Skip to contentSkip to conversation

VFP save/restore for OMAP3

PATCH Superseded Tero Kristo 4 participants 10 messages Nov 24, 2009 – Nov 27, 2009

Subsystems (5)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover VFP save/restore for OMAP3

From: Tero Kristo <[email protected]>

Hi,

The following two patches are needed by OMAP3 to save/restore VFP context during off-mode. Patch 1 adds generic support inside ARM VFP code, and the second one adds the necessary hooks into OMAP3 power management code.

--Tero

#1 lore raw

PATCH 1/2 ARM: Implemented support for VFP PM context saving

From: Tero Kristo <[email protected]>

In some ARM architectures, like OMAP3, the VFP context can be lost during dynamic sleep cycle. For this purpose, there is now a function vfp_pm_save_context() that should be called before the VFP is assumed to lose context. Next VFP trap will then restore context automatically.

We need to have the last_VFP_context[cpu] cleared after the save in idle, else the restore would fail to restore when it sees that the last_VFP_context is same as the current threads vfp_state. This happens when the same process/thread traps an exception post idle.

Main work for this patch was done by Peter and Rajendra. Some cleanup and optimization by Tero.

Cc Vishwanath Sripathy <[email protected]>
Cc Rajendra Nayak <[email protected]>
Cc Richard Woodruff <[email protected]>
Cc Peter 'p2' De Schrijver <[email protected]>
---
 arch/arm/vfp/vfpmodule.c |   25 +++++++++++++++++++++++++
 1 files changed, 25 insertions(+), 0 deletions(-)
arch/arm/vfp/vfpmodule.c +25 -0
index 2d7423a..80a08bd 100644
@@ -329,6 +329,31 @@ static void vfp_enable(void *unused)
 #ifdef CONFIG_PM
 #include <linux/sysdev.h>
 
+void vfp_pm_save_context(void)
+{
+	struct thread_info *thread = current_thread_info();
+	u32 fpexc = fmrx(FPEXC);
+	__u32 cpu = thread->cpu;
+
+	if (last_VFP_context[cpu]) {
+		if (!(fpexc & FPEXC_EN)) {
+			/* enable vfp now to save context */
+			vfp_enable(NULL);
+			fmxr(FPEXC, fmrx(FPEXC) | FPEXC_EN);
+		}
+		vfp_save_state(last_VFP_context[cpu], fpexc);
+
+		/* Disable vfp. The next inst traps an exception and restores*/
+		fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
+
+		/*
+		 * This is needed else the restore might fail if it sees
+		 * last_VFP_context if same as the current threads vfp_state.
+		 */
+		last_VFP_context[cpu] = NULL;
+	}
+}
+
 static int vfp_pm_suspend(struct sys_device *dev, pm_message_t state)
 {
 	struct thread_info *ti = current_thread_info();
Signature · 1 line
1.5.4.3
#2 lore raw

PATCH 2/2 OMAP3: Implemented VFP restore/save context

From: Tero Kristo <[email protected]>

VFP save context is called before MPU/NEON off. Restore is not needed as the next VFP trap will restore context automatically. Uses the support routine implemented in arch/arm/vfp/vfpmodule.c.

Cc Vishwanath Sripathy <[email protected]>
Cc Rajendra Nayak <[email protected]>
Cc Richard Woodruff <[email protected]>
Cc Peter 'p2' De Schrijver <[email protected]>
---
 arch/arm/mach-omap2/pm.h     |    1 +
 arch/arm/mach-omap2/pm34xx.c |   21 ++++++++++++++++++++-
 2 files changed, 21 insertions(+), 1 deletions(-)
arch/arm/mach-omap2/pm.h +1 -0
index 4f22107..dd5bbaf 100644
@@ -18,6 +18,7 @@ extern u32 sleep_while_idle;
 extern u32 voltage_off_while_idle;
 
 extern void *omap3_secure_ram_storage;
+extern void vfp_pm_save_context(void);
 extern void omap3_pm_off_mode_enable(int);
 extern void omap_sram_idle(void);
 extern int omap3_can_sleep(void);
arch/arm/mach-omap2/pm34xx.c +20 -1
index b26ae5b..4b01303 100644
@@ -321,6 +321,18 @@ static void restore_control_register(u32 val)
 	__asm__ __volatile__ ("mcr p15, 0, %0, c1, c0, 0" : : "r" (val));
 }
 
