1
0
Fork 0
alistair23-linux/arch/x86/include/asm/fpu/internal.h

596 lines
15 KiB
C
Raw Normal View History

/*
* Copyright (C) 1994 Linus Torvalds
*
* Pentium III FXSR, SSE support
* General FPU state handling cleanups
* Gareth Hughes <gareth@valinux.com>, May 2000
* x86-64 work by Andi Kleen 2002
*/
#ifndef _ASM_X86_FPU_INTERNAL_H
#define _ASM_X86_FPU_INTERNAL_H
#include <linux/regset.h>
#include <linux/compat.h>
#include <linux/sched.h>
#include <linux/slab.h>
#include <asm/user.h>
#include <asm/fpu/api.h>
#include <asm/fpu/xsave.h>
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
#ifdef CONFIG_X86_64
# include <asm/sigcontext32.h>
# include <asm/user32.h>
struct ksignal;
int ia32_setup_rt_frame(int sig, struct ksignal *ksig,
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
compat_sigset_t *set, struct pt_regs *regs);
int ia32_setup_frame(int sig, struct ksignal *ksig,
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
compat_sigset_t *set, struct pt_regs *regs);
#else
# define user_i387_ia32_struct user_i387_struct
# define user32_fxsr_struct user_fxsr_struct
# define ia32_setup_frame __setup_frame
# define ia32_setup_rt_frame __setup_rt_frame
#endif
#define MXCSR_DEFAULT 0x1f80
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
extern unsigned int mxcsr_feature_mask;
extern void fpu__init_cpu(void);
extern void eager_fpu_init(void);
extern void fpu__init_system_xstate(void);
extern void fpu__init_cpu_xstate(void);
extern void fpu__init_system(struct cpuinfo_x86 *c);
extern void fpu__activate_curr(struct fpu *fpu);
extern void fpstate_init(struct fpu *fpu);
extern void fpu__clear(struct task_struct *tsk);
extern int dump_fpu(struct pt_regs *, struct user_i387_struct *);
extern void fpu__restore(void);
extern void fpu__init_check_bugs(void);
extern void fpu__resume_cpu(void);
DECLARE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
extern void convert_from_fxsr(struct user_i387_ia32_struct *env,
struct task_struct *tsk);
extern void convert_to_fxsr(struct task_struct *tsk,
const struct user_i387_ia32_struct *env);
extern user_regset_active_fn regset_fpregs_active, regset_xregset_fpregs_active;
extern user_regset_get_fn fpregs_get, xfpregs_get, fpregs_soft_get,
xstateregs_get;
extern user_regset_set_fn fpregs_set, xfpregs_set, fpregs_soft_set,
xstateregs_set;
/*
* xstateregs_active == regset_fpregs_active. Please refer to the comment
* at the definition of regset_fpregs_active.
*/
#define xstateregs_active regset_fpregs_active
#ifdef CONFIG_MATH_EMULATION
extern void finit_soft_fpu(struct i387_soft_struct *soft);
#else
static inline void finit_soft_fpu(struct i387_soft_struct *soft) {}
#endif
/*
* Must be run with preemption disabled: this clears the fpu_fpregs_owner_ctx,
* on this CPU.
*
* This will disable any lazy FPU state restore of the current FPU state,
* but if the current thread owns the FPU, it will still be saved by.
*/
static inline void __cpu_disable_lazy_restore(unsigned int cpu)
{
per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
}
static inline int fpu_want_lazy_restore(struct fpu *fpu, unsigned int cpu)
{
return fpu == this_cpu_read_stable(fpu_fpregs_owner_ctx) && cpu == fpu->last_cpu;
}
static inline int is_ia32_compat_frame(void)
{
return config_enabled(CONFIG_IA32_EMULATION) &&
test_thread_flag(TIF_IA32);
}
static inline int is_ia32_frame(void)
{
return config_enabled(CONFIG_X86_32) || is_ia32_compat_frame();
}
static inline int is_x32_frame(void)
{
return config_enabled(CONFIG_X86_X32_ABI) && test_thread_flag(TIF_X32);
}
#define X87_FSW_ES (1 << 7) /* Exception Summary */
static __always_inline __pure bool use_eager_fpu(void)
{
return static_cpu_has_safe(X86_FEATURE_EAGER_FPU);
}
static __always_inline __pure bool use_xsaveopt(void)
{
return static_cpu_has_safe(X86_FEATURE_XSAVEOPT);
}
static __always_inline __pure bool use_xsave(void)
{
return static_cpu_has_safe(X86_FEATURE_XSAVE);
}
static __always_inline __pure bool use_fxsr(void)
{
return static_cpu_has_safe(X86_FEATURE_FXSR);
}
static inline void fx_finit(struct i387_fxsave_struct *fx)
{
fx->cwd = 0x37f;
fx->mxcsr = MXCSR_DEFAULT;
}
extern void __sanitize_i387_state(struct task_struct *);
static inline void sanitize_i387_state(struct task_struct *tsk)
{
if (!use_xsaveopt())
return;
__sanitize_i387_state(tsk);
}
#define user_insn(insn, output, input...) \
({ \
int err; \
asm volatile(ASM_STAC "\n" \
"1:" #insn "\n\t" \
"2: " ASM_CLAC "\n" \
".section .fixup,\"ax\"\n" \
"3: movl $-1,%[err]\n" \
" jmp 2b\n" \
".previous\n" \
_ASM_EXTABLE(1b, 3b) \
: [err] "=r" (err), output \
: "0"(0), input); \
err; \
})
#define check_insn(insn, output, input...) \
({ \
int err; \
asm volatile("1:" #insn "\n\t" \
"2:\n" \
".section .fixup,\"ax\"\n" \
"3: movl $-1,%[err]\n" \
" jmp 2b\n" \
".previous\n" \
_ASM_EXTABLE(1b, 3b) \
: [err] "=r" (err), output \
: "0"(0), input); \
err; \
})
static inline int fsave_user(struct i387_fsave_struct __user *fx)
{
return user_insn(fnsave %[fx]; fwait, [fx] "=m" (*fx), "m" (*fx));
}
static inline int fxsave_user(struct i387_fxsave_struct __user *fx)
{
if (config_enabled(CONFIG_X86_32))
return user_insn(fxsave %[fx], [fx] "=m" (*fx), "m" (*fx));
else if (config_enabled(CONFIG_AS_FXSAVEQ))
return user_insn(fxsaveq %[fx], [fx] "=m" (*fx), "m" (*fx));
/* See comment in fpu_fxsave() below. */
return user_insn(rex64/fxsave (%[fx]), "=m" (*fx), [fx] "R" (fx));
}
static inline int fxrstor_checking(struct i387_fxsave_struct *fx)
{
if (config_enabled(CONFIG_X86_32))
return check_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
else if (config_enabled(CONFIG_AS_FXSAVEQ))
return check_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
/* See comment in fpu_fxsave() below. */
return check_insn(rex64/fxrstor (%[fx]), "=m" (*fx), [fx] "R" (fx),
"m" (*fx));
}
static inline int fxrstor_user(struct i387_fxsave_struct __user *fx)
{
if (config_enabled(CONFIG_X86_32))
