Blackfin arch: Faster C implementation of no-MPU CPLB handler

This is a mixture ofcMichael McTernan's patch and the existing cplb-mpu code.

We ditch the old cplb-nompu implementation, which is a good example of
why a good algorithm in a HLL is preferrable to a bad algorithm written in
assembly.  Rather than try to construct a table of all posible CPLBs and
search it, we just create a (smaller) table of memory regions and
their attributes.  Some of the data structures are now unified for both
the mpu and nompu cases.  A lot of needless complexity in cplbinit.c is
removed.

Further optimizations:
  * compile cplbmgr.c with a lot of -ffixed-reg options, and omit saving
    these registers on the stack when entering a CPLB exception.
  * lose cli/nop/nop/sti sequences for some workarounds - these don't
  * make
    sense in an exception context

Additional code unification should be possible after this.

[Mike Frysinger <vapier.adi@gmail.com>:
 - convert CPP if statements to C if statements
 - remove redundant statements
 - use a do...while loop rather than a for loop to get slightly better
   optimization and to avoid gcc "may be used uninitialized" warnings ...
   we know that the [id]cplb_nr_bounds variables will never be 0, so this
   is OK
 - the no-mpu code was the last user of MAX_MEM_SIZE and with that rewritten,
   we can punt it
 - add some BUG_ON() checks to make sure we dont overflow the small
   cplb_bounds array
 - add i/d cplb entries for the bootrom because there is functions/data in
   there we want to access
 - we do not need a NULL trailing entry as any time we access the bounds
   arrays, we use the nr_bounds variable
]

Signed-off-by: Michael McTernan <mmcternan@airvana.com>
Signed-off-by: Mike Frysinger <vapier.adi@gmail.com>
Signed-off-by: Bernd Schmidt <bernds_cb1@t-online.de>
Signed-off-by: Bryan Wu <cooloney@kernel.org>

1 files changed, 112 insertions, 386 deletions

diff --git a/arch/blackfin/kernel/cplb-nompu/cplbinit.c b/arch/blackfin/kernel/cplb-nompu/cplbinit.c
index 4c010ba50a80..0e28f7595733 100644
--- a/arch/blackfin/kernel/cplb-nompu/cplbinit.c
+++ b/arch/blackfin/kernel/cplb-nompu/cplbinit.c
@@ -29,417 +29,143 @@
 #include <asm/cplbinit.h>
 #include <asm/mem_map.h>
 
-u_long icplb_tables[NR_CPUS][CPLB_TBL_ENTRIES+1];
-u_long dcplb_tables[NR_CPUS][CPLB_TBL_ENTRIES+1];
+struct cplb_entry icplb_tbl[NR_CPUS][MAX_CPLBS] PDT_ATTR;
+struct cplb_entry dcplb_tbl[NR_CPUS][MAX_CPLBS] PDT_ATTR;
 
-#ifdef CONFIG_CPLB_SWITCH_TAB_L1
-#define PDT_ATTR __attribute__((l1_data))
-#else
-#define PDT_ATTR
-#endif
-
-u_long ipdt_tables[NR_CPUS][MAX_SWITCH_I_CPLBS+1] PDT_ATTR;
-u_long dpdt_tables[NR_CPUS][MAX_SWITCH_D_CPLBS+1] PDT_ATTR;
-#ifdef CONFIG_CPLB_INFO
-u_long ipdt_swapcount_tables[NR_CPUS][MAX_SWITCH_I_CPLBS] PDT_ATTR;
-u_long dpdt_swapcount_tables[NR_CPUS][MAX_SWITCH_D_CPLBS] PDT_ATTR;
-#endif
+int first_switched_icplb PDT_ATTR;
+int first_switched_dcplb PDT_ATTR;
 
-struct s_cplb {
-	struct cplb_tab init_i;
-	struct cplb_tab init_d;
-	struct cplb_tab switch_i;
-	struct cplb_tab switch_d;
-};
+struct cplb_boundary dcplb_bounds[9] PDT_ATTR;
+struct cplb_boundary icplb_bounds[7] PDT_ATTR;
 