+static inline void omap3_save_neon_context(void)
+{
+#ifdef CONFIG_VFP
+	vfp_pm_save_context();
+#endif
+}
+
+static inline void omap3_restore_neon_context(void)
+{
+	return;
+}
+
 /* Function to restore the table entry that was modified for enabling MMU */
 static void restore_table_entry(void)
 {
@@ -365,6 +377,7 @@ void omap_sram_idle(void)
 	/* save_state = 3 => L1, L2 and logic lost */
 	int save_state = 0;
 	int mpu_next_state = PWRDM_POWER_ON;
+	int neon_next_state = PWRDM_POWER_ON;
 	int per_next_state = PWRDM_POWER_ON;
 	int core_next_state = PWRDM_POWER_ON;
 	int core_prev_state, per_prev_state;
@@ -398,8 +411,12 @@ void omap_sram_idle(void)
 	pwrdm_pre_transition();
 
 	/* NEON control */
-	if (pwrdm_read_pwrst(neon_pwrdm) == PWRDM_POWER_ON)
+	if (pwrdm_read_pwrst(neon_pwrdm) == PWRDM_POWER_ON) {
 		pwrdm_set_next_pwrst(neon_pwrdm, mpu_next_state);
+		neon_next_state = mpu_next_state;
+		if (neon_next_state == PWRDM_POWER_OFF)
+			omap3_save_neon_context();
+	}
 
 	/* PER */
 	per_next_state = pwrdm_read_next_pwrst(per_pwrdm);
@@ -537,6 +554,8 @@ void omap_sram_idle(void)
 		omap3_disable_io_chain();
 	}
 
+	if (neon_next_state == PWRDM_POWER_OFF)
+		omap3_restore_neon_context();
 
 	pwrdm_post_transition();
 }
Signature · 1 line
1.5.4.3
#3 lore raw

PATCH 2/2 OMAP3: Implemented VFP restore/save context

Quoted 85 lines

-----Original Message-----
From: linux-omap-owner at vger.kernel.org [mailto:linux-omap- owner at vger.kernel.org] On Behalf Of Tero Kristo
Sent: Tuesday, November 24, 2009 4:07 PM
To: linux-arm-kernel at lists.infradead.org
Cc: linux-omap at vger.kernel.org
Subject: [PATCH 2/2] OMAP3: Implemented VFP restore/save context

From: Tero Kristo <[email protected]>

VFP save context is called before MPU/NEON off. Restore is not needed as the next VFP trap will restore context automatically. Uses the support routine implemented in arch/arm/vfp/vfpmodule.c.

Cc Vishwanath Sripathy <[email protected]>
Cc Rajendra Nayak <[email protected]>
Cc Richard Woodruff <[email protected]>
Cc Peter 'p2' De Schrijver <[email protected]>
---
 arch/arm/mach-omap2/pm.h     |    1 +
 arch/arm/mach-omap2/pm34xx.c |   21 ++++++++++++++++++++-
 2 files changed, 21 insertions(+), 1 deletions(-)
arch/arm/mach-omap2/pm.h +1 -0
index 4f22107..dd5bbaf 100644
@@ -18,6 +18,7 @@ extern u32 sleep_while_idle;
 extern u32 voltage_off_while_idle;
 
 extern void *omap3_secure_ram_storage;
+extern void vfp_pm_save_context(void);
 extern void omap3_pm_off_mode_enable(int);
 extern void omap_sram_idle(void);
 extern int omap3_can_sleep(void);
arch/arm/mach-omap2/pm34xx.c +18 -1
index b26ae5b..4b01303 100644
@@ -321,6 +321,18 @@ static void restore_control_register(u32 val)
 	__asm__ __volatile__ ("mcr p15, 0, %0, c1, c0, 0" : : "r" (val));
 }
 