return user_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
else if (config_enabled(CONFIG_AS_FXSAVEQ))
return user_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
/* See comment in fpu_fxsave() below. */
return user_insn(rex64/fxrstor (%[fx]), "=m" (*fx), [fx] "R" (fx),
"m" (*fx));
}
static inline int frstor_checking(struct i387_fsave_struct *fx)
{
return check_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
}
static inline int frstor_user(struct i387_fsave_struct __user *fx)
{
return user_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
}
static inline void fpu_fxsave(struct fpu *fpu)
{
if (config_enabled(CONFIG_X86_32))
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
asm volatile( "fxsave %[fx]" : [fx] "=m" (fpu->state.fxsave));
else if (config_enabled(CONFIG_AS_FXSAVEQ))
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
asm volatile("fxsaveq %[fx]" : [fx] "=m" (fpu->state.fxsave));
else {
/* Using "rex64; fxsave %0" is broken because, if the memory
* operand uses any extended registers for addressing, a second
* REX prefix will be generated (to the assembler, rex64
* followed by semicolon is a separate instruction), and hence
* the 64-bitness is lost.
*
* Using "fxsaveq %0" would be the ideal choice, but is only
* supported starting with gas 2.16.
*
* Using, as a workaround, the properly prefixed form below
* isn't accepted by any binutils version so far released,
* complaining that the same type of prefix is used twice if
* an extended register is needed for addressing (fix submitted
* to mainline 2005-11-21).
*
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
* asm volatile("rex64/fxsave %0" : "=m" (fpu->state.fxsave));
*
* This, however, we can work around by forcing the compiler to
* select an addressing mode that doesn't require extended
* registers.
*/
asm volatile( "rex64/fxsave (%[fx])"
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
: "=m" (fpu->state.fxsave)
: [fx] "R" (&fpu->state.fxsave));
}
}
/*
* These must be called with preempt disabled. Returns
x86/fpu: Rename fpu_save_init() to copy_fpregs_to_fpstate() So fpu_save_init() is a historic name that got its name when the only way the FPU state was FNSAVE, which cleared (well, destroyed) the FPU state after saving it. Nowadays the name is misleading, because ever since the introduction of FXSAVE (and more modern FPU saving instructions) the 'we need to reload the FPU state' part is only true if there's a pending FPU exception [*], which is almost never the case. So rename it to copy_fpregs_to_fpstate() to make it clear what's happening. Also add a few comments about why we cannot keep registers in certain cases. Also clean up the control flow a bit, to make it more apparent when we are dropping/keeping FP registers, and to optimize the common case (of keeping fpregs) some more. [*] Probably not true anymore, modern instructions always leave the FPU state intact, even if exceptions are pending: because pending FP exceptions are posted on the next FP instruction, not asynchronously. They were truly asynchronous back in the IRQ13 case, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. But a cleanup patch is obviously not the place to change subtle behavior. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 18:53:16 -06:00
* 'true' if the FPU state is still intact and we can
* keep registers active.
*
* The legacy FNSAVE instruction cleared all FPU state
* unconditionally, so registers are essentially destroyed.
* Modern FPU state can be kept in registers, if there are
x86/fpu: Optimize copy_fpregs_to_fpstate() by removing the FNCLEX synchronization with FP exceptions So we have the following ancient code in copy_fpregs_to_fpstate(): if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) { asm volatile("fnclex"); goto drop_fpregs; } which clears pending FPU exceptions and then drops registers, which causes the next FP instruction of the saved context to re-load the saved FPU state, with all pending exceptions marked properly, and will re-start the exception handling mechanism in the hardware. Since FPU exceptions are always issued on instruction boundaries, in particular on the next FP instruction following the exception generating instruction, there's no fear of getting an FP exception asynchronously. They were truly asynchronous back in the IRQ13 days, when the FPU was a weird and expensive co-processor that did its own processing, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. With the introduction of optimized XSAVE support there's a new complication: if the xstate features bit indicates that a particular state component is unused (in 'init state'), then the hardware does not guarantee that the XSAVE (et al) instruction keeps the underlying FPU state image in memory valid and current. In practice this means that the hardware won't write it, and the exceptions flag in the state might be an older version, with it still being set. This meant that we had to check the xfeatures flag as well, adding another memory load and branch to a critical hot path of the scheduler. So optimize all this by removing both the old quirk and the new check, and straight-line optimizing the most common cases with likely() hints. Quite a bit of code gets removed this way: arch/x86/kernel/process_64.o: text data bss dec filename 5484 8 0 5492 process_64.o.before 5416 8 0 5424 process_64.o.after Now there's also a chance that some weird behavior or erratum was masked by our IRQ13 handling quirk (or that I misunderstood the nature of the quirk), and that this change triggers some badness. There's no real good way to protect against that possibility other than keeping this change well isolated, well commented and well bisectable. If you bisect a weird (or not so weird) breakage to this commit then please let us know! Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 19:32:18 -06:00
* no pending FP exceptions.
