[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>

arch/arm/Kconfig

@@ -374,6 +374,14 @@ config PLAT_IOP
 
 source arch/arm/mm/Kconfig
 
+config IWMMXT
+	bool "Enable iWMMXt support"
+	depends CPU_XSCALE || CPU_XSC3
+	default y if PXA27x
+	help
+	  Enable support for iWMMXt context switching at run time if
+	  running on a CPU that supports it.
+
 #  bool 'Use XScale PMU as timer source' CONFIG_XSCALE_PMU_TIMER
 config XSCALE_PMU
 	bool

arch/arm/kernel/Makefile

@@ -24,7 +24,9 @@ obj-$(CONFIG_OABI_COMPAT)	+= sys_oabi-compat.o
 obj-$(CONFIG_CRUNCH)		+= crunch.o crunch-bits.o
 AFLAGS_crunch-bits.o		:= -Wa,-mcpu=ep9312
 
-obj-$(CONFIG_IWMMXT)		+= iwmmxt.o iwmmxt-notifier.o
+obj-$(CONFIG_CPU_XSCALE)	+= xscale-cp0.o
+obj-$(CONFIG_CPU_XSC3)		+= xscale-cp0.o
+obj-$(CONFIG_IWMMXT)		+= iwmmxt.o
 AFLAGS_iwmmxt.o			:= -Wa,-mcpu=iwmmxt
 
 ifneq ($(CONFIG_ARCH_EBSA110),y)

arch/arm/kernel/entry-armv.S

@@ -589,10 +589,6 @@ ENTRY(__switch_to)
 	strex	r5, r4, [ip]			@ Clear exclusive monitor
 #endif
 #endif
-#if defined(CONFIG_CPU_XSCALE) && !defined(CONFIG_IWMMXT)
-	mra	r4, r5, acc0
-	stmia   ip, {r4, r5}
-#endif
 #if defined(CONFIG_HAS_TLS_REG)
 	mcr	p15, 0, r3, c13, c0, 3		@ set TLS register
 #elif !defined(CONFIG_TLS_REG_EMUL)
@@ -601,11 +597,6 @@ ENTRY(__switch_to)
 #endif
 #ifdef CONFIG_MMU
 	mcr	p15, 0, r6, c3, c0, 0		@ Set domain register
-#endif
-#if defined(CONFIG_CPU_XSCALE) && !defined(CONFIG_IWMMXT)
-	add	r4, r2, #TI_CPU_DOMAIN + 40	@ cpu_context_save->extra
-	ldmib	r4, {r4, r5}
-	mar	acc0, r4, r5
 #endif
 	mov	r5, r0
 	add	r4, r2, #TI_CPU_SAVE

arch/arm/kernel/iwmmxt-notifier.c

-/*
- *  linux/arch/arm/kernel/iwmmxt-notifier.c
- *
- *  XScale iWMMXt (Concan) context switching and handling
- *
- *  Initial code:
- *  Copyright (c) 2003, Intel Corporation
- *
- *  Full lazy switching support, optimizations and more, by Nicolas Pitre
- *  Copyright (c) 2003-2004, MontaVista Software, Inc.
- *
- * 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/module.h>
-#include <linux/types.h>
-#include <linux/kernel.h>
-#include <linux/signal.h>
-#include <linux/sched.h>
-#include <linux/init.h>
-#include <asm/thread_notify.h>
-#include <asm/io.h>
-
-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.
-		 */
-
-	case THREAD_NOTIFY_RELEASE:
-		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,
-};
-
-static int __init iwmmxt_init(void)
-{
-	thread_register_notifier(&iwmmxt_notifier_block);
-
-	return 0;
-}
-
-late_initcall(iwmmxt_init);

arch/arm/kernel/setup.c

@@ -357,9 +357,6 @@ static void __init setup_processor(void)
 #ifndef CONFIG_VFP
 	elf_hwcap &= ~HWCAP_VFP;
 #endif
-#ifndef CONFIG_IWMMXT
-	elf_hwcap &= ~HWCAP_IWMMXT;
-#endif
 
 	cpu_proc_init();
 }

arch/arm/kernel/xscale-cp0.c

+/*
+ * 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/module.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <linux/signal.h>
+#include <linux/sched.h>
+#include <linux/init.h>
+#include <asm/thread_notify.h>
+#include <asm/io.h>
+
+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.
+		 */
+
+	case THREAD_NOTIFY_RELEASE:
+		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;
+
+	cp_access = xscale_cp_access_read() & ~3;
+	xscale_cp_access_write(cp_access | 1);
+
+	if (cpu_has_iwmmxt()) {
+#ifndef CONFIG_IWMMXT
+		printk(KERN_WARNING "CAUTION: XScale iWMMXt coprocessor "
+			"detected, but kernel support is missing.\n");
+#else
+		printk(KERN_INFO "XScale iWMMXt coprocessor detected.\n");
+		elf_hwcap |= HWCAP_IWMMXT;
+		thread_register_notifier(&iwmmxt_notifier_block);
+#endif
+	} else {
+		printk(KERN_INFO "XScale DSP coprocessor detected.\n");
+		thread_register_notifier(&dsp_notifier_block);
+		cp_access |= 1;
+	}
+
+	xscale_cp_access_write(cp_access);
+
+	return 0;
+}
+
+late_initcall(xscale_cp0_init);

arch/arm/mach-pxa/Kconfig

@@ -13,12 +13,10 @@ config ARCH_LUBBOCK
 config MACH_LOGICPD_PXA270
 	bool "LogicPD PXA270 Card Engine Development Platform"
 	select PXA27x
-	select IWMMXT
 
 config MACH_MAINSTONE
 	bool "Intel HCDDBBVA0 Development Platform"
 	select PXA27x
-	select IWMMXT
 
 config ARCH_PXA_IDP
 	bool "Accelent Xscale IDP"
@@ -53,7 +51,6 @@ config PXA_SHARPSL_25x
 config PXA_SHARPSL_27x
 	bool "Sharp PXA270 models (SL-Cxx00)"
 	select PXA27x
-	select IWMMXT
 
 endchoice
 
@@ -129,11 +126,6 @@ config PXA27x
 	help
 	  Select code specific to PXA27x variants
 