+static inline void omap3_save_neon_context(void)
+{
+#ifdef CONFIG_VFP
+	vfp_pm_save_context();
+#endif
+}
+
+static inline void omap3_restore_neon_context(void)
+{
+	return;
+}
+
 /* Function to restore the table entry that was modified for enabling MMU */
 static void restore_table_entry(void)
 {
@@ -365,6 +377,7 @@ void omap_sram_idle(void)
 	/* save_state = 3 => L1, L2 and logic lost */
 	int save_state = 0;
 	int mpu_next_state = PWRDM_POWER_ON;
+	int neon_next_state = PWRDM_POWER_ON;
 	int per_next_state = PWRDM_POWER_ON;
 	int core_next_state = PWRDM_POWER_ON;
 	int core_prev_state, per_prev_state;
@@ -398,8 +411,12 @@ void omap_sram_idle(void)
 	pwrdm_pre_transition();
 
 	/* NEON control */
-	if (pwrdm_read_pwrst(neon_pwrdm) == PWRDM_POWER_ON)
+	if (pwrdm_read_pwrst(neon_pwrdm) == PWRDM_POWER_ON) {
 		pwrdm_set_next_pwrst(neon_pwrdm, mpu_next_state);
+		neon_next_state = mpu_next_state;
+		if (neon_next_state == PWRDM_POWER_OFF)
+			omap3_save_neon_context();
+	}
 
 	/* PER */
 	per_next_state = pwrdm_read_next_pwrst(per_pwrdm);

arch/arm/mach-omap2/pm34xx.c void omap_sram_idle(void) · line 559

Quoted 6 lines
arch/arm/mach-omap2/pm34xx.c +2 -0
@@ -537,6 +554,8 @@ void omap_sram_idle(void)
 		omap3_disable_io_chain();
 	}
 
+	if (neon_next_state == PWRDM_POWER_OFF)
+		omap3_restore_neon_context();
 

Why do call restore here when restore is done automatically upon context switch?

Quoted 9 lines
arch/arm/mach-omap2/pm34xx.c +0 -1
 	pwrdm_post_transition();
 }
--

1.5.4.3

--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo at vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html

#4 lore raw

PATCH 2/2 OMAP3: Implemented VFP restore/save context

Quoted 102 lines

-----Original Message-----
From: ext Sripathy, Vishwanath [mailto:vishwanath.bs at ti.com]
Sent: 24 November, 2009 13:34
To: Kristo Tero (Nokia-D/Tampere); linux-arm-kernel at lists.infradead.org
Cc: linux-omap at vger.kernel.org
Subject: RE: [PATCH 2/2] OMAP3: Implemented VFP restore/save context

Quoted 11 lines

-----Original Message-----
From: linux-omap-owner at vger.kernel.org [mailto:linux-omap- owner at vger.kernel.org] On Behalf Of Tero Kristo
Sent: Tuesday, November 24, 2009 4:07 PM
To: linux-arm-kernel at lists.infradead.org
Cc: linux-omap at vger.kernel.org
Subject: [PATCH 2/2] OMAP3: Implemented VFP restore/save context

From: Tero Kristo <[email protected]>

VFP save context is called before MPU/NEON off. Restore is

not needed as

Quoted 1 line

the next VFP trap will restore context automatically. Uses

the support

Review

Quoted 26 lines

routine implemented in arch/arm/vfp/vfpmodule.c.

