source: rtems/c/src/lib/libbsp/powerpc/mvme5500/start/preload.S @ ee732739

/*
 * Mini-loader for the SVGM BSP.
 *
 * $Id$
 *
 * Author: Till Straumann, 10/2001 <strauman@slac.stanford.edu>
 *
 * Some ideas are borrowed from the powerpc/shared/bootloader
 * by
 *  Copyright (C) 1998, 1999 Gabriel Paubert, paubert@iram.es
 *  Copyright (C) 1999 Eric Valette. valette@crf.canon.fr
 *
 * The SMON firmware is unable to load the RTEMS image below
 * 0x2000 (I believe their stack is growing below 0x1000).
 * 
 * The code provided by this file is responsible for the performing
 * the following steps:
 *
 *  1) Save commandline parameters to an area that is 
 *       a) not covered by the downloaded image
 *       b) will not be overwritten by the moved image
 *          nor the final BSS segment (rtems clears BSS
 *          before saving the command line).
 *  2) Move the entire image (including this very file) to
 *     its final location starting at 0x0000.
 *     It is important to note that _NO_STACK_ is available
 *     during this step. Also, there is probably no return to
 *     SMON because relocating RTEMS will destroy vital SMON
 *     data (such as its stack).
 *  3) Flush the cache to make sure the relocated image is actually
 *     in memory.
 *  4) setup RTEMS environment (initial register values), most 
 *     notably an initial STACK. The initial stack may be small and
 *     is used by RTEMS only at a very early stage.
 *     A safe place for the stack seems to be the 00..0x7f area.
 *     NOTE: we should respect the MAILBOX area 0x80..0xff!
 *  5) switch the MMU off (because that's what RTEMS is expecting
 *     it to be at startup).
 *  6) fire up rtems...
 *
 *
 *  Calling convention:
 *     R1: SMON SP
 *     R3: command line string start
 *     R4: command line string end + 1
 *     R5: where SMON put the image
 *         if R5 is 0, the preloader will use its entry point
 *         as the image starting address.
 *         See NOTE below.
 *     R6: end of the image (i.e. R6-R5 is the image length)
 *         if R6 is 0, _edata will be used as the image length
 *         See NOTE below.
 *
 *     NOTE: if the symbol DONT_USE_R5_ENTRY is defined,
 *         R5/R6 are never used and the necessary parameters are
 *         determined at runtime (R5) / linkage (R6) [_edata]
 *
 *  ASSUMPTIONS:
 *    The code RELIES on the assumption that the image will be
 *    moved DOWNWARDS in memory and that the this loader is
 *    prepended to the image, i.e. it is safe to do
 *        codemove(codemove,0,codemove_end - codemove);
 *        (*0)(codemove_end, codemove_end-codemove, __rtems_end-codemove_end);
 *    where codemove(from, to, nbytes) is defined as
 *        codemove(from, to, nbytes) { while (nbytes--) *(to++)=*(from++); }
 *    Implicit to these assumptions is the assumption that the destination
 *    address is cache block aligned.
 *    Furthermore, the byte count is assumed to be a multiple
 *    of four
 *
 */
#if 0
#include <rtems/score/powerpc.h>
#else
#ifndef PPC_CACHE_ALIGNMENT
#define PPC_CACHE_ALIGNMENT     32
#endif
#endif

#include <rtems/score/cpu.h>
#include <rtems/asm.h>

/* Note that major modifications may be needed
 * if DESTINATION_ADDR is not 0
 */
#define KERNELBASE                      0x0
#define INITIAL_STACK           0x78                                    /* 8-byte aligned */
#define CACHE_LINE_SIZE         PPC_CACHE_ALIGNMENT     /* autodetect doesn't work, see below */
#define ASSUME_RTEMS_INSTALLS_VECTORS                           /* assume we need not load vectors */
#define DONT_USE_R5_ENTRY                                                       /* always dynamically determine the address we're running from */

        /* put this into its own section which we want to
         * be loaded at the very beginning. We should probably
         * not use more than 255 bytes.
         */
        PUBLIC_VAR(__rtems_start)
        PUBLIC_VAR(__rtems_entry_point)
        PUBLIC_VAR(__rtems_end)
        .section .entry_point_section,"awx",@progbits
preload:
        /* find out where we are */
        bl      here
here:
        xor             r0,r0,r0
        mtmsr   r0      /* clear MSR to known state */
        mflr    r5
        addi    r5,r5,-(here-preload)
        lis             r27,_edata@h
        ori             r27,r27,_edata@l

        /* at this point the register contents are
         * R3:  command line start
         * R4:  R3 + command line length
         * R5:  address we are running from / loaded to 
         * R27: image end
         */

