/*
 * Copyright 2011 Stefan Lankes, Chair for Operating Systems,
 *                               RWTH Aachen University
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *
 * This file is part of MetalSVM. 
 */

#include <metalsvm/stddef.h>
#include <metalsvm/stdio.h>
#include <metalsvm/stdlib.h>
#include <metalsvm/mmu.h>
#include <metalsvm/tasks.h>
#include <metalsvm/page.h>
#include <metalsvm/errno.h>
#include <asm/irqflags.h>
#include <asm/processor.h>
#include <asm/page.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/RCCE.h>
#include <asm/RCCE_lib.h>
#include <asm/iRCCE.h>
#include <asm/SCC_API.h>
#include <asm/icc.h>
#include <asm/svm.h>

#define USE_PERFCOUNTERS	1
#define USE_RESP_MAIL		1

#define SHARED_PAGES		(4*(RCCE_SHM_SIZE_MAX >> PAGE_SHIFT))
#define OWNER_SIZE		((SHARED_PAGES * sizeof(uint8_t) + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1))

#define AIREG1 0
#define AIREG2 (AIREG1 + 1)
#define LOCK_ID 0

#define ABS(a) (((a) < 0) ? -(a) : (a))

t_vcharp RC_SHM_BUFFER_START();

typedef struct {
	int counter;
	int initializer;
} atomic_increg_t;

static volatile atomic_increg_t *incregs = NULL;
static RCCE_FLAG release;

/*
 *  Details on L2 cache (nedded for flushing) 
 */

#define OWN_MPB     0xd8000000

#define L2_LINESIZE 32UL
#define L2_WAYS     4UL
#define L2_CAPACITY (256*1024UL)
#define L2_WBSTRIDE (L2_CAPACITY/L2_WAYS)

#ifdef SVM_WB
/* Helper function to read data into all 4 ways of L2 cache */ 

__attribute__((always_inline)) static inline void svm_purge_set(const size_t set)
{
	register char tmp;

	/* Translate the set to a kernel space virtual address */
	const volatile char*  dummyData = (volatile char*)set;

	/* Now read new data into all four ways, and then reread the first */
	tmp = *dummyData;
	tmp = *(dummyData + L2_WBSTRIDE);
	tmp = *(dummyData + L2_WBSTRIDE * 2);
	tmp = *(dummyData + L2_WBSTRIDE * 3);
}

static size_t dummy_base = OWN_MPB + L2_CAPACITY;
static size_t dummy_offset = 0;
#endif

/*
 * This array describes the owner of a specific page.
 * Only the owner of a page is able to change the possession.
 * => No lock is needded.
 */
static volatile uint8_t*	page_owner = NULL;

// helper array to convert a physical to a virtual address
static size_t phys2virt[SHARED_PAGES] = {[0 ... SHARED_PAGES-1] = 0};
static const size_t shmbegin = SHM_ADDR;
static uint32_t emit[RCCE_MAXNP] = {[0 ... RCCE_MAXNP-1] = 0};
static uint32_t request[RCCE_MAXNP] = {[0 ... RCCE_MAXNP-1] = 0};
static uint32_t forward[RCCE_MAXNP] = {[0 ... RCCE_MAXNP-1] = 0};
static uint32_t alloc_page = 0;
static uint32_t map_page = 0;
#if USE_PERFCOUNTERS
static uint64_t alloc_ticks = 0;
static uint64_t request_ticks = 0;
static uint64_t emit_ticks = 0;
static uint64_t wait_ticks = 0;
static uint64_t max_wait = 0;
static uint64_t min_wait = (uint64_t) -1;
#endif

int svm_init(void)
{
	uint32_t i, flags;

	kprintf("Shared memory starts at the physical address 0x%x\n", shmbegin);

	page_owner = (uint8_t*) map_region(0, shmbegin, OWNER_SIZE >> PAGE_SHIFT, MAP_KERNEL_SPACE|MAP_NO_CACHE);
	if (BUILTIN_EXPECT(!page_owner, 0))
		return -ENOMEM;

	if (!RCCE_IAM) {
		memset((void*)page_owner, 0xFF, OWNER_SIZE);

