#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/init.h>
#include <linux/ptrace.h>
#include <linux/slab.h>
#include <linux/vmalloc.h>
#include <linux/mm.h>
#include <linux/interrupt.h>
#ifdef CONFIG_MIPS
#include <asm/mipsregs.h>
#include <asm/mipsmtregs.h>
#include <asm/bitops.h>
#endif /*--- #ifdef CONFIG_MIPS ---*/
#include <asm/mach_avm.h>

#ifdef CONFIG_LANTIQ
#include <ifx_gptu.h>
#endif
/*------------------------------------------------------------------------------------------*\
\*------------------------------------------------------------------------------------------*/
#ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD
#include <asm/yield_context.h>
#ifndef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT
int yield_thread_freq = 20 * 1000; /* --- freq in Hz -> this is 20 kHz  --- */
int yield_freq_randomizer_percent = 25;
module_param(yield_thread_freq, int, S_IRUSR | S_IWUSR);
module_param(yield_freq_randomizer_percent , int, S_IRUSR | S_IWUSR);
#endif
#endif
#if defined(CONFIG_AVM_SIMPLE_PROFILING) 
#include <linux/avm_profile.h>
#include <linux/skbuff.h>

unsigned int profile_BlockNeeded = (CONFIG_AVM_PROFILING_TRACE_MODE * 10);

unsigned int profile_current_mask = 0;
unsigned int profile_current_start_handle = 0;
extern unsigned long gCycle_per_usec;

struct _simple_profiling simple_profiling;
EXPORT_SYMBOL(simple_profiling);
/*--------------------------------------------------------------------------------*\
 * time und pos-Reservierung muessen atomar passieren!
\*--------------------------------------------------------------------------------*/
static unsigned int lavm_simple_profiling_incpos(unsigned int *time) {
    unsigned int pos = (unsigned int)-1;
    unsigned long flags;
    
    if(simple_profiling.enabled == 1) {
        spin_lock_irqsave(&simple_profiling.lock, flags);
        *time = avm_profile_counter();
         pos  = simple_profiling.pos++;
        if(pos >= simple_profiling.len - 1) {
            simple_profiling.enabled  = 0;
            simple_profiling.pos      = simple_profiling.len - 1;
            spin_unlock_irqrestore(&simple_profiling.lock, flags);
            simple_profiling.end_time = jiffies | 1;
            printk(KERN_INFO"[simple_profiling] buffer full: %u entries recorded\n", simple_profiling.len);
            return (unsigned int)-1;
        } else {
            spin_unlock_irqrestore(&simple_profiling.lock, flags);
        }
    } else if(simple_profiling.enabled == 2) {
        spin_lock_irqsave(&simple_profiling.lock, flags);
        *time = avm_profile_counter();
         pos  = simple_profiling.pos++;
         if(pos >= simple_profiling.len) {        
            if(simple_profiling.end_time == 0) simple_profiling.end_time = jiffies | 1;
            simple_profiling.wraparround = 1;
            pos                          = 0;
            simple_profiling.pos         = 0;
            spin_unlock_irqrestore(&simple_profiling.lock, flags);
            printk(KERN_INFO"[simple_profiling] wraparround: %u entries recorded\n", simple_profiling.len);
        } else {
            spin_unlock_irqrestore(&simple_profiling.lock, flags);
        }
    }
    return pos;
}
/*------------------------------------------------------------------------------------------*\
 * addr =  interrupted epc 
\*------------------------------------------------------------------------------------------*/
unsigned int __avm_simple_profiling_irqcontext(unsigned int epc) {
    unsigned int pos, time;
    struct _avm_profile_data *data;
    if(!(simple_profiling.mask & (1 << avm_profile_data_type_code_address_info))) return 0;

    pos  = lavm_simple_profiling_incpos(&time);

    if(pos == (unsigned int)-1) return 0;

    data = &((struct _avm_profile_data *)(simple_profiling.data[pos / profile_DataSetsPerBlock]))[pos % profile_DataSetsPerBlock];
    data->id             = task_pid_nr(current) & 0xFFFF;
    data->type           = avm_profile_data_type_code_address_info;
    data->cpu_id         = smp_processor_id();
    data->addr           = epc;
    data->time           = time;
#if defined(AVM_PROFILE_ACTIVITY_INCLUDED)
    data->total_access   = avm_profile_sdramacess();
    data->total_activate = avm_profile_sdramactivate();
#endif/*--- #if defined(AVM_PROFILE_ACTIVITY_INCLUDED) ---*/
    return time;
}
/*--- EXPORT_SYMBOL(__avm_simple_profiling_irqcontext); ---*/

