[ARM] 3881/4: xscale: clean up cp0/cp1 handling
XScale cores either have a DSP coprocessor (which contains a single
40 bit accumulator register), or an iWMMXt coprocessor (which contains
eight 64 bit registers.)
Because of the small amount of state in the DSP coprocessor, access to
the DSP coprocessor (CP0) is always enabled, and DSP context switching
is done unconditionally on every task switch. Access to the iWMMXt
coprocessor (CP0/CP1) is enabled only when an iWMMXt instruction is
first issued, and iWMMXt context switching is done lazily.
CONFIG_IWMMXT is supposed to mean 'the cpu we will be running on will
have iWMMXt support', but boards are supposed to select this config
symbol by hand, and at least one pxa27x board doesn't get this right,
so on that board, proc-xscale.S will incorrectly assume that we have a
DSP coprocessor, enable CP0 on boot, and we will then only save the
first iWMMXt register (wR0) on context switches, which is Bad.
This patch redefines CONFIG_IWMMXT as 'the cpu we will be running on
might have iWMMXt support, and we will enable iWMMXt context switching
if it does.' This means that with this patch, running a CONFIG_IWMMXT=n
kernel on an iWMMXt-capable CPU will no longer potentially corrupt iWMMXt
state over context switches, and running a CONFIG_IWMMXT=y kernel on a
non-iWMMXt capable CPU will still do DSP context save/restore.
These changes should make iWMMXt work on PXA3xx, and as a side effect,
enable proper acc0 save/restore on non-iWMMXt capable xsc3 cores such
as IOP13xx and IXP23xx (which will not have CONFIG_CPU_XSCALE defined),
as well as setting and using HWCAP_IWMMXT properly.
Signed-off-by: Lennert Buytenhek <buytenh@wantstofly.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2006-12-03 18:51:14 +01:00
|
|
|
/*
|
|
|
|
* linux/arch/arm/kernel/xscale-cp0.c
|
|
|
|
*
|
|
|
|
* XScale DSP and iWMMXt coprocessor context switching and handling
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or modify
|
|
|
|
* it under the terms of the GNU General Public License version 2 as
|
|
|
|
* published by the Free Software Foundation.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/types.h>
|
|
|
|
#include <linux/kernel.h>
|
|
|
|
#include <linux/signal.h>
|
|
|
|
#include <linux/sched.h>
|
|
|
|
#include <linux/init.h>
|
2008-09-06 13:10:45 +02:00
|
|
|
#include <linux/io.h>
|
[ARM] 3881/4: xscale: clean up cp0/cp1 handling
XScale cores either have a DSP coprocessor (which contains a single
40 bit accumulator register), or an iWMMXt coprocessor (which contains
eight 64 bit registers.)
Because of the small amount of state in the DSP coprocessor, access to
the DSP coprocessor (CP0) is always enabled, and DSP context switching
is done unconditionally on every task switch. Access to the iWMMXt
coprocessor (CP0/CP1) is enabled only when an iWMMXt instruction is
first issued, and iWMMXt context switching is done lazily.
CONFIG_IWMMXT is supposed to mean 'the cpu we will be running on will
have iWMMXt support', but boards are supposed to select this config
symbol by hand, and at least one pxa27x board doesn't get this right,
so on that board, proc-xscale.S will incorrectly assume that we have a
DSP coprocessor, enable CP0 on boot, and we will then only save the
first iWMMXt register (wR0) on context switches, which is Bad.
This patch redefines CONFIG_IWMMXT as 'the cpu we will be running on
might have iWMMXt support, and we will enable iWMMXt context switching
if it does.' This means that with this patch, running a CONFIG_IWMMXT=n
kernel on an iWMMXt-capable CPU will no longer potentially corrupt iWMMXt
state over context switches, and running a CONFIG_IWMMXT=y kernel on a
non-iWMMXt capable CPU will still do DSP context save/restore.
These changes should make iWMMXt work on PXA3xx, and as a side effect,
enable proper acc0 save/restore on non-iWMMXt capable xsc3 cores such
as IOP13xx and IXP23xx (which will not have CONFIG_CPU_XSCALE defined),
as well as setting and using HWCAP_IWMMXT properly.
