* [PATCH Dovetail] arm64: fpsimd: dovetail: Allow preemption of in-band fpsimd kernel sections
@ 2025-02-19 14:17 Florian Bezdeka
2025-02-20 11:36 ` Jan Kiszka
2025-03-02 15:07 ` Philippe Gerum
0 siblings, 2 replies; 5+ messages in thread
From: Florian Bezdeka @ 2025-02-19 14:17 UTC (permalink / raw)
To: xenomai; +Cc: Florian Bezdeka, powertree, Jan Kiszka, Philippe Gerum
Preemption of in-band in-kernel fpsimd sections was not allowed /
properly implemented in Dovetail up to now.
That piled up again as some crypto code was optimized for arm64 in
Linux 6.14. A possible call stack that makes clear that the FPU was
corrupted follows. The inband stage was using crypto/FPU code inside
the kernel and got interrupted by OOB.
[ 331.703695] task_fpsimd_load+0xd0/0x214 (P)
[ 331.703713] fpsimd_restore_current_state+0x9c/0xb8
[ 331.703732] fpsimd_restore_current_oob+0x1c/0x2c
[ 331.703751] dovetail_context_switch+0x90/0x180
[ 331.703769] pipeline_switch_to+0x10/0x1c
[ 331.703792] ___xnsched_run+0x1bc/0x29c
[ 331.703811] run_oob_call+0x98/0x17c
[ 331.703830] handle_irq_pipelined_finish+0x1e0/0x1ec
[ 331.703848] handle_irq_pipelined+0x58/0x6c
[ 331.703865] call_on_irq_stack+0x24/0x4c
[ 331.703884] do_interrupt_handler+0x150/0x178
[ 331.703902] el1_interrupt+0x44/0x11c
[ 331.703921] el1h_64_irq_handler+0x18/0x24
[ 331.703941] el1h_64_irq+0x6c/0x70 <- oob
[ 331.703957] crc32c_le_arm64_4way+0x4c/0xc0 (P) <- inband
[ 331.703978] ext4_block_bitmap_csum_set+0x38/0x60
[ 331.704002] ext4_mb_mark_context+0x1b0/0x398
[ 331.704025] ext4_mb_mark_diskspace_used+0xd8/0x1f8
To fix that we have to hook into the task switch as we do on x86 and
invoke the FPU (re)store logic.
Cc: powertree <powertree@163.com>
Cc: Jan Kiszka <jan.kiszka@siemens.com>
Cc: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Florian Bezdeka <florian.bezdeka@siemens.com>
---
arch/arm64/include/asm/dovetail.h | 10 +++++++--
arch/arm64/include/asm/fpsimd.h | 7 +++++-
arch/arm64/include/asm/thread_info.h | 1 +
arch/arm64/kernel/fpsimd.c | 33 ++++++++++++++++++++++++++++
4 files changed, 48 insertions(+), 3 deletions(-)
diff --git a/arch/arm64/include/asm/dovetail.h b/arch/arm64/include/asm/dovetail.h
index dc46f131e3eb..8cd359710353 100644
--- a/arch/arm64/include/asm/dovetail.h
+++ b/arch/arm64/include/asm/dovetail.h
@@ -26,11 +26,17 @@ static inline void arch_dovetail_exec_prepare(void)
{ }
static inline void arch_dovetail_switch_prepare(bool leave_inband)
-{ }
+{
+ if (leave_inband)
+ fpsimd_suspend_inband();
+}
static inline void arch_dovetail_switch_finish(bool enter_inband)
{
- fpsimd_restore_current_oob();
+ if (enter_inband)
+ fpsimd_resume_inband();
+ else
+ fpsimd_restore_current_oob();
}
/*
diff --git a/arch/arm64/include/asm/fpsimd.h b/arch/arm64/include/asm/fpsimd.h
index 8e12d5be076e..1ba7b75b2277 100644
--- a/arch/arm64/include/asm/fpsimd.h
+++ b/arch/arm64/include/asm/fpsimd.h
@@ -79,10 +79,15 @@ extern void fpsimd_flush_thread(void);
extern void fpsimd_signal_preserve_current_state(void);
extern void fpsimd_preserve_current_state(void);
extern void fpsimd_restore_current_state(void);
-extern void fpsimd_restore_current_oob(void);