/*------------------------------------------------------------------------------------------*\
\*------------------------------------------------------------------------------------------*/

#ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD
unsigned int avm_simple_profiling_yield(unsigned int epc, unsigned int cpu_id, unsigned int irq_num) {
    unsigned int pos, time;
    struct _avm_profile_data *data;
    if(simple_profiling.enabled == 0) return 0;
    if(!(simple_profiling.mask & (1 << avm_profile_data_type_code_address_yield))) return 0;

    pos  = lavm_simple_profiling_incpos(&time);

    if(pos == (unsigned int)-1) return 0;

    data = &((struct _avm_profile_data *)(simple_profiling.data[pos / profile_DataSetsPerBlock]))[pos % profile_DataSetsPerBlock];
    /*--- data->curr_pid = task_pid_nr(current); ---*/
    data->id     = task_pid_nr(current) & 0XFFFF;
    data->cpu_id = cpu_id;
    data->type   = avm_profile_data_type_code_address_info;
    data->addr   = epc;
    data->time   = time;
#if defined(AVM_PROFILE_ACTIVITY_INCLUDED)
#ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT
    data->total_access   = read_c0_perfcnt(0,2);
    data->total_activate = read_c0_perfcnt(1,2);
#else
    data->total_access   = avm_profile_sdramacess();
    data->total_activate = avm_profile_sdramactivate();
#endif /*--- #ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT ---*/
#endif/*--- #if defined(AVM_PROFILE_ACTIVITY_INCLUDED) ---*/
    return time;
}
#endif
/*------------------------------------------------------------------------------------------*\
\*------------------------------------------------------------------------------------------*/
void __avm_simple_profiling_skb(unsigned int addr, unsigned int where, struct sk_buff *skb) {
    unsigned int pos, time;
    struct _avm_profile_data *data;
    if(!(simple_profiling.mask & (1 << avm_profile_data_type_trace_skb))) return;

    pos  = lavm_simple_profiling_incpos(&time);
    if(pos == (unsigned int)-1) return;

    data = (struct _avm_profile_data *)(simple_profiling.data[pos / profile_DataSetsPerBlock]) + (pos % profile_DataSetsPerBlock);
    /*--- data->curr_pid = task_pid_nr(current); ---*/

    data->id     = skb_cloned(skb) & 0xFFFF;
    data->cpu_id = smp_processor_id();
    data->type   = avm_profile_data_type_trace_skb;
    data->addr   = addr;
    data->time   = time;
#if defined(AVM_PROFILE_ACTIVITY_INCLUDED)
    data->total_access   = skb->uniq_id & 0xFFFFFF;
    data->total_activate = where;
#endif/*--- #if defined(AVM_PROFILE_ACTIVITY_INCLUDED) ---*/
}
EXPORT_SYMBOL(__avm_simple_profiling_skb);

/*------------------------------------------------------------------------------------------*\
\*------------------------------------------------------------------------------------------*/
void avm_simple_profiling_irq_onoff(unsigned int switch_on __attribute__((unused)), unsigned int addr __attribute__((unused)), unsigned int id __attribute__((unused))) {
#if defined(CONFIG_AVM_SIMPLE_PROFILING_IRQ_ON_OFF)
    if(!(simple_profiling.mask & ((1 << avm_profile_data_type_irq_on) | (1 << avm_profile_data_type_irq_off))))
            return;

    avm_simple_profiling_log(switch_on ? avm_profile_data_type_irq_on : avm_profile_data_type_irq_off, addr, id);
#endif /*--- #if defined(CONFIG_AVM_SIMPLE_PROFILING_IRQ_ON_OFF) ---*/
}
EXPORT_SYMBOL(avm_simple_profiling_irq_onoff);


