#include <stdint.h>
#include <stdlib.hfa>

#include <unistd.h>								// STDERR_FILENO
#include <inttypes.h>
#include "bits/debug.hfa"
#include "stats.hfa"

#if !defined(__CFA_NO_STATISTICS__)
	void __init_stats( struct __stats_t * stats ) {
		stats->ready.pick.push.attempt  = 0;
		stats->ready.pick.push.success  = 0;
		stats->ready.pick.push.local    = 0;
		stats->ready.pick.push.lsuccess = 0;
		stats->ready.pick.ext.attempt  = 0;
		stats->ready.pick.ext.success  = 0;
		stats->ready.pick.ext.local    = 0;
		stats->ready.pick.ext.lsuccess = 0;
		stats->ready.pick.pop .probe    = 0;
		stats->ready.pick.pop .attempt  = 0;
		stats->ready.pick.pop .success  = 0;
		stats->ready.pick.pop .local    = 0;
		stats->ready.pick.pop .lsuccess = 0;
		stats->ready.threads.migration = 0;
		stats->ready.threads.threads   = 0;
		stats->ready.sleep.halts   = 0;
		stats->ready.sleep.cancels = 0;
		stats->ready.sleep.wakes   = 0;
		stats->ready.sleep.exits   = 0;

		#if defined(CFA_HAVE_LINUX_IO_URING_H)
			stats->io.alloc.fast        = 0;
			stats->io.alloc.slow        = 0;
			stats->io.alloc.fail        = 0;
			stats->io.alloc.revoke      = 0;
			stats->io.alloc.block       = 0;
			stats->io.submit.fast       = 0;
			stats->io.submit.slow       = 0;
			stats->io.flush.external    = 0;
			stats->io.calls.flush       = 0;
			stats->io.calls.submitted   = 0;
			stats->io.calls.drain       = 0;
			stats->io.calls.completed   = 0;
			stats->io.calls.errors.busy = 0;
			stats->io.poller.sleeps     = 0;
		#endif
	}

	void __tally_stats( struct __stats_t * cltr, struct __stats_t * proc ) {
		__atomic_fetch_add( &cltr->ready.pick.push.attempt , proc->ready.pick.push.attempt , __ATOMIC_SEQ_CST ); proc->ready.pick.push.attempt  = 0;
		__atomic_fetch_add( &cltr->ready.pick.push.success , proc->ready.pick.push.success , __ATOMIC_SEQ_CST ); proc->ready.pick.push.success  = 0;
		__atomic_fetch_add( &cltr->ready.pick.push.local   , proc->ready.pick.push.local   , __ATOMIC_SEQ_CST ); proc->ready.pick.push.local    = 0;
		__atomic_fetch_add( &cltr->ready.pick.push.lsuccess, proc->ready.pick.push.lsuccess, __ATOMIC_SEQ_CST ); proc->ready.pick.push.lsuccess = 0;
		__atomic_fetch_add( &cltr->ready.pick.ext.attempt  , proc->ready.pick.ext.attempt  , __ATOMIC_SEQ_CST ); proc->ready.pick.ext.attempt   = 0;
		__atomic_fetch_add( &cltr->ready.pick.ext.success  , proc->ready.pick.ext.success  , __ATOMIC_SEQ_CST ); proc->ready.pick.ext.success   = 0;
		__atomic_fetch_add( &cltr->ready.pick.ext.local    , proc->ready.pick.ext.local    , __ATOMIC_SEQ_CST ); proc->ready.pick.ext.local     = 0;
		__atomic_fetch_add( &cltr->ready.pick.ext.lsuccess , proc->ready.pick.ext.lsuccess , __ATOMIC_SEQ_CST ); proc->ready.pick.ext.lsuccess  = 0;
		__atomic_fetch_add( &cltr->ready.pick.pop .probe   , proc->ready.pick.pop .probe   , __ATOMIC_SEQ_CST ); proc->ready.pick.pop .probe    = 0;
		__atomic_fetch_add( &cltr->ready.pick.pop .attempt , proc->ready.pick.pop .attempt , __ATOMIC_SEQ_CST ); proc->ready.pick.pop .attempt  = 0;
		__atomic_fetch_add( &cltr->ready.pick.pop .success , proc->ready.pick.pop .success , __ATOMIC_SEQ_CST ); proc->ready.pick.pop .success  = 0;
		__atomic_fetch_add( &cltr->ready.pick.pop .local   , proc->ready.pick.pop .local   , __ATOMIC_SEQ_CST ); proc->ready.pick.pop .local    = 0;
		__atomic_fetch_add( &cltr->ready.pick.pop .lsuccess, proc->ready.pick.pop .lsuccess, __ATOMIC_SEQ_CST ); proc->ready.pick.pop .lsuccess = 0;
		__atomic_fetch_add( &cltr->ready.threads.migration , proc->ready.threads.migration , __ATOMIC_SEQ_CST ); proc->ready.threads.migration  = 0;
		__atomic_fetch_add( &cltr->ready.threads.threads   , proc->ready.threads.threads   , __ATOMIC_SEQ_CST ); proc->ready.threads.threads    = 0;
		__atomic_fetch_add( &cltr->ready.sleep.halts       , proc->ready.sleep.halts       , __ATOMIC_SEQ_CST ); proc->ready.sleep.halts        = 0;
		__atomic_fetch_add( &cltr->ready.sleep.cancels     , proc->ready.sleep.cancels     , __ATOMIC_SEQ_CST ); proc->ready.sleep.cancels      = 0;
		__atomic_fetch_add( &cltr->ready.sleep.wakes       , proc->ready.sleep.wakes       , __ATOMIC_SEQ_CST ); proc->ready.sleep.wakes        = 0;
		__atomic_fetch_add( &cltr->ready.sleep.exits       , proc->ready.sleep.exits       , __ATOMIC_SEQ_CST ); proc->ready.sleep.exits        = 0;