extern void fpsimd_update_current_state(struct user_fpsimd_state const *state);
extern void fpsimd_kvm_prepare(void);
+#ifdef CONFIG_DOVETAIL
+extern void fpsimd_restore_current_oob(void);
+extern void fpsimd_suspend_inband(void);
+extern void fpsimd_resume_inband(void);
+#endif
+
struct cpu_fp_state {
struct user_fpsimd_state *st;
void *sve_state;
diff --git a/arch/arm64/include/asm/thread_info.h b/arch/arm64/include/asm/thread_info.h
index 3028a8f294c2..8ee08fd30d8c 100644
--- a/arch/arm64/include/asm/thread_info.h
+++ b/arch/arm64/include/asm/thread_info.h
@@ -76,6 +76,7 @@ void arch_setup_new_exec(void);
#define TIF_SECCOMP 11 /* syscall secure computing */
#define TIF_SYSCALL_EMU 12 /* syscall emulation active */
#define TIF_MAYDAY 13 /* Emergency trap pending */
+#define TIF_KERNEL_FP_PREEMPTED 14 /* Dovetail: kernel mode FPSIMD section preempted by OOB */
#define TIF_MEMDIE 18 /* is terminating due to OOM killer */
#define TIF_FREEZE 19
#define TIF_RESTORE_SIGMASK 20
diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
index 96f2f60b4caa..032234593d6e 100644
--- a/arch/arm64/kernel/fpsimd.c
+++ b/arch/arm64/kernel/fpsimd.c
@@ -2240,6 +2240,39 @@ void cpu_enable_fpsimd(const struct arm64_cpu_capabilities *__always_unused p)
isb();
}
+#ifdef CONFIG_DOVETAIL
+
+/*
+ * Holds the in-kernel fpu state when preempted by a task running on the
+ * out-of-band stage.
+ */
+static DEFINE_PER_CPU(struct user_fpsimd_state, in_kernel_fpstate);
+
+void fpsimd_suspend_inband(void)
+{
+ struct user_fpsimd_state *kfpu = this_cpu_ptr(&in_kernel_fpstate);
+
+ /*
+ * If TIF_KERNEL_FPSTATE is set, we are dealing with the preemption of an
+ * inband kernel context currently using the fpu by a thread which resumes
+ * on the oob stage.
+ */
+ if (test_thread_flag(TIF_KERNEL_FPSTATE)) {
+ fpsimd_save_state(kfpu);
+ set_thread_flag(TIF_KERNEL_FP_PREEMPTED);
+ }
+}
+
+void fpsimd_resume_inband(void)
+{
+ struct user_fpsimd_state *kfpu = this_cpu_ptr(&in_kernel_fpstate);
+
+ if (test_and_clear_thread_flag(TIF_KERNEL_FP_PREEMPTED))
+ fpsimd_load_state(kfpu);
+}
+
+#endif
+
/*
* FP/SIMD support code initialisation.
*/
--
2.39.5
^ permalink raw reply related [flat|nested] 5+ messages in thread
* Re: [PATCH Dovetail] arm64: fpsimd: dovetail: Allow preemption of in-band fpsimd kernel sections
2025-02-19 14:17 [PATCH Dovetail] arm64: fpsimd: dovetail: Allow preemption of in-band fpsimd kernel sections Florian Bezdeka
@ 2025-02-20 11:36 ` Jan Kiszka
2025-03-13 8:55 ` Florian Bezdeka
2025-03-02 15:07 ` Philippe Gerum
1 sibling, 1 reply; 5+ messages in thread
From: Jan Kiszka @ 2025-02-20 11:36 UTC (permalink / raw)
To: Florian Bezdeka, xenomai; +Cc: powertree, Philippe Gerum
On 19.02.25 15:17, Florian Bezdeka wrote:
> Preemption of in-band in-kernel fpsimd sections was not allowed /
> properly implemented in Dovetail up to now.
>
> That piled up again as some crypto code was optimized for arm64 in
> Linux 6.14. A possible call stack that makes clear that the FPU was
> corrupted follows. The inband stage was using crypto/FPU code inside
> the kernel and got interrupted by OOB.