/*------------------------------------------------------------------------------------------*\
\*------------------------------------------------------------------------------------------*/
void __avm_simple_profiling_log(enum _avm_profile_data_type type, unsigned int addr, unsigned int id) {
    unsigned int pos, time;
    struct _avm_profile_data *data;
    /*--- if(simple_profiling.enabled == 0) return; ---*/
    if(!(simple_profiling.mask & (1 << type))) return;

    pos  = lavm_simple_profiling_incpos(&time);
    if(pos == (unsigned int)-1) return;

    data = (struct _avm_profile_data *)(simple_profiling.data[pos / profile_DataSetsPerBlock]) + (pos % profile_DataSetsPerBlock);
    /*--- data->curr_pid = task_pid_nr(current); ---*/

    switch(type) {
        default:
        case avm_profile_data_type_unknown:
            data->type = type;
            break;
        case avm_profile_data_type_code_address_info:
            if(id != AVM_PROFILE_IDLE_ID) {
                /*--- if no reserved id used - remark with pid ---*/
                id = task_pid_nr(current);
            }
            /*--- kein break; ---*/
        case avm_profile_data_type_hw_irq_begin:
        case avm_profile_data_type_hw_irq_end:
        case avm_profile_data_type_sw_irq_begin:
        case avm_profile_data_type_sw_irq_end:
        case avm_profile_data_type_timer_begin:
        case avm_profile_data_type_timer_end:
        case avm_profile_data_type_tasklet_begin:
        case avm_profile_data_type_tasklet_end:
        case avm_profile_data_type_hi_tasklet_begin:
        case avm_profile_data_type_hi_tasklet_end:
        case avm_profile_data_type_workitem_begin:
        case avm_profile_data_type_workitem_end:
        case avm_profile_data_type_func_begin:
        case avm_profile_data_type_func_end:
        case avm_profile_data_type_trigger_tasklet_begin:
        case avm_profile_data_type_trigger_tasklet_end:
        case avm_profile_data_type_trigger_user_begin:
        case avm_profile_data_type_trigger_user_end:
            data->id             = id & 0xFFFF;
            data->cpu_id         = smp_processor_id();
            data->type           = type;
            data->addr           = addr;
            data->time           = time;
#if defined(AVM_PROFILE_ACTIVITY_INCLUDED)
            data->total_access   = avm_profile_sdramacess();
            data->total_activate = avm_profile_sdramactivate();
#endif/*--- #if defined(AVM_PROFILE_ACTIVITY_INCLUDED) ---*/
            break;
    }
}
EXPORT_SYMBOL(__avm_simple_profiling_log);

/*------------------------------------------------------------------------------------------*\
\*------------------------------------------------------------------------------------------*/
#ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD
static int avm_simple_profiling_yield_handler( int signal, void *ref ){

    unsigned int irq = 0;
	unsigned long vpflags;
	unsigned int ntc = 2; //currently we are just interested in our SMP VPE contexts
	unsigned int tcs_to_profile_mask = ( 1 << 1 ) | ( 1 << 0 );  //default - will be replaced in YIELD_PCNT-Mode
	int tc;

#ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT
    tcs_to_profile_mask = ack_irqs_reload_counter();
#else
    ifx_gptu_timer_yield_ack( simple_profiling.yield_timer_no, yield_freq_randomizer_percent );
#endif

	vpflags = dvpe();

	for (tc = 0; tc < ntc; tc++) {
	    if ( tcs_to_profile_mask & (1 << tc) ){
            unsigned int tcrestart_val;
            unsigned int cpu_id = tc;

            settc(tc);
            tcrestart_val = read_tc_c0_tcrestart();
            avm_simple_profiling_yield( tcrestart_val, cpu_id, irq );
	    }
	}

	evpe(vpflags);

