//
// 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         /* See feature_test_macros(7) */

#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 __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) {}

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

#else
	#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 "kernel_private.hfa"
	#include "thread.hfa"

	void ?{}(io_context_params & this) {
		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 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
		volatile     bool stopped; // Whether the poller has finished running
		volatile uint64_t epoch;   // Epoch used for memory reclamation
	} 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.stack   = __create_pthread( &iopoll.thrd, iopoll_loop, 0p );
		iopoll.run     = true;
		iopoll.stopped = false;
		iopoll.epoch   = 0;
	}

	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");

			// increment the epoch to notify any deleters we are starting a new cycle
			__atomic_fetch_add(&iopoll.epoch, 1, __ATOMIC_SEQ_CST);

			// 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) {
				$io_context * io_ctx = ($io_context *)(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->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->efd, &v);

				post( io_ctx->sem );
			}
		}

		__atomic_store_n(&iopoll.stopped, true, __ATOMIC_SEQ_CST);

		__cfadbg_print_safe(io_core, "Kernel : IO poller thread stopping\n" );
		unregister(&id);
		return 0p;
	}

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

	static void __io_uring_setup ( $io_context & this, const io_context_params & params_in );
	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, struct cluster & cl) {
		(this.self){ "IO Poller", cl };
		this.ext_sq.empty = true;
		__io_uring_setup( this, cl.io.params );
		__cfadbg_print_safe(io_core, "Kernel I/O : Created ring for io_context %u (%p)\n", this.fd, &this);

		__epoll_register(this);

		__ioarbiter_register(*cl.io.arbiter, this);

		__thrd_start( this, main );
		__cfadbg_print_safe(io_core, "Kernel I/O : Started poller thread for io_context %u\n", this.fd);
	}

	void ^?{}($io_context & mutex this) {
		__cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %u\n", this.fd);

		^(this.self){};
		__cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %u\n", this.fd);

		__ioarbiter_unregister(*this.arbiter, this);

		__epoll_unregister(this);

		__io_uring_teardown( this );
		__cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %u\n", this.fd);
	}

	void ?{}(io_context & this, struct cluster & cl) {
		// this.ctx = new(cl);
		this.ctx = alloc();
		(*this.ctx){ cl };

		__cfadbg_print_safe(io_core, "Kernel I/O : io_context %u ready\n", this.ctx->fd);
	}

	void ^?{}(io_context & this) {
		post( this.ctx->sem );

		delete(this.ctx);
	}

	extern void __disable_interrupts_hard();
	extern void __enable_interrupts_hard();

	static void __io_uring_setup( $io_context & this, const io_context_params & params_in ) {
		// 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\n");
		}

		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.ready = 0;
		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.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);

		// Step 4 : eventfd
		// io_uring_register is so f*cking slow on some machine that it
		// will never succeed if preemption isn't hard blocked
		__disable_interrupts_hard();

		int efd = eventfd(0, 0);
		if (efd < 0) {
			abort("KERNEL ERROR: IO_URING EVENTFD - %s\n", strerror(errno));
		}

		int ret = syscall( __NR_io_uring_register, fd, IORING_REGISTER_EVENTFD, &efd, 1);
		if (ret < 0) {
			abort("KERNEL ERROR: IO_URING EVENTFD REGISTER - %s\n", strerror(errno));
		}

		__enable_interrupts_hard();

		// 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;
		this.efd        = efd;
	}

	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;

		// 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);
		close(this.efd);

		free( this.sq.free_ring.array ); // Maybe null, doesn't matter
	}

//=============================================================================================
// 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.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