diff options
Diffstat (limited to 'arch/microblaze/kernel/head.S')
| -rw-r--r-- | arch/microblaze/kernel/head.S | 204 |
1 files changed, 161 insertions, 43 deletions
diff --git a/arch/microblaze/kernel/head.S b/arch/microblaze/kernel/head.S index 42434008209..4655ff342c6 100644 --- a/arch/microblaze/kernel/head.S +++ b/arch/microblaze/kernel/head.S @@ -39,7 +39,7 @@ #include <asm/mmu.h> #include <asm/processor.h> -.data +.section .data .global empty_zero_page .align 12 empty_zero_page: @@ -50,6 +50,11 @@ swapper_pg_dir: #endif /* CONFIG_MMU */ +.section .rodata +.align 4 +endian_check: + .word 1 + __HEAD ENTRY(_start) #if CONFIG_KERNEL_BASE_ADDR == 0 @@ -58,34 +63,43 @@ ENTRY(_start) real_start: #endif - mfs r1, rmsr - andi r1, r1, ~2 - mts rmsr, r1 + mts rmsr, r0 +/* Disable stack protection from bootloader */ + mts rslr, r0 + addi r8, r0, 0xFFFFFFFF + mts rshr, r8 /* - * Here is checking mechanism which check if Microblaze has msr instructions - * We load msr and compare it with previous r1 value - if is the same, - * msr instructions works if not - cpu don't have them. + * According to Xilinx, msrclr instruction behaves like 'mfs rX,rpc' + * if the msrclr instruction is not enabled. We use this to detect + * if the opcode is available, by issuing msrclr and then testing the result. + * r8 == 0 - msr instructions are implemented + * r8 != 0 - msr instructions are not implemented */ - /* r8=0 - I have msr instr, 1 - I don't have them */ - rsubi r0, r0, 1 /* set the carry bit */ - msrclr r0, 0x4 /* try to clear it */ - /* read the carry bit, r8 will be '0' if msrclr exists */ - addik r8, r0, 0 + mfs r1, rmsr + msrclr r8, 0 /* clear nothing - just read msr for test */ + cmpu r8, r8, r1 /* r1 must contain msr reg content */ /* r7 may point to an FDT, or there may be one linked in. if it's in r7, we've got to save it away ASAP. We ensure r7 points to a valid FDT, just in case the bootloader is broken or non-existent */ beqi r7, no_fdt_arg /* NULL pointer? don't copy */ - lw r11, r0, r7 /* Does r7 point to a */ - rsubi r11, r11, OF_DT_HEADER /* valid FDT? */ +/* Does r7 point to a valid FDT? Load HEADER magic number */ + /* Run time Big/Little endian platform */ + /* Save 1 as word and load byte - 0 - BIG, 1 - LITTLE */ + lbui r11, r0, TOPHYS(endian_check) + beqid r11, big_endian /* DO NOT break delay stop dependency */ + lw r11, r0, r7 /* Big endian load in delay slot */ + lwr r11, r0, r7 /* Little endian load */ +big_endian: + rsubi r11, r11, OF_DT_HEADER /* Check FDT header */ beqi r11, _prepare_copy_fdt or r7, r0, r0 /* clear R7 when not valid DTB */ bnei r11, no_fdt_arg /* No - get out of here */ _prepare_copy_fdt: or r11, r0, r0 /* incremment */ ori r4, r0, TOPHYS(_fdt_start) - ori r3, r0, (0x4000 - 4) + ori r3, r0, (0x8000 - 4) _copy_fdt: lw r12, r7, r11 /* r12 = r7 + r11 */ sw r12, r4, r11 /* addr[r4 + r11] = r12 */ @@ -99,31 +113,35 @@ no_fdt_arg: #ifndef CONFIG_CMDLINE_BOOL /* * handling command line - * copy command line to __init_end. There is space for storing command line. + * copy command line directly to cmd_line placed in data section. */ - or r6, r0, r0 /* incremment */ - ori r4, r0, __init_end /* load address of command line */ + beqid r5, skip /* Skip if NULL pointer */ + or r11, r0, r0 /* incremment */ + ori r4, r0, cmd_line /* load address of command line */ tophys(r4,r4) /* convert to phys