    return YIELD_HANDLED ;
}
/*--------------------------------------------------------------------------------*\
\*--------------------------------------------------------------------------------*/
static void avm_simple_profiling_yield_enable(unsigned int on) {
    if(on) {
        simple_profiling.yield_handler = NO_YIELD_HANDLER;
        simple_profiling.yield_ref = &simple_profiling;
#ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT
        simple_profiling.yield_signal = YIELD_SIGNAL_PERFORMANE_COUNTER;
#else
        simple_profiling.yield_signal = YIELD_SIGNAL_TIMER1;
#endif
        printk(KERN_ERR "[%s] start yield profiling callback=%p ref_addr=%#x \n", __FUNCTION__, avm_simple_profiling_yield_handler, simple_profiling.yield_ref );
        simple_profiling.yield_handler = request_yield_handler( simple_profiling.yield_signal, avm_simple_profiling_yield_handler, simple_profiling.yield_ref );

#ifndef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT
            if ( simple_profiling.yield_handler  >= 0 ) {

                printk(KERN_ERR "[%s] request_yield_handler successful  req_res =%d\n", __FUNCTION__,  simple_profiling.yield_handler);

                /* --- freq in 1/1000 Hz  --- */
                simple_profiling.yield_timer_no = ifx_gptu_timer_set( 0, yield_thread_freq * 1000, 1, 0, TIMER_FLAG_YIELD_MODE , 0, 0 );

                switch( simple_profiling.yield_timer_no ) {
                    case TIMER1A: YIELDEN_GPCT_1A(1); break;
                    case TIMER1B: YIELDEN_GPCT_1B(1); break;
                    case TIMER2A: YIELDEN_GPCT_2A(1); break;
                    case TIMER2B: YIELDEN_GPCT_2B(1); break;
                    case TIMER3A: YIELDEN_GPCT_3A(1); break;
                    case TIMER3B: YIELDEN_GPCT_3B(1); break;
                    default:
                        printk(KERN_ERR "[%s] could not reserve valid HW Timer: %d\n", __FUNCTION__, simple_profiling.yield_timer_no);
                        ifx_gptu_timer_free( simple_profiling.yield_timer_no );
                        simple_profiling.yield_handler = NO_YIELD_HANDLER;
                        return;
                }
                ifx_gptu_timer_start( simple_profiling.yield_timer_no, 0 );
                printk(KERN_ERR "[%s] request_yield_handler successful timer_no= %d\n", __FUNCTION__, simple_profiling.yield_timer_no);
            }
#endif/*--- #ifndef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT ---*/
            return;
    }
#ifndef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT
    if ( simple_profiling.yield_handler >= 0 ){
        ifx_gptu_timer_stop( simple_profiling.yield_timer_no );
        ifx_gptu_timer_free( simple_profiling.yield_timer_no );
        free_yield_handler( simple_profiling.yield_signal, simple_profiling.yield_ref );