-#if defined(CONFIG_BFIN_DCACHE) || defined(CONFIG_BFIN_ICACHE)
-static struct cplb_desc cplb_data[] = {
-	{
-		.start = 0,
-		.end = SIZE_1K,
-		.psize = SIZE_1K,
-		.attr = INITIAL_T | SWITCH_T | I_CPLB | D_CPLB,
-		.i_conf = SDRAM_OOPS,
-		.d_conf = SDRAM_OOPS,
-#if defined(CONFIG_DEBUG_HUNT_FOR_ZERO)
-		.valid = 1,
-#else
-		.valid = 0,
-#endif
-		.name = "Zero Pointer Guard Page",
-	},
-	{
-		.start = 0,	/* dyanmic */
-		.end = 0,	/* dynamic */
-		.psize = SIZE_4M,
-		.attr = INITIAL_T | SWITCH_T | I_CPLB,
-		.i_conf = L1_IMEMORY,
-		.d_conf = 0,
-		.valid = 1,
-		.name = "L1 I-Memory",
-	},
-	{
-		.start = 0,	/* dynamic */
-		.end = 0,	/* dynamic */
-		.psize = SIZE_4M,
-		.attr = INITIAL_T | SWITCH_T | D_CPLB,
-		.i_conf = 0,
-		.d_conf = L1_DMEMORY,
-#if ((L1_DATA_A_LENGTH > 0) || (L1_DATA_B_LENGTH > 0))
-		.valid = 1,
-#else
-		.valid = 0,
-#endif
-		.name = "L1 D-Memory",
-	},
-	{
-		.start = L2_START,
-		.end = L2_START + L2_LENGTH,
-		.psize = SIZE_1M,
-		.attr = L2_ATTR,
-		.i_conf = L2_IMEMORY,
-		.d_conf = L2_DMEMORY,
-		.valid = (L2_LENGTH > 0),
-		.name = "L2 Memory",
-	},
-	{
-		.start = 0,
-		.end = 0,  /* dynamic */
-		.psize = 0,
-		.attr = INITIAL_T | SWITCH_T | I_CPLB | D_CPLB,
-		.i_conf = SDRAM_IGENERIC,
-		.d_conf = SDRAM_DGENERIC,
-		.valid = 1,
-		.name = "Kernel Memory",
-	},
-	{
-		.start = 0, /* dynamic */
-		.end = 0, /* dynamic */
-		.psize = 0,
-		.attr = INITIAL_T | SWITCH_T | D_CPLB,
-		.i_conf = SDRAM_IGENERIC,
-		.d_conf = SDRAM_DNON_CHBL,
-		.valid = 1,
-		.name = "uClinux MTD Memory",
-	},
-	{
-		.start = 0, /* dynamic */
-		.end = 0,   /* dynamic */
-		.psize = SIZE_1M,
-		.attr = INITIAL_T | SWITCH_T | D_CPLB,
-		.d_conf = SDRAM_DNON_CHBL,
-		.valid = 1,
-		.name = "Uncached DMA Zone",
-	},
-	{
-		.start = 0, /* dynamic */
-		.end = 0, /* dynamic */
-		.psize = 0,
-		.attr = SWITCH_T | D_CPLB,
-		.i_conf = 0, /* dynamic */
-		.d_conf = 0, /* dynamic */
-		.valid = 1,
-		.name = "Reserved Memory",
-	},
-	{
-		.start = ASYNC_BANK0_BASE,
-		.end = ASYNC_BANK3_BASE + ASYNC_BANK3_SIZE,
-		.psize = 0,
-		.attr = SWITCH_T | D_CPLB,
-		.d_conf = SDRAM_EBIU,
-		.valid = 1,
-		.name = "Asynchronous Memory Banks",
-	},
-	{
-		.start = BOOT_ROM_START,
-		.end = BOOT_ROM_START + BOOT_ROM_LENGTH,
-		.psize = SIZE_1M,
-		.attr = SWITCH_T | I_CPLB | D_CPLB,
-		.i_conf = SDRAM_IGENERIC,
-		.d_conf = SDRAM_DGENERIC,
-		.valid = 1,
-		.name = "On-Chip BootROM",
-	},
-};
+int icplb_nr_bounds PDT_ATTR;
+int dcplb_nr_bounds PDT_ATTR;
 