		#if defined(CFA_HAVE_LINUX_IO_URING_H)
			__atomic_fetch_add( &cltr->io.alloc.fast       , proc->io.alloc.fast       , __ATOMIC_SEQ_CST ); proc->io.alloc.fast        = 0;
			__atomic_fetch_add( &cltr->io.alloc.slow       , proc->io.alloc.slow       , __ATOMIC_SEQ_CST ); proc->io.alloc.slow        = 0;
			__atomic_fetch_add( &cltr->io.alloc.fail       , proc->io.alloc.fail       , __ATOMIC_SEQ_CST ); proc->io.alloc.fail        = 0;
			__atomic_fetch_add( &cltr->io.alloc.revoke     , proc->io.alloc.revoke     , __ATOMIC_SEQ_CST ); proc->io.alloc.revoke      = 0;
			__atomic_fetch_add( &cltr->io.alloc.block      , proc->io.alloc.block      , __ATOMIC_SEQ_CST ); proc->io.alloc.block       = 0;
			__atomic_fetch_add( &cltr->io.submit.fast      , proc->io.submit.fast      , __ATOMIC_SEQ_CST ); proc->io.submit.fast       = 0;
			__atomic_fetch_add( &cltr->io.submit.slow      , proc->io.submit.slow      , __ATOMIC_SEQ_CST ); proc->io.submit.slow       = 0;
			__atomic_fetch_add( &cltr->io.flush.external   , proc->io.flush.external   , __ATOMIC_SEQ_CST ); proc->io.flush.external    = 0;
			__atomic_fetch_add( &cltr->io.calls.flush      , proc->io.calls.flush      , __ATOMIC_SEQ_CST ); proc->io.calls.flush       = 0;
			__atomic_fetch_add( &cltr->io.calls.submitted  , proc->io.calls.submitted  , __ATOMIC_SEQ_CST ); proc->io.calls.submitted   = 0;
			__atomic_fetch_add( &cltr->io.calls.drain      , proc->io.calls.drain      , __ATOMIC_SEQ_CST ); proc->io.calls.drain       = 0;
			__atomic_fetch_add( &cltr->io.calls.completed  , proc->io.calls.completed  , __ATOMIC_SEQ_CST ); proc->io.calls.completed   = 0;
			__atomic_fetch_add( &cltr->io.calls.errors.busy, proc->io.calls.errors.busy, __ATOMIC_SEQ_CST ); proc->io.calls.errors.busy = 0;
			__atomic_fetch_add( &cltr->io.poller.sleeps    , proc->io.poller.sleeps    , __ATOMIC_SEQ_CST ); proc->io.poller.sleeps     = 0;
		#endif
	}