		// owner vector is owned by core 0
		for(i=0; i<(OWNER_SIZE >> PAGE_SHIFT); i++)
			page_owner[i] = 0;
	}

	// initialize svm barrier
	incregs = (volatile atomic_increg_t*) map_region(0, 0xF900E000, 2, MAP_KERNEL_SPACE|MAP_NO_CACHE);
	if (BUILTIN_EXPECT(!incregs, 0))
		return -ENOMEM;
	kprintf("Map atomic counters at 0x%x\n", incregs);

	if (!RCCE_IAM) {
		incregs[AIREG1].initializer = 0;
		incregs[AIREG2].initializer = 0;
	}

	// iRCCE is not thread save => disable interrupts
	flags = irq_nested_disable();
	RCCE_flag_alloc(&release);
	irq_nested_enable(flags);

	RCCE_barrier(&RCCE_COMM_WORLD);

	return 0;
}

static size_t get_shpages(uint32_t n)
{
	int x = X_PID(RC_MY_COREID);
	int y = Y_PID(RC_MY_COREID);
	size_t i, j = 0, k = 0, start = SHM_X0_Y0;
	int diff, min = x + y;

	diff = ABS(5 - x) + ABS(0 - y);
	if (diff < min) {
		min = diff;
		start = SHM_X5_Y0;
	}

	diff = ABS(0 - x) + ABS(2 - y);
	if (diff < min) {
		min = diff;
		start = SHM_X0_Y2;
	}

	diff = ABS(5 - x) + ABS(2 - y);
	if (diff < min) {
		min = diff;
		start = SHM_X5_Y2;
	}

	for(i=0; (i < SHARED_PAGES) && (k < n); i++) {
		k = 0;
		j = (((start - shmbegin) >> PAGE_SHIFT) + i) % SHARED_PAGES;

		while((k<n) && (i < SHARED_PAGES) && (page_owner[j+k] >= RCCE_MAXNP)) {
			k++; i++;
		}
	}

	if (BUILTIN_EXPECT(i >= SHARED_PAGES, 0))
		return 0;

	memset((void*) (page_owner+j), RCCE_IAM, sizeof(uint8_t)*n);

	return shmbegin + (j << PAGE_SHIFT);
}

size_t shmalloc(uint32_t n)
{
	size_t ret;

	RCCE_acquire_lock(RC_COREID[LOCK_ID]);
	ret = get_shpages(n);
	RCCE_release_lock(RC_COREID[LOCK_ID]);

	return ret;
}

/*
 * This function is called by the pagefault handler
 * => the interrupt flags is already cleared 
 */
int svm_alloc_page(size_t addr, page_table_t* pgt)
{
#if USE_PERFCOUNTERS
	uint64_t start = rdtsc();
#endif
	uint32_t index2 = (addr >> 12) & 0x3FF;
	size_t phyaddr;
	t_vcharp mpb = (t_vcharp) ((size_t)(virt_to_phys(addr) >> PAGE_SHIFT) | ((size_t) RCCE_comm_buffer[RCCE_IAM] - RCCE_LINE_SIZE));
	uint16_t offset = 0xFFFF;

	addr &= PAGE_MASK; // align address to the page boundary
	
	RCCE_acquire_lock(RC_COREID[LOCK_ID]);

	iRCCE_get((t_vcharp) &offset, mpb, sizeof(uint16_t), RCCE_IAM);

	if (!offset) {
		int i;

		phyaddr = get_shpages(1);
		offset = (uint16_t) ((phyaddr - shmbegin) >> PAGE_SHIFT);
		for(i=0; i<RCCE_NP; i++)
			iRCCE_put(mpb, (t_vcharp) &offset, sizeof(uint16_t), i);