-static bool __init lock_kernel_check(u32 start, u32 end)
+void __init generate_cplb_tables_cpu(unsigned int cpu)
 {
-	if (start >= (u32)__init_begin || end <= (u32)_stext)
-		return false;
-
-	/* This cplb block overlapped with kernel area. */
-	return true;
-}
+	int i_d, i_i;
+	unsigned long addr;
 
-static void __init
-fill_cplbtab(struct cplb_tab *table,
-	     unsigned long start, unsigned long end,
-	     unsigned long block_size, unsigned long cplb_data)
-{
-	int i;
+	struct cplb_entry *d_tbl = dcplb_tbl[cpu];
+	struct cplb_entry *i_tbl = icplb_tbl[cpu];
 
-	switch (block_size) {
-	case SIZE_4M:
-		i = 3;
-		break;
-	case SIZE_1M:
-		i = 2;
-		break;
-	case SIZE_4K:
-		i = 1;
-		break;
-	case SIZE_1K:
-	default:
-		i = 0;
-		break;
-	}
+	printk(KERN_INFO "NOMPU: setting up cplb tables\n");
 
-	cplb_data = (cplb_data & ~(3 << 16)) | (i << 16);
+	i_d = i_i = 0;
 
-	while ((start < end) && (table->pos < table->size)) {
+	/* Set up the zero page.  */
+	d_tbl[i_d].addr = 0;
+	d_tbl[i_d++].data = SDRAM_OOPS | PAGE_SIZE_1KB;
 
-		table->tab[table->pos++] = start;
+	/* Cover kernel memory with 4M pages.  */
+	addr = 0;
 
-		if (lock_kernel_check(start, start + block_size))
-			table->tab[table->pos++] =
-			    cplb_data | CPLB_LOCK | CPLB_DIRTY;
-		else
-			table->tab[table->pos++] = cplb_data;
+	for (; addr < memory_start; addr += 4 * 1024 * 1024) {
+		d_tbl[i_d].addr = addr;
+		d_tbl[i_d++].data = SDRAM_DGENERIC | PAGE_SIZE_4MB;
+		i_tbl[i_i].addr = addr;
+		i_tbl[i_i++].data = SDRAM_IGENERIC | PAGE_SIZE_4MB;
+	}
 
-		start += block_size;
+	/* Cover L1 memory.  One 4M area for code and data each is enough.  */
+	if (L1_DATA_A_LENGTH || L1_DATA_B_LENGTH) {
+		d_tbl[i_d].addr = L1_DATA_A_START;
+		d_tbl[i_d++].data = L1_DMEMORY | PAGE_SIZE_4MB;
 	}
-}
+	i_tbl[i_i].addr = L1_CODE_START;
+	i_tbl[i_i++].data = L1_IMEMORY | PAGE_SIZE_4MB;
 
-static void __init close_cplbtab(struct cplb_tab *table)
-{
-	while (table->pos < table->size)
-		table->tab[table->pos++] = 0;
-}
+	first_switched_dcplb = i_d;
+	first_switched_icplb = i_i;
 
-/* helper function */
-static void __init
-__fill_code_cplbtab(struct cplb_tab *t, int i, u32 a_start, u32 a_end)
-{
-	if (cplb_data[i].psize) {
-		fill_cplbtab(t,
-				cplb_data[i].start,
-				cplb_data[i].end,
-				cplb_data[i].psize,
-				cplb_data[i].i_conf);
-	} else {
-#if defined(CONFIG_BFIN_ICACHE)
-		if (ANOMALY_05000263 && i == SDRAM_KERN) {
-			fill_cplbtab(t,
-					cplb_data[i].start,
-					cplb_data[i].end,
-					SIZE_4M,
-					cplb_data[i].i_conf);
-		} else
-#endif
-		{
-			fill_cplbtab(t,
-					cplb_data[i].start,
-					a_start,
-					SIZE_1M,
-					cplb_data[i].i_conf);
-			fill_cplbtab(t,
-					a_start,
-					a_end,
-					SIZE_4M,
-					cplb_data[i].i_conf);
-			fill_cplbtab(t, a_end,
-					cplb_data[i].end,
-					SIZE_1M,
-					cplb_data[i].i_conf);
-		}
-	}
-}
+	BUG_ON(first_switched_dcplb > MAX_CPLBS);
+	BUG_ON(first_switched_icplb > MAX_CPLBS);
 