>
> [ 331.703695] task_fpsimd_load+0xd0/0x214 (P)
> [ 331.703713] fpsimd_restore_current_state+0x9c/0xb8
> [ 331.703732] fpsimd_restore_current_oob+0x1c/0x2c
> [ 331.703751] dovetail_context_switch+0x90/0x180
> [ 331.703769] pipeline_switch_to+0x10/0x1c
> [ 331.703792] ___xnsched_run+0x1bc/0x29c
> [ 331.703811] run_oob_call+0x98/0x17c
> [ 331.703830] handle_irq_pipelined_finish+0x1e0/0x1ec
> [ 331.703848] handle_irq_pipelined+0x58/0x6c
> [ 331.703865] call_on_irq_stack+0x24/0x4c
> [ 331.703884] do_interrupt_handler+0x150/0x178
> [ 331.703902] el1_interrupt+0x44/0x11c
> [ 331.703921] el1h_64_irq_handler+0x18/0x24
> [ 331.703941] el1h_64_irq+0x6c/0x70 <- oob
> [ 331.703957] crc32c_le_arm64_4way+0x4c/0xc0 (P) <- inband
> [ 331.703978] ext4_block_bitmap_csum_set+0x38/0x60
> [ 331.704002] ext4_mb_mark_context+0x1b0/0x398
> [ 331.704025] ext4_mb_mark_diskspace_used+0xd8/0x1f8
>
> To fix that we have to hook into the task switch as we do on x86 and
> invoke the FPU (re)store logic.
>
> Cc: powertree <powertree@163.com>
> Cc: Jan Kiszka <jan.kiszka@siemens.com>
> Cc: Philippe Gerum <rpm@xenomai.org>
>
Great!
Did you already check how easily it applies to older (but maintained)
dovetail kernels?
Jan
> Signed-off-by: Florian Bezdeka <florian.bezdeka@siemens.com>
> ---
> arch/arm64/include/asm/dovetail.h | 10 +++++++--
> arch/arm64/include/asm/fpsimd.h | 7 +++++-
> arch/arm64/include/asm/thread_info.h | 1 +
> arch/arm64/kernel/fpsimd.c | 33 ++++++++++++++++++++++++++++
> 4 files changed, 48 insertions(+), 3 deletions(-)
>
> diff --git a/arch/arm64/include/asm/dovetail.h b/arch/arm64/include/asm/dovetail.h
> index dc46f131e3eb..8cd359710353 100644
> --- a/arch/arm64/include/asm/dovetail.h
> +++ b/arch/arm64/include/asm/dovetail.h
> @@ -26,11 +26,17 @@ static inline void arch_dovetail_exec_prepare(void)
> { }
>
> static inline void arch_dovetail_switch_prepare(bool leave_inband)
> -{ }
> +{
> + if (leave_inband)
> + fpsimd_suspend_inband();
> +}
>
> static inline void arch_dovetail_switch_finish(bool enter_inband)
> {
> - fpsimd_restore_current_oob();
> + if (enter_inband)
> + fpsimd_resume_inband();
> + else
> + fpsimd_restore_current_oob();
> }
>
> /*
> diff --git a/arch/arm64/include/asm/fpsimd.h b/arch/arm64/include/asm/fpsimd.h
> index 8e12d5be076e..1ba7b75b2277 100644
> --- a/arch/arm64/include/asm/fpsimd.h
> +++ b/arch/arm64/include/asm/fpsimd.h
> @@ -79,10 +79,15 @@ extern void fpsimd_flush_thread(void);
> extern void fpsimd_signal_preserve_current_state(void);
> extern void fpsimd_preserve_current_state(void);
> extern void fpsimd_restore_current_state(void);
> -extern void fpsimd_restore_current_oob(void);
> extern void fpsimd_update_current_state(struct user_fpsimd_state const *state);
> extern void fpsimd_kvm_prepare(void);
>
> +#ifdef CONFIG_DOVETAIL
> +extern void fpsimd_restore_current_oob(void);
> +extern void fpsimd_suspend_inband(void);
> +extern void fpsimd_resume_inband(void);
> +#endif
> +
> struct cpu_fp_state {
> struct user_fpsimd_state *st;
> void *sve_state;
> diff --git a/arch/arm64/include/asm/thread_info.h b/arch/arm64/include/asm/thread_info.h
> index 3028a8f294c2..8ee08fd30d8c 100644
> --- a/arch/arm64/include/asm/thread_info.h
> +++ b/arch/arm64/include/asm/thread_info.h
> @@ -76,6 +76,7 @@ void arch_setup_new_exec(void);
> #define TIF_SECCOMP 11 /* syscall secure computing */
> #define TIF_SYSCALL_EMU 12 /* syscall emulation active */
> #define TIF_MAYDAY 13 /* Emergency trap pending */
> +#define TIF_KERNEL_FP_PREEMPTED 14 /* Dovetail: kernel mode FPSIMD section preempted by OOB */
> #define TIF_MEMDIE 18 /* is terminating due to OOM killer */
> #define TIF_FREEZE 19
> #define TIF_RESTORE_SIGMASK 20
> diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
> index 96f2f60b4caa..032234593d6e 100644
> --- a/arch/arm64/kernel/fpsimd.c
> +++ b/arch/arm64/kernel/fpsimd.c
> @@ -2240,6 +2240,39 @@ void cpu_enable_fpsimd(const struct arm64_cpu_capabilities *__always_unused p)
> isb();
> }
>
> +#ifdef CONFIG_DOVETAIL
> +
> +/*
> + * Holds the in-kernel fpu state when preempted by a task running on the
> + * out-of-band stage.