	void __print_stats( struct __stats_t * stats, int flags, const char * type, const char * name, void * id ) with( *stats ) {

		if( flags & CFA_STATS_READY_Q ) {
			double push_len = ((double)ready.pick.push.attempt) / ready.pick.push.success;
			double ext_len  = ((double)ready.pick.ext .attempt) / ready.pick.ext .success;
			double pop_len  = ((double)ready.pick.pop .attempt) / ready.pick.pop .success;

			double lpush_len = ((double)ready.pick.push.local) / ready.pick.push.lsuccess;
			double lext_len  = ((double)ready.pick.ext .local) / ready.pick.ext .lsuccess;
			double lpop_len  = ((double)ready.pick.pop .local) / ready.pick.pop .lsuccess;

			__cfaabi_bits_print_safe( STDOUT_FILENO,
				"----- %s \"%s\" (%p) - Ready Q Stats -----\n"
				"- total threads  : %'15" PRIu64 "run, %'15" PRIu64 "schd (%'" PRIu64 "ext, %'" PRIu64 "mig, %'" PRId64 " )\n"
				"- push avg probe : %'3.2lf, %'3.2lfl (%'15" PRIu64 " attempts, %'15" PRIu64 " locals)\n"
				"- ext  avg probe : %'3.2lf, %'3.2lfl (%'15" PRIu64 " attempts, %'15" PRIu64 " locals)\n"
				"- pop  avg probe : %'3.2lf, %'3.2lfl (%'15" PRIu64 " attempts, %'15" PRIu64 " locals)\n"
				"- Idle Sleep     : %'15" PRIu64 "h, %'15" PRIu64 "c, %'15" PRIu64 "w, %'15" PRIu64 "e\n"
				"\n"
				, type, name, id
				, ready.pick.pop.success
				, ready.pick.push.success + ready.pick.ext.success
				, ready.pick.ext.success, ready.threads.migration, ready.threads.threads
				, push_len, lpush_len, ready.pick.push.attempt, ready.pick.push.local
				, ext_len , lext_len , ready.pick.ext .attempt, ready.pick.ext .local
				, pop_len , lpop_len , ready.pick.pop .attempt, ready.pick.pop .local
				, ready.sleep.halts, ready.sleep.cancels, ready.sleep.wakes, ready.sleep.exits
			);
		}

		#if defined(CFA_HAVE_LINUX_IO_URING_H)
			if( flags & CFA_STATS_IO ) {
				uint64_t total_allocs = io.alloc.fast + io.alloc.slow;
				double avgfasta = ((double)io.alloc.fast) / total_allocs;

				uint64_t total_submits = io.submit.fast + io.submit.slow;
				double avgfasts = ((double)io.submit.fast) / total_submits;

				double avgsubs = ((double)io.calls.submitted) / io.calls.flush;
				double avgcomp = ((double)io.calls.completed) / io.calls.drain;

				__cfaabi_bits_print_safe( STDOUT_FILENO,
					"----- %s \"%s\" (%p) - I/O Stats -----\n"
					"- total allocations : %'" PRIu64 "f, %'" PRIu64 "s (%'2.2lff) \n"
					"-     failures      : %'" PRIu64 "oom, %'" PRIu64 "rvk, %'" PRIu64 "blk\n"
					"- total submits     : %'" PRIu64 "f, %'" PRIu64 "s (%'2.2lf) \n"
					"- flush external    : %'" PRIu64 "\n"
					"- io_uring_enter    : %'" PRIu64 " (%'" PRIu64 ", %'" PRIu64 " EBUSY)\n"
					"-     submits       : %'" PRIu64 " (%'.2lf) \n"
					"-     completes     : %'" PRIu64 " (%'.2lf) \n"
					"- poller sleeping   : %'" PRIu64 "\n"
					"\n"
					, type,  name, id
					, io.alloc.fast, io.alloc.slow, avgfasta
					, io.alloc.fail, io.alloc.revoke, io.alloc.block
					, io.submit.fast, io.submit.slow, avgfasts
					, io.flush.external
					, io.calls.flush, io.calls.drain, io.calls.errors.busy
					, io.calls.submitted, avgsubs
					, io.calls.completed, avgcomp
					, io.poller.sleeps
				);
			}
		#endif
	}
#endif