-static void __init
-__fill_data_cplbtab(struct cplb_tab *t, int i, u32 a_start, u32 a_end)
-{
-	if (cplb_data[i].psize) {
-		fill_cplbtab(t,
-				cplb_data[i].start,
-				cplb_data[i].end,
-				cplb_data[i].psize,
-				cplb_data[i].d_conf);
-	} else {
-		fill_cplbtab(t,
-				cplb_data[i].start,
-				a_start, SIZE_1M,
-				cplb_data[i].d_conf);
-		fill_cplbtab(t, a_start,
-				a_end, SIZE_4M,
-				cplb_data[i].d_conf);
-		fill_cplbtab(t, a_end,
-				cplb_data[i].end,
-				SIZE_1M,
-				cplb_data[i].d_conf);
-	}
+	while (i_d < MAX_CPLBS)
+		d_tbl[i_d++].data = 0;
+	while (i_i < MAX_CPLBS)
+		i_tbl[i_i++].data = 0;
 }
 
-void __init generate_cplb_tables_cpu(unsigned int cpu)
+void __init generate_cplb_tables_all(void)
 {
+	int i_d, i_i;
 
-	u16 i, j, process;
-	u32 a_start, a_end, as, ae, as_1m;
-
-	struct cplb_tab *t_i = NULL;
-	struct cplb_tab *t_d = NULL;
-	struct s_cplb cplb;
-
-	printk(KERN_INFO "NOMPU: setting up cplb tables for global access\n");
-
-	cplb.init_i.size = CPLB_TBL_ENTRIES;
-	cplb.init_d.size = CPLB_TBL_ENTRIES;
-	cplb.switch_i.size = MAX_SWITCH_I_CPLBS;
-	cplb.switch_d.size = MAX_SWITCH_D_CPLBS;
-
-	cplb.init_i.pos = 0;
-	cplb.init_d.pos = 0;
-	cplb.switch_i.pos = 0;
-	cplb.switch_d.pos = 0;
-
-	cplb.init_i.tab = icplb_tables[cpu];
-	cplb.init_d.tab = dcplb_tables[cpu];
-	cplb.switch_i.tab = ipdt_tables[cpu];
-	cplb.switch_d.tab = dpdt_tables[cpu];
-
-	cplb_data[L1I_MEM].start = get_l1_code_start_cpu(cpu);
-	cplb_data[L1I_MEM].end = cplb_data[L1I_MEM].start + L1_CODE_LENGTH;
-	cplb_data[L1D_MEM].start = get_l1_data_a_start_cpu(cpu);
-	cplb_data[L1D_MEM].end = get_l1_data_b_start_cpu(cpu) + L1_DATA_B_LENGTH;
-	cplb_data[SDRAM_KERN].end = memory_end;
-
+	i_d = 0;
+	/* Normal RAM, including MTD FS.  */
 #ifdef CONFIG_MTD_UCLINUX
-	cplb_data[SDRAM_RAM_MTD].start = memory_mtd_start;
-	cplb_data[SDRAM_RAM_MTD].end = memory_mtd_start + mtd_size;
-	cplb_data[SDRAM_RAM_MTD].valid = mtd_size > 0;
-# if defined(CONFIG_ROMFS_FS)
-	cplb_data[SDRAM_RAM_MTD].attr |= I_CPLB;
