///////////////////////////////////////////////////////////////////////////////////
// File     : fbf_driver.c
// Date     : 23/05/2013
// Author   : alain greiner
// Copyright (c) UPMC-LIP6
///////////////////////////////////////////////////////////////////////////////////
// The fbf_driver.c and fbf_driver.h files are part ot the GIET-VM kernel.
// This driver supports the SoCLib vci_framebuffer component.
//
// It can exist only one frame buffer in the architecture.
//
// There exist two methods to access the VciFrameBuffer device:
//  
// 1) The _fb_sync_write() and _fb_sync_read() functions use a memcpy strategy 
// to implement the transfer between a data buffer (user space) and the frame
// buffer (kernel space). They are blocking until completion of the transfer.
//
// 2) The _fb_cma_init(), _fb_cma_write() and _fb_cma_stop() functions use
// the VciChbufDma component (non replicated) to transfer a flow of images from
// an user space chained buffer (two buffers) to the frame buffer.
// A CMA channel must be allocated to the task requesting it in the mapping_info,
// and stored in the task context.
///////////////////////////////////////////////////////////////////////////////////
// The seg_fbf_base virtual address must be defined in giet_vsegs.ld file.
///////////////////////////////////////////////////////////////////////////////////

#include <giet_config.h>
#include <fbf_driver.h>
#include <utils.h>
#include <dma_driver.h>
#include <cma_driver.h>
#include <tty_driver.h>
#include <ctx_handler.h>
#include <mmc_driver.h>
#include <vmem.h>

#if !defined(GIET_USE_IOMMU) 
# error: You must define GIET_USE_IOMMU in the giet_config.h file
#endif

#if !defined( USE_IOB )
# error: You must define USE_IOB in the hard_config.h file
#endif

#define in_unckdata __attribute__((section (".unckdata")))

////////////// memcpy approach //////////////////////////////////////////////////

//////////////////////////////////////////////////////////////////////////////////
// _fb_sync_write()
// Transfer data from an memory buffer to the frame_buffer device using a memcpy. 
// - offset : offset (in bytes) in the frame buffer.
// - buffer : base address of the memory buffer.
// - length : number of bytes to be transfered.
//////////////////////////////////////////////////////////////////////////////////

unsigned int _fb_sync_write(unsigned int offset, 
                            const void * buffer, 
                            unsigned int length) 
{
    char* fb_address = (char *)&seg_fbf_base + offset;
    _memcpy( fb_address, buffer, length);
    return 0;
}
//////////////////////////////////////////////////////////////////////////////////
// _fb_sync_read()
// Transfer data from the frame_buffer device to a memory buffer using a memcpy.
// - offset : offset (in bytes) in the frame buffer.
// - buffer : base address of the memory buffer.
// - length : number of bytes to be transfered.
//////////////////////////////////////////////////////////////////////////////////
unsigned int _fb_sync_read( unsigned int   offset, 
                            void*          buffer, 
                            unsigned int   length) 
{
    char* fb_address = (char *)&seg_fbf_base + offset;
    _memcpy( buffer, fb_address, length);
    return 0;
}


///////////////// CDMA approach //////////////////////////////////////////////////

// This structure contains two chbuf descriptors that can be used by 
// the VciChbufDma component to tranfer a flow of images:
// - The SRC chbuf descriptor contain two slots (two user buffers)
// - The DST chbuf descriptor contains only one slot (frame buffer)

typedef struct fb_cma_channel_s
{
    paddr_t       buf0;     // physical address + status for user buffer 0 
    paddr_t       buf1;     // physical address + status for user buffer 1
    paddr_t       fbf;      // physical address + status for frame buffer
    unsigned int  length;   // buffer length (number of bytes)
    unsigned int  padding;  // unused (just to have channel size = 32 bytes)
} fb_cma_channel_t;

// array of FB_CMA channels descriptors (32 bytes per entry)
// each entry contains one SRC and one DST chbuf descriptors.
in_unckdata volatile fb_cma_channel_t 
_fb_cma_channel[NB_CMA_CHANNELS] __attribute__((aligned(64)));

// array of physical addresses for the FB_CMA channels descriptors
in_unckdata volatile paddr_t          
_fb_cma_desc_paddr[NB_CMA_CHANNELS];