Cc Vishwanath Sripathy <[email protected]>
Cc Rajendra Nayak <[email protected]>
Cc Richard Woodruff <[email protected]>
Cc Peter 'p2' De Schrijver <[email protected]>
---
 arch/arm/mach-omap2/pm.h     |    1 +
 arch/arm/mach-omap2/pm34xx.c |   21 ++++++++++++++++++++-
 2 files changed, 21 insertions(+), 1 deletions(-)
arch/arm/mach-omap2/pm.h +1 -0
index 4f22107..dd5bbaf 100644
@@ -18,6 +18,7 @@ extern u32 sleep_while_idle;
 extern u32 voltage_off_while_idle;
 
 extern void *omap3_secure_ram_storage;
+extern void vfp_pm_save_context(void);
 extern void omap3_pm_off_mode_enable(int);
 extern void omap_sram_idle(void);
 extern int omap3_can_sleep(void);
(quoted diff) +0 -0

b/arch/arm/mach-omap2/pm34xx.c

arch/arm/mach-omap2/pm34xx.c static void restore_control_register(u32 val) · line 321

Quoted 5 lines

index b26ae5b..4b01303 100644

arch/arm/mach-omap2/pm34xx.c +0 -0
@@ -321,6 +321,18 @@ static void restore_control_register(u32 val)
 	__asm__ __volatile__ ("mcr p15, 0, %0, c1, c0, 0" : : 

"r" (val));

Quoted 5 lines
arch/arm/mach-omap2/pm34xx.c +3 -0
 }
 
+static inline void omap3_save_neon_context(void)
+{
+#ifdef CONFIG_VFP

arch/arm/mach-omap2/pm34xx.c static void restore_control_register(u32 val)

Quoted 10 lines
arch/arm/mach-omap2/pm34xx.c +9 -0
+	vfp_pm_save_context();
+#endif
+}
+
+static inline void omap3_restore_neon_context(void)
+{
+	return;
+}
+
 /* Function to restore the table entry that was modified 

for enabling MMU */

Quoted 28 lines
arch/arm/mach-omap2/pm34xx.c +6 -1
 static void restore_table_entry(void)
 {
@@ -365,6 +377,7 @@ void omap_sram_idle(void)
 	/* save_state = 3 => L1, L2 and logic lost */
 	int save_state = 0;
 	int mpu_next_state = PWRDM_POWER_ON;
+	int neon_next_state = PWRDM_POWER_ON;
 	int per_next_state = PWRDM_POWER_ON;
 	int core_next_state = PWRDM_POWER_ON;
 	int core_prev_state, per_prev_state;
@@ -398,8 +411,12 @@ void omap_sram_idle(void)
 	pwrdm_pre_transition();
 
 	/* NEON control */
-	if (pwrdm_read_pwrst(neon_pwrdm) == PWRDM_POWER_ON)
+	if (pwrdm_read_pwrst(neon_pwrdm) == PWRDM_POWER_ON) {
 		pwrdm_set_next_pwrst(neon_pwrdm, mpu_next_state);
+		neon_next_state = mpu_next_state;
+		if (neon_next_state == PWRDM_POWER_OFF)
+			omap3_save_neon_context();
+	}
 
 	/* PER */
 	per_next_state = pwrdm_read_next_pwrst(per_pwrdm);
@@ -537,6 +554,8 @@ void omap_sram_idle(void)
 		omap3_disable_io_chain();
 	}
 
Quoted 5 lines

arch/arm/mach-omap2/pm34xx.c void omap_sram_idle(void) · line 559

Quoted 3 lines
arch/arm/mach-omap2/pm34xx.c +2 -0
+	if (neon_next_state == PWRDM_POWER_OFF)
+		omap3_restore_neon_context();
 

Why do call restore here when restore is done automatically upon context switch?

That is just a stub if we would need to add something there at some point. restore_neon_context() currently does nothing.

Quoted 12 lines
Quoted 8 lines
arch/arm/mach-omap2/pm34xx.c +0 -1
 	pwrdm_post_transition();
 }
--

1.5.4.3

--
To unsubscribe from this list: send the line "unsubscribe

linux-omap" in

Quoted 2 lines

the body of a message to majordomo at vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html

#5 lore raw

PATCH 1/2 ARM: Implemented support for VFP PM context saving

In reply to Tero Kristo
Quoted 5 lines · Tero Kristo

In some ARM architectures, like OMAP3, the VFP context can be lost during dynamic sleep cycle. For this purpose, there is now a function vfp_pm_save_context() that should be called before the VFP is assumed to lose context. Next VFP trap will then restore context automatically.

