[GIT PULL] x86/fpu changes for v4.6

From: Ingo Molnar
Date: Mon Mar 14 2016 - 09:17:04 EST


Linus,

Please pull the latest x86-fpu-for-linus git tree from:

git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git x86-fpu-for-linus

# HEAD: 14ddde78c787248ada9da2f2d04900548ffcf342 Merge branch 'linus' into x86/fpu, to pick up fixes

The biggest change in terms of impact is the changing of the FPU context switch
model to 'eagerfpu' for all CPU types, via:

58122bf1d856 x86/fpu: Default eagerfpu=on on all CPUs

This makes all FPU saves and restores synchronous and makes the FPU code a lot
more obvious to read. In the next cycle, if this change is problem free, we'll
remove the old lazy FPU restore code altogether.

This change flushed out some old bugs, which should all be fixed by now, BYMMV.

Thanks,

Ingo

------------------>
Andy Lutomirski (5):
x86/fpu: Fix math emulation in eager fpu mode
x86/fpu: Fix FNSAVE usage in eagerfpu mode
x86/fpu: Fold fpu_copy() into fpu__copy()
x86/fpu: Speed up lazy FPU restores slightly
x86/fpu: Default eagerfpu=on on all CPUs


arch/x86/include/asm/fpu/internal.h | 3 ++-
arch/x86/kernel/fpu/core.c | 52 +++++++++++++++++++------------------
arch/x86/kernel/fpu/init.c | 13 ++++------
arch/x86/kernel/traps.c | 3 +--
4 files changed, 35 insertions(+), 36 deletions(-)

diff --git a/arch/x86/include/asm/fpu/internal.h b/arch/x86/include/asm/fpu/internal.h
index 0fd440df63f1..a1f78a9fbf41 100644
--- a/arch/x86/include/asm/fpu/internal.h
+++ b/arch/x86/include/asm/fpu/internal.h
@@ -589,7 +589,8 @@ switch_fpu_prepare(struct fpu *old_fpu, struct fpu *new_fpu, int cpu)
* If the task has used the math, pre-load the FPU on xsave processors
* or if the past 5 consecutive context-switches used math.
*/
- fpu.preload = new_fpu->fpstate_active &&
+ fpu.preload = static_cpu_has(X86_FEATURE_FPU) &&
+ new_fpu->fpstate_active &&
(use_eager_fpu() || new_fpu->counter > 5);

if (old_fpu->fpregs_active) {
diff --git a/arch/x86/kernel/fpu/core.c b/arch/x86/kernel/fpu/core.c
index d25097c3fc1d..299b58bb975b 100644
--- a/arch/x86/kernel/fpu/core.c
+++ b/arch/x86/kernel/fpu/core.c
@@ -114,6 +114,10 @@ void __kernel_fpu_begin(void)
kernel_fpu_disable();

if (fpu->fpregs_active) {
+ /*
+ * Ignore return value -- we don't care if reg state
+ * is clobbered.
+ */
copy_fpregs_to_fpstate(fpu);
} else {
this_cpu_write(fpu_fpregs_owner_ctx, NULL);
@@ -189,8 +193,12 @@ void fpu__save(struct fpu *fpu)

preempt_disable();
if (fpu->fpregs_active) {
- if (!copy_fpregs_to_fpstate(fpu))
- fpregs_deactivate(fpu);
+ if (!copy_fpregs_to_fpstate(fpu)) {
+ if (use_eager_fpu())
+ copy_kernel_to_fpregs(&fpu->state);
+ else
+ fpregs_deactivate(fpu);
+ }
}
preempt_enable();
}
@@ -223,14 +231,15 @@ void fpstate_init(union fpregs_state *state)
}
EXPORT_SYMBOL_GPL(fpstate_init);

-/*
- * Copy the current task's FPU state to a new task's FPU context.
- *
- * In both the 'eager' and the 'lazy' case we save hardware registers
- * directly to the destination buffer.
- */
-static void fpu_copy(struct fpu *dst_fpu, struct fpu *src_fpu)
+int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
{
+ dst_fpu->counter = 0;
+ dst_fpu->fpregs_active = 0;
+ dst_fpu->last_cpu = -1;
+
+ if (!src_fpu->fpstate_active || !cpu_has_fpu)
+ return 0;
+
WARN_ON_FPU(src_fpu != &current->thread.fpu);

