Re: [RFC PATCH] x86, fpu: Use eagerfpu by default on all CPUs

[Borislav Petkov]

On Sat, Feb 21, 2015 at 08:23:52PM +0100, Ingo Molnar wrote:
> to switch between the modes?

I went all out and did a debugfs file, see patch at the end, which
counts FPU saves. Then I ran this script:

---
#!/bin/bash

D="/sys/kernel/debug/fpu/eager"

echo "Lazy FPU: "
echo 0 > $D
echo -n "  FPU saves before: "; cat $D

perf stat -a -e task-clock,cycles,instructions,branch-misses,cache-misses,faults,context-switches,migrations --sync --pre ~/bin/pre-build-kernel.sh make -s -j12

echo -n "  FPU saves after: "; cat $D

echo ""
echo "Eager FPU: "
echo 1 > $D
echo -n "  FPU saves before: "; cat $D

perf stat -a -e task-clock,cycles,instructions,branch-misses,cache-misses,faults,context-switches,migrations --sync --pre ~/bin/pre-build-kernel.sh make -s -j12

echo -n "  FPU saves after: "; cat $D
---

which spit this:

Lazy FPU:
  FPU saves before: 3
Setup is 16252 bytes (padded to 16384 bytes).
System is 4222 kB
CRC c79a13ab
Kernel: arch/x86/boot/bzImage is ready  (#41)

 Performance counter stats for 'system wide':

    1315527.989020      task-clock (msec)         #    6.003 CPUs utilized           [100.00%]
 3,042,312,057,208      cycles                    #    2.313 GHz                     [100.00%]
 2,790,807,863,402      instructions              #    0.92  insns per cycle         [100.00%]
    31,658,299,111      branch-misses             #   24.065 M/sec                   [100.00%]
    27,504,255,277      cache-misses              #   20.907 M/sec
        26,802,015      faults                    #    0.020 M/sec                   [100.00%]
         1,248,899      context-switches          #    0.949 K/sec                   [100.00%]
            69,553      migrations                #    0.053 K/sec

     219.127929718 seconds time elapsed

  FPU saves after: 704186

Eager FPU:
  FPU saves before: 4
Setup is 16252 bytes (padded to 16384 bytes).
System is 4222 kB
CRC 6767bb2e
Kernel: arch/x86/boot/bzImage is ready  (#42)

 Performance counter stats for 'system wide':

    1321651.543922      task-clock (msec)         #    6.003 CPUs utilized           [100.00%]
 3,044,403,437,364      cycles                    #    2.303 GHz                     [100.00%]
 2,790,835,886,565      instructions              #    0.92  insns per cycle         [100.00%]
    31,638,090,259      branch-misses             #   23.938 M/sec                   [100.00%]
    27,491,643,095      cache-misses              #   20.801 M/sec
        26,869,732      faults                    #    0.020 M/sec                   [100.00%]
         1,252,034      context-switches          #    0.947 K/sec                   [100.00%]
            69,247      migrations                #    0.052 K/sec

     220.148034331 seconds time elapsed

  FPU saves after: 901638

---
so we have a second slowdown and 200K FPU saves more in eager mode.

Provided I've not done a mistake, looks like the increase in cycles gets
mirrored in 1 second time longer. I've not done the --repeat 10 thing
again, maybe I should do it too, just to be fair as this is a single
run.

---
 arch/x86/include/asm/fpu-internal.h |  4 ++++
 arch/x86/kernel/xsave.c             | 47 ++++++++++++++++++++++++++++++++++++-
 2 files changed, 50 insertions(+), 1 deletion(-)

diff --git a/arch/x86/include/asm/fpu-internal.h b/arch/x86/include/asm/fpu-internal.h
index e97622f57722..7141f353e960 100644
--- a/arch/x86/include/asm/fpu-internal.h
+++ b/arch/x86/include/asm/fpu-internal.h
@@ -38,6 +38,8 @@ int ia32_setup_frame(int sig, struct ksignal *ksig,
 # define ia32_setup_rt_frame	__setup_rt_frame
 #endif
 
+
+extern unsigned long fpu_saved;
 extern unsigned int mxcsr_feature_mask;
 extern void fpu_init(void);
 extern void eager_fpu_init(void);
@@ -242,6 +244,8 @@ static inline void fpu_fxsave(struct fpu *fpu)
  */
 static inline int fpu_save_init(struct fpu *fpu)
 {
+	fpu_saved++;
+
 	if (use_xsave()) {
 		fpu_xsave(fpu);
 
diff --git a/arch/x86/kernel/xsave.c b/arch/x86/kernel/xsave.c
index 0de1fae2bdf0..029de8b629d0 100644
--- a/arch/x86/kernel/xsave.c
+++ b/arch/x86/kernel/xsave.c
@@ -14,6 +14,8 @@
 #include <asm/sigframe.h>
 #include <asm/xcr.h>
 
+#include <linux/debugfs.h>
+
 /*
  * Supported feature mask by the CPU and the kernel.
  */
@@ -638,7 +640,7 @@ static void __init xstate_enable_boot_cpu(void)
 	setup_init_fpu_buf();
 
 	/* Auto enable eagerfpu for xsaveopt */
-	if (cpu_has_xsaveopt && eagerfpu != DISABLE)
+	if (eagerfpu != DISABLE)
 		eagerfpu = ENABLE;
 
 	if (pcntxt_mask & XSTATE_EAGER) {
@@ -739,3 +741,46 @@ void *get_xsave_addr(struct xsave_struct *xsave, int xstate)
 	return (void *)xsave + xstate_comp_offsets[feature];
 }
 EXPORT_SYMBOL_GPL(get_xsave_addr);
+
+unsigned long fpu_saved;
+
+static int eager_get(void *data, u64 *val)
+{
+	*val = fpu_saved;
+
+	return 0;
+}
+
+static int eager_set(void *data, u64 val)
+{
+	if (val)
+		setup_force_cpu_cap(X86_FEATURE_EAGER_FPU);
+	else
+		setup_clear_cpu_cap(X86_FEATURE_EAGER_FPU);
+
+	fpu_saved = 0;
+
+	return 0;
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(eager_fops, eager_get, eager_set, "%llu\n");
+
+static int __init setup_eagerfpu_knob(void)
+{
+	static struct dentry *d_eager, *f_eager;
+
+	d_eager = debugfs_create_dir("fpu", NULL);
+	if (!d_eager) {
+		pr_err("Error creating fpu debugfs dir\n");
+		return -ENOMEM;
+	}
+
+	f_eager = debugfs_create_file("eager", 0644, d_eager, NULL, &eager_fops);
+	if (!f_eager) {
+		pr_err("Error creating fpu debugfs node\n");
+		return -ENOMEM;
+	}
+
+	return 0;
+}
+late_initcall(setup_eagerfpu_knob);
-- 
2.2.0.33.gc18b867

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/