![Russell King](/assets/img/avatar_default.png)
The existing memory barrier macro causes a significant amount of code to be inserted inline at every call site. For example, in gpio_set_irq_type(), we have this for mb(): c0344c08: f57ff04e dsb st c0344c0c: e59f8190 ldr r8, [pc, #400] ; c0344da4 <gpio_set_irq_type+0x230> c0344c10: e3590004 cmp r9, #4 c0344c14: e5983014 ldr r3, [r8, #20] c0344c18: 0a000054 beq c0344d70 <gpio_set_irq_type+0x1fc> c0344c1c: e3530000 cmp r3, #0 c0344c20: 0a000004 beq c0344c38 <gpio_set_irq_type+0xc4> c0344c24: e50b2030 str r2, [fp, #-48] ; 0xffffffd0 c0344c28: e50bc034 str ip, [fp, #-52] ; 0xffffffcc c0344c2c: e12fff33 blx r3 c0344c30: e51bc034 ldr ip, [fp, #-52] ; 0xffffffcc c0344c34: e51b2030 ldr r2, [fp, #-48] ; 0xffffffd0 c0344c38: e5963004 ldr r3, [r6, #4] Moving the outer_cache_sync() call out of line reduces the impact of the barrier: c0344968: f57ff04e dsb st c034496c: e35a0004 cmp sl, #4 c0344970: e50b2030 str r2, [fp, #-48] ; 0xffffffd0 c0344974: 0a000044 beq c0344a8c <gpio_set_irq_type+0x1b8> c0344978: ebf363dd bl c001d8f4 <arm_heavy_mb> c034497c: e5953004 ldr r3, [r5, #4] This should reduce the cache footprint of this code. Overall, this results in a reduction of around 20K in the kernel size: text data bss dec hex filename 10773970 667392 10369656 21811018 14ccf4a ../build/imx6/vmlinux-old 10754219 667392 10369656 21791267 14c8223 ../build/imx6/vmlinux-new Another advantage to this approach is that we can finally resolve the issue of SoCs which have their own memory barrier requirements within multiplatform kernels (such as OMAP.) Here, the bus interconnects need additional handling to ensure that writes become visible in the correct order (eg, between dma_map() operations, writes to DMA coherent memory, and MMIO accesses.) Acked-by: Tony Lindgren <tony@atomide.com> Acked-by: Richard Woodruff <r-woodruff2@ti.com> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
68 lines
1.6 KiB
C
68 lines
1.6 KiB
C
/*
|
|
* Copyright (C) ST-Ericsson SA 2011
|
|
*
|
|
* License terms: GNU General Public License (GPL) version 2
|
|
*/
|
|
|
|
#include <linux/io.h>
|
|
#include <linux/of.h>
|
|
#include <linux/of_address.h>
|
|
|
|
#include <asm/outercache.h>
|
|
#include <asm/hardware/cache-l2x0.h>
|
|
|
|
#include "db8500-regs.h"
|
|
#include "id.h"
|
|
|
|
static int __init ux500_l2x0_unlock(void)
|
|
{
|
|
int i;
|
|
struct device_node *np;
|
|
void __iomem *l2x0_base;
|
|
|
|
np = of_find_compatible_node(NULL, NULL, "arm,pl310-cache");
|
|
l2x0_base = of_iomap(np, 0);
|
|
of_node_put(np);
|
|
if (!l2x0_base)
|
|
return -ENODEV;
|
|
|
|
/*
|
|
* Unlock Data and Instruction Lock if locked. Ux500 U-Boot versions
|
|
* apparently locks both caches before jumping to the kernel. The
|
|
* l2x0 core will not touch the unlock registers if the l2x0 is
|
|
* already enabled, so we do it right here instead. The PL310 has
|
|
* 8 sets of registers, one per possible CPU.
|
|
*/
|
|
for (i = 0; i < 8; i++) {
|
|
writel_relaxed(0x0, l2x0_base + L2X0_LOCKDOWN_WAY_D_BASE +
|
|
i * L2X0_LOCKDOWN_STRIDE);
|
|
writel_relaxed(0x0, l2x0_base + L2X0_LOCKDOWN_WAY_I_BASE +
|
|
i * L2X0_LOCKDOWN_STRIDE);
|
|
}
|
|
iounmap(l2x0_base);
|
|
return 0;
|
|
}
|
|
|
|
static void ux500_l2c310_write_sec(unsigned long val, unsigned reg)
|
|
{
|
|
/*
|
|
* We can't write to secure registers as we are in non-secure
|
|
* mode, until we have some SMI service available.
|
|
*/
|
|
}
|
|
|
|
static int __init ux500_l2x0_init(void)
|
|
{
|
|
/* Multiplatform guard */
|
|
if (!((cpu_is_u8500_family() || cpu_is_ux540_family())))
|
|
return -ENODEV;
|
|
|
|
/* Unlock before init */
|
|
ux500_l2x0_unlock();
|
|
outer_cache.write_sec = ux500_l2c310_write_sec;
|
|
l2x0_of_init(0, ~0);
|
|
|
|
return 0;
|
|
}
|
|
early_initcall(ux500_l2x0_init);
|