		RCCE_release_lock(RC_COREID[LOCK_ID]);

		pgt->entries[index2] &= 0xFFF;
		pgt->entries[index2] &= ~PG_SVM_INIT;
		pgt->entries[index2] |= phyaddr|PG_PRESENT;
		phys2virt[(phyaddr - shmbegin) >> PAGE_SHIFT] = addr;
		tlb_flush_one_page(addr);

		alloc_page++;
		//kprintf("map new page frame 0x%x at 0x%x, flags0x%x, offset 0x%x, mpb 0x%x\n", phyaddr, addr, pgt->entries[index2] & 0xFFF, (int) offset, mpb);

#if USE_PERFCOUNTERS
		alloc_ticks += rdtsc() - start;
#endif
		return 0;
	} else {
		RCCE_release_lock(RC_COREID[LOCK_ID]);

		phyaddr = shmbegin + ((size_t)offset << PAGE_SHIFT);
		pgt->entries[index2] &= 0xFFF;
		pgt->entries[index2] &= ~PG_SVM_INIT;
		if (pgt->entries[index2] & PG_SVM_LAZYRELEASE)
			pgt->entries[index2] |= phyaddr|PG_PRESENT;
		else
			pgt->entries[index2] |= phyaddr;
		phys2virt[(phyaddr - shmbegin) >> PAGE_SHIFT] = addr;
		tlb_flush_one_page(addr);

		map_page++;
		//kprintf("map existing page frame 0x%x at 0x%x, offset 0x%x, mpb 0x%x\n", phyaddr, addr, offset, mpb);

#if USE_PERFCOUNTERS
		alloc_ticks += rdtsc() - start;
#endif
		if (pgt->entries[index2] & PG_SVM_LAZYRELEASE)
			return 0;
		if (pgt->entries[index2] & PG_RW)
			return svm_access_request(addr);

		return 0;
	}
}

/*
 * This function is called by the pagefault handler
 * => the interrupt flags is already cleared 
 */
int svm_access_request(size_t addr)
{
#if USE_PERFCOUNTERS
	uint64_t start = rdtsc();
#endif
	size_t phyaddr = virt_to_phys(addr);
	uint32_t pageid;
	int remote_rank;
	uint8_t payload[iRCCE_MAIL_HEADER_PAYLOAD];
	int ret;

	if (phyaddr < shmbegin)
		return -EINVAL;
	if (phyaddr >= shmbegin + RCCE_SHM_SIZE_MAX)
		return -EINVAL;
	pageid = (phyaddr-shmbegin) >> PAGE_SHIFT;

	remote_rank = page_owner[pageid];
	if (remote_rank == RCCE_IAM)
		return 0;

	((size_t*) payload)[0] = RCCE_IAM;
	((size_t*) payload)[1] = phyaddr;

	//kprintf("send request (0x%x) to %d\n", addr, remote_rank);
	/* send ping request */
	iRCCE_mail_send(2*sizeof(size_t), SVM_REQ, 0, (char*) payload, remote_rank);
	NOP4;
	icc_send_gic_irq(remote_rank);	
	request[remote_rank]++;

#if USE_RESP_MAIL
#if USE_PERFCOUNTERS
	uint64_t wait_start = rdtsc();
#endif
	// wait for response
	icc_wait(SVM_RESP);
#if USE_PERFCOUNTERS
	uint64_t res = rdtsc() - wait_start;
	wait_ticks += res;
	if (min_wait > res)
		min_wait = res;
	if (max_wait < res)
		max_wait = res;
#endif

#else
	NOP8;
	while (page_owner[pageid] != RCCE_IAM) {
		icc_mail_check();
		NOP8;
	}
#endif

	addr &= PAGE_MASK; // align address to page boundary
	ret = change_page_permissions(addr, addr + PAGE_SIZE, VMA_READ|VMA_WRITE|VMA_CACHEABLE);