*/
x86/fpu: Rename fpu_save_init() to copy_fpregs_to_fpstate() So fpu_save_init() is a historic name that got its name when the only way the FPU state was FNSAVE, which cleared (well, destroyed) the FPU state after saving it. Nowadays the name is misleading, because ever since the introduction of FXSAVE (and more modern FPU saving instructions) the 'we need to reload the FPU state' part is only true if there's a pending FPU exception [*], which is almost never the case. So rename it to copy_fpregs_to_fpstate() to make it clear what's happening. Also add a few comments about why we cannot keep registers in certain cases. Also clean up the control flow a bit, to make it more apparent when we are dropping/keeping FP registers, and to optimize the common case (of keeping fpregs) some more. [*] Probably not true anymore, modern instructions always leave the FPU state intact, even if exceptions are pending: because pending FP exceptions are posted on the next FP instruction, not asynchronously. They were truly asynchronous back in the IRQ13 case, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. But a cleanup patch is obviously not the place to change subtle behavior. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 18:53:16 -06:00
static inline int copy_fpregs_to_fpstate(struct fpu *fpu)
{
x86/fpu: Optimize copy_fpregs_to_fpstate() by removing the FNCLEX synchronization with FP exceptions So we have the following ancient code in copy_fpregs_to_fpstate(): if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) { asm volatile("fnclex"); goto drop_fpregs; } which clears pending FPU exceptions and then drops registers, which causes the next FP instruction of the saved context to re-load the saved FPU state, with all pending exceptions marked properly, and will re-start the exception handling mechanism in the hardware. Since FPU exceptions are always issued on instruction boundaries, in particular on the next FP instruction following the exception generating instruction, there's no fear of getting an FP exception asynchronously. They were truly asynchronous back in the IRQ13 days, when the FPU was a weird and expensive co-processor that did its own processing, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. With the introduction of optimized XSAVE support there's a new complication: if the xstate features bit indicates that a particular state component is unused (in 'init state'), then the hardware does not guarantee that the XSAVE (et al) instruction keeps the underlying FPU state image in memory valid and current. In practice this means that the hardware won't write it, and the exceptions flag in the state might be an older version, with it still being set. This meant that we had to check the xfeatures flag as well, adding another memory load and branch to a critical hot path of the scheduler. So optimize all this by removing both the old quirk and the new check, and straight-line optimizing the most common cases with likely() hints. Quite a bit of code gets removed this way: arch/x86/kernel/process_64.o: text data bss dec filename 5484 8 0 5492 process_64.o.before 5416 8 0 5424 process_64.o.after Now there's also a chance that some weird behavior or erratum was masked by our IRQ13 handling quirk (or that I misunderstood the nature of the quirk), and that this change triggers some badness. There's no real good way to protect against that possibility other than keeping this change well isolated, well commented and well bisectable. If you bisect a weird (or not so weird) breakage to this commit then please let us know! Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 19:32:18 -06:00
if (likely(use_xsave())) {
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
xsave_state(&fpu->state.xsave);
x86/fpu: Optimize copy_fpregs_to_fpstate() by removing the FNCLEX synchronization with FP exceptions So we have the following ancient code in copy_fpregs_to_fpstate(): if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) { asm volatile("fnclex"); goto drop_fpregs; } which clears pending FPU exceptions and then drops registers, which causes the next FP instruction of the saved context to re-load the saved FPU state, with all pending exceptions marked properly, and will re-start the exception handling mechanism in the hardware. Since FPU exceptions are always issued on instruction boundaries, in particular on the next FP instruction following the exception generating instruction, there's no fear of getting an FP exception asynchronously. They were truly asynchronous back in the IRQ13 days, when the FPU was a weird and expensive co-processor that did its own processing, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. With the introduction of optimized XSAVE support there's a new complication: if the xstate features bit indicates that a particular state component is unused (in 'init state'), then the hardware does not guarantee that the XSAVE (et al) instruction keeps the underlying FPU state image in memory valid and current. In practice this means that the hardware won't write it, and the exceptions flag in the state might be an older version, with it still being set. This meant that we had to check the xfeatures flag as well, adding another memory load and branch to a critical hot path of the scheduler. So optimize all this by removing both the old quirk and the new check, and straight-line optimizing the most common cases with likely() hints. Quite a bit of code gets removed this way: arch/x86/kernel/process_64.o: text data bss dec filename 5484 8 0 5492 process_64.o.before 5416 8 0 5424 process_64.o.after Now there's also a chance that some weird behavior or erratum was masked by our IRQ13 handling quirk (or that I misunderstood the nature of the quirk), and that this change triggers some badness. There's no real good way to protect against that possibility other than keeping this change well isolated, well commented and well bisectable. If you bisect a weird (or not so weird) breakage to this commit then please let us know! Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 19:32:18 -06:00
return 1;
}
