Re: [Xenomai] [PATCHv2 4/4] cobalt/x86: add ipipe-4.14 eager fpu support

[Jan Kiszka]

On 18.09.18 18:16, Henning Schild wrote:

Am Mon, 17 Sep 2018 21:18:48 +0200
schrieb "[ext] Henning Schild" <henning.sch...@siemens.com>:

Am Mon, 17 Sep 2018 20:23:16 +0200
schrieb Jan Kiszka <jan.kis...@siemens.com>:

On 17.09.18 20:17, Jan Kiszka wrote:

On 14.09.18 17:10, Henning Schild wrote:

4.14 is always eager, unfortunately we will need a few ifdefs
inside the eager fpu support as well.

Signed-off-by: Henning Schild <henning.sch...@siemens.com>
---
   .../arch/x86/include/asm/xenomai/wrappers.h   |  4 ++++
   kernel/cobalt/arch/x86/thread.c               | 21
++++++++++++++++++- 2 files changed, 24 insertions(+), 1
deletion(-)

diff --git
a/kernel/cobalt/arch/x86/include/asm/xenomai/wrappers.h
b/kernel/cobalt/arch/x86/include/asm/xenomai/wrappers.h index
a47730106..a4cc368a5 100644 ---
a/kernel/cobalt/arch/x86/include/asm/xenomai/wrappers.h +++
b/kernel/cobalt/arch/x86/include/asm/xenomai/wrappers.h @@ -32,6
+32,10 @@ LINUX_VERSION_CODE < KERNEL_VERSION(4,5,0)
   #define IPIPE_X86_FPU_EAGER
   #endif
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(4,14,0)
+#define IPIPE_X86_FPU_EAGER
+#endif
+
   #if LINUX_VERSION_CODE < KERNEL_VERSION(4,2,0)
   #include <asm/i387.h>
diff --git a/kernel/cobalt/arch/x86/thread.c
b/kernel/cobalt/arch/x86/thread.c index 18cf636e5..2668ce274
100644 --- a/kernel/cobalt/arch/x86/thread.c
+++ b/kernel/cobalt/arch/x86/thread.c
@@ -42,6 +42,7 @@ static struct kmem_cache *xstate_cache;
   #define cpu_has_xsave boot_cpu_has(X86_FEATURE_XSAVE)
   #endif
+#ifndef IPIPE_X86_FPU_EAGER
   #if LINUX_VERSION_CODE < KERNEL_VERSION(4,2,0)
   #include <asm/i387.h>
   #include <asm/fpu-internal.h>
@@ -72,6 +73,9 @@ static inline void x86_fpregs_activate(struct
task_struct *t) #define x86_xstate_alignment
__alignof__(union fpregs_state) #endif
+#else /* IPIPE_X86_FPU_EAGER */
+#define x86_xstate_alignment        __alignof__(union
fpregs_state) +#endif /* ! IPIPE_X86_FPU_EAGER */
   #if LINUX_VERSION_CODE < KERNEL_VERSION(4,8,0)
   /*
@@ -465,9 +469,15 @@ void xnarch_leave_root(struct xnthread
*root) // save fpregs from in-kernel use
       copy_fpregs_to_fpstate(rootcb->kfpu);
       kernel_fpu_enable();
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(4,14,0)
+    // restore current's fpregs
+    __cpu_invalidate_fpregs_state();
+    switch_fpu_finish(&current->thread.fpu, smp_processor_id());
+#else
       // mark current thread as not owning the FPU anymore
       if (&current->thread.fpu.fpstate_active)
           fpregs_deactivate(&current->thread.fpu);
+#endif
   }
   void xnarch_switch_fpu(struct xnthread *from, struct xnthread
*to) @@ -528,7 +538,11 @@ void xnarch_init_shadow_tcb(struct
xnthread *thread) #else /* IPIPE_X86_FPU_EAGER */
       /* XNFPU is always set */
       xnthread_set_state(thread, XNFPU);
+#if LINUX_VERSION_CODE < KERNEL_VERSION(4,14,0)
       fpu__activate_fpstate_read(&p->thread.fpu);
+#else
+    fpu__initialize(&p->thread.fpu);
+#endif
   #endif /* ! IPIPE_X86_FPU_EAGER */
   }
@@ -537,7 +551,12 @@ int mach_x86_thread_init(void)
       xstate_cache = kmem_cache_create("cobalt_x86_xstate",
                        fpu_kernel_xstate_size,
                        x86_xstate_alignment,
-                     SLAB_NOTRACK, NULL);
+#if LINUX_VERSION_CODE < KERNEL_VERSION(4,14,0)
+                     SLAB_NOTRACK,
+#else
+                     0,
+#endif
+                     NULL);
       if (xstate_cache == NULL)
           return -ENOMEM;

Thanks, taken the whole series to next so that we can move
forward. I've just changed "//" (C++) style comments to
"old-style" C. Kernel style. Please avoid them in the future.

Sorry missed the // part of the review.

Forgot: Did you happen to try the series with stable-3.0.x as well?
I consider them relevant for that branch as well, so that users can
update the kernel. But I plan to leave them in next until we have
ipipe kernels available and tested them a bit broader.

Did not do that yet. I guess the ifdefs should all resolve to
lazy-fpu, making the series a no-op for lazy kernels. Will test
against 4.9 and 4.4 and confirm that my guess is right.

Might have gotten the question wrong in the first place. I did test
xenomai (next) against ipipe 4.(4|9).y. After that i backported
(cherry-picked) the changes right onto stable-3.0.x and tested against
a lazy 4.4 and a 4.14. So i think all combinations should be tested.

Applied all four to stable - looks good.

Jan

--
Siemens AG, Corporate Technology, CT RDA IOT SES-DE
Corporate Competence Center Embedded Linux

_______________________________________________
Xenomai mailing list
Xenomai@xenomai.org
https://xenomai.org/mailman/listinfo/xenomai