kernel_optimize_test/include/asm-blackfin/delay.h
Mike Frysinger 9e83b98a79 Blackfin arch: add support for working around anomaly 05000312
Anomaly 05000312 - Errors When SSYNC, CSYNC, or Loads to LT, LB and LC Registers Are Interrupted:

DESCRIPTION:
When instruction cache is enabled, erroneous behavior may occur when any of the following instructions are interrupted:

. CSYNC
• SSYNC
• LCx =
• LTx = (only when LCx is non-zero)
• LBx = (only when LCx is non-zero)

When this problem occurs, a variety of incorrect things could happen, including an illegal instruction exception. Additional errors could
show up as an exception, a hardware error, or an instruction that is valid but different than the one that was expected.

WORKAROUND:
Place a cli before all SSYNC, CSYNC, "LCx =", "LTx =", and "LBx =" instructions to disable interrupts, and place an sti after each of these
instructions to re-enable interrupts. When these instructions are executed in code that is already non-interruptible, the problem will not
occur.

Signed-off-by: Mike Frysinger <michael.frysinger@analog.com>
Signed-off-by: Bryan Wu <bryan.wu@analog.com>
2007-11-21 16:08:58 +08:00

63 lines
1.3 KiB
C

/*
* delay.h - delay functions
*
* Copyright (c) 2004-2007 Analog Devices Inc.
*
* Licensed under the GPL-2 or later.
*/
#ifndef __ASM_DELAY_H__
#define __ASM_DELAY_H__
#include <asm/mach/anomaly.h>
static inline void __delay(unsigned long loops)
{
if (ANOMALY_05000312) {
/* Interrupted loads to loop registers -> bad */
unsigned long tmp;
__asm__ __volatile__(
"[--SP] = LC0;"
"[--SP] = LT0;"
"[--SP] = LB0;"
"LSETUP (1f,1f) LC0 = %1;"
"1: NOP;"
/* We take advantage of the fact that LC0 is 0 at
* the end of the loop. Otherwise we'd need some
* NOPs after the CLI here.
*/
"CLI %0;"
"LB0 = [SP++];"
"LT0 = [SP++];"
"LC0 = [SP++];"
"STI %0;"
: "=d" (tmp)
: "a" (loops)
);
} else
__asm__ __volatile__ (
"LSETUP(1f, 1f) LC0 = %0;"
"1: NOP;"
:
: "a" (loops)
: "LT0", "LB0", "LC0"
);
}
#include <linux/param.h> /* needed for HZ */
/*
* Use only for very small delays ( < 1 msec). Should probably use a
* lookup table, really, as the multiplications take much too long with
* short delays. This is a "reasonable" implementation, though (and the
* first constant multiplications gets optimized away if the delay is
* a constant)
*/
static inline void udelay(unsigned long usecs)
{
extern unsigned long loops_per_jiffy;
__delay(usecs * loops_per_jiffy / (1000000 / HZ));
}
#endif