x86/fpu: Optimize copy_fpregs_to_fpstate() by removing the FNCLEX synchronization with FP exceptions So we have the following ancient code in copy_fpregs_to_fpstate(): if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) { asm volatile("fnclex"); goto drop_fpregs; } which clears pending FPU exceptions and then drops registers, which causes the next FP instruction of the saved context to re-load the saved FPU state, with all pending exceptions marked properly, and will re-start the exception handling mechanism in the hardware. Since FPU exceptions are always issued on instruction boundaries, in particular on the next FP instruction following the exception generating instruction, there's no fear of getting an FP exception asynchronously. They were truly asynchronous back in the IRQ13 days, when the FPU was a weird and expensive co-processor that did its own processing, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. With the introduction of optimized XSAVE support there's a new complication: if the xstate features bit indicates that a particular state component is unused (in 'init state'), then the hardware does not guarantee that the XSAVE (et al) instruction keeps the underlying FPU state image in memory valid and current. In practice this means that the hardware won't write it, and the exceptions flag in the state might be an older version, with it still being set. This meant that we had to check the xfeatures flag as well, adding another memory load and branch to a critical hot path of the scheduler. So optimize all this by removing both the old quirk and the new check, and straight-line optimizing the most common cases with likely() hints. Quite a bit of code gets removed this way: arch/x86/kernel/process_64.o: text data bss dec filename 5484 8 0 5492 process_64.o.before 5416 8 0 5424 process_64.o.after Now there's also a chance that some weird behavior or erratum was masked by our IRQ13 handling quirk (or that I misunderstood the nature of the quirk), and that this change triggers some badness. There's no real good way to protect against that possibility other than keeping this change well isolated, well commented and well bisectable. If you bisect a weird (or not so weird) breakage to this commit then please let us know! Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 19:32:18 -06:00
if (likely(use_fxsr())) {
fpu_fxsave(fpu);
return 1;
}
/*
x86/fpu: Optimize copy_fpregs_to_fpstate() by removing the FNCLEX synchronization with FP exceptions So we have the following ancient code in copy_fpregs_to_fpstate(): if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) { asm volatile("fnclex"); goto drop_fpregs; } which clears pending FPU exceptions and then drops registers, which causes the next FP instruction of the saved context to re-load the saved FPU state, with all pending exceptions marked properly, and will re-start the exception handling mechanism in the hardware. Since FPU exceptions are always issued on instruction boundaries, in particular on the next FP instruction following the exception generating instruction, there's no fear of getting an FP exception asynchronously. They were truly asynchronous back in the IRQ13 days, when the FPU was a weird and expensive co-processor that did its own processing, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. With the introduction of optimized XSAVE support there's a new complication: if the xstate features bit indicates that a particular state component is unused (in 'init state'), then the hardware does not guarantee that the XSAVE (et al) instruction keeps the underlying FPU state image in memory valid and current. In practice this means that the hardware won't write it, and the exceptions flag in the state might be an older version, with it still being set. This meant that we had to check the xfeatures flag as well, adding another memory load and branch to a critical hot path of the scheduler. So optimize all this by removing both the old quirk and the new check, and straight-line optimizing the most common cases with likely() hints. Quite a bit of code gets removed this way: arch/x86/kernel/process_64.o: text data bss dec filename 5484 8 0 5492 process_64.o.before 5416 8 0 5424 process_64.o.after Now there's also a chance that some weird behavior or erratum was masked by our IRQ13 handling quirk (or that I misunderstood the nature of the quirk), and that this change triggers some badness. There's no real good way to protect against that possibility other than keeping this change well isolated, well commented and well bisectable. If you bisect a weird (or not so weird) breakage to this commit then please let us know! Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 19:32:18 -06:00
* Legacy FPU register saving, FNSAVE always clears FPU registers,
* so we have to mark them inactive:
*/
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
asm volatile("fnsave %[fx]; fwait" : [fx] "=m" (fpu->state.fsave));
x86/fpu: Rename fpu_save_init() to copy_fpregs_to_fpstate() So fpu_save_init() is a historic name that got its name when the only way the FPU state was FNSAVE, which cleared (well, destroyed) the FPU state after saving it. Nowadays the name is misleading, because ever since the introduction of FXSAVE (and more modern FPU saving instructions) the 'we need to reload the FPU state' part is only true if there's a pending FPU exception [*], which is almost never the case. So rename it to copy_fpregs_to_fpstate() to make it clear what's happening. Also add a few comments about why we cannot keep registers in certain cases. Also clean up the control flow a bit, to make it more apparent when we are dropping/keeping FP registers, and to optimize the common case (of keeping fpregs) some more. [*] Probably not true anymore, modern instructions always leave the FPU state intact, even if exceptions are pending: because pending FP exceptions are posted on the next FP instruction, not asynchronously. They were truly asynchronous back in the IRQ13 case, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. But a cleanup patch is obviously not the place to change subtle behavior. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 18:53:16 -06:00
return 0;
}
extern void fpu__save(struct fpu *fpu);
static inline int fpu_restore_checking(struct fpu *fpu)
{
if (use_xsave())
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
return fpu_xrstor_checking(&fpu->state.xsave);
else if (use_fxsr())
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
return fxrstor_checking(&fpu->state.fxsave);
else
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
return frstor_checking(&fpu->state.fsave);
}
static inline int restore_fpu_checking(struct fpu *fpu)
{
/*
* AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception is
* pending. Clear the x87 state here by setting it to fixed values.
* "m" is a random variable that should be in L1.