Quoted 7 lines

We need to have the last_VFP_context[cpu] cleared after the save in idle, else the restore would fail to restore when it sees that the last_VFP_context is same as the current threads vfp_state. This happens when the same process/thread traps an exception post idle.

Main work for this patch was done by Peter and Rajendra. Some cleanup and optimization by Tero.

Why not re-use vfp_pm_suspend() ? Haven't you shown that vfp_pm_suspend may be buggy since it doesn't save in the VFP-disabled case?

#6 lore raw

PATCH 1/2 ARM: Implemented support for VFP PM context saving

Quoted 11 lines · Russell King - ARM Linux
Quoted 10 lines · Tero Kristo

In some ARM architectures, like OMAP3, the VFP context can be lost during dynamic sleep cycle. For this purpose, there is now a function vfp_pm_save_context() that should be called before the VFP is assumed to lose context. Next VFP trap will then restore context automatically.

We need to have the last_VFP_context[cpu] cleared after the save in idle, else the restore would fail to restore when it sees that the last_VFP_context is same as the current threads vfp_state. This happens when the same process/thread traps an exception post idle.

Quoted 5 lines
Quoted 2 lines

Main work for this patch was done by Peter and Rajendra. Some cleanup and optimization by Tero.

Why not re-use vfp_pm_suspend() ? Haven't you shown that vfp_pm_suspend may be buggy since it doesn't save in the VFP-disabled case?

BTW, the two patches below were mentioned to me some time ago but I haven't got the time to look at them:

[ARM] vfp: Fix bug in vfp_pm_suspend
https://www.codeaurora.org/gitweb/quic/le/?p=kernel/msm.git;a=commit;h=88984c9b2d69c222ee1e2afc948ca73f597d40ff

[ARM] vfp: Add additional vfp interfaces
https://www.codeaurora.org/gitweb/quic/le/?p=kernel/msm.git;a=commit;h=393e4bfaaf79377d29cd6bb2228f87601aeca668

Signature · 1 line
Catalin
#7 lore raw

PATCH 1/2 ARM: Implemented support for VFP PM context saving

Quoted 39 lines

-----Original Message-----
From: ext Catalin Marinas [mailto:catalin.marinas at arm.com]
Sent: 24 November, 2009 15:20
To: Russell King - ARM Linux
Cc: Kristo Tero (Nokia-D/Tampere); linux-omap at vger.kernel.org; linux-arm-kernel at lists.infradead.org; Dave Estes
Subject: Re: [PATCH 1/2] ARM: Implemented support for VFP PM context saving

Quoted 2 lines · Russell King - ARM Linux
Quoted 1 line · Tero Kristo

In some ARM architectures, like OMAP3, the VFP context can

be lost during

Quoted 2 lines
Quoted 2 lines

dynamic sleep cycle. For this purpose, there is now a function vfp_pm_save_context() that should be called before the VFP

is assumed to

Quoted 1 line
Quoted 1 line

lose context. Next VFP trap will then restore context

automatically.

Quoted 2 lines
Quoted 2 lines

We need to have the last_VFP_context[cpu] cleared after

the save in idle,

Quoted 1 line
Quoted 1 line

else the restore would fail to restore when it sees that

the last_VFP_context

Quoted 1 line
Quoted 1 line

is same as the current threads vfp_state. This happens

when the same

Quoted 3 lines
Quoted 3 lines

process/thread traps an exception post idle.

Main work for this patch was done by Peter and Rajendra.

Some cleanup and

Quoted 3 lines
Quoted 1 line

optimization by Tero.

Why not re-use vfp_pm_suspend() ? Haven't you shown that

vfp_pm_suspend

Quoted 1 line

may be buggy since it doesn't save in the VFP-disabled case?