        switch( simple_profiling.yield_timer_no ) {
            case TIMER1A: YIELDEN_GPCT_1A(0); break;
            case TIMER1B: YIELDEN_GPCT_1B(0); break;
            case TIMER2A: YIELDEN_GPCT_2A(0); break;
            case TIMER2B: YIELDEN_GPCT_2B(0); break;
            case TIMER3A: YIELDEN_GPCT_3A(0); break;
            case TIMER3B: YIELDEN_GPCT_3B(0); break;
        }
#endif /*--- #ifndef CONFIG_AVM_SIMPLE_PROFILING_YIELD_PCNT ---*/
        simple_profiling.yield_handler = NO_YIELD_HANDLER;
    }
}
#endif/*--- #ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD ---*/
/*--------------------------------------------------------------------------------*\
\*--------------------------------------------------------------------------------*/
static int alloc_profiler_memory(void) {
    unsigned int i;
    simple_profiling.data = (void **)kmalloc(sizeof(void *) * profile_BlockNeeded, GFP_ATOMIC);
    if(simple_profiling.data == NULL) {
        return -ENOMEM;
    }
    memset(simple_profiling.data, 0, sizeof(void *) * profile_BlockNeeded);
    simple_profiling.len = profile_DataSetsPerBlock * profile_BlockNeeded;
    for(i = 0 ; i < profile_BlockNeeded ; i++) {
        struct page *ptr = alloc_pages(__GFP_NORETRY, 0);
        if(ptr == NULL) {
            simple_profiling.len = profile_DataSetsPerBlock * i;
            printk(KERN_INFO"[%s] resize %d pages instead %d pages\n", __func__, i, profile_BlockNeeded);
            profile_BlockNeeded = i;
            break;
        }
        simple_profiling.data[i] = (void *)lowmem_page_address(ptr);
        /*--- printk("\t%d: %x ==> %x\n", i, ptr, simple_profiling.data[i]); ---*/
    }
    printk(KERN_INFO"[%s] need %d pages for %d entries Buffer %d samples per block\n", __func__, i, simple_profiling.len, profile_DataSetsPerBlock);
    return 0;
}
/*--------------------------------------------------------------------------------*\
\*--------------------------------------------------------------------------------*/
static void free_profiler_memory(void) {
    unsigned int i;
    if(simple_profiling.data == NULL) {
        return;
    }
    for(i = 0 ; i < profile_BlockNeeded ; i++) {
        free_pages((unsigned int)simple_profiling.data[i], 0);
    }
    kfree(simple_profiling.data);
    simple_profiling.data = NULL;
}
/*--------------------------------------------------------------------------------*\
\*--------------------------------------------------------------------------------*/
void avm_simple_profiling_memresize(unsigned int BlockNeeded) {
    if(profile_BlockNeeded != BlockNeeded) {
        avm_simple_profiling_enable(0, 0, NULL, NULL, 1);
        free_profiler_memory();
        atomic_set(&simple_profiling.busy, 0);
        profile_BlockNeeded = BlockNeeded;
    }
}

/*--------------------------------------------------------------------------------*\
\*--------------------------------------------------------------------------------*/
static inline __u32 CLK_TO_MSEC(__u64 cycles) {
    do_div(cycles, gCycle_per_usec * 1000UL); 
    return (__u32)cycles;
}
/*--------------------------------------------------------------------------------*\
\*--------------------------------------------------------------------------------*/
unsigned int timediff_msec(unsigned int count) {
    unsigned long i, msec = 0, last_time = 0;
    unsigned long long cycle_time = 0; 

    for(i = 0 ; i < count ; i += 100 /*--- not every entry ---*/ ) {
        struct _avm_profile_data *pdata = avm_simple_profiling_by_idx(i);
        if(pdata) {
            if(likely(last_time)) {
                unsigned long diff_time   = (pdata->time - last_time);
                cycle_time += (__u64)diff_time;
            }
            last_time = pdata->time;
        }
    }
    if(gCycle_per_usec) {
        msec = CLK_TO_MSEC(cycle_time);
    }
    return msec;
}
/*------------------------------------------------------------------------------------------*\
 * on = 1: normal mode
 * on = 2: wraparround
 * setbusy: busy-Flag setzen
 * timediff in msec
\*------------------------------------------------------------------------------------------*/
int avm_simple_profiling_enable(unsigned int on, unsigned int mask, unsigned int *count, unsigned long *timediff, unsigned int set_busy) {
    unsigned int _count;
    if(on) {

        if(atomic_add_return(1, &simple_profiling.busy) > 1) {
            printk(KERN_ERR"simple_profiling-analysing busy - can't enable profiler, try it later\n");
            return -1;
        }
        if(mask) {
            profile_current_mask = mask;
            profile_current_start_handle = on;
        }
        if(simple_profiling.data == NULL) {
            if(alloc_profiler_memory()) {
                printk(KERN_ERR"can't enable profiler (memory-error)\n");
                return -1;
            }
        }
        simple_profiling.pos         = 0;
        simple_profiling.start_time  = jiffies;
        simple_profiling.end_time    = 0;
        simple_profiling.wraparround = 0;
        simple_profiling.mask        = mask;
        atomic_set(&simple_profiling.busy, 0);
        switch (on) {
            default:
                printk(KERN_ERR "[%s]Unknown profiling mode. Assuming normal profiling without waparound.\n", __func__);
            case 1:
                simple_profiling.enabled = 1;
                break;
            case 2:
                simple_profiling.enabled = 2; /*--- wrap-arround mode ---*/
                break;
        }
        if(count) *count = 0;
        if(timediff)*timediff = 0;
        avm_profile_counter_init();
#ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD
        if(mask == ( 1 << avm_profile_data_type_code_address_yield )){
            avm_simple_profiling_yield_enable(on);
        }
#endif
        return 0;
    }
    if(set_busy) {
        atomic_set(&simple_profiling.busy, 1);
    }
    simple_profiling.enabled = 0;
    avm_profile_counter_exit();
#ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD
    avm_simple_profiling_yield_enable(on);
#endif/*--- #ifdef CONFIG_AVM_SIMPLE_PROFILING_YIELD ---*/