*/
if (unlikely(static_cpu_has_bug_safe(X86_BUG_FXSAVE_LEAK))) {
asm volatile(
"fnclex\n\t"
"emms\n\t"
"fildl %P[addr]" /* set F?P to defined value */
: : [addr] "m" (fpu->fpregs_active));
}
return fpu_restore_checking(fpu);
}
/*
* Wrap lazy FPU TS handling in a 'hw fpregs activation/deactivation'
* idiom, which is then paired with the sw-flag (fpregs_active) later on:
*/
static inline void __fpregs_activate_hw(void)
{
if (!use_eager_fpu())
clts();
}
static inline void __fpregs_deactivate_hw(void)
{
if (!use_eager_fpu())
stts();
}
/* Must be paired with an 'stts' (fpregs_deactivate_hw()) after! */
static inline void __fpregs_deactivate(struct fpu *fpu)
{
fpu->fpregs_active = 0;
this_cpu_write(fpu_fpregs_owner_ctx, NULL);
}
/* Must be paired with a 'clts' (fpregs_activate_hw()) before! */
static inline void __fpregs_activate(struct fpu *fpu)
{
fpu->fpregs_active = 1;
this_cpu_write(fpu_fpregs_owner_ctx, fpu);
}
/*
* The question "does this thread have fpu access?"
* is slightly racy, since preemption could come in
* and revoke it immediately after the test.
*
* However, even in that very unlikely scenario,
* we can just assume we have FPU access - typically
* to save the FP state - we'll just take a #NM
* fault and get the FPU access back.
*/
static inline int user_has_fpu(void)
{
return current->thread.fpu.fpregs_active;
}
/*
* Encapsulate the CR0.TS handling together with the
* software flag.
*
* These generally need preemption protection to work,
* do try to avoid using these on their own.
*/
static inline void fpregs_activate(struct fpu *fpu)
{
__fpregs_activate_hw();
__fpregs_activate(fpu);
}
static inline void fpregs_deactivate(struct fpu *fpu)
{
__fpregs_deactivate(fpu);
__fpregs_deactivate_hw();
}
static inline void drop_fpu(struct fpu *fpu)
x86, fpu: use non-lazy fpu restore for processors supporting xsave Fundamental model of the current Linux kernel is to lazily init and restore FPU instead of restoring the task state during context switch. This changes that fundamental lazy model to the non-lazy model for the processors supporting xsave feature. Reasons driving this model change are: i. Newer processors support optimized state save/restore using xsaveopt and xrstor by tracking the INIT state and MODIFIED state during context-switch. This is faster than modifying the cr0.TS bit which has serializing semantics. ii. Newer glibc versions use SSE for some of the optimized copy/clear routines. With certain workloads (like boot, kernel-compilation etc), application completes its work with in the first 5 task switches, thus taking upto 5 #DNA traps with the kernel not getting a chance to apply the above mentioned pre-load heuristic. iii. Some xstate features (like AMD's LWP feature) don't honor the cr0.TS bit and thus will not work correctly in the presence of lazy restore. Non-lazy state restore is needed for enabling such features. Some data on a two socket SNB system: * Saved 20K DNA exceptions during boot on a two socket SNB system. * Saved 50K DNA exceptions during kernel-compilation workload. * Improved throughput of the AVX based checksumming function inside the kernel by ~15% as xsave/xrstor is faster than the serializing clts/stts pair. Also now kernel_fpu_begin/end() relies on the patched alternative instructions. So move check_fpu() which uses the kernel_fpu_begin/end() after alternative_instructions(). Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1345842782-24175-7-git-send-email-suresh.b.siddha@intel.com Merge 32-bit boot fix from, Link: http://lkml.kernel.org/r/1347300665-6209-4-git-send-email-suresh.b.siddha@intel.com Cc: Jim Kukunas <james.t.kukunas@linux.intel.com> Cc: NeilBrown <neilb@suse.de> Cc: Avi Kivity <avi@redhat.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-08-24 15:13:02 -06:00
{
/*
* Forget coprocessor state..
*/
preempt_disable();
fpu->counter = 0;
if (fpu->fpregs_active) {
x86, fpu: use non-lazy fpu restore for processors supporting xsave Fundamental model of the current Linux kernel is to lazily init and restore FPU instead of restoring the task state during context switch. This changes that fundamental lazy model to the non-lazy model for the processors supporting xsave feature. Reasons driving this model change are: i. Newer processors support optimized state save/restore using xsaveopt and xrstor by tracking the INIT state and MODIFIED state during context-switch. This is faster than modifying the cr0.TS bit which has serializing semantics. ii. Newer glibc versions use SSE for some of the optimized copy/clear routines. With certain workloads (like boot, kernel-compilation etc), application completes its work with in the first 5 task switches, thus taking upto 5 #DNA traps with the kernel not getting a chance to apply the above mentioned pre-load heuristic. iii. Some xstate features (like AMD's LWP feature) don't honor the cr0.TS bit and thus will not work correctly in the presence of lazy restore. Non-lazy state restore is needed for enabling such features. Some data on a two socket SNB system: * Saved 20K DNA exceptions during boot on a two socket SNB system. * Saved 50K DNA exceptions during kernel-compilation workload. * Improved throughput of the AVX based checksumming function inside the kernel by ~15% as xsave/xrstor is faster than the serializing clts/stts pair. Also now kernel_fpu_begin/end() relies on the patched alternative instructions. So move check_fpu() which uses the kernel_fpu_begin/end() after alternative_instructions(). Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1345842782-24175-7-git-send-email-suresh.b.siddha@intel.com Merge 32-bit boot fix from, Link: http://lkml.kernel.org/r/1347300665-6209-4-git-send-email-suresh.b.siddha@intel.com Cc: Jim Kukunas <james.t.kukunas@linux.intel.com> Cc: NeilBrown <neilb@suse.de> Cc: Avi Kivity <avi@redhat.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-08-24 15:13:02 -06:00
/* Ignore delayed exceptions from user space */
asm volatile("1: fwait\n"
"2:\n"
_ASM_EXTABLE(1b, 2b));
fpregs_deactivate(fpu);