Signed-off-by: Lennert Buytenhek <buytenh@wantstofly.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2006-12-03 18:51:14 +01:00
|
|
|
#include <asm/thread_notify.h>
|
ARM: make xscale iwmmxt code multiplatform aware
In a multiplatform configuration, we may end up building a kernel for
both Marvell PJ1 and an ARMv4 CPU implementation. In that case, the
xscale-cp0 code is built with gcc -march=armv4{,t}, which results in a
build error from the coprocessor instructions.
Since we know this code will only have to run on an actual xscale
processor, we can simply build the entire file for ARMv5TE.
Related to this, we need to handle the iWMMXT initialization sequence
differently during boot, to ensure we don't try to touch xscale
specific registers on other CPUs from the xscale_cp0_init initcall.
cpu_is_xscale() used to be hardcoded to '1' in any configuration that
enables any XScale-compatible core, but this breaks once we can have a
combined kernel with MMP1 and something else.
In this patch, I replace the existing cpu_is_xscale() macro with a new
cpu_is_xscale_family() macro that evaluates true for xscale, xsc3 and
mohawk, which makes the behavior more deterministic.
The two existing users of cpu_is_xscale() are modified accordingly,
but slightly change behavior for kernels that enable CPU_MOHAWK without
also enabling CPU_XSCALE or CPU_XSC3. Previously, these would leave leave
PMD_BIT4 in the page tables untouched, now they clear it as we've always
done for kernels that enable both MOHAWK and the support for the older
CPU types.
Since the previous behavior was inconsistent, I assume it was
unintentional.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
2014-04-15 15:38:39 +02:00
|
|
|
#include <asm/cputype.h>
|
|
|
|
|
|
|
|
asm(" .arch armv5te\n");
|
[ARM] 3881/4: xscale: clean up cp0/cp1 handling
XScale cores either have a DSP coprocessor (which contains a single
40 bit accumulator register), or an iWMMXt coprocessor (which contains
eight 64 bit registers.)
Because of the small amount of state in the DSP coprocessor, access to
the DSP coprocessor (CP0) is always enabled, and DSP context switching
is done unconditionally on every task switch. Access to the iWMMXt
coprocessor (CP0/CP1) is enabled only when an iWMMXt instruction is
first issued, and iWMMXt context switching is done lazily.
CONFIG_IWMMXT is supposed to mean 'the cpu we will be running on will
have iWMMXt support', but boards are supposed to select this config
symbol by hand, and at least one pxa27x board doesn't get this right,
so on that board, proc-xscale.S will incorrectly assume that we have a
DSP coprocessor, enable CP0 on boot, and we will then only save the
first iWMMXt register (wR0) on context switches, which is Bad.
This patch redefines CONFIG_IWMMXT as 'the cpu we will be running on
might have iWMMXt support, and we will enable iWMMXt context switching
if it does.' This means that with this patch, running a CONFIG_IWMMXT=n
kernel on an iWMMXt-capable CPU will no longer potentially corrupt iWMMXt
state over context switches, and running a CONFIG_IWMMXT=y kernel on a
non-iWMMXt capable CPU will still do DSP context save/restore.
These changes should make iWMMXt work on PXA3xx, and as a side effect,
enable proper acc0 save/restore on non-iWMMXt capable xsc3 cores such
as IOP13xx and IXP23xx (which will not have CONFIG_CPU_XSCALE defined),
as well as setting and using HWCAP_IWMMXT properly.