address */ ori r3, r0, COMMAND_LINE_SIZE - 1 /* number of loops */ _copy_command_line: - lbu r2, r5, r6 /* r2=r5+r6 - r5 contain pointer to command line */ - sb r2, r4, r6 /* addr[r4+r6]= r2*/ - addik r6, r6, 1 /* increment counting */ + /* r2=r5+r6 - r5 contain pointer to command line */ + lbu r2, r5, r11 + beqid r2, skip /* Skip if no data */ + sb r2, r4, r11 /* addr[r4+r6]= r2 */ + addik r11, r11, 1 /* increment counting */ bgtid r3, _copy_command_line /* loop for all entries */ - addik r3, r3, -1 /* descrement loop */ + addik r3, r3, -1 /* decrement loop */ addik r5, r4, 0 /* add new space for command line */ tovirt(r5,r5) +skip: #endif /* CONFIG_CMDLINE_BOOL */ #ifdef NOT_COMPILE /* save bram context */ - or r6, r0, r0 /* incremment */ + or r11, r0, r0 /* incremment */ ori r4, r0, TOPHYS(_bram_load_start) /* save bram context */ ori r3, r0, (LMB_SIZE - 4) _copy_bram: - lw r7, r0, r6 /* r7 = r0 + r6 */ - sw r7, r4, r6 /* addr[r4 + r6] = r7*/ - addik r6, r6, 4 /* increment counting */ + lw r7, r0, r11 /* r7 = r0 + r6 */ + sw r7, r4, r11 /* addr[r4 + r6] = r7 */ + addik r11, r11, 4 /* increment counting */ bgtid r3, _copy_bram /* loop for all entries */ addik r3, r3, -4 /* descrement loop */ #endif @@ -139,6 +157,7 @@ _copy_bram: _invalidate: mts rtlbx, r3 mts rtlbhi, r0 /* flush: ensure V is clear */ + mts rtlblo, r0 bgtid r3, _invalidate /* loop for all entries */ addik r3, r3, -1 /* sync */ @@ -158,6 +177,53 @@ _invalidate: addik r3,r0, CONFIG_KERNEL_START /* Load the kernel virtual address */ tophys(r4,r3) /* Load the kernel physical address */ + /* start to do TLB calculation */ + addik r12, r0, _end + rsub r12, r3, r12 + addik r12, r12, CONFIG_LOWMEM_SIZE >> PTE_SHIFT /* that's the pad */ + + or r9, r0, r0 /* TLB0 = 0 */ + or r10, r0, r0 /* TLB1 = 0 */ + + addik r11, r12, -0x1000000 + bgei r11, GT16 /* size is greater than 16MB */ + addik r11, r12, -0x0800000 + bgei r11, GT8 /* size is greater than 8MB */ + addik r11, r12, -0x0400000 + bgei r11, GT4 /* size is greater than 4MB */ + /* size is less than 4MB */ + addik r11, r12, -0x0200000 + bgei r11, GT2 /* size is greater than 2MB */ + addik r9, r0, 0x0100000 /* TLB0 must be 1MB */ + addik r11, r12, -0x0100000 + bgei r11, GT1 /* size is greater than 1MB */ + /* TLB1 is 0 which is setup above */ + bri tlb_end +GT4: /* r11 contains the rest - will be either 1 or 4 */ + ori r9, r0, 0x400000 /* TLB0 is 4MB */ + bri TLB1 +GT16: /* TLB0 is 16MB */ + addik r9, r0, 0x1000000 /* means TLB0 is 16MB */ +TLB1: + /* must be used r2 because of subtract if failed */ + addik r2, r11, -0x0400000 + bgei r2, GT20 /* size is greater than 16MB */ + /* size is >16MB and <20MB */ + addik r11, r11, -0x0100000 + bgei r11, GT17 /* size is greater than 17MB */ + /* kernel is >16MB and < 17MB */ +GT1: + addik r10, r0, 0x0100000 /* means TLB1 is 1MB */ + bri tlb_end +GT2: /* TLB0 is 0 and TLB1 will be 4MB */ +GT17: /* TLB1 is 4MB - kernel size <20MB */ + addik r10, r0, 0x0400000 /* means TLB1 is 4MB */ + bri tlb_end +GT8: /* TLB0 is still zero that's why I can use only TLB1 */ +GT20: /* TLB1 is 16MB - kernel size >20MB */ + addik r10, r0, 0x1000000 /* means TLB1 is 16MB */ +tlb_end: + /* * Configure and load two entries into TLB slots 0 and 1. * In case we are pinning TLBs, these are reserved in by the @@ -167,29 +233,82 @@ _invalidate: andi r4,r4,0xfffffc00 /* Mask off the real