Take 3. Remove unnecessary .mips0 and make rollback_handler a bit
faster.
------------------------------------------------------------------------
Subject: Fix potential latency problem due to non-atomic cpu_wait.
If an interrupt happened between checking of NEED_RESCHED and WAIT
instruction, adjust EPC to restart from checking of NEED_RESCHED.
Signed-off-by: Atsushi Nemoto <anemo@xxxxxxxxxxxxx>
---
arch/mips/kernel/cpu-probe.c | 16 ++--------------
arch/mips/kernel/genex.S | 37 +++++++++++++++++++++++++++++++++++++
arch/mips/kernel/traps.c | 22 ++++++++++++++++------
3 files changed, 55 insertions(+), 20 deletions(-)
diff --git a/arch/mips/kernel/cpu-probe.c b/arch/mips/kernel/cpu-probe.c
index 5c27943..1f71fec 100644
--- a/arch/mips/kernel/cpu-probe.c
+++ b/arch/mips/kernel/cpu-probe.c
@@ -45,18 +45,7 @@ static void r39xx_wait(void)
local_irq_enable();
}
-/*
- * There is a race when WAIT instruction executed with interrupt
- * enabled.
- * But it is implementation-dependent wheter the pipelie restarts when
- * a non-enabled interrupt is requested.
- */
-static void r4k_wait(void)
-{
- __asm__(" .set mips3 \n"
- " wait \n"
- " .set mips0 \n");
-}
+extern void r4k_wait(void);
/*
* This variant is preferable as it allows testing need_resched and going to
@@ -128,7 +117,7 @@ static int __init wait_disable(char *s)
__setup("nowait", wait_disable);
-static inline void check_wait(void)
+void __init check_wait(void)
{
struct cpuinfo_mips *c = ¤t_cpu_data;
@@ -239,7 +228,6 @@ static inline void check_errata(void)
void __init check_bugs32(void)
{
- check_wait();
check_errata();
}
diff --git a/arch/mips/kernel/genex.S b/arch/mips/kernel/genex.S
index e76a76b..96f46b3 100644
--- a/arch/mips/kernel/genex.S
+++ b/arch/mips/kernel/genex.S
@@ -20,6 +20,7 @@
#include <asm/stackframe.h>
#include <asm/war.h>
#include <asm/page.h>
+#include <asm/thread_info.h>
#define PANIC_PIC(msg) \
.set push; \
@@ -126,7 +127,42 @@ handle_vcei:
__FINIT
+ .align 5 /* 32 byte rollback region */
+LEAF(r4k_wait)
+ .set push
+ .set noreorder
+ /* start of rollback region */
+ LONG_L t0, TI_FLAGS($28)
+ nop
+ andi t0, _TIF_NEED_RESCHED
+ bnez t0, 1f
+ nop
+ nop
+ nop
+ .set mips3
+ wait
+ /* end of rollback region (the region size must be power of two) */
+ .set pop
+1:
+ jr ra
+ END(r4k_wait)
+
+ .macro BUILD_ROLLBACK_PROLOGUE handler
+ FEXPORT(rollback_\handler)
+ .set push
+ .set noat
+ MFC0 k0, CP0_EPC
+ PTR_LA k1, r4k_wait
+ ori k0, 0x1f /* 32 byte rollback region */
+ xori k0, 0x1f
+ bne k0, k1, 9f
+ MTC0 k0, CP0_EPC
+9:
+ .set pop
+ .endm
+
.align 5
+BUILD_ROLLBACK_PROLOGUE handle_int
NESTED(handle_int, PT_SIZE, sp)
#ifdef CONFIG_TRACE_IRQFLAGS
/*
@@ -201,6 +237,7 @@ NESTED(except_vec_ejtag_debug, 0, sp)
* This prototype is copied to ebase + n*IntCtl.VS and patched
* to invoke the handler
*/
+BUILD_ROLLBACK_PROLOGUE except_vec_vi
NESTED(except_vec_vi, 0, sp)
SAVE_SOME
SAVE_AT
diff --git a/arch/mips/kernel/traps.c b/arch/mips/kernel/traps.c
index 23e73d0..23807c6 100644
--- a/arch/mips/kernel/traps.c
+++ b/arch/mips/kernel/traps.c
@@ -43,6 +43,9 @@
#include <asm/types.h>
#include <asm/stacktrace.h>
+extern void check_wait(void);
+extern asmlinkage void r4k_wait(void);
+extern asmlinkage void rollback_handle_int(void);
extern asmlinkage void handle_int(void);
extern asmlinkage void handle_tlbm(void);
extern asmlinkage void handle_tlbl(void);
@@ -1146,6 +1149,9 @@ static void *set_vi_srs_handler(int n, vi_handler_t addr, int srs)
extern char except_vec_vi, except_vec_vi_lui;
extern char except_vec_vi_ori, except_vec_vi_end;
+ extern char rollback_except_vec_vi;
+ char *vec_start = (cpu_wait == r4k_wait) ?
+ &rollback_except_vec_vi : &except_vec_vi;
#ifdef CONFIG_MIPS_MT_SMTC
/*
* We need to provide the SMTC vectored interrupt handler
@@ -1153,11 +1159,11 @@ static void *set_vi_srs_handler(int n, vi_handler_t addr, int srs)
* Status.IM bit to be masked before going there.
*/
extern char except_vec_vi_mori;
- const int mori_offset = &except_vec_vi_mori - &except_vec_vi;
+ const int mori_offset = &except_vec_vi_mori - vec_start;
#endif /* CONFIG_MIPS_MT_SMTC */
- const int handler_len = &except_vec_vi_end - &except_vec_vi;
- const int lui_offset = &except_vec_vi_lui - &except_vec_vi;
- const int ori_offset = &except_vec_vi_ori - &except_vec_vi;
+ const int handler_len = &except_vec_vi_end - vec_start;
+ const int lui_offset = &except_vec_vi_lui - vec_start;
+ const int ori_offset = &except_vec_vi_ori - vec_start;
if (handler_len > VECTORSPACING) {
/*
@@ -1167,7 +1173,7 @@ static void *set_vi_srs_handler(int n, vi_handler_t addr, int srs)
panic("VECTORSPACING too small");
}
- memcpy(b, &except_vec_vi, handler_len);
+ memcpy(b, vec_start, handler_len);
#ifdef CONFIG_MIPS_MT_SMTC
BUG_ON(n > 7); /* Vector index %d exceeds SMTC maximum. */
@@ -1437,6 +1443,10 @@ void __init trap_init(void)
extern char except_vec3_generic, except_vec3_r4000;
extern char except_vec4;
unsigned long i;
+ int rollback;
+
+ check_wait();
+ rollback = (cpu_wait == r4k_wait);
if (cpu_has_veic || cpu_has_vint)
ebase = (unsigned long) alloc_bootmem_low_pages(0x200 + VECTORSPACING*64);
@@ -1496,7 +1506,7 @@ void __init trap_init(void)
if (board_be_init)
board_be_init();
- set_except_vector(0, handle_int);
+ set_except_vector(0, rollback ? rollback_handle_int : handle_int);
set_except_vector(1, handle_tlbm);
set_except_vector(2, handle_tlbl);
set_except_vector(3, handle_tlbs);
