source: libcfa/src/concurrency/io.cfa@ a33c113

//
// Cforall Version 1.0.0 Copyright (C) 2020 University of Waterloo
//
// The contents of this file are covered under the licence agreement in the
// file "LICENCE" distributed with Cforall.
//
// io.cfa --
//
// Author           : Thierry Delisle
// Created On       : Thu Apr 23 17:31:00 2020
// Last Modified By :
// Last Modified On :
// Update Count     :
//

#define __cforall_thread__

#if defined(__CFA_DEBUG__)
        // #define __CFA_DEBUG_PRINT_IO__
        // #define __CFA_DEBUG_PRINT_IO_CORE__
#endif


#if defined(CFA_HAVE_LINUX_IO_URING_H)
        #define _GNU_SOURCE         /* See feature_test_macros(7) */
        #include <errno.h>
        #include <signal.h>
        #include <stdint.h>
        #include <string.h>
        #include <unistd.h>

        extern "C" {
                #include <sys/syscall.h>
                #include <sys/eventfd.h>

                #include <linux/io_uring.h>
        }

        #include "stats.hfa"
        #include "kernel.hfa"
        #include "kernel/fwd.hfa"
        #include "io/types.hfa"

        __attribute__((unused)) static const char * opcodes[] = {
                "OP_NOP",
                "OP_READV",
                "OP_WRITEV",
                "OP_FSYNC",
                "OP_READ_FIXED",
                "OP_WRITE_FIXED",
                "OP_POLL_ADD",
                "OP_POLL_REMOVE",
                "OP_SYNC_FILE_RANGE",
                "OP_SENDMSG",
                "OP_RECVMSG",
                "OP_TIMEOUT",
                "OP_TIMEOUT_REMOVE",
                "OP_ACCEPT",
                "OP_ASYNC_CANCEL",
                "OP_LINK_TIMEOUT",
                "OP_CONNECT",
                "OP_FALLOCATE",
                "OP_OPENAT",
                "OP_CLOSE",
                "OP_FILES_UPDATE",
                "OP_STATX",
                "OP_READ",
                "OP_WRITE",
                "OP_FADVISE",
                "OP_MADVISE",
                "OP_SEND",
                "OP_RECV",
                "OP_OPENAT2",
                "OP_EPOLL_CTL",
                "OP_SPLICE",
                "OP_PROVIDE_BUFFERS",
                "OP_REMOVE_BUFFERS",
                "OP_TEE",
                "INVALID_OP"
        };

        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 );
//=============================================================================================
// 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() );
                /* paranoid */ verify( proc );
                /* paranoid */ verify( proc->io.ctx );

                // 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;

                __u32 count = tail - head;
                __STATS__( false, io.calls.drain++; io.calls.completed += count; )

                for(i; count) {
                        unsigned idx = (head + i) & mask;
                        volatile struct io_uring_cqe & cqe = ctx->cq.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 );

                        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_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;

                __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:
                                // 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;
        }

//=============================================================================================
// I/O Submissions
//=============================================================================================

// Submition steps :
// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
//     listed in sq.array are visible by the kernel. For those not listed, the kernel does not
//     offer any assurance that an entry is not being filled by multiple flags. Therefore, we
//     need to write an allocator that allows allocating concurrently.
//
// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
//
// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
//     needs to arrive to two concensus at the same time:
//     A - The order in which entries are listed in the array: no two threads must pick the
//         same index for their entries
//     B - When can the tail be update for the kernel. EVERY entries in the array between
//         head and tail must be fully filled and shouldn't ever be touched again.
//
        //=============================================================================================
        // 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.
        // The kernel cannot see these entries until they are submitted, but other threads must be
        // able to see which entries can be used and which are already un used by an other thread
        // 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( __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, 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?
                {
                        __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, idxs, have, lazy);
        }

        //=============================================================================================
        // Flushing
        // 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;

                __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

                // the 3 fields are organized like this diagram
                // except it's are ring
                // ---+--------+--------+----
                // ---+--------+--------+----
                //    ^        ^        ^
                // phead    chead    ctail

                // make sure ctail doesn't wrap around and reach phead
                /* paranoid */ verify(
                           (ctail >= chead && chead >= phead)
                        || (chead >= phead && phead >= ctail)
                        || (phead >= ctail && ctail >= chead)
                );

                // find the range we need to clear
                __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);

                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 );
                }
        }
#endif