Signed-off-by: Lennert Buytenhek <buytenh@wantstofly.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2006-12-03 18:51:14 +01:00
|
|
|
|
|
|
|
static inline void dsp_save_state(u32 *state)
|
|
|
|
{
|
|
|
|
__asm__ __volatile__ (
|
|
|
|
"mrrc p0, 0, %0, %1, c0\n"
|
|
|
|
: "=r" (state[0]), "=r" (state[1]));
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void dsp_load_state(u32 *state)
|
|
|
|
{
|
|
|
|
__asm__ __volatile__ (
|
|
|
|
"mcrr p0, 0, %0, %1, c0\n"
|
|
|
|
: : "r" (state[0]), "r" (state[1]));
|
|
|
|
}
|
|
|
|
|
|
|
|
static int dsp_do(struct notifier_block *self, unsigned long cmd, void *t)
|
|
|
|
{
|
|
|
|
struct thread_info *thread = t;
|
|
|
|
|
|
|
|
switch (cmd) {
|
|
|
|
case THREAD_NOTIFY_FLUSH:
|
|
|
|
thread->cpu_context.extra[0] = 0;
|
|
|
|
thread->cpu_context.extra[1] = 0;
|
|
|
|
break;
|
|
|
|
|
|
|
|
case THREAD_NOTIFY_SWITCH:
|
|
|
|
dsp_save_state(current_thread_info()->cpu_context.extra);
|
|
|
|
dsp_load_state(thread->cpu_context.extra);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NOTIFY_DONE;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct notifier_block dsp_notifier_block = {
|
|
|
|
.notifier_call = dsp_do,
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
|
|
#ifdef CONFIG_IWMMXT
|
|
|
|
static int iwmmxt_do(struct notifier_block *self, unsigned long cmd, void *t)
|
|
|
|
{
|
|
|
|
struct thread_info *thread = t;
|
|
|
|
|
|
|
|
switch (cmd) {
|
|
|
|
case THREAD_NOTIFY_FLUSH:
|
|
|
|
/*
|
|
|
|
* flush_thread() zeroes thread->fpstate, so no need
|
|
|
|
* to do anything here.
|
|
|
|
*
|
|
|
|
* FALLTHROUGH: Ensure we don't try to overwrite our newly
|
|
|
|
* initialised state information on the first fault.
|
|
|
|
*/
|
|
|
|
|
2009-12-18 15:34:43 +01:00
|
|
|
case THREAD_NOTIFY_EXIT:
|
[ARM] 3881/4: xscale: clean up cp0/cp1 handling
XScale cores either have a DSP coprocessor (which contains a single
40 bit accumulator register), or an iWMMXt coprocessor (which contains
eight 64 bit registers.)
Because of the small amount of state in the DSP coprocessor, access to
the DSP coprocessor (CP0) is always enabled, and DSP context switching
is done unconditionally on every task switch. Access to the iWMMXt
coprocessor (CP0/CP1) is enabled only when an iWMMXt instruction is
first issued, and iWMMXt context switching is done lazily.
CONFIG_IWMMXT is supposed to mean 'the cpu we will be running on will
have iWMMXt support', but boards are supposed to select this config
symbol by hand, and at least one pxa27x board doesn't get this right,
so on that board, proc-xscale.S will incorrectly assume that we have a
DSP coprocessor, enable CP0 on boot, and we will then only save the
first iWMMXt register (wR0) on context switches, which is Bad.
This patch redefines CONFIG_IWMMXT as 'the cpu we will be running on
might have iWMMXt support, and we will enable iWMMXt context switching
if it does.' This means that with this patch, running a CONFIG_IWMMXT=n
kernel on an iWMMXt-capable CPU will no longer potentially corrupt iWMMXt
state over context switches, and running a CONFIG_IWMMXT=y kernel on a
non-iWMMXt capable CPU will still do DSP context save/restore.
These changes should make iWMMXt work on PXA3xx, and as a side effect,
enable proper acc0 save/restore on non-iWMMXt capable xsc3 cores such
as IOP13xx and IXP23xx (which will not have CONFIG_CPU_XSCALE defined),
as well as setting and using HWCAP_IWMMXT properly.