> + */
> +static DEFINE_PER_CPU(struct user_fpsimd_state, in_kernel_fpstate);
> +
> +void fpsimd_suspend_inband(void)
> +{
> + struct user_fpsimd_state *kfpu = this_cpu_ptr(&in_kernel_fpstate);
> +
> + /*
> + * If TIF_KERNEL_FPSTATE is set, we are dealing with the preemption of an
> + * inband kernel context currently using the fpu by a thread which resumes
> + * on the oob stage.
> + */
> + if (test_thread_flag(TIF_KERNEL_FPSTATE)) {
> + fpsimd_save_state(kfpu);
> + set_thread_flag(TIF_KERNEL_FP_PREEMPTED);
> + }
> +}
> +
> +void fpsimd_resume_inband(void)
> +{
> + struct user_fpsimd_state *kfpu = this_cpu_ptr(&in_kernel_fpstate);
> +
> + if (test_and_clear_thread_flag(TIF_KERNEL_FP_PREEMPTED))
> + fpsimd_load_state(kfpu);
> +}
> +
> +#endif
> +
> /*
> * FP/SIMD support code initialisation.
> */
--
Siemens AG, Foundational Technologies
Linux Expert Center
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH Dovetail] arm64: fpsimd: dovetail: Allow preemption of in-band fpsimd kernel sections
2025-02-19 14:17 [PATCH Dovetail] arm64: fpsimd: dovetail: Allow preemption of in-band fpsimd kernel sections Florian Bezdeka
2025-02-20 11:36 ` Jan Kiszka
@ 2025-03-02 15:07 ` Philippe Gerum
1 sibling, 0 replies; 5+ messages in thread
From: Philippe Gerum @ 2025-03-02 15:07 UTC (permalink / raw)
To: Florian Bezdeka; +Cc: xenomai, powertree, Jan Kiszka
Florian Bezdeka <florian.bezdeka@siemens.com> writes:
> Preemption of in-band in-kernel fpsimd sections was not allowed /
> properly implemented in Dovetail up to now.
>
> That piled up again as some crypto code was optimized for arm64 in
> Linux 6.14. A possible call stack that makes clear that the FPU was
> corrupted follows. The inband stage was using crypto/FPU code inside
> the kernel and got interrupted by OOB.
>
> [ 331.703695] task_fpsimd_load+0xd0/0x214 (P)
> [ 331.703713] fpsimd_restore_current_state+0x9c/0xb8
> [ 331.703732] fpsimd_restore_current_oob+0x1c/0x2c
> [ 331.703751] dovetail_context_switch+0x90/0x180
> [ 331.703769] pipeline_switch_to+0x10/0x1c
> [ 331.703792] ___xnsched_run+0x1bc/0x29c
> [ 331.703811] run_oob_call+0x98/0x17c
> [ 331.703830] handle_irq_pipelined_finish+0x1e0/0x1ec
> [ 331.703848] handle_irq_pipelined+0x58/0x6c
> [ 331.703865] call_on_irq_stack+0x24/0x4c
> [ 331.703884] do_interrupt_handler+0x150/0x178
> [ 331.703902] el1_interrupt+0x44/0x11c
> [ 331.703921] el1h_64_irq_handler+0x18/0x24
> [ 331.703941] el1h_64_irq+0x6c/0x70 <- oob
> [ 331.703957] crc32c_le_arm64_4way+0x4c/0xc0 (P) <- inband
> [ 331.703978] ext4_block_bitmap_csum_set+0x38/0x60
> [ 331.704002] ext4_mb_mark_context+0x1b0/0x398
> [ 331.704025] ext4_mb_mark_diskspace_used+0xd8/0x1f8
>
> To fix that we have to hook into the task switch as we do on x86 and
> invoke the FPU (re)store logic.