-
-	/*
-	 * The ROMFS_FS size is often not multiple of 1MB.
-	 * This can cause multiple CPLB sets covering the same memory area.
-	 * This will then cause multiple CPLB hit exceptions.
-	 * Workaround: We ensure a contiguous memory area by extending the kernel
-	 * memory section over the mtd section.
-	 * For ROMFS_FS memory must be covered with ICPLBs anyways.
-	 * So there is no difference between kernel and mtd memory setup.
-	 */
-
-	cplb_data[SDRAM_KERN].end = memory_mtd_start + mtd_size;;
-	cplb_data[SDRAM_RAM_MTD].valid = 0;
-
-# endif
+	dcplb_bounds[i_d].eaddr = memory_mtd_start + mtd_size;
 #else
-	cplb_data[SDRAM_RAM_MTD].valid = 0;
+	dcplb_bounds[i_d].eaddr = memory_end;
 #endif
+	dcplb_bounds[i_d++].data = SDRAM_DGENERIC;
+	/* DMA uncached region.  */
+	if (DMA_UNCACHED_REGION) {
+		dcplb_bounds[i_d].eaddr = _ramend;
+		dcplb_bounds[i_d++].data = SDRAM_DNON_CHBL;
+	}
+	if (_ramend != physical_mem_end) {
+		/* Reserved memory.  */
+		dcplb_bounds[i_d].eaddr = physical_mem_end;
+		dcplb_bounds[i_d++].data = (reserved_mem_dcache_on ?
+					    SDRAM_DGENERIC : SDRAM_DNON_CHBL);
+	}
+	/* Addressing hole up to the async bank.  */
+	dcplb_bounds[i_d].eaddr = ASYNC_BANK0_BASE;
+	dcplb_bounds[i_d++].data = 0;
+	/* ASYNC banks.  */
+	dcplb_bounds[i_d].eaddr = ASYNC_BANK3_BASE + ASYNC_BANK3_SIZE;
+	dcplb_bounds[i_d++].data = SDRAM_EBIU;
+	/* Addressing hole up to BootROM.  */
+	dcplb_bounds[i_d].eaddr = BOOT_ROM_START;
+	dcplb_bounds[i_d++].data = 0;
+	/* BootROM -- largest one should be less than 1 meg.  */
+	dcplb_bounds[i_d].eaddr = BOOT_ROM_START + (1 * 1024 * 1024);
+	dcplb_bounds[i_d++].data = SDRAM_DGENERIC;
+	if (L2_LENGTH) {
+		/* Addressing hole up to L2 SRAM.  */
+		dcplb_bounds[i_d].eaddr = L2_START;
+		dcplb_bounds[i_d++].data = 0;
+		/* L2 SRAM.  */
+		dcplb_bounds[i_d].eaddr = L2_START + L2_LENGTH;
+		dcplb_bounds[i_d++].data = L2_DMEMORY;
+	}
+	dcplb_nr_bounds = i_d;
+	BUG_ON(dcplb_nr_bounds > ARRAY_SIZE(dcplb_bounds));
 