page number */ ori r4,r4,(TLB_WR | TLB_EX) /* Set the write and execute bits */ + /* + * TLB0 is always used - check if is not zero (r9 stores TLB0 value) + * if is use TLB1 value and clear it (r10 stores TLB1 value) + */ + bnei r9, tlb0_not_zero + add r9, r10, r0 + add r10, r0, r0 +tlb0_not_zero: + + /* look at the code below */ + ori r30, r0, 0x200 + andi r29, r9, 0x100000 + bneid r29, 1f + addik r30, r30, 0x80 + andi r29, r9, 0x400000 + bneid r29, 1f + addik r30, r30, 0x80 + andi r29, r9, 0x1000000 + bneid r29, 1f + addik r30, r30, 0x80 +1: andi r3,r3,0xfffffc00 /* Mask off the effective page number */ - ori r3,r3,(TLB_VALID | TLB_PAGESZ(PAGESZ_16M)) + ori r3,r3,(TLB_VALID) + or r3, r3, r30 - mts rtlbx,r0 /* TLB slow 0 */ + /* Load tlb_skip size value which is index to first unused TLB entry */ + lwi r11, r0, TOPHYS(tlb_skip) + mts rtlbx,r11 /* TLB slow 0 */ mts rtlblo,r4 /* Load the data portion of the entry */ mts rtlbhi,r3 /* Load the tag portion of the entry */ - addik r4, r4, 0x01000000 /* Map next 16 M entries */ - addik r3, r3, 0x01000000 + /* Increase tlb_skip size */ + addik r11, r11, 1 + swi r11, r0, TOPHYS(tlb_skip) + + /* TLB1 can be zeroes that's why we not setup it */ + beqi r10, jump_over2 + + /* look at the code below */ + ori r30, r0, 0x200 + andi r29, r10, 0x100000 + bneid r29, 1f + addik r30, r30, 0x80 + andi r29, r10, 0x400000 + bneid r29, 1f + addik r30, r30, 0x80 + andi r29, r10, 0x1000000 + bneid r29, 1f + addik r30, r30, 0x80 +1: + addk r4, r4, r9 /* previous addr + TLB0 size */ + addk r3, r3, r9 + + andi r3,r3,0xfffffc00 /* Mask off the effective page number */ + ori r3,r3,(TLB_VALID) + or r3, r3, r30 - ori r6,r0,1 /* TLB slot 1 */ - mts rtlbx,r6 + lwi r11, r0, TOPHYS(tlb_skip) + mts rtlbx, r11 /* r11 is used from TLB0 */ mts rtlblo,r4 /* Load the data portion of the entry */ mts rtlbhi,r3 /* Load the tag portion of the entry */ + /* Increase tlb_skip size */ + addik r11, r11, 1 + swi r11, r0, TOPHYS(tlb_skip) + +jump_over2: /* * Load a TLB entry for LMB, since we need access to * the exception vectors, using a 4k real==virtual mapping. */ - ori r6,r0,3 /* TLB slot 3 */ - mts rtlbx,r6 + /* Use temporary TLB_ID for LMB - clear this temporary mapping later */ + ori r11, r0, MICROBLAZE_LMB_TLB_ID + mts rtlbx,r11 ori r4,r0,(TLB_WR | TLB_EX) ori r3,r0,(TLB_VALID | TLB_PAGESZ(PAGESZ_4K)) @@ -213,26 +332,26 @@ start_here: #endif /* CONFIG_MMU */ /* Initialize small data anchors */ - la r13, r0, _KERNEL_SDA_BASE_ - la r2, r0, _KERNEL_SDA2_BASE_ + addik r13, r0, _KERNEL_SDA_BASE_ + addik r2, r0, _KERNEL_SDA2_BASE_ /* Initialize stack pointer */ - la r1, r0, init_thread_union + THREAD_SIZE - 4 + addik r1, r0, init_thread_union + THREAD_SIZE - 4 /* Initialize r31 with current task address */ - la r31, r0, init_task + addik r31, r0, init_task /* * Call platform dependent initialize function. * Please see $(ARCH)/mach-$(SUBARCH)/setup.c for * the function. */ - la r9, r0, machine_early_init - brald r15, r9 + addik r11, r0, machine_early_init + brald r15, r11 nop #ifndef CONFIG_MMU - la r15, r0, machine_halt + addik r15, r0, machine_halt braid start_kernel nop #else @@ -257,8 +376,7 @@ start_here: /* Load up the kernel context */ kernel_load_context: - # Keep entry 0 and 1 valid. Entry 3 mapped to LMB can go away. - ori r5,r0,3 + ori r5, r0, MICROBLAZE_LMB_TLB_ID mts rtlbx,r5 nop mts rtlbhi,r0 |