        /* save command line start */
        mr              r6, r3
        /* save the command line parameters if they are to be overwritten */
        sub.    r17, r4, r3             /* string length */
        ble             leaveparms              /* <=0 -> no parameters */
        /* copy has to be out of the way of the bss; therefore we must
         * put the string out of the way of both, the current end of
         * the image (without bss) AND the end of the loaded image
         * (including bss):
         * |......image.........|  downloaded image
         * |image_bss...........|  loaded image with bss appended
         *
         *             ^ safe place for string
         *
         * the alternative scenario looks like this:
         * |..image.............|  downloaded image
         * |image_bss...........|  loaded image with bss appended
         *           ^ safe place for string
         */
        lis             r18, __rtems_end+0x10000@h      /* round up, save one instruction */
        add             r16, r5, r27    /* image end + 1 */
        cmpw    r16, r18
        bge             ishighenough
        mr              r16,r18                 /* __rtems_end is higher than the image end
                                                         * (without bss)
                                                         */
ishighenough:
        cmpw    r16, r3             /* destination start > current string start ? */
        ble             leaveparms              /* string already after dst, leave it */
        /* copy string from the last byte downwards */
        add             r6, r16, r17    /* last byte of destination + 1 */
        mtctr   r17
1:
        lbzu    r3, -1(r4)
        stbu    r3,     -1(r6)
        bdnz    1b
leaveparms:
        add             r7, r6, r17             /* destination + strlen */

#ifndef CACHE_LINE_SIZE
        /* Oh well, SMON has inhibited the cache, so this
         * nice routine doesn't work...
         */
        /* figure out the cache line size */
        li              r16, 0x80
        cmpw    r5, r16                 /* 'from' must be > 0x80 */
        blt             panic
        
1:      /* store some arbitrary, nonzero stuff in 0..0x7c */
        stwu    r16,-4(r16)
        cmpwi   r16,0
        bne             1b
        dcbz    0,r16                   /* zero out one cache line */
        subi    r16,r16,4
2:      lwzu    r0,4(r16)               /* search for a non-zero word */
        cmpwi   r0,0
        beq             2b
        /* OK, r16 now hold the size of a cache line in bytes */
#else
        li              r16,CACHE_LINE_SIZE
#endif
        
    lis         r3,preload@h
        ori             r3,r3,preload@l
        mr              r4,r5                   /* from-addr */
        li              r5,_preload_size/* this is never > 16k */
        /* now move ourselves to the link address ('preload').
         * We set up the LR, so domove() 'returns' to the
         * relocated copy
         */
        lis             r0,return_here@h
        ori             r0,r0,return_here@l
        mtlr    r0      
        b               domove                  /* move the preloader itself */
return_here:
        /* now we move the entire rest of the image */
#ifdef ASSUME_RTEMS_INSTALLS_VECTORS
        lis             r3,__rtems_start@h
        ori             r3,r3,__rtems_start@l
        lis             r0,preload@h    /* calculate/adjust from address */
        ori             r0,r0,preload@l
        sub             r0,r3,r0
        add             r4,r4,r0
        sub             r5,r27,r3
#else
        add             r3,r3,r5                /* add preloader size to destination */
        add             r4,r4,r5                /* and source addresses */
        sub             r5,r27,r5               /* length of the remaining rest */
#endif
        bl              domove
        /* OK, now everything should be in place.
     * we are ready to start...
         */

        /* setup initial stack for rtems early boot */
        li              r1,INITIAL_STACK
        /* disable the MMU and fire up rtems */
        mfmsr   r0
        ori     r0,r0,MSR_IR|MSR_DR|MSR_IP|MSR_ME
        xori    r0,r0,MSR_IR|MSR_DR
        mtsrr1  r0
        lis             r0,__rtems_entry_point@h
        ori             r0,r0,__rtems_entry_point@l
        mtsrr0  r0
        /* R6: start of command line */
        /* R7: end of command line +1 */
        rfi

    /* domove(to, from, nbytes):
     *
     * move a R5 bytes from R4 to R3 and flush
     * the caches for the destination memory
     * region. R16 provides the cache line size.
         * DESTROYS: R0, R17, R18, CTR, CR
     */
domove:
        addi    r0,r5,3                 /* convert to word count */
        srwi.   r0,r0,2
        beq     3f                                      /* nothing to do */
        cmpw    r3,r4                   /* from == to ?  */
        beq 3f
        mtctr   r0
        la              r18,-4(r4)
        la              r17,-4(r3)
1:      lwzu    r0,4(r18)
        stwu    r0,4(r17)
        bdnz    1b                              /* move data */
        /* now, we must flush the destination cache region */
#ifndef CACHE_LINE_SIZE
        cmpwi   r16,0
        beq             3f                              /* nothing to do */
#endif
#if defined(CACHE_LINE_SIZE) && CACHE_LINE_SIZE > 0
        add             r17,r3,r5               /* target end pointer */
        subi    r0,r16,1
        add             r17,r17,r0
        andc    r17,r17,r0              /* cache aligned target end pointer */
        mr              r18,r3
2:      cmpw    r18,r17
        dcbst   0,r18                   /* write out data cache line */
        icbi    0,r18                   /* invalidate corresponding i-cache line */
        add             r18,r18,r16
        blt             2b
        sync                                    /* make sure data is written back */
        isync                                   /* invalidate possibly preloaded instructions */
#endif
3:
        blr

#if !defined(CACHE_LINE_SIZE)
panic:
        li              r10,0x63
        mfmsr   r0
        ori             r0,r0,MSR_IP
        mtmsr   r0
        sc
#endif

/* DONT PUT ANY CODE BELOW HERE */
_preload_size = . - preload