#if USE_PERFCOUNTERS
	request_ticks += rdtsc() - start;
#endif

	return ret;
}

//static atomic_int32_t size_counter = ATOMIC_INIT(0);

void* svm_malloc(size_t size, uint32_t consistency)
{
	size_t viraddr, phyaddr, i, j;
	t_vcharp mpb_addr;
	uint32_t flags;
	task_t* task = per_core(current_task);
	uint32_t map_flags = MAP_KERNEL_SPACE|MAP_SVM_INIT;
	uint8_t buffer[RCCE_LINE_SIZE]= {[0 ... RCCE_LINE_SIZE-1] = 0};

	if(!(consistency & SVM_L2))
		map_flags |= MAP_MPE;
	else
		task->flags |= TASK_L2;

	if (consistency & SVM_STRONG)
		map_flags |= MAP_SVM_STRONG;
	else if (consistency & SVM_LAZYRELEASE)
		map_flags |= MAP_SVM_LAZYRELEASE;
	else return NULL;

	// currently, we allocate memory in page size granulation
	size = (size + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1);

#if 0 	// Workaround for our MARC paper
	// iRCCE is not thread save => disable interrupts
	flags = irq_nested_disable();

	kprintf("Entering shmmalloc: size 0x%x, owner_size 0x%x\n", size, OWNER_SIZE);
	if (RCCE_IAM && (consistency & SVM_STRONG))
		map_flags |= MAP_NO_ACCESS;

	viraddr = vm_alloc(size >> PAGE_SHIFT, map_flags);
	kprintf("vm_alloc returns 0x%x\n", viraddr);

	static uint32_t last = 0;	
	// get memory on MC0
	if (last)
		phyaddr = last + size/4;
	else 
		last = phyaddr = (size_t) RCCE_shmalloc(size/4);
	map_region(viraddr, phyaddr, (size/4) >> PAGE_SHIFT, map_flags|MAP_REMAP);
	for(i=0; i<size/4; i+=PAGE_SIZE)
		phys2virt[(phyaddr + i - shmbegin) >> PAGE_SHIFT] = viraddr + i;
	kprintf("svmmalloc on MC0: phyaddr 0x%x, viraddr 0x%x, size 0x%x\n", phyaddr, viraddr, size);

	// get memory on MC1
	phyaddr =  shmbegin + 0x1000000 + atomic_int32_read(&size_counter);
	map_region(viraddr + size/4, phyaddr, (size/4) >> PAGE_SHIFT, map_flags|MAP_REMAP);
	for(i=0; i<size/4; i+=PAGE_SIZE)
		phys2virt[(phyaddr + i - shmbegin) >> PAGE_SHIFT] = viraddr + size/4 + i;
	kprintf("svmmalloc on MC1: phyaddr 0x%x, viraddr 0x%x, size 0x%x\n", phyaddr, viraddr+size/4, size);

	// get memory on MC2
	phyaddr =  shmbegin + 0x2000000 + atomic_int32_read(&size_counter);
	map_region(viraddr + 2 * size/4, phyaddr, (size/4) >> PAGE_SHIFT, map_flags|MAP_REMAP);
	for(i=0; i<size/4; i+=PAGE_SIZE)
		phys2virt[(phyaddr + i - shmbegin) >> PAGE_SHIFT] = viraddr + 2 * size/4 + i;
	kprintf("svmmalloc on MC2: phyaddr 0x%x, viraddr 0x%x, size 0x%x\n", phyaddr, viraddr+2*size/4, size);