>
> Cc: powertree <powertree@163.com>
> Cc: Jan Kiszka <jan.kiszka@siemens.com>
> Cc: Philippe Gerum <rpm@xenomai.org>
>
> Signed-off-by: Florian Bezdeka <florian.bezdeka@siemens.com>
> ---
> arch/arm64/include/asm/dovetail.h | 10 +++++++--
> arch/arm64/include/asm/fpsimd.h | 7 +++++-
> arch/arm64/include/asm/thread_info.h | 1 +
> arch/arm64/kernel/fpsimd.c | 33 ++++++++++++++++++++++++++++
> 4 files changed, 48 insertions(+), 3 deletions(-)
LGTM, stable under significant stress load on imx8mp.
Acked-by: Philippe Gerum <rpm@xenomai.org>
--
Philippe.
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH Dovetail] arm64: fpsimd: dovetail: Allow preemption of in-band fpsimd kernel sections
2025-02-20 11:36 ` Jan Kiszka
@ 2025-03-13 8:55 ` Florian Bezdeka
2025-03-13 10:12 ` Jan Kiszka
0 siblings, 1 reply; 5+ messages in thread
From: Florian Bezdeka @ 2025-03-13 8:55 UTC (permalink / raw)
To: Jan Kiszka, xenomai; +Cc: powertree, Philippe Gerum
On Thu, 2025-02-20 at 12:36 +0100, Jan Kiszka wrote:
> On 19.02.25 15:17, Florian Bezdeka wrote:
> > Preemption of in-band in-kernel fpsimd sections was not allowed /
> > properly implemented in Dovetail up to now.
> >
> > That piled up again as some crypto code was optimized for arm64 in
> > Linux 6.14. A possible call stack that makes clear that the FPU was
> > corrupted follows. The inband stage was using crypto/FPU code inside
> > the kernel and got interrupted by OOB.
> >
> > [ 331.703695] task_fpsimd_load+0xd0/0x214 (P)
> > [ 331.703713] fpsimd_restore_current_state+0x9c/0xb8
> > [ 331.703732] fpsimd_restore_current_oob+0x1c/0x2c
> > [ 331.703751] dovetail_context_switch+0x90/0x180
> > [ 331.703769] pipeline_switch_to+0x10/0x1c
> > [ 331.703792] ___xnsched_run+0x1bc/0x29c
> > [ 331.703811] run_oob_call+0x98/0x17c
> > [ 331.703830] handle_irq_pipelined_finish+0x1e0/0x1ec
> > [ 331.703848] handle_irq_pipelined+0x58/0x6c
> > [ 331.703865] call_on_irq_stack+0x24/0x4c
> > [ 331.703884] do_interrupt_handler+0x150/0x178
> > [ 331.703902] el1_interrupt+0x44/0x11c
> > [ 331.703921] el1h_64_irq_handler+0x18/0x24
> > [ 331.703941] el1h_64_irq+0x6c/0x70 <- oob
> > [ 331.703957] crc32c_le_arm64_4way+0x4c/0xc0 (P) <- inband
> > [ 331.703978] ext4_block_bitmap_csum_set+0x38/0x60
> > [ 331.704002] ext4_mb_mark_context+0x1b0/0x398
> > [ 331.704025] ext4_mb_mark_diskspace_used+0xd8/0x1f8
> >
> > To fix that we have to hook into the task switch as we do on x86 and
> > invoke the FPU (re)store logic.
> >
> > Cc: powertree <powertree@163.com>
> > Cc: Jan Kiszka <jan.kiszka@siemens.com>
> > Cc: Philippe Gerum <rpm@xenomai.org>
> >
>
> Great!
>
> Did you already check how easily it applies to older (but maintained)
> dovetail kernels?
Philippe already picked it into v6.1.y-cip-dovetail-rebase, so I think
the only remaining branch would be 5.10, right?