//////////////////////////////////////////////////////////////////////////////////////
// _fb_cma_init()
// This function uses the _fb_cma_channel[] and _fb_cma_desc_paddr[] arrays,
// (that are both indexed by the channel index), and does four things:
//
// 1) computes the physical addresses for the two source user buffers, for 
//    the destination frame buffer. It initialises the channel descriptor 
//    _fb_cma_channel[i], containing the SRC chbuf descriptor (two buffers), 
//    the DST chbuf descriptor (one single buffer), and the buffer length. 
// 
// 2) computes the physical address for the channel descriptor and register it
//    in the _fb_cma_desc_paddr[i]. 
//    
// 3) makes a SYNC request to L2 cache for channel descriptor, because the
//    channel descriptor is directly accessed in XRAM by the CMA component.
//
// 4) Starts the CMA hardware channel, that will poll the channel descriptor
//    to fransfer an user buffer to the frame buffer as soon as the source 
//    user buffer is marked valid.
//
// Returns 0 if success, > 0 if error 
//////////////////////////////////////////////////////////////////////////////////////
unsigned int _fb_cma_init( const void*  vbase0,  // first user buffer vbase address
                           const void*  vbase1,  // second user buffer vbase address
                           unsigned int length ) // buffer length (number of bytes)
{
#if NB_CMA_CHANNELS > 0

    unsigned int  channel_id;          // CMA channel index
    unsigned int  user_ptab;           // page table virtual address
    unsigned int  ko;                  // unsuccessfull V2P translation
    unsigned int  vaddr;               // virtual address
    unsigned int  flags;               // protection flags
    unsigned int  ppn;                 // physical page number
    paddr_t       desc_paddr;          // physical address of channel descriptor

    // get CMA channel index
    channel_id = _get_context_slot(CTX_CMA_ID);
    if ( channel_id >= NB_CMA_CHANNELS )
    {
        _tty_get_lock( 0 );
        _puts("\n[GIET ERROR] in _fb_cma_init() : CMA channel index too large\n");
        _tty_release_lock( 0 );
        return 1;
    }

    // checking size for channel descriptor
    if ( sizeof(fb_cma_channel_t) != 32 )
    {
        _tty_get_lock( 0 );
        _puts("\n[GIET ERROR] in _fb_cma_init() : bad fb_cma_channel size\n");
        _tty_release_lock( 0 );
        return 1;
    }

    // checking channel descriptor alignment (32 bytes)
    if ( (unsigned int)(&_fb_cma_channel[channel_id]) & 0x1F ) 
    {
        _tty_get_lock( 0 );
        _puts("\n[GIET ERROR] in _fb_cma_init() : bad fb_cma_channel alignment\n");
        _tty_release_lock( 0 );
        return 1;
    }

    // checking user buffer virtual addresses and length alignment
    if ( ((unsigned int)vbase0 & 0x3) || ((unsigned int)vbase1 & 0x3) || (length & 0x3) ) 
    {
        _tty_get_lock( 0 );
        _puts("\n[GIET ERROR] in _fb_cma_init() : user buffer not word aligned\n");
        _tty_release_lock( 0 );
        return 1;
    }

    // get page table virtual address
    user_ptab = _get_context_slot(CTX_PTAB_ID);

    // compute and register frame buffer physical address
    vaddr = ((unsigned int)&seg_fbf_base);
    ko    = _v2p_translate( (page_table_t*) user_ptab, 
                         (vaddr >> 12),
                         &ppn, 
                         &flags );
    if (ko) 
    {
        _tty_get_lock( 0 );
        _puts("\n[GIET ERROR] in _fb_cma_init() : frame buffer unmapped\n");
        _tty_release_lock( 0 );
        return 1;
    }
    _fb_cma_channel[channel_id].fbf = ((paddr_t)ppn << 12) | (vaddr & 0x00000FFF);

    // Compute and register first user buffer physical address
    vaddr = (unsigned int)vbase0; 
    ko = _v2p_translate( (page_table_t*) user_ptab, 
                         (vaddr >> 12),
                         &ppn, 
                         &flags );
    if (ko) 
    {
        _tty_get_lock( 0 );
        _puts("\n[GIET ERROR] in _fb_cma_init() : user buffer 0 unmapped\n");
        _tty_release_lock( 0 );
        return 1;
    } 
    if ((flags & PTE_U) == 0) 
    {
        _tty_get_lock( 0 );
        _puts("[GIET ERROR] in _fb_cma_init() : user buffer 0 not in user space\n");
        _tty_release_lock( 0 );
        return 1; 
    }
    _fb_cma_channel[channel_id].buf0 = ((paddr_t)ppn << 12) | (vaddr & 0x00000FFF);

