 1; )
-
-                        __cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
-
-                        lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
-                        {
-                                while( !empty(ctx.ext_sq.queue) ) {
-                                        __external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
-
-                                        __submit(&ctx, ei.idxs, ei.have, ei.lazy);
-
-                                        post( ei.sem );
-                                }
-
-                                ctx.ext_sq.empty = true;
-                        }
-                        unlock(ctx.ext_sq.lock );
-                }
+        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);
         }
 #endif