Signed-off-by: Lennert Buytenhek <buytenh@wantstofly.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2006-12-03 18:51:14 +01:00
|
|
|
iwmmxt_task_release(thread);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case THREAD_NOTIFY_SWITCH:
|
|
|
|
iwmmxt_task_switch(thread);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NOTIFY_DONE;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct notifier_block iwmmxt_notifier_block = {
|
|
|
|
.notifier_call = iwmmxt_do,
|
|
|
|
};
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
|
|
static u32 __init xscale_cp_access_read(void)
|
|
|
|
{
|
|
|
|
u32 value;
|
|
|
|
|
|
|
|
__asm__ __volatile__ (
|
|
|
|
"mrc p15, 0, %0, c15, c1, 0\n\t"
|
|
|
|
: "=r" (value));
|
|
|
|
|
|
|
|
return value;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __init xscale_cp_access_write(u32 value)
|
|
|
|
{
|
|
|
|
u32 temp;
|
|
|
|
|
|
|
|
__asm__ __volatile__ (
|
|
|
|
"mcr p15, 0, %1, c15, c1, 0\n\t"
|
|
|
|
"mrc p15, 0, %0, c15, c1, 0\n\t"
|
|
|
|
"mov %0, %0\n\t"
|
|
|
|
"sub pc, pc, #4\n\t"
|
|
|
|
: "=r" (temp) : "r" (value));
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Detect whether we have a MAC coprocessor (40 bit register) or an
|
|
|
|
* iWMMXt coprocessor (64 bit registers) by loading 00000100:00000000
|
|
|
|
* into a coprocessor register and reading it back, and checking
|
|
|
|
* whether the upper word survived intact.
|
|
|
|
*/
|
|
|
|
static int __init cpu_has_iwmmxt(void)
|
|
|
|
{
|
|
|
|
u32 lo;
|
|
|
|
u32 hi;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This sequence is interpreted by the DSP coprocessor as:
|
|
|
|
* mar acc0, %2, %3
|
|
|
|
* mra %0, %1, acc0
|
|
|
|
*
|
|
|
|
* And by the iWMMXt coprocessor as:
|
|
|
|
* tmcrr wR0, %2, %3
|
|
|
|
* tmrrc %0, %1, wR0
|
|
|
|
*/
|
|
|
|
__asm__ __volatile__ (
|
|
|
|
"mcrr p0, 0, %2, %3, c0\n"
|
|
|
|
"mrrc p0, 0, %0, %1, c0\n"
|
|
|
|
: "=r" (lo), "=r" (hi)
|
|
|
|
: "r" (0), "r" (0x100));
|
|
|
|
|
|
|
|
return !!hi;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we detect that the CPU has iWMMXt (and CONFIG_IWMMXT=y), we
|
|
|
|
* disable CP0/CP1 on boot, and let call_fpe() and the iWMMXt lazy
|
|
|
|
* switch code handle iWMMXt context switching. If on the other
|
|
|
|
* hand the CPU has a DSP coprocessor, we keep access to CP0 enabled
|
|
|
|
* all the time, and save/restore acc0 on context switch in non-lazy
|
|
|
|
* fashion.
|
|
|
|
*/
|
|
|
|
static int __init xscale_cp0_init(void)
|
|
|
|
{
|
|
|
|
u32 cp_access;
|
|
|
|
|
ARM: make xscale iwmmxt code multiplatform aware
In a multiplatform configuration, we may end up building a kernel for
both Marvell PJ1 and an ARMv4 CPU implementation. In that case, the
xscale-cp0 code is built with gcc -march=armv4{,t}, which results in a
build error from the coprocessor instructions.
Since we know this code will only have to run on an actual xscale
processor, we can simply build the entire file for ARMv5TE.
Related to this, we need to handle the iWMMXT initialization sequence
differently during boot, to ensure we don't try to touch xscale
specific registers on other CPUs from the xscale_cp0_init initcall.
cpu_is_xscale() used to be hardcoded to '1' in any configuration that
enables any XScale-compatible core, but this breaks once we can have a
combined kernel with MMP1 and something else.
In this patch, I replace the existing cpu_is_xscale() macro with a new
cpu_is_xscale_family() macro that evaluates true for xscale, xsc3 and
mohawk, which makes the behavior more deterministic.
The two existing users of cpu_is_xscale() are modified accordingly,
but slightly change behavior for kernels that enable CPU_MOHAWK without
also enabling CPU_XSCALE or CPU_XSC3. Previously, these would leave leave
PMD_BIT4 in the page tables untouched, now they clear it as we've always
done for kernels that enable both MOHAWK and the support for the older
CPU types.
Since the previous behavior was inconsistent, I assume it was
unintentional.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
2014-04-15 15:38:39 +02:00
|
|
|
/* do not attempt to probe iwmmxt on non-xscale family CPUs */
|
|
|
|
if (!cpu_is_xscale_family())
|
|
|
|
return 0;
|
|
|
|
|
[ARM] 3881/4: xscale: clean up cp0/cp1 handling
XScale cores either have a DSP coprocessor (which contains a single
40 bit accumulator register), or an iWMMXt coprocessor (which contains
eight 64 bit registers.)
