source: libcfa/src/concurrency/io/setup.cfa @ 01ba701

//
// 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/setup.cfa --
//
// Author           : Thierry Delisle
// Created On       : Fri Jul 31 16:25:51 2020
// Last Modified By :
// Last Modified On :
// Update Count     :
//

#define __cforall_thread__
#define _GNU_SOURCE

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

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

#if !defined(CFA_HAVE_LINUX_IO_URING_H)
        void ?{}(io_context_params & this) libcfa_public {}

        void  ?{}(io_context$ & this, struct cluster & cl) {}
        void ^?{}(io_context$ & this) {}

        void __cfa_io_start( processor * proc ) {}
        bool __cfa_io_flush( processor * proc ) { return false; }
        bool __cfa_io_drain( processor * proc ) __attribute__((nonnull (1))) { return false; }
        void __cfa_io_idle ( processor * ) __attribute__((nonnull (1))) {}
        void __cfa_io_stop ( processor * proc ) {}

        io_arbiter$ * create(void) { return 0p; }
        void destroy(io_arbiter$ *) {}

#else
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Waddress-of-packed-member"
        #include <errno.h>
        #include <stdint.h>
        #include <string.h>
        #include <signal.h>
        #include <unistd.h>

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

                #include <linux/io_uring.h>
        }

        #include "bitmanip.hfa"
        #include "fstream.hfa"
        #include "kernel/private.hfa"
        #include "limits.hfa"
        #include "thread.hfa"
#pragma GCC diagnostic pop

        void ?{}(io_context_params & this) libcfa_public {
                this.num_entries = 256;
        }

        static void * __io_poller_slow( void * arg );

        // Weirdly, some systems that do support io_uring don't actually define these
        #ifdef __alpha__
                /*
                * alpha is the only exception, all other architectures
                * have common numbers for new system calls.
                */
                #ifndef __NR_io_uring_setup
                        #define __NR_io_uring_setup           535
                #endif
                #ifndef __NR_io_uring_enter
                        #define __NR_io_uring_enter           536
                #endif
                #ifndef __NR_io_uring_register
                        #define __NR_io_uring_register        537
                #endif
        #else /* !__alpha__ */
                #ifndef __NR_io_uring_setup
                        #define __NR_io_uring_setup           425
                #endif
                #ifndef __NR_io_uring_enter
                        #define __NR_io_uring_enter           426
                #endif
                #ifndef __NR_io_uring_register
                        #define __NR_io_uring_register        427
                #endif
        #endif

//=============================================================================================
// I/O Context Constrution/Destruction
//=============================================================================================



        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.queue){};
                __io_uring_setup( this, cl.io.params, proc->idle_wctx.evfd );
                __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);
        }

        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;

                __u32 nentries = params_in.num_entries != 0 ? params_in.num_entries : 256;
                if( !is_pow2(nentries) ) {
                        abort("ERROR: I/O setup 'num_entries' must be a power of 2, was %u\n", nentries);
                }

                int fd = syscall(__NR_io_uring_setup, nentries, &params );
                if(fd < 0) {
                        abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno));
                }

                // Step 2 : mmap result
                struct __sub_ring_t & sq = this.sq;
                struct __cmp_ring_t & cq = this.cq;

                // calculate the right ring size
                sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned)           );
                cq.ring_sz = params.cq_off.cqes  + (params.cq_entries * sizeof(struct io_uring_cqe));

                // Requires features
                #if defined(IORING_FEAT_SINGLE_MMAP)
                        // adjust the size according to the parameters
                        if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
                                cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz);
                        }
                #endif

                // mmap the Submit Queue into existence
                sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING);
                if (sq.ring_ptr == (void*)MAP_FAILED) {
                        abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno));
                }

                // Requires features
                #if defined(IORING_FEAT_SINGLE_MMAP)
                        // mmap the Completion Queue into existence (may or may not be needed)
                        if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) {
                                cq.ring_ptr = sq.ring_ptr;
                        }
                        else
                #endif
                {
                        // We need multiple call to MMAP
                        cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING);
                        if (cq.ring_ptr == (void*)MAP_FAILED) {
                                munmap(sq.ring_ptr, sq.ring_sz);
                                abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno));
                        }
                }