-config IWMMXT
-	bool
-	help
-	  Enable support for iWMMXt
-
 config PXA_SHARP_C7xx
 	bool
 	select PXA_SSP

arch/arm/mach-pxa/pm.c

@@ -83,7 +83,8 @@ int pxa_pm_enter(suspend_state_t state)
 
 #ifdef CONFIG_IWMMXT
 	/* force any iWMMXt context to ram **/
-	iwmmxt_task_disable(NULL);
+	if (elf_hwcap & HWCAP_IWMMXT)
+		iwmmxt_task_disable(NULL);
 #endif
 
 	/* preserve current time */

arch/arm/mm/proc-xscale.S

@@ -491,12 +491,7 @@ __xscale_setup:
 	mcr	p15, 0, ip, c7, c7, 0		@ invalidate I, D caches & BTB
 	mcr	p15, 0, ip, c7, c10, 4		@ Drain Write (& Fill) Buffer
 	mcr	p15, 0, ip, c8, c7, 0		@ invalidate I, D TLBs
-#ifdef CONFIG_IWMMXT
-	mov	r0, #0				@ initially disallow access to CP0/CP1
-#else
-	mov	r0, #1				@ Allow access to CP0
-#endif
-	orr     r0, r0, #1 << 6			@ cp6 for IOP3xx and Bulverde
+	mov	r0, #1 << 6			@ cp6 for IOP3xx and Bulverde
 	orr	r0, r0, #1 << 13		@ Its undefined whether this
 	mcr	p15, 0, r0, c15, c1, 0		@ affects USR or SVC modes
 
@@ -909,7 +904,7 @@ __pxa270_proc_info:
 	b	__xscale_setup
 	.long	cpu_arch_name
 	.long	cpu_elf_name
-	.long	HWCAP_SWP|HWCAP_HALF|HWCAP_THUMB|HWCAP_FAST_MULT|HWCAP_EDSP|HWCAP_IWMMXT
+	.long	HWCAP_SWP|HWCAP_HALF|HWCAP_THUMB|HWCAP_FAST_MULT|HWCAP_EDSP
 	.long	cpu_pxa270_name
 	.long	xscale_processor_functions
 	.long	v4wbi_tlb_fns

include/asm-arm/elf.h

@@ -114,40 +114,24 @@ extern char elf_platform[];
    have no such handler.  */
 #define ELF_PLAT_INIT(_r, load_addr)	(_r)->ARM_r0 = 0
 
-#ifndef CONFIG_IWMMXT
-
-/* Old NetWinder binaries were compiled in such a way that the iBCS
-   heuristic always trips on them.  Until these binaries become uncommon
-   enough not to care, don't trust the `ibcs' flag here.  In any case
-   there is no other ELF system currently supported by iBCS.
-   @@ Could print a warning message to encourage users to upgrade.  */
-#define SET_PERSONALITY(ex,ibcs2) \
-	set_personality(((ex).e_flags & EF_ARM_APCS26 ? PER_LINUX : PER_LINUX_32BIT))
-
-#else
-
 /*
- * All iWMMXt capable CPUs don't support 26-bit mode.  Yet they can run
- * legacy binaries which used to contain FPA11 floating point instructions
- * that have always been emulated by the kernel.  PFA11 and iWMMXt overlap
- * on coprocessor 1 space though.  We therefore must decide if given task
- * is allowed to use CP 0 and 1 for iWMMXt, or if they should be blocked
- * at all times for the prefetch exception handler to catch FPA11 opcodes
- * and emulate them.  The best indication to discriminate those two cases
- * is the SOFT_FLOAT flag in the ELF header.
+ * Since the FPA coprocessor uses CP1 and CP2, and iWMMXt uses CP0
+ * and CP1, we only enable access to the iWMMXt coprocessor if the
+ * binary is EABI or softfloat (and thus, guaranteed not to use
+ * FPA instructions.)
  */
-
-#define SET_PERSONALITY(ex,ibcs2) \
-do { \
-	set_personality(PER_LINUX_32BIT); \
-	if (((ex).e_flags & EF_ARM_EABI_MASK) || \
-	    ((ex).e_flags & EF_ARM_SOFT_FLOAT)) \
-		set_thread_flag(TIF_USING_IWMMXT); \
-	else \
-		clear_thread_flag(TIF_USING_IWMMXT); \
-} while (0)
-
-#endif
+#define SET_PERSONALITY(ex, ibcs2)					\
+	do {								\
+		if ((ex).e_flags & EF_ARM_APCS26) {			\
+			set_personality(PER_LINUX);			\
+		} else {						\
+			set_personality(PER_LINUX_32BIT);		\
+			if (elf_hwcap & HWCAP_IWMMXT && (ex).e_flags & (EF_ARM_EABI_MASK | EF_ARM_SOFT_FLOAT)) \
+				set_thread_flag(TIF_USING_IWMMXT);	\
+			else						\
+				clear_thread_flag(TIF_USING_IWMMXT);	\
+		}							\
+	} while (0)
 
 #endif