Because of the small amount of state in the DSP coprocessor, access to
the DSP coprocessor (CP0) is always enabled, and DSP context switching
is done unconditionally on every task switch. Access to the iWMMXt
coprocessor (CP0/CP1) is enabled only when an iWMMXt instruction is
first issued, and iWMMXt context switching is done lazily.
CONFIG_IWMMXT is supposed to mean 'the cpu we will be running on will
have iWMMXt support', but boards are supposed to select this config
symbol by hand, and at least one pxa27x board doesn't get this right,
so on that board, proc-xscale.S will incorrectly assume that we have a
DSP coprocessor, enable CP0 on boot, and we will then only save the
first iWMMXt register (wR0) on context switches, which is Bad.
This patch redefines CONFIG_IWMMXT as 'the cpu we will be running on
might have iWMMXt support, and we will enable iWMMXt context switching
if it does.' This means that with this patch, running a CONFIG_IWMMXT=n
kernel on an iWMMXt-capable CPU will no longer potentially corrupt iWMMXt
state over context switches, and running a CONFIG_IWMMXT=y kernel on a
non-iWMMXt capable CPU will still do DSP context save/restore.
These changes should make iWMMXt work on PXA3xx, and as a side effect,
enable proper acc0 save/restore on non-iWMMXt capable xsc3 cores such
as IOP13xx and IXP23xx (which will not have CONFIG_CPU_XSCALE defined),
as well as setting and using HWCAP_IWMMXT properly.
Signed-off-by: Lennert Buytenhek <buytenh@wantstofly.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2006-12-03 18:51:14 +01:00
|
|
|
cp_access = xscale_cp_access_read() & ~3;
|
|
|
|
xscale_cp_access_write(cp_access | 1);
|
|
|
|
|
|
|
|
if (cpu_has_iwmmxt()) {
|
|
|
|
#ifndef CONFIG_IWMMXT
|
2014-10-28 12:26:42 +01:00
|
|
|
pr_warn("CAUTION: XScale iWMMXt coprocessor detected, but kernel support is missing.\n");
|
[ARM] 3881/4: xscale: clean up cp0/cp1 handling
XScale cores either have a DSP coprocessor (which contains a single
40 bit accumulator register), or an iWMMXt coprocessor (which contains
eight 64 bit registers.)
Because of the small amount of state in the DSP coprocessor, access to
the DSP coprocessor (CP0) is always enabled, and DSP context switching
is done unconditionally on every task switch. Access to the iWMMXt
coprocessor (CP0/CP1) is enabled only when an iWMMXt instruction is
first issued, and iWMMXt context switching is done lazily.
CONFIG_IWMMXT is supposed to mean 'the cpu we will be running on will
have iWMMXt support', but boards are supposed to select this config
symbol by hand, and at least one pxa27x board doesn't get this right,
so on that board, proc-xscale.S will incorrectly assume that we have a
DSP coprocessor, enable CP0 on boot, and we will then only save the
first iWMMXt register (wR0) on context switches, which is Bad.
This patch redefines CONFIG_IWMMXT as 'the cpu we will be running on
might have iWMMXt support, and we will enable iWMMXt context switching
if it does.' This means that with this patch, running a CONFIG_IWMMXT=n
kernel on an iWMMXt-capable CPU will no longer potentially corrupt iWMMXt
state over context switches, and running a CONFIG_IWMMXT=y kernel on a
non-iWMMXt capable CPU will still do DSP context save/restore.
These changes should make iWMMXt work on PXA3xx, and as a side effect,
enable proper acc0 save/restore on non-iWMMXt capable xsc3 cores such
as IOP13xx and IXP23xx (which will not have CONFIG_CPU_XSCALE defined),
as well as setting and using HWCAP_IWMMXT properly.
Signed-off-by: Lennert Buytenhek <buytenh@wantstofly.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2006-12-03 18:51:14 +01:00
|
|
|
#else
|
2014-10-28 12:26:42 +01:00
|
|
|
pr_info("XScale iWMMXt coprocessor detected.\n");
|
[ARM] 3881/4: xscale: clean up cp0/cp1 handling
XScale cores either have a DSP coprocessor (which contains a single
40 bit accumulator register), or an iWMMXt coprocessor (which contains
eight 64 bit registers.)