Florian
>
> Jan
>
> > Signed-off-by: Florian Bezdeka <florian.bezdeka@siemens.com>
> > ---
> > arch/arm64/include/asm/dovetail.h | 10 +++++++--
> > arch/arm64/include/asm/fpsimd.h | 7 +++++-
> > arch/arm64/include/asm/thread_info.h | 1 +
> > arch/arm64/kernel/fpsimd.c | 33 ++++++++++++++++++++++++++++
> > 4 files changed, 48 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/arm64/include/asm/dovetail.h b/arch/arm64/include/asm/dovetail.h
> > index dc46f131e3eb..8cd359710353 100644
> > --- a/arch/arm64/include/asm/dovetail.h
> > +++ b/arch/arm64/include/asm/dovetail.h
> > @@ -26,11 +26,17 @@ static inline void arch_dovetail_exec_prepare(void)
> > { }
> >
> > static inline void arch_dovetail_switch_prepare(bool leave_inband)
> > -{ }
> > +{
> > + if (leave_inband)
> > + fpsimd_suspend_inband();
> > +}
> >
> > static inline void arch_dovetail_switch_finish(bool enter_inband)
> > {
> > - fpsimd_restore_current_oob();
> > + if (enter_inband)
> > + fpsimd_resume_inband();
> > + else
> > + fpsimd_restore_current_oob();
> > }
> >
> > /*
> > diff --git a/arch/arm64/include/asm/fpsimd.h b/arch/arm64/include/asm/fpsimd.h
> > index 8e12d5be076e..1ba7b75b2277 100644
> > --- a/arch/arm64/include/asm/fpsimd.h
> > +++ b/arch/arm64/include/asm/fpsimd.h
> > @@ -79,10 +79,15 @@ extern void fpsimd_flush_thread(void);
> > extern void fpsimd_signal_preserve_current_state(void);
> > extern void fpsimd_preserve_current_state(void);
> > extern void fpsimd_restore_current_state(void);
> > -extern void fpsimd_restore_current_oob(void);
> > extern void fpsimd_update_current_state(struct user_fpsimd_state const *state);
> > extern void fpsimd_kvm_prepare(void);
> >
> > +#ifdef CONFIG_DOVETAIL
> > +extern void fpsimd_restore_current_oob(void);
> > +extern void fpsimd_suspend_inband(void);
> > +extern void fpsimd_resume_inband(void);
> > +#endif
> > +
> > struct cpu_fp_state {
> > struct user_fpsimd_state *st;
> > void *sve_state;
> > diff --git a/arch/arm64/include/asm/thread_info.h b/arch/arm64/include/asm/thread_info.h
> > index 3028a8f294c2..8ee08fd30d8c 100644
> > --- a/arch/arm64/include/asm/thread_info.h
> > +++ b/arch/arm64/include/asm/thread_info.h
> > @@ -76,6 +76,7 @@ void arch_setup_new_exec(void);
> > #define TIF_SECCOMP 11 /* syscall secure computing */
> > #define TIF_SYSCALL_EMU 12 /* syscall emulation active */
> > #define TIF_MAYDAY 13 /* Emergency trap pending */
> > +#define TIF_KERNEL_FP_PREEMPTED 14 /* Dovetail: kernel mode FPSIMD section preempted by OOB */
> > #define TIF_MEMDIE 18 /* is terminating due to OOM killer */
> > #define TIF_FREEZE 19
> > #define TIF_RESTORE_SIGMASK 20
> > diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
> > index 96f2f60b4caa..032234593d6e 100644
> > --- a/arch/arm64/kernel/fpsimd.c
> > +++ b/arch/arm64/kernel/fpsimd.c
> > @@ -2240,6 +2240,39 @@ void cpu_enable_fpsimd(const struct arm64_cpu_capabilities *__always_unused p)
> > isb();
> > }
> >
> > +#ifdef CONFIG_DOVETAIL
> > +
> > +/*
> > + * Holds the in-kernel fpu state when preempted by a task running on the
> > + * out-of-band stage.
> > + */
> > +static DEFINE_PER_CPU(struct user_fpsimd_state, in_kernel_fpstate);
> > +
> > +void fpsimd_suspend_inband(void)
> > +{
> > + struct user_fpsimd_state *kfpu = this_cpu_ptr(&in_kernel_fpstate);
> > +
> > + /*
> > + * If TIF_KERNEL_FPSTATE is set, we are dealing with the preemption of an
> > + * inband kernel context currently using the fpu by a thread which resumes
> > + * on the oob stage.