x86, fpu: use non-lazy fpu restore for processors supporting xsave Fundamental model of the current Linux kernel is to lazily init and restore FPU instead of restoring the task state during context switch. This changes that fundamental lazy model to the non-lazy model for the processors supporting xsave feature. Reasons driving this model change are: i. Newer processors support optimized state save/restore using xsaveopt and xrstor by tracking the INIT state and MODIFIED state during context-switch. This is faster than modifying the cr0.TS bit which has serializing semantics. ii. Newer glibc versions use SSE for some of the optimized copy/clear routines. With certain workloads (like boot, kernel-compilation etc), application completes its work with in the first 5 task switches, thus taking upto 5 #DNA traps with the kernel not getting a chance to apply the above mentioned pre-load heuristic. iii. Some xstate features (like AMD's LWP feature) don't honor the cr0.TS bit and thus will not work correctly in the presence of lazy restore. Non-lazy state restore is needed for enabling such features. Some data on a two socket SNB system: * Saved 20K DNA exceptions during boot on a two socket SNB system. * Saved 50K DNA exceptions during kernel-compilation workload. * Improved throughput of the AVX based checksumming function inside the kernel by ~15% as xsave/xrstor is faster than the serializing clts/stts pair. Also now kernel_fpu_begin/end() relies on the patched alternative instructions. So move check_fpu() which uses the kernel_fpu_begin/end() after alternative_instructions(). Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1345842782-24175-7-git-send-email-suresh.b.siddha@intel.com Merge 32-bit boot fix from, Link: http://lkml.kernel.org/r/1347300665-6209-4-git-send-email-suresh.b.siddha@intel.com Cc: Jim Kukunas <james.t.kukunas@linux.intel.com> Cc: NeilBrown <neilb@suse.de> Cc: Avi Kivity <avi@redhat.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-08-24 15:13:02 -06:00
}
fpu->fpstate_active = 0;
x86, fpu: use non-lazy fpu restore for processors supporting xsave Fundamental model of the current Linux kernel is to lazily init and restore FPU instead of restoring the task state during context switch. This changes that fundamental lazy model to the non-lazy model for the processors supporting xsave feature. Reasons driving this model change are: i. Newer processors support optimized state save/restore using xsaveopt and xrstor by tracking the INIT state and MODIFIED state during context-switch. This is faster than modifying the cr0.TS bit which has serializing semantics. ii. Newer glibc versions use SSE for some of the optimized copy/clear routines. With certain workloads (like boot, kernel-compilation etc), application completes its work with in the first 5 task switches, thus taking upto 5 #DNA traps with the kernel not getting a chance to apply the above mentioned pre-load heuristic. iii. Some xstate features (like AMD's LWP feature) don't honor the cr0.TS bit and thus will not work correctly in the presence of lazy restore. Non-lazy state restore is needed for enabling such features. Some data on a two socket SNB system: * Saved 20K DNA exceptions during boot on a two socket SNB system. * Saved 50K DNA exceptions during kernel-compilation workload. * Improved throughput of the AVX based checksumming function inside the kernel by ~15% as xsave/xrstor is faster than the serializing clts/stts pair. Also now kernel_fpu_begin/end() relies on the patched alternative instructions. So move check_fpu() which uses the kernel_fpu_begin/end() after alternative_instructions(). Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1345842782-24175-7-git-send-email-suresh.b.siddha@intel.com Merge 32-bit boot fix from, Link: http://lkml.kernel.org/r/1347300665-6209-4-git-send-email-suresh.b.siddha@intel.com Cc: Jim Kukunas <james.t.kukunas@linux.intel.com> Cc: NeilBrown <neilb@suse.de> Cc: Avi Kivity <avi@redhat.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-08-24 15:13:02 -06:00
preempt_enable();
}
static inline void restore_init_xstate(void)
{
if (use_xsave())
xrstor_state(&init_xstate_ctx, -1);
else
fxrstor_checking(&init_xstate_ctx.i387);
}
/*
* Reset the FPU state in the eager case and drop it in the lazy case (later use
* will reinit it).
*/
static inline void fpu_reset_state(struct fpu *fpu)
x86, fpu: use non-lazy fpu restore for processors supporting xsave Fundamental model of the current Linux kernel is to lazily init and restore FPU instead of restoring the task state during context switch. This changes that fundamental lazy model to the non-lazy model for the processors supporting xsave feature. Reasons driving this model change are: i. Newer processors support optimized state save/restore using xsaveopt and xrstor by tracking the INIT state and MODIFIED state during context-switch. This is faster than modifying the cr0.TS bit which has serializing semantics. ii. Newer glibc versions use SSE for some of the optimized copy/clear routines. With certain workloads (like boot, kernel-compilation etc), application completes its work with in the first 5 task switches, thus taking upto 5 #DNA traps with the kernel not getting a chance to apply the above mentioned pre-load heuristic. iii. Some xstate features (like AMD's LWP feature) don't honor the cr0.TS bit and thus will not work correctly in the presence of lazy restore. Non-lazy state restore is needed for enabling such features. Some data on a two socket SNB system: * Saved 20K DNA exceptions during boot on a two socket SNB system. * Saved 50K DNA exceptions during kernel-compilation workload. * Improved throughput of the AVX based checksumming function inside the kernel by ~15% as xsave/xrstor is faster than the serializing clts/stts pair. Also now kernel_fpu_begin/end() relies on the patched alternative instructions. So move check_fpu() which uses the kernel_fpu_begin/end() after alternative_instructions(). Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1345842782-24175-7-git-send-email-suresh.b.siddha@intel.com Merge 32-bit boot fix from, Link: http://lkml.kernel.org/r/1347300665-6209-4-git-send-email-suresh.b.siddha@intel.com Cc: Jim Kukunas <james.t.kukunas@linux.intel.com> Cc: NeilBrown <neilb@suse.de> Cc: Avi Kivity <avi@redhat.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-08-24 15:13:02 -06:00
{
if (!use_eager_fpu())
drop_fpu(fpu);
else
restore_init_xstate();