BTW, the two patches below were mentioned to me some time ago but I haven't got the time to look at them:

Quoted 7 lines

[ARM] vfp: Fix bug in vfp_pm_suspend
https://www.codeaurora.org/gitweb/quic/le/?p=kernel/msm.git;a=c
ommit;h=88984c9b2d69c222ee1e2afc948ca73f597d40ff

[ARM] vfp: Add additional vfp interfaces
https://www.codeaurora.org/gitweb/quic/le/?p=kernel/msm.git;a=c
ommit;h=393e4bfaaf79377d29cd6bb2228f87601aeca668

These look pretty much like the same thing we have attempted on the omap3 patches, I could try these out and check if they work for omap3 also. They fix the potentially broken suspend also.

-Tero

#8 lore raw

PATCH 1/2 ARM: Implemented support for VFP PM context saving

In reply to Catalin Marinas
Quoted 5 lines · Catalin Marinas

BTW, the two patches below were mentioned to me some time ago but I haven't got the time to look at them:

[ARM] vfp: Fix bug in vfp_pm_suspend
https://www.codeaurora.org/gitweb/quic/le/?p=kernel/msm.git;a=commit;h=88984c9b2d69c222ee1e2afc948ca73f597d40ff

This one is bad - it gets the current CPU by directly referencing ti->cpu. Too bad if you have kernel preemption enabled; the value obtained that way is effectively meaningless. The only way to get a meaningful value is via 'get_cpu()' and after you've done with using it, 'put_cpu()'. That ensures you can't be preempted onto a different CPU mid-operation.

It's safe in vfp_notifier because we're called in an already atomic context.

Quoted 2 lines

[ARM] vfp: Add additional vfp interfaces
https://www.codeaurora.org/gitweb/quic/le/?p=kernel/msm.git;a=commit;h=393e4bfaaf79377d29cd6bb2228f87601aeca668

I don't like what's in this one. Lack of explaination in the commit log doesn't help.

#9 lore raw

PATCH 1/2 ARM: Implemented support for VFP PM context saving

Quoted 27 lines

-----Original Message-----
From: Russell King - ARM Linux [mailto:linux at arm.linux.org.uk]
Sent: 24 November, 2009 17:19
To: Catalin Marinas
Cc: Kristo Tero (Nokia-D/Tampere); linux-omap at vger.kernel.org; linux-arm-kernel at lists.infradead.org; Dave Estes
Subject: Re: [PATCH 1/2] ARM: Implemented support for VFP PM context saving

Quoted 5 lines · Catalin Marinas

BTW, the two patches below were mentioned to me some time ago but I haven't got the time to look at them:

[ARM] vfp: Fix bug in vfp_pm_suspend

https://www.codeaurora.org/gitweb/quic/le/?p=kernel/msm.git;a=c
ommit;h=88984c9b2d69c222ee1e2afc948ca73f597d40ff

This one is bad - it gets the current CPU by directly referencing ti->cpu.
Too bad if you have kernel preemption enabled; the value obtained that way is effectively meaningless. The only way to get a meaningful value is via 'get_cpu()' and after you've done with using it, 'put_cpu()'. That ensures you can't be preempted onto a different CPU mid-operation.

Quoted 2 lines

It's safe in vfp_notifier because we're called in an already atomic context.

I investigated this issue a bit more, and indeed there is a potential bug in the vfp_pm_suspend(). Most of the time it works fine as apparently shell process has VFP enabled at least on my system, and it saves the state. The issue is different with dynamic idle, we are calling the code from init thread which does not need VFP for anything, and thus VFP is always disabled if we try to call vfp_pm_suspend(). For OMAP3, I found a way to fix the dynamic idle part to work properly by just simply calling vfp_sync_state() from idle. This functionality is supposed to be used by ptrace, but I guess it could be used for this also?

A proper fix for suspend is bit more difficult, as I don't know too well how SMP is supposed to work in this case. I guess there is a separate VFP co-processor available for each ARM core, but vfp_pm_suspend() is only called once for the whole system?

Quoted 10 lines
Quoted 3 lines

[ARM] vfp: Add additional vfp interfaces

https://www.codeaurora.org/gitweb/quic/le/?p=kernel/msm.git;a=c
ommit;h=393e4bfaaf79377d29cd6bb2228f87601aeca668

I don't like what's in this one. Lack of explaination in the commit log
doesn't help.