    if(simple_profiling.end_time  == 0) simple_profiling.end_time = jiffies;
    if(simple_profiling.wraparround == 0) {
        _count = simple_profiling.pos;
    } else {
        _count = simple_profiling.len;
    }
    printk(KERN_INFO"[%s] off: %u entries\n", __func__, _count);
    if(count) {
        *count = _count;
    }
    if(timediff) {
        *timediff = timediff_msec(_count);
    }
    return 0;
}

/*------------------------------------------------------------------------------------------*\
\*------------------------------------------------------------------------------------------*/
void avm_simple_profiling_restart(void) {
    if(profile_current_start_handle || profile_current_mask) {
        /*--- printk("[%s] Push-Button triggering profile %d mask=0x%x\nUse:\n" ---*/
               /*--- "  'echo profiler 0 >/dev/debug' to stop profiling\n" ---*/
               /*--- "  'echo profiler 2 >/dev/debug' to write results to /var/profile.csv\n", ---*/ 
               /*--- __FUNCTION__, profile_current_start_handle, profile_current_mask); ---*/
        avm_simple_profiling_enable(profile_current_start_handle, profile_current_mask, NULL, NULL, 0);
    }
    else {
        printk(KERN_INFO"[%s] Push-Button for profiling ignored (not initialized)\n", __func__);
    }
    return;
}
EXPORT_SYMBOL(avm_simple_profiling_restart);

/*--------------------------------------------------------------------------------*\
\*--------------------------------------------------------------------------------*/
struct _avm_profile_data *avm_simple_profiling_by_idx(unsigned int idx) {
    unsigned int pos;
    struct _avm_profile_data *data;
    if(simple_profiling.wraparround == 0) {
        if(idx < simple_profiling.pos) {
            data = ((struct _avm_profile_data *)(simple_profiling.data[idx / profile_DataSetsPerBlock]) + (idx % profile_DataSetsPerBlock));
            return data;
        }
        return NULL;
    }
    pos = simple_profiling.pos + idx;
    if(pos >= simple_profiling.len) {
        pos -= simple_profiling.len;
        if(pos >= simple_profiling.pos) {
            return NULL;
        }
    }
    data = ((struct _avm_profile_data *)(simple_profiling.data[pos / profile_DataSetsPerBlock]) + (pos % profile_DataSetsPerBlock));
    return data;
}
/*------------------------------------------------------------------------------------------*\
\*------------------------------------------------------------------------------------------*/
int avm_simple_profiling_init(void) {
    spin_lock_init(&simple_profiling.lock);
    simple_profiling.len     = 0;
    simple_profiling.enabled = 0;
    simple_profiling.pos     = 0;
    atomic_set(&simple_profiling.busy, 0);
    gCycle_per_usec = avm_get_clock(avm_clock_id_cpu) / 1000000;
#if !defined(CONFIG_X86)
    gCycle_per_usec /= 2;
#endif/*--- #if !defined(CONFIG_X86)  ---*/
    /*--- printk(KERN_INFO "[simple_profiling] %u entries %u min\n", simple_profiling.len, CONFIG_AVM_PROFILING_TRACE_MODE); ---*/
    return 0;
}

module_init(avm_simple_profiling_init);
#endif /*--- #if defined(CONFIG_AVM_SIMPLE_PROFILING) ---*/