x86, fpu: use non-lazy fpu restore for processors supporting xsave Fundamental model of the current Linux kernel is to lazily init and restore FPU instead of restoring the task state during context switch. This changes that fundamental lazy model to the non-lazy model for the processors supporting xsave feature. Reasons driving this model change are: i. Newer processors support optimized state save/restore using xsaveopt and xrstor by tracking the INIT state and MODIFIED state during context-switch. This is faster than modifying the cr0.TS bit which has serializing semantics. ii. Newer glibc versions use SSE for some of the optimized copy/clear routines. With certain workloads (like boot, kernel-compilation etc), application completes its work with in the first 5 task switches, thus taking upto 5 #DNA traps with the kernel not getting a chance to apply the above mentioned pre-load heuristic. iii. Some xstate features (like AMD's LWP feature) don't honor the cr0.TS bit and thus will not work correctly in the presence of lazy restore. Non-lazy state restore is needed for enabling such features. Some data on a two socket SNB system: * Saved 20K DNA exceptions during boot on a two socket SNB system. * Saved 50K DNA exceptions during kernel-compilation workload. * Improved throughput of the AVX based checksumming function inside the kernel by ~15% as xsave/xrstor is faster than the serializing clts/stts pair. Also now kernel_fpu_begin/end() relies on the patched alternative instructions. So move check_fpu() which uses the kernel_fpu_begin/end() after alternative_instructions(). Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1345842782-24175-7-git-send-email-suresh.b.siddha@intel.com Merge 32-bit boot fix from, Link: http://lkml.kernel.org/r/1347300665-6209-4-git-send-email-suresh.b.siddha@intel.com Cc: Jim Kukunas <james.t.kukunas@linux.intel.com> Cc: NeilBrown <neilb@suse.de> Cc: Avi Kivity <avi@redhat.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-08-24 15:13:02 -06:00
}
/*
* FPU state switching for scheduling.
*
* This is a two-stage process:
*
* - switch_fpu_prepare() saves the old state and
* sets the new state of the CR0.TS bit. This is
* done within the context of the old process.
*
* - switch_fpu_finish() restores the new state as
* necessary.
*/
typedef struct { int preload; } fpu_switch_t;
static inline fpu_switch_t
switch_fpu_prepare(struct fpu *old_fpu, struct fpu *new_fpu, int cpu)
{
fpu_switch_t fpu;
x86, fpu: use non-lazy fpu restore for processors supporting xsave Fundamental model of the current Linux kernel is to lazily init and restore FPU instead of restoring the task state during context switch. This changes that fundamental lazy model to the non-lazy model for the processors supporting xsave feature. Reasons driving this model change are: i. Newer processors support optimized state save/restore using xsaveopt and xrstor by tracking the INIT state and MODIFIED state during context-switch. This is faster than modifying the cr0.TS bit which has serializing semantics. ii. Newer glibc versions use SSE for some of the optimized copy/clear routines. With certain workloads (like boot, kernel-compilation etc), application completes its work with in the first 5 task switches, thus taking upto 5 #DNA traps with the kernel not getting a chance to apply the above mentioned pre-load heuristic. iii. Some xstate features (like AMD's LWP feature) don't honor the cr0.TS bit and thus will not work correctly in the presence of lazy restore. Non-lazy state restore is needed for enabling such features. Some data on a two socket SNB system: * Saved 20K DNA exceptions during boot on a two socket SNB system. * Saved 50K DNA exceptions during kernel-compilation workload. * Improved throughput of the AVX based checksumming function inside the kernel by ~15% as xsave/xrstor is faster than the serializing clts/stts pair. Also now kernel_fpu_begin/end() relies on the patched alternative instructions. So move check_fpu() which uses the kernel_fpu_begin/end() after alternative_instructions(). Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1345842782-24175-7-git-send-email-suresh.b.siddha@intel.com Merge 32-bit boot fix from, Link: http://lkml.kernel.org/r/1347300665-6209-4-git-send-email-suresh.b.siddha@intel.com Cc: Jim Kukunas <james.t.kukunas@linux.intel.com> Cc: NeilBrown <neilb@suse.de> Cc: Avi Kivity <avi@redhat.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-08-24 15:13:02 -06:00
/*
* 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 &&
(use_eager_fpu() || new_fpu->counter > 5);
if (old_fpu->fpregs_active) {
x86/fpu: Rename fpu_save_init() to copy_fpregs_to_fpstate() So fpu_save_init() is a historic name that got its name when the only way the FPU state was FNSAVE, which cleared (well, destroyed) the FPU state after saving it. Nowadays the name is misleading, because ever since the introduction of FXSAVE (and more modern FPU saving instructions) the 'we need to reload the FPU state' part is only true if there's a pending FPU exception [*], which is almost never the case. So rename it to copy_fpregs_to_fpstate() to make it clear what's happening. Also add a few comments about why we cannot keep registers in certain cases. Also clean up the control flow a bit, to make it more apparent when we are dropping/keeping FP registers, and to optimize the common case (of keeping fpregs) some more. [*] Probably not true anymore, modern instructions always leave the FPU state intact, even if exceptions are pending: because pending FP exceptions are posted on the next FP instruction, not asynchronously. They were truly asynchronous back in the IRQ13 case, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. But a cleanup patch is obviously not the place to change subtle behavior. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 18:53:16 -06:00
if (!copy_fpregs_to_fpstate(old_fpu))
old_fpu->last_cpu = -1;
else
old_fpu->last_cpu = cpu;
/* But leave fpu_fpregs_owner_ctx! */
old_fpu->fpregs_active = 0;
/* Don't change CR0.TS if we just switch! */
if (fpu.preload) {
new_fpu->counter++;
__fpregs_activate(new_fpu);
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
prefetch(&new_fpu->state);
} else {
__fpregs_deactivate_hw();
}
} else {
old_fpu->counter = 0;
old_fpu->last_cpu = -1;
if (fpu.preload) {
new_fpu->counter++;
if (fpu_want_lazy_restore(new_fpu, cpu))
fpu.preload = 0;
else
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
prefetch(&new_fpu->state);
fpregs_activate(new_fpu);
}
}
return fpu;
}
/*
* By the time this gets called, we've already cleared CR0.TS and
* given the process the FPU if we are going to preload the FPU
* state - all we need to do is to conditionally restore the register
* state itself.