    // Compute and register second user buffer physical address
    vaddr = (unsigned int)vbase1; 
    ko = _v2p_translate( (page_table_t*) user_ptab, 
                         (vaddr >> 12),
                         &ppn, 
                         &flags );
    if (ko) 
    {
        _tty_get_lock( 0 );
        _puts("\n[GIET ERROR] in _fb_cma_init() : user buffer 1 unmapped\n");
        _tty_release_lock( 0 );
        return 1;
    } 
    if ((flags & PTE_U) == 0) 
    {
        _tty_get_lock( 0 );
        _puts("[GIET ERROR] in _fb_cma_init() : user buffer 1 not in user space\n");
        _tty_release_lock( 0 );
        return 1; 
    }
    _fb_cma_channel[channel_id].buf1 = ((paddr_t)ppn << 12) | (vaddr & 0x00000FFF);

    // register buffer length in channel descriptor
    _fb_cma_channel[channel_id].length = length;

    // Compute and register physical adress of the channel descriptor
    vaddr = (unsigned int)(&_fb_cma_channel[channel_id]);
    ko = _v2p_translate( (page_table_t*) user_ptab, 
                         (vaddr >> 12),
                         &ppn, 
                         &flags );
    if (ko) 
    {
        _tty_get_lock( 0 );
        _puts("\n[GIET ERROR] in _fb_cma_init() : channel descriptor unmapped\n");
        _tty_release_lock( 0 );
        return 1;
    } 
    _fb_cma_desc_paddr[channel_id] = (((paddr_t)ppn) << 12) | (vaddr & 0x00000FFF);

    desc_paddr                     = (((paddr_t)ppn) << 12) | (vaddr & 0x00000FFF);
    

#if GIET_DEBUG_FBF_DRIVER
_tty_get_lock( 0 );
_puts("\n[CMA DEBUG] fb_cma_init()");
_puts("\n - fbf       pbase = ");
_putl( _fb_cma_channel[channel_id].fbf );
_puts("\n - buf0      pbase = ");
_putl( _fb_cma_channel[channel_id].buf0 );
_puts("\n - buf1      pbase = ");
_putl( _fb_cma_channel[channel_id].buf1 );
_puts("\n - channel   pbase = ");
_putl( _fb_cma_desc_paddr[channel_id] );
_puts("\n");
_tty_release_lock( 0 );
#endif

    // SYNC request for channel descriptor
    _memc_sync( desc_paddr, 32 );

    // CMA channel activation
    unsigned int* cma_vbase = (unsigned int *)&seg_cma_base;
    unsigned int  offset    = channel_id * CHBUF_CHANNEL_SPAN;

    cma_vbase[offset + CHBUF_SRC_DESC]  = (unsigned int)(desc_paddr & 0xFFFFFFFF);
    cma_vbase[offset + CHBUF_SRC_EXT]   = (unsigned int)(desc_paddr >> 32);
    cma_vbase[offset + CHBUF_SRC_NBUFS] = 2;
    cma_vbase[offset + CHBUF_DST_DESC]  = (unsigned int)(desc_paddr & 0xFFFFFFFF) + 16;
    cma_vbase[offset + CHBUF_DST_EXT]   = (unsigned int)(desc_paddr >> 32);
    cma_vbase[offset + CHBUF_DST_NBUFS] = 1;
    cma_vbase[offset + CHBUF_BUF_SIZE]  = length;
    cma_vbase[offset + CHBUF_PERIOD]    = 300;
    cma_vbase[offset + CHBUF_RUN]       = 1;

    return 0;

#else

    _tty_get_lock( 0 );
    _puts("\n[GIET ERROR] in _fb_cma_init() : no CMA channel allocated\n");
    _tty_release_lock( 0 );

    return 1;
#endif
}
////////////////////////////////////////////////////////////////////////////////////
// _fb_cma_write()
//
// It updates the status of the SRC and DST chbuf descriptors, to allow the CMA
// component to transfer the source user buffer to the frame buffer.
//
// If the IO Bridge component is used:
// 1) it makes an INVAL request for the channel descriptor, before testing the
//    source buffer status, because it is modified in XRAM by the CMA component.
// 2) it makes a SYNC request for the source user buffer before activating the CMA 
//    transfer, because the data will be read from XRAM by the CMA component.
// 3) it makes a SYNC request for the channel descriptor after modification 
//    of the SRC and DST status, because these descriptors will be read from XRAM 
//    by the CMA component.
//
// The buffer_id argument is the user buffer index (0 => buf0 / not 0 => buf1)
// Returns 0 if success, > 0 if error 
////////////////////////////////////////////////////////////////////////////////////
unsigned int _fb_cma_write( unsigned int buffer_id )
{
#if NB_CMA_CHANNELS > 0

    volatile paddr_t buf_paddr;
    unsigned int     buf_length;
    unsigned int     full = 1;

    unsigned int     count = 0;