	// get memory on MC3
	phyaddr =  shmbegin + 0x3000000 + atomic_int32_read(&size_counter);
	map_region(viraddr + 3 * size/4, phyaddr, (size/4) >> PAGE_SHIFT, map_flags|MAP_REMAP);
	for(i=0; i<size/4; i+=PAGE_SIZE)
		phys2virt[(phyaddr + i - shmbegin) >> PAGE_SHIFT] = viraddr + 3 * size/4 + i;
	kprintf("svmmalloc on MC3: phyaddr 0x%x, viraddr 0x%x, size 0x%x\n", phyaddr, viraddr+3*size/4, size);

	atomic_int32_add(&size_counter, size/4);
	irq_nested_enable(flags);

	kprintf("shmmalloc returns 0x%x\n", viraddr);

	return (void*) viraddr;
#endif
#if 0
	// iRCCE is not thread save => disable interrupts
	flags = irq_nested_disable();
	phyaddr = (size_t) RCCE_shmalloc(size);

	if (RCCE_IAM && (consistency & SVM_STRONG))
		map_flags |= MAP_NO_ACCESS;
	irq_nested_enable(flags);

	if (BUILTIN_EXPECT(!phyaddr, 0))
		return NULL;
	if (BUILTIN_EXPECT(phyaddr & 0xFFF, 0)) {
		kprintf("RCCE_shmalloc returns not a page aligned physical address: 0x%x\n", phyaddr);
		return NULL;
	}

	viraddr = map_region(0, phyaddr, size >> PAGE_SHIFT, map_flags);
	for(i=0; i<size; i+=PAGE_SIZE)
		phys2virt[(phyaddr + i - shmbegin) >> PAGE_SHIFT] = viraddr + i;

	kprintf("svmmalloc: phyaddr 0x%x, viraddr 0x%x, size 0x%x\n", phyaddr, viraddr, size);

	return (void*) viraddr;
#endif

	map_flags |= MAP_NO_ACCESS;
#ifndef SVM_WB
	map_flags |= MAP_MPE;
#endif

	viraddr = map_region(0, 0, size >> PAGE_SHIFT, map_flags);
	kprintf("svmmalloc: viraddr 0x%x, size 0x%x, flags 0x%x\n", viraddr, size, map_flags);

	map_flags |= MAP_REMAP;
	for(i=0, j=0, mpb_addr=0; i<size; i+=PAGE_SIZE, j++) {
		if (j % (RCCE_LINE_SIZE/sizeof(uint16_t)) == 0) {
			flags = irq_nested_disable();
			mpb_addr = RCCE_malloc(RCCE_LINE_SIZE);

			if (BUILTIN_EXPECT(!mpb_addr, 0)) {
				irq_nested_enable(flags);
				kputs("RCCE_malloc failed\n");
				goto out;
			}

			//kprintf("mpb_addr 0x%x\n", mpb_addr);
			iRCCE_put(mpb_addr, buffer, RCCE_LINE_SIZE, RCCE_IAM);

			irq_nested_enable(flags);
		}

		phyaddr = (size_t)mpb_addr + (j % (RCCE_LINE_SIZE/sizeof(uint16_t))) * sizeof(uint16_t);
		phyaddr <<= PAGE_SHIFT;
		//kprintf("viraddr 0x%x, phyaddr 0x%x, flags 0x%x\n", viraddr+i, phyaddr, map_flags);
		map_region(viraddr+i, phyaddr, 1, map_flags);
	}

	return (void*) viraddr;

out:
	// TODO: error handling
	return NULL;
}

void svm_free(void* addr, size_t size)
{
	//size_t phyaddr, i;
	//uint32_t flags;

	if (BUILTIN_EXPECT(!addr || !size, 0))
		return;

	// TODO: free is currently not implemented
	return;

#if 0
	phyaddr = virt_to_phys((size_t) addr);

	// currently, we allocate memory in page size granulation
	size = (size + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1);

	kprintf("svmfree: phyaddr 0x%x, viraddr 0x%x, size 0x%x\n", phyaddr, addr, size);

	unmap_region((size_t) addr, size >> PAGE_SHIFT);
	for(i=0; i<size; i+=PAGE_SIZE)
		phys2virt[(phyaddr + i - shmbegin) >> PAGE_SHIFT] = 0;

	// iRCCE is not thread save => disable interrupts
	flags = irq_nested_disable();
	RCCE_shfree((t_vcharp) phyaddr);
	irq_nested_enable(flags);
#endif
}