/*
@@ -243,10 +252,9 @@ static void fpu_copy(struct fpu *dst_fpu, struct fpu *src_fpu)
/*
* Save current FPU registers directly into the child
* FPU context, without any memory-to-memory copying.
- *
- * If the FPU context got destroyed in the process (FNSAVE
- * done on old CPUs) then copy it back into the source
- * context and mark the current task for lazy restore.
+ * In lazy mode, if the FPU context isn't loaded into
+ * fpregs, CR0.TS will be set and do_device_not_available
+ * will load the FPU context.
*
* We have to do all this with preemption disabled,
* mostly because of the FNSAVE case, because in that
@@ -259,19 +267,13 @@ static void fpu_copy(struct fpu *dst_fpu, struct fpu *src_fpu)
preempt_disable();
if (!copy_fpregs_to_fpstate(dst_fpu)) {
memcpy(&src_fpu->state, &dst_fpu->state, xstate_size);
- fpregs_deactivate(src_fpu);
+
+ if (use_eager_fpu())
+ copy_kernel_to_fpregs(&src_fpu->state);
+ else
+ fpregs_deactivate(src_fpu);
}
preempt_enable();
-}
-
-int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
-{
- dst_fpu->counter = 0;
- dst_fpu->fpregs_active = 0;
- dst_fpu->last_cpu = -1;
-
- if (src_fpu->fpstate_active && cpu_has_fpu)
- fpu_copy(dst_fpu, src_fpu);

return 0;
}
@@ -423,7 +425,7 @@ void fpu__clear(struct fpu *fpu)
{
WARN_ON_FPU(fpu != &current->thread.fpu); /* Almost certainly an anomaly */

- if (!use_eager_fpu()) {
+ if (!use_eager_fpu() || !static_cpu_has(X86_FEATURE_FPU)) {
/* FPU state will be reallocated lazily at the first use. */
fpu__drop(fpu);
} else {
diff --git a/arch/x86/kernel/fpu/init.c b/arch/x86/kernel/fpu/init.c
index 6d9f0a7ef4c8..471fe277ff40 100644
--- a/arch/x86/kernel/fpu/init.c
+++ b/arch/x86/kernel/fpu/init.c
@@ -260,7 +260,10 @@ static void __init fpu__init_system_xstate_size_legacy(void)
* not only saved the restores along the way, but we also have the
* FPU ready to be used for the original task.
*
- * 'eager' switching is used on modern CPUs, there we switch the FPU
+ * 'lazy' is deprecated because it's almost never a performance win
+ * and it's much more complicated than 'eager'.
+ *
+ * 'eager' switching is by default on all CPUs, there we switch the FPU
* state during every context switch, regardless of whether the task
* has used FPU instructions in that time slice or not. This is done
* because modern FPU context saving instructions are able to optimize
@@ -271,7 +274,7 @@ static void __init fpu__init_system_xstate_size_legacy(void)
* to use 'eager' restores, if we detect that a task is using the FPU
* frequently. See the fpu->counter logic in fpu/internal.h for that. ]
*/
-static enum { AUTO, ENABLE, DISABLE } eagerfpu = AUTO;
+static enum { ENABLE, DISABLE } eagerfpu = ENABLE;

/*
* Find supported xfeatures based on cpu features and command-line input.
@@ -348,15 +351,9 @@ static void __init fpu__init_system_ctx_switch(void)
*/
static void __init fpu__init_parse_early_param(void)
{
- /*
- * No need to check "eagerfpu=auto" again, since it is the
- * initial default.
- */
if (cmdline_find_option_bool(boot_command_line, "eagerfpu=off")) {
eagerfpu = DISABLE;
fpu__clear_eager_fpu_features();
- } else if (cmdline_find_option_bool(boot_command_line, "eagerfpu=on")) {
- eagerfpu = ENABLE;
}

if (cmdline_find_option_bool(boot_command_line, "no387"))
diff --git a/arch/x86/kernel/traps.c b/arch/x86/kernel/traps.c
index ade185a46b1d..36a9c017540e 100644
--- a/arch/x86/kernel/traps.c
+++ b/arch/x86/kernel/traps.c
@@ -750,10 +750,9 @@ dotraplinkage void
do_device_not_available(struct pt_regs *regs, long error_code)
{
RCU_LOCKDEP_WARN(!rcu_is_watching(), "entry code didn't wake RCU");
- BUG_ON(use_eager_fpu());

#ifdef CONFIG_MATH_EMULATION
- if (read_cr0() & X86_CR0_EM) {
+ if (!boot_cpu_has(X86_FEATURE_FPU) && (read_cr0() & X86_CR0_EM)) {
struct math_emu_info info = { };

conditional_sti(regs);