-	cplb_data[SDRAM_DMAZ].start = _ramend - DMA_UNCACHED_REGION;
-	cplb_data[SDRAM_DMAZ].end = _ramend;
-
-	cplb_data[RES_MEM].start = _ramend;
-	cplb_data[RES_MEM].end = physical_mem_end;
-
-	if (reserved_mem_dcache_on)
-		cplb_data[RES_MEM].d_conf = SDRAM_DGENERIC;
-	else
-		cplb_data[RES_MEM].d_conf = SDRAM_DNON_CHBL;
-
-	if (reserved_mem_icache_on)
-		cplb_data[RES_MEM].i_conf = SDRAM_IGENERIC;
-	else
-		cplb_data[RES_MEM].i_conf = SDRAM_INON_CHBL;
-
-	for (i = ZERO_P; i < ARRAY_SIZE(cplb_data); ++i) {
-		if (!cplb_data[i].valid)
-			continue;
-
-		as_1m = cplb_data[i].start % SIZE_1M;
-
-		/* We need to make sure all sections are properly 1M aligned
-		 * However between Kernel Memory and the Kernel mtd section, depending on the
-		 * rootfs size, there can be overlapping memory areas.
-		 */
-
-		if (as_1m && i != L1I_MEM && i != L1D_MEM) {
+	i_i = 0;
+	/* Normal RAM, including MTD FS.  */
 #ifdef CONFIG_MTD_UCLINUX
-			if (i == SDRAM_RAM_MTD) {
-				if ((cplb_data[SDRAM_KERN].end + 1) > cplb_data[SDRAM_RAM_MTD].start)
-					cplb_data[SDRAM_RAM_MTD].start = (cplb_data[i].start & (-2*SIZE_1M)) + SIZE_1M;
-				else
-					cplb_data[SDRAM_RAM_MTD].start = (cplb_data[i].start & (-2*SIZE_1M));
-			} else
+	icplb_bounds[i_i].eaddr = memory_mtd_start + mtd_size;
+#else
+	icplb_bounds[i_i].eaddr = memory_end;
 #endif
-				printk(KERN_WARNING "Unaligned Start of %s at 0x%X\n",
-				       cplb_data[i].name, cplb_data[i].start);
-		}
-
-		as = cplb_data[i].start % SIZE_4M;
-		ae = cplb_data[i].end % SIZE_4M;
-
-		if (as)
-			a_start = cplb_data[i].start + (SIZE_4M - (as));
-		else
-			a_start = cplb_data[i].start;
-
-		a_end = cplb_data[i].end - ae;
-
-		for (j = INITIAL_T; j <= SWITCH_T; j++) {
-
-			switch (j) {
-			case INITIAL_T:
-				if (cplb_data[i].attr & INITIAL_T) {
-					t_i = &cplb.init_i;
-					t_d = &cplb.init_d;
-					process = 1;
-				} else
-					process = 0;
-				break;
-			case SWITCH_T:
-				if (cplb_data[i].attr & SWITCH_T) {
-					t_i = &cplb.switch_i;
-					t_d = &cplb.switch_d;
-					process = 1;
-				} else
-					process = 0;
-				break;
-			default:
-					process = 0;
-				break;
-			}
-
-			if (!process)
-				continue;
-			if (cplb_data[i].attr & I_CPLB)
-				__fill_code_cplbtab(t_i, i, a_start, a_end);
-
-			if (cplb_data[i].attr & D_CPLB)
-				__fill_data_cplbtab(t_d, i, a_start, a_end);
-		}
+	icplb_bounds[i_i++].data = SDRAM_IGENERIC;
+	/* DMA uncached region.  */
+	if (DMA_UNCACHED_REGION) {
+		icplb_bounds[i_i].eaddr = _ramend;
+		icplb_bounds[i_i++].data = 0;
 	}
-
-	/* make sure we locked the kernel start */
-	BUG_ON(cplb.init_i.pos < 2 + cplb_data[ZERO_P].valid);
-	BUG_ON(cplb.init_d.pos < 1 + cplb_data[ZERO_P].valid + cplb_data[L1D_MEM].valid);
-
-	/* make sure we didnt overflow the table */
-	BUG_ON(cplb.init_i.size < cplb.init_i.pos);
-	BUG_ON(cplb.init_d.size < cplb.init_d.pos);
-	BUG_ON(cplb.switch_i.size < cplb.switch_i.pos);
-	BUG_ON(cplb.switch_d.size < cplb.switch_d.pos);
-
-	/* close tables */
-	close_cplbtab(&cplb.init_i);
-	close_cplbtab(&cplb.init_d);
-
-	cplb.init_i.tab[cplb.init_i.pos] = -1;
-	cplb.init_d.tab[cplb.init_d.pos] = -1;
-	cplb.switch_i.tab[cplb.switch_i.pos] = -1;
-	cplb.switch_d.tab[cplb.switch_d.pos] = -1;
-
+	if (_ramend != physical_mem_end) {
+		/* Reserved memory.  */
+		icplb_bounds[i_i].eaddr = physical_mem_end;
+		icplb_bounds[i_i++].data = (reserved_mem_icache_on ?
+					    SDRAM_IGENERIC : SDRAM_INON_CHBL);
+	}
+	/* Addressing hole up to BootROM.  */
+	icplb_bounds[i_i].eaddr = BOOT_ROM_START;
+	icplb_bounds[i_i++].data = 0;
+	/* BootROM -- largest one should be less than 1 meg.  */
+	icplb_bounds[i_i].eaddr = BOOT_ROM_START + (1 * 1024 * 1024);
+	icplb_bounds[i_i++].data = SDRAM_IGENERIC;
+	if (L2_LENGTH) {
+		/* Addressing hole up to L2 SRAM, including the async bank.  */
+		icplb_bounds[i_i].eaddr = L2_START;
+		icplb_bounds[i_i++].data = 0;
+		/* L2 SRAM.  */
+		icplb_bounds[i_i].eaddr = L2_START + L2_LENGTH;
+		icplb_bounds[i_i++].data = L2_IMEMORY;
+	}
+	icplb_nr_bounds = i_i;
+	BUG_ON(icplb_nr_bounds > ARRAY_SIZE(icplb_bounds));
 }
-
-#endif