    // get CMA channel index 
    unsigned int channel_id = _get_context_slot(CTX_CMA_ID);

#if GIET_DEBUG_FBF_DRIVER
_tty_get_lock( 0 );
_puts("\n[CMA DEBUG] fb_cma_write() for CMA channel ");
_putd( channel_id );
_puts(" / buf_id = ");
_putd( buffer_id );
_puts("\n");
_tty_release_lock( 0 );
#endif

    // waiting buffer empty
    while ( full )
    {  
        if ( USE_IOB )
        {
            // INVAL L2 cache for the channel descriptor,
            _memc_inval( _fb_cma_desc_paddr[channel_id], 32 );

            // INVAL L1 cache for the channel descriptor,
            _dcache_buf_invalidate( &_fb_cma_channel[channel_id], 32 );
        }

        // read SRC buffer descriptor
        if ( buffer_id == 0 ) buf_paddr = _fb_cma_channel[channel_id].buf0;
        else                  buf_paddr = _fb_cma_channel[channel_id].buf1;
        full = ( (unsigned int)(buf_paddr>>63) );

        count++;
        if ( count == 10 ) _exit();

#if GIET_DEBUG_FBF_DRIVER
_tty_get_lock( 0 );
_puts(" - buffer descriptor = ");
_putl( buf_paddr );
_puts(" at cycle ");
_putd( _get_proctime() );
_puts("\n");
_tty_release_lock( 0 );
#endif

    }

    if ( USE_IOB )
    {
        // SYNC request for the user buffer because 
        // this buffer will be read from XRAM by the CMA component
        _memc_sync( buf_paddr, _fb_cma_channel[channel_id].length );
    }

    // set SRC full 
    if ( buffer_id == 0 )
    _fb_cma_channel[channel_id].buf0 = buf_paddr | 0x8000000000000000ULL;
    else
    _fb_cma_channel[channel_id].buf1 = buf_paddr | 0x8000000000000000ULL;

    // set DST empty
    _fb_cma_channel[channel_id].fbf  = _fb_cma_channel[channel_id].fbf
                                       & 0x7FFFFFFFFFFFFFFFULL;

    if ( USE_IOB )
    {
        // SYNC request for the channel descriptor, because
        // it will be read in XRAM by the CMA component
        _memc_sync( _fb_cma_desc_paddr[channel_id], 32 );
    }

    return 0;

#else

    _tty_get_lock( 0 );
    _puts("\n[GIET ERROR] in _fb_cma_channel() : no CMA channel allocated\n");
    _tty_release_lock( 0 );
    return 1;

#endif
}
//////////////////////////////////////////////////////////////////////////////////
// _fb_cma_stop()
// This function desactivates the CMA channel allocated to the calling task.
// Returns 0 if success, > 0 if error 
//////////////////////////////////////////////////////////////////////////////////
unsigned int _fb_cma_stop( unsigned int buffer_id )
{
#if NB_CMA_CHANNELS > 0

    // get CMA channel allocated 
    unsigned int channel_id = _get_context_slot(CTX_CMA_ID);

    // CMA channel desactivation
    unsigned int* cma_vbase = (unsigned int *)&seg_cma_base;
    unsigned int  offset     = channel_id * CHBUF_CHANNEL_SPAN;
    cma_vbase[offset + CHBUF_RUN] = 0;
    return 0;

#else

    _tty_get_lock( 0 );
    _puts("\n[GIET ERROR] in _fb_cma_stop() : no CMA channel allocated\n");
    _tty_release_lock( 0 );
    return 1;

#endif
}
    

// Local Variables:
// tab-width: 4
// c-basic-offset: 4
// c-file-offsets:((innamespace . 0)(inline-open . 0))
// indent-tabs-mode: nil
// End:
// vim: filetype=c:expandtab:shiftwidth=4:tabstop=4:softtabstop=4