microblaze: Improve TLB calculation for small systems

Systems with small amount of memory need to be handled
differently. Linux can't allocate the whole 32MB with two TLBs
because then there is no MMU protection.
Signed-off-by: NMichal Simek <monstr@monstr.eu>

arch/microblaze/Kconfig

@@ -259,6 +259,10 @@ config MICROBLAZE_32K_PAGES
 
 endchoice
 
+config KERNEL_PAD
+	hex "Kernel PAD for unpacking" if ADVANCED_OPTIONS
+	default "0x80000" if MMU
+
 endmenu
 
 source "mm/Kconfig"

arch/microblaze/include/asm/setup.h

@@ -39,7 +39,8 @@ extern void of_platform_reset_gpio_probe(void);
 void time_init(void);
 void init_IRQ(void);
 void machine_early_init(const char *cmdline, unsigned int ram,
-			unsigned int fdt, unsigned int msr);
+		unsigned int fdt, unsigned int msr, unsigned int tlb0,
+		unsigned int tlb1);
 
 void machine_restart(char *cmd);
 void machine_shutdown(void);

arch/microblaze/include/asm/system.h

@@ -83,6 +83,7 @@ void default_idle(void);
 void free_init_pages(char *what, unsigned long begin, unsigned long end);
 void free_initmem(void);
 extern char *klimit;
+extern unsigned long kernel_tlb;
 extern void ret_from_fork(void);
 
 extern void *alloc_maybe_bootmem(size_t size, gfp_t mask);

arch/microblaze/kernel/head.S

@@ -168,6 +168,53 @@ _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_KERNEL_PAD /* 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 substract 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
@@ -177,16 +224,56 @@ _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 can be zeroes that's why we not setup it */
+	beqi	r9, jump_over
+
+	/* 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:
+	ori r11, r30, 0
+
 	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, r11
 
 	mts     rtlbx,r0		/* 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
+jump_over:
+	/* 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:
+	ori r12, r30, 0
+
+	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, r12
 
 	ori	r6,r0,1			/* TLB slot 1 */
 	mts     rtlbx,r6
@@ -194,6 +281,7 @@ _invalidate:
 	mts	rtlblo,r4		/* Load the data portion of the entry */
 	mts	rtlbhi,r3		/* Load the tag portion of the entry */
 
+jump_over2:
 	/*
 	 * Load a TLB entry for LMB, since we need access to
 	 * the exception vectors, using a 4k real==virtual mapping.
@@ -237,8 +325,8 @@ start_here:
 	 * Please see $(ARCH)/mach-$(SUBARCH)/setup.c for
 	 * the function.
 	 */
-	addik	r9, r0, machine_early_init
-	brald	r15, r9
+	addik	r11, r0, machine_early_init
+	brald	r15, r11
 	nop
 
 #ifndef CONFIG_MMU

arch/microblaze/kernel/setup.c

@@ -97,8 +97,11 @@ inline unsigned get_romfs_len(unsigned *addr)
 }
 #endif	/* CONFIG_MTD_UCLINUX_EBSS */
 
+unsigned long kernel_tlb;
+
 void __init machine_early_init(const char *cmdline, unsigned int ram,
-		unsigned int fdt, unsigned int msr)
+		unsigned int fdt, unsigned int msr, unsigned int tlb0,
+		unsigned int tlb1)
 {
 	unsigned long *src, *dst;
 	unsigned int offset = 0;
@@ -145,6 +148,12 @@ void __init machine_early_init(const char *cmdline, unsigned int ram,
 	setup_early_printk(NULL);
 #endif
 
+	/* setup kernel_tlb after BSS cleaning
+	 * Maybe worth to move to asm code */
+	kernel_tlb = tlb0 + tlb1;
+	/* printk("TLB1 0x%08x, TLB0 0x%08x, tlb 0x%x\n", tlb0,
+							tlb1, kernel_tlb); */
+
 	printk("Ramdisk addr 0x%08x, ", ram);
 	if (fdt)
 		printk("FDT at 0x%08x\n", fdt);

arch/microblaze/mm/init.c

@@ -398,10 +398,16 @@ asmlinkage void __init mmu_init(void)
 		machine_restart(NULL);
 	}
 
-	if ((u32) memblock.memory.regions[0].size < 0x1000000) {
-		printk(KERN_EMERG "Memory must be greater than 16MB\n");
+	if ((u32) memblock.memory.regions[0].size < 0x400000) {
+		printk(KERN_EMERG "Memory must be greater than 4MB\n");
 		machine_restart(NULL);
 	}
+
+	if ((u32) memblock.memory.regions[0].size < kernel_tlb) {
+		printk(KERN_EMERG "Kernel size is greater than memory node\n");
+		machine_restart(NULL);
+	}
+
 	/* Find main memory where the kernel is */
 	memory_start = (u32) memblock.memory.regions[0].base;
 	lowmem_size = memory_size = (u32) memblock.memory.regions[0].size;
@@ -462,11 +468,11 @@ void __init *early_get_page(void)
 		p = alloc_bootmem_pages(PAGE_SIZE);
 	} else {
 		/*
-		 * Mem start + 32MB -> here is limit
+		 * Mem start + kernel_tlb -> here is limit
 		 * because of mem mapping from head.S
 		 */
 		p = __va(memblock_alloc_base(PAGE_SIZE, PAGE_SIZE,
-					memory_start + 0x2000000));
+					memory_start + kernel_tlb));
 	}
 	return p;
 }