                // mmap the submit queue entries
                size_t size = params.sq_entries * sizeof(struct io_uring_sqe);
                sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES);
                if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) {
                        munmap(sq.ring_ptr, sq.ring_sz);
                        if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz);
                        abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno));
                }

                // Step 3 : Initialize the data structure
                // Get the pointers from the kernel to fill the structure
                // submit queue
                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
                cq.lock      = false;
                cq.id        = MAX;
                cq.ts        = rdtscl();
                cq.head      = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.head);
                cq.tail      = (volatile __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail);
                cq.mask      = (   const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask);
                cq.num       = (   const __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries);
                cq.overflow  = (         __u32 *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow);
                cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes);

                #if !defined(CFA_WITH_IO_URING_IDLE)
                {
                        // Step 4 : eventfd
                        __cfadbg_print_safe(io_core, "Kernel I/O : registering %d for completion with ring %d\n", procfd, fd);

                        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));
                        }

                        __cfadbg_print_safe(io_core, "Kernel I/O : registered %d for completion with ring %d\n", procfd, fd);
                }
                #endif

                // TODO: implement a proper version of this.
                // I have not found a better maximum that works in general but users should be able to configure it
                // the same way they configure other I/O options
                // #if defined(CFA_HAVE_IORING_REGISTER_IOWQ_MAX_WORKERS)
                // {
                //      // Step 5 : max worker count
                //      __cfadbg_print_safe(io_core, "Kernel I/O : lmiting max workers for ring %d\n", fd);

                //      unsigned int maxes[2];
                //      maxes[0] = 64; // max number of bounded workers (Regular files / block)
                //      maxes[1] = 64;  // max number of unbounded workers (IOSQE_ASYNC)
                //      int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_IOWQ_MAX_WORKERS, maxes, 2);
                //      if (ret < 0) {
                //              abort("KERNEL ERROR: IO_URING MAX WORKER REGISTER - %s\n", strerror(errno));
                //      }

                //      __cfadbg_print_safe(io_core, "Kernel I/O : lmited max workers for ring %d\n", fd);
                // }
                // #endif

                // some paranoid checks
                /* paranoid */ verifyf( (*cq.mask) == ((*cq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (*cq.num) - 1ul32, *cq.num, *cq.mask  );
                /* paranoid */ verifyf( (*cq.num)  >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num );
                /* paranoid */ verifyf( (*cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head );
                /* paranoid */ verifyf( (*cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail );

                /* paranoid */ verifyf( (*sq.mask) == ((*sq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (*sq.num) - 1ul32, *sq.num, *sq.mask );
                /* paranoid */ verifyf( (*sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num );
                /* paranoid */ verifyf( (*sq.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 = 0;
                this.fd         = fd;
        }

        static void __io_uring_teardown( io_context$ & this ) {
                // Shutdown the io rings
                struct __sub_ring_t & sq = this.sq;
                struct __cmp_ring_t & cq = this.cq;
                {
                        __u32 fhead = sq.free_ring.head;
                        __u32 ftail = sq.free_ring.tail;

                        __u32 total = *sq.num;
                        __u32 avail = ftail - fhead;

                        if(avail != total) abort | "Processor (" | (void*)this.proc | ") tearing down ring with" | (total - avail) | "entries allocated but not submitted, out of" | total;
                }

                // unmap the submit queue entries
                munmap(sq.sqes, (*sq.num) * sizeof(struct io_uring_sqe));

                // unmap the Submit Queue ring
                munmap(sq.ring_ptr, sq.ring_sz);

                // unmap the Completion Queue ring, if it is different
                if (cq.ring_ptr != sq.ring_ptr) {
                        munmap(cq.ring_ptr, cq.ring_sz);
                }

                // close the file descriptor
                close(this.fd);

                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
//=============================================================================================
        // 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 ?{}( io_arbiter$ & this ) {
                this.pending.empty = true;
        }

        void ^?{}( io_arbiter$ & mutex this ) {}

        io_arbiter$ * create(void) {
                return new();
        }
        void destroy(io_arbiter$ * arbiter) {
                delete(arbiter);
        }

//=============================================================================================
// I/O Context Misc Setup
//=============================================================================================

#endif