Because of the small amount of state in the DSP coprocessor, access to
the DSP coprocessor (CP0) is always enabled, and DSP context switching
is done unconditionally on every task switch. Access to the iWMMXt
coprocessor (CP0/CP1) is enabled only when an iWMMXt instruction is
first issued, and iWMMXt context switching is done lazily.
CONFIG_IWMMXT is supposed to mean 'the cpu we will be running on will
have iWMMXt support', but boards are supposed to select this config
symbol by hand, and at least one pxa27x board doesn't get this right,
so on that board, proc-xscale.S will incorrectly assume that we have a
DSP coprocessor, enable CP0 on boot, and we will then only save the
first iWMMXt register (wR0) on context switches, which is Bad.
This patch redefines CONFIG_IWMMXT as 'the cpu we will be running on
might have iWMMXt support, and we will enable iWMMXt context switching
if it does.' This means that with this patch, running a CONFIG_IWMMXT=n
kernel on an iWMMXt-capable CPU will no longer potentially corrupt iWMMXt
state over context switches, and running a CONFIG_IWMMXT=y kernel on a
non-iWMMXt capable CPU will still do DSP context save/restore.
These changes should make iWMMXt work on PXA3xx, and as a side effect,
enable proper acc0 save/restore on non-iWMMXt capable xsc3 cores such
as IOP13xx and IXP23xx (which will not have CONFIG_CPU_XSCALE defined),
as well as setting and using HWCAP_IWMMXT properly.
Signed-off-by: Lennert Buytenhek <buytenh@wantstofly.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2006-12-03 18:51:14 +01:00
|
|
|
elf_hwcap |= HWCAP_IWMMXT;
|
|
|
|
thread_register_notifier(&iwmmxt_notifier_block);
|
|
|
|
#endif
|
|
|
|
} else {
|
2014-10-28 12:26:42 +01:00
|
|
|
pr_info("XScale DSP coprocessor detected.\n");
|
[ARM] 3881/4: xscale: clean up cp0/cp1 handling
XScale cores either have a DSP coprocessor (which contains a single
40 bit accumulator register), or an iWMMXt coprocessor (which contains
eight 64 bit registers.)
Because of the small amount of state in the DSP coprocessor, access to
the DSP coprocessor (CP0) is always enabled, and DSP context switching
is done unconditionally on every task switch. Access to the iWMMXt
coprocessor (CP0/CP1) is enabled only when an iWMMXt instruction is
first issued, and iWMMXt context switching is done lazily.
CONFIG_IWMMXT is supposed to mean 'the cpu we will be running on will
have iWMMXt support', but boards are supposed to select this config
symbol by hand, and at least one pxa27x board doesn't get this right,
so on that board, proc-xscale.S will incorrectly assume that we have a
DSP coprocessor, enable CP0 on boot, and we will then only save the
first iWMMXt register (wR0) on context switches, which is Bad.
This patch redefines CONFIG_IWMMXT as 'the cpu we will be running on
might have iWMMXt support, and we will enable iWMMXt context switching
if it does.' This means that with this patch, running a CONFIG_IWMMXT=n
kernel on an iWMMXt-capable CPU will no longer potentially corrupt iWMMXt
state over context switches, and running a CONFIG_IWMMXT=y kernel on a
non-iWMMXt capable CPU will still do DSP context save/restore.
These changes should make iWMMXt work on PXA3xx, and as a side effect,
enable proper acc0 save/restore on non-iWMMXt capable xsc3 cores such
as IOP13xx and IXP23xx (which will not have CONFIG_CPU_XSCALE defined),
as well as setting and using HWCAP_IWMMXT properly.
Signed-off-by: Lennert Buytenhek <buytenh@wantstofly.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2006-12-03 18:51:14 +01:00
|
|
|
thread_register_notifier(&dsp_notifier_block);
|
|
|
|
cp_access |= 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
xscale_cp_access_write(cp_access);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
late_initcall(xscale_cp0_init);
|