/*
 * This function is called by icc_mail_check.
 * => Interrupt flag is alread cleared.
 */
int svm_emit_page(size_t phyaddr, int ue)
{
#if USE_PERFCOUNTERS
	uint64_t start = rdtsc();	
#endif
	uint32_t pageid;
	int remote_rank;

	//kprintf("Try to emit page 0x%x to %d\n", phyaddr, ue);

	if (phyaddr < shmbegin)
		return -EINVAL;
	if (phyaddr >= shmbegin + RCCE_SHM_SIZE_MAX)
		return -EINVAL;
	pageid = (phyaddr-shmbegin) >> PAGE_SHIFT;

	remote_rank = page_owner[pageid];	
	if (remote_rank != RCCE_IAM) {
		// Core is nor owner => forward request to new owner
		uint8_t payload[iRCCE_MAIL_HEADER_PAYLOAD];

		kprintf("Ups, core %d is not owner of page 0x%x\n", RCCE_IAM, phyaddr);

		((size_t*) payload)[0] = ue;
		((size_t*) payload)[1] = phyaddr;

		/* send ping request */
		iRCCE_mail_send(2*sizeof(size_t), SVM_REQ, 0, (char*)payload, remote_rank);
		/* send interrupt */
                icc_send_gic_irq(remote_rank);

		forward[remote_rank]++;
	} else {
		size_t viraddr;

		svm_flush(phyaddr);

#if USE_RESP_MAIL
		// send response back to ue
		// ue is polling for the response => no irq is needed
		iRCCE_mail_send(0, SVM_RESP, 0, NULL, ue);
#endif

		emit[ue]++;
		page_owner[pageid] = ue;
		viraddr = phys2virt[pageid];
		change_page_permissions(viraddr, viraddr+PAGE_SIZE, VMA_NOACCESS|VMA_READ|VMA_CACHEABLE);
	}

#if USE_PERFCOUNTERS
	emit_ticks += rdtsc() - start;
#endif

	return 0;
}

#if 0
void svm_flush(void)
{
	int z, tmp;

        // need to write to another line to make sure the write combine buffer gets flushed
        *(int *)RCCE_fool_write_combine_buffer = 1;
        flush_cache();

#error Currently not supported
#if 0
	// try to flush L2 cache
	z = Z_PID(RC_COREID[my_ue]);
	tmp=ReadConfigReg(CRB_OWN + (z==0 ? GLCFG0 : GLCFG1));
	tmp &= ~(1 << GLCFG_XFLSHNN_BIT);
	SetConfigReg(CRB_OWN + (z==0 ? GLCFG0 : GLCFG1), tmp);

	while(!(ReadConfigReg(CRB_OWN + (z==0 ? GLCFG0 : GLCFG1)) & (1 << GLCFG_XFLSHNN_BIT))) {
		NOP8;
	}
#endif
}
#endif

/*
 * Function to flush one page or entire cache.
 */

#ifdef SVM_WB

void svm_invalidate(void)
{
	task_t* task = per_core(current_task);

	if(task->flags & TASK_L2) {
 		asm volatile ( ".byte 0x0f; .byte 0x0a;\n" ); // CL1FLUSHMB
	} else {
		/* no action needed svm_flush already invalidates cache */
		return;
	}

}

void svm_flush(size_t phyaddr)
{
	task_t* task = per_core(current_task);
	page_dir_t* pgd = task->pgd;
	page_table_t* pgt = NULL;

	size_t step = 0;
	size_t stride = L2_LINESIZE;
	size_t range = L2_WBSTRIDE;

	size_t viraddr;

	uint32_t index1, index2;

	uint32_t flags;

	/* flush entire Cache if phyaddr == 0 */
	if(!phyaddr) {
		if( task->flags & TASK_L2 ){
			goto flush_l2;
		} else {
			goto flush_l1;
		}
	/* flush one page */
	} else {
		/* align the address to page boundaries */
		phyaddr &= ~(PAGE_SIZE-1);
		