> > + */
> > + if (test_thread_flag(TIF_KERNEL_FPSTATE)) {
> > + fpsimd_save_state(kfpu);
> > + set_thread_flag(TIF_KERNEL_FP_PREEMPTED);
> > + }
> > +}
> > +
> > +void fpsimd_resume_inband(void)
> > +{
> > + struct user_fpsimd_state *kfpu = this_cpu_ptr(&in_kernel_fpstate);
> > +
> > + if (test_and_clear_thread_flag(TIF_KERNEL_FP_PREEMPTED))
> > + fpsimd_load_state(kfpu);
> > +}
> > +
> > +#endif
> > +
> > /*
> > * FP/SIMD support code initialisation.
> > */
>
>
> --
> Siemens AG, Foundational Technologies
> Linux Expert Center
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH Dovetail] arm64: fpsimd: dovetail: Allow preemption of in-band fpsimd kernel sections
2025-03-13 8:55 ` Florian Bezdeka
@ 2025-03-13 10:12 ` Jan Kiszka
0 siblings, 0 replies; 5+ messages in thread
From: Jan Kiszka @ 2025-03-13 10:12 UTC (permalink / raw)
To: Florian Bezdeka, xenomai; +Cc: powertree, Philippe Gerum
On 13.03.25 09:55, Florian Bezdeka wrote:
> On Thu, 2025-02-20 at 12:36 +0100, Jan Kiszka wrote:
>> On 19.02.25 15:17, Florian Bezdeka wrote:
>>> Preemption of in-band in-kernel fpsimd sections was not allowed /
>>> properly implemented in Dovetail up to now.
>>>
>>> That piled up again as some crypto code was optimized for arm64 in
>>> Linux 6.14. A possible call stack that makes clear that the FPU was
>>> corrupted follows. The inband stage was using crypto/FPU code inside
>>> the kernel and got interrupted by OOB.
>>>
>>> [ 331.703695] task_fpsimd_load+0xd0/0x214 (P)
>>> [ 331.703713] fpsimd_restore_current_state+0x9c/0xb8
>>> [ 331.703732] fpsimd_restore_current_oob+0x1c/0x2c
>>> [ 331.703751] dovetail_context_switch+0x90/0x180
>>> [ 331.703769] pipeline_switch_to+0x10/0x1c
>>> [ 331.703792] ___xnsched_run+0x1bc/0x29c
>>> [ 331.703811] run_oob_call+0x98/0x17c
>>> [ 331.703830] handle_irq_pipelined_finish+0x1e0/0x1ec
>>> [ 331.703848] handle_irq_pipelined+0x58/0x6c
>>> [ 331.703865] call_on_irq_stack+0x24/0x4c
>>> [ 331.703884] do_interrupt_handler+0x150/0x178
>>> [ 331.703902] el1_interrupt+0x44/0x11c
>>> [ 331.703921] el1h_64_irq_handler+0x18/0x24
>>> [ 331.703941] el1h_64_irq+0x6c/0x70 <- oob
>>> [ 331.703957] crc32c_le_arm64_4way+0x4c/0xc0 (P) <- inband
>>> [ 331.703978] ext4_block_bitmap_csum_set+0x38/0x60
>>> [ 331.704002] ext4_mb_mark_context+0x1b0/0x398
>>> [ 331.704025] ext4_mb_mark_diskspace_used+0xd8/0x1f8
>>>
>>> To fix that we have to hook into the task switch as we do on x86 and
>>> invoke the FPU (re)store logic.
>>>
>>> Cc: powertree <powertree@163.com>
>>> Cc: Jan Kiszka <jan.kiszka@siemens.com>
>>> Cc: Philippe Gerum <rpm@xenomai.org>
>>>
>>
>> Great!
>>
>> Did you already check how easily it applies to older (but maintained)
>> dovetail kernels?
>
> Philippe already picked it into v6.1.y-cip-dovetail-rebase, so I think
> the only remaining branch would be 5.10, right?
>
Ok, will have a look at that (and any other pending backports) later.
Jan
--
Siemens AG, Foundational Technologies
Linux Expert Center
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2025-03-13 10:12 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2025-02-19 14:17 [PATCH Dovetail] arm64: fpsimd: dovetail: Allow preemption of in-band fpsimd kernel sections Florian Bezdeka
2025-02-20 11:36 ` Jan Kiszka
2025-03-13 8:55 ` Florian Bezdeka
2025-03-13 10:12 ` Jan Kiszka
2025-03-02 15:07 ` Philippe Gerum
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.