*/
static inline void switch_fpu_finish(struct fpu *new_fpu, fpu_switch_t fpu_switch)
{
if (fpu_switch.preload) {
if (unlikely(restore_fpu_checking(new_fpu)))
fpu_reset_state(new_fpu);
}
}
/*
* Signal frame handlers...
*/
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
extern int save_xstate_sig(void __user *buf, void __user *fx, int size);
extern int __restore_xstate_sig(void __user *buf, void __user *fx, int size);
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
static inline int xstate_sigframe_size(void)
{
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
return use_xsave() ? xstate_size + FP_XSTATE_MAGIC2_SIZE : xstate_size;
}
static inline int restore_xstate_sig(void __user *buf, int ia32_frame)
{
void __user *buf_fx = buf;
int size = xstate_sigframe_size();
if (ia32_frame && use_fxsr()) {
buf_fx = buf + sizeof(struct i387_fsave_struct);
size += sizeof(struct i387_fsave_struct);
}
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
return __restore_xstate_sig(buf, buf_fx, size);
}
/*
* Needs to be preemption-safe.
*
* NOTE! user_fpu_begin() must be used only immediately before restoring
* the save state. It does not do any saving/restoring on its own. In
* lazy FPU mode, it is just an optimization to avoid a #NM exception,
* the task can lose the FPU right after preempt_enable().
*/
static inline void user_fpu_begin(void)
{
struct fpu *fpu = &current->thread.fpu;
preempt_disable();
if (!user_has_fpu())
fpregs_activate(fpu);
preempt_enable();
}
/*
* i387 state interaction
*/
static inline unsigned short get_fpu_cwd(struct task_struct *tsk)
{
if (cpu_has_fxsr) {
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
return tsk->thread.fpu.state.fxsave.cwd;
} else {
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
return (unsigned short)tsk->thread.fpu.state.fsave.cwd;
}
}
static inline unsigned short get_fpu_swd(struct task_struct *tsk)
{
if (cpu_has_fxsr) {
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
return tsk->thread.fpu.state.fxsave.swd;
} else {
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
return (unsigned short)tsk->thread.fpu.state.fsave.swd;
}
}
static inline unsigned short get_fpu_mxcsr(struct task_struct *tsk)
{
if (cpu_has_xmm) {
x86/fpu: Simplify FPU handling by embedding the fpstate in task_struct (again) So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49276e ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b5386 ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by: Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-26 20:19:39 -06:00
return tsk->thread.fpu.state.fxsave.mxcsr;
} else {
return MXCSR_DEFAULT;
}
}
extern int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu);
x86, fpu: Unify signal handling code paths for x86 and x86_64 kernels Currently for x86 and x86_32 binaries, fpstate in the user sigframe is copied to/from the fpstate in the task struct. And in the case of signal delivery for x86_64 binaries, if the fpstate is live in the CPU registers, then the live state is copied directly to the user sigframe. Otherwise fpstate in the task struct is copied to the user sigframe. During restore, fpstate in the user sigframe is restored directly to the live CPU registers. Historically, different code paths led to different bugs. For example, x86_64 code path was not preemption safe till recently. Also there is lot of code duplication for support of new features like xsave etc. Unify signal handling code paths for x86 and x86_64 kernels. New strategy is as follows: Signal delivery: Both for 32/64-bit frames, align the core math frame area to 64bytes as needed by xsave (this where the main fpu/extended state gets copied to and excludes the legacy compatibility fsave header for the 32-bit [f]xsave frames). If the state is live, copy the register state directly to the user frame. If not live, copy the state in the thread struct to the user frame. And for 32-bit [f]xsave frames, construct the fsave header separately before the actual [f]xsave area. Signal return: As the 32-bit frames with [f]xstate has an additional 'fsave' header, copy everything back from the user sigframe to the fpstate in the task structure and reconstruct the fxstate from the 'fsave' header (Also user passed pointers may not be correctly aligned for any attempt to directly restore any partial state). At the next fpstate usage, everything will be restored to the live CPU registers. For all the 64-bit frames and the 32-bit fsave frame, restore the state from the user sigframe directly to the live CPU registers. 64-bit signals always restored the math frame directly, so we can expect the math frame pointer to be correctly aligned. For 32-bit fsave frames, there are no alignment requirements, so we can restore the state directly. "lat_sig catch" microbenchmark numbers (for x86, x86_64, x86_32 binaries) are with in the noise range with this change. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1343171129-2747-4-git-send-email-suresh.b.siddha@intel.com [ Merged in compilation fix ] Link: http://lkml.kernel.org/r/1344544736.8326.17.camel@sbsiddha-desk.sc.intel.com Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2012-07-24 17:05:29 -06:00
static inline unsigned long
alloc_mathframe(unsigned long sp, int ia32_frame, unsigned long *buf_fx,
unsigned long *size)
{
unsigned long frame_size = xstate_sigframe_size();
*buf_fx = sp = round_down(sp - frame_size, 64);
if (ia32_frame && use_fxsr()) {
frame_size += sizeof(struct i387_fsave_struct);
sp -= sizeof(struct i387_fsave_struct);
}
*size = frame_size;
return sp;
}
#endif /* _ASM_X86_FPU_INTERNAL_H */