		/* lookup pgt to check if L2 is enabled */
		viraddr = phys2virt[(phyaddr - shmbegin) >> PAGE_SHIFT];
		index1 = viraddr >> 22;		
		index2 = (viraddr >> 12) & 0x3FF;

		/* check if pgt is present */
		if (!pgd || !(pgd->entries[index1] & PAGE_MASK))
			goto wrong_addr;

		pgt = (page_table_t*)((KERNEL_SPACE - 1024 * PAGE_SIZE + index1 * PAGE_SIZE) & PAGE_MASK);
		
		if( pgt->entries[index2] & PG_MPE ) {
			goto flush_l1;
		} else {
			phyaddr = phyaddr % L2_WBSTRIDE;
			range = PAGE_SIZE;
			goto flush_l2;
		}
	}

	/*
	 *  FLUSH L1 CACHE:
	 */

flush_l1:
	kputs("flush L1\n");
	*(int *)RCCE_fool_write_combine_buffer = 1;
	//__asm__ volatile ( "wbinvd;\n\t" );
	flush_cache();
	return;

flush_l2:
	/*
	 *  FLUSH L2 CACHE:
	 *  disable iterrupts due to pseudo LRU behavior of L2 cache
	 */
	flags = irq_nested_disable();
	/* toggle between dummy areas */
	phyaddr += dummy_base + dummy_offset;
	kprintf("flush-l2: phyaddr 0x%x\n", phyaddr);
	if(dummy_offset)
		dummy_offset = 0;
	else
		dummy_offset = L2_CAPACITY;
	
	flush_cache();
	for( step = 0; step < range; step += stride )
		svm_purge_set( phyaddr + step );
	irq_nested_enable(flags);
	return;

wrong_addr:
	kputs("svm flush error: address not valid!\n");
	return;

}
#endif

int svm_barrier(uint32_t flags)
{
	int i;
	RCCE_COMM *comm = &RCCE_COMM_WORLD;
	static int index = 0;

	if (flags & SVM_LAZYRELEASE) {
		svm_flush(0);
		svm_invalidate();
	}

#if 1
	// Lubachevsky barrier with flags
	index = !index;
	if (incregs[AIREG1].counter > (comm->size - 2)) {
		incregs[AIREG1].initializer = 0;
		while(incregs[AIREG1].initializer);
		for (i = 0; i < comm->size; i++)
			RCCE_flag_write(&release, index, comm->member[i]);
	} else RCCE_wait_until(release, index);
#else
	RCCE_barrier(&RCCE_COMM_WORLD);
#endif

	return 0;
}

//extern uint64_t check_ticks;
//extern uint64_t recv_ticks;

int svm_statistics(void)
{
	uint32_t i;

	kprintf("emit\t:");
	for(i=0; i<RCCE_MAXNP; i++)
		kprintf("\t%u", emit[i]);
	kprintf("\nrequest\t:");
	for(i=0; i<RCCE_MAXNP; i++) 
		kprintf("\t%u", request[i]);
	kprintf("\nforward\t:");
	for(i=0; i<RCCE_MAXNP; i++)
		kprintf("\t%u", forward[i]);
	kputs("\n");
	kprintf("allocate page frame: %u\n", alloc_page);
	kprintf("map page frame: %d\n", map_page);
#if USE_PERFCOUNTERS
	kprintf("alloc ticks: %llu\n", alloc_ticks);
	kprintf("request ticks: %llu\n", request_ticks);
	kprintf("wait ticks: %llu\n", wait_ticks);
	kprintf("emit ticks: %llu\n", emit_ticks);
	kprintf("max wait: %llu\n", max_wait);
	kprintf("min wait: %llu\n", min_wait);
	//kprintf("check_ticks: %llu\n", check_ticks);
	//kprintf("recv_tick: %llu\n", recv_ticks);
#endif

	return 0;
}

#endif