2011-04-27 14:26:52

by Hans Rosenfeld

[permalink] [raw]
Subject: [PATCH 1/2] x86, xsave: fix non-lazy allocation of the xsave area

A single static xsave area just for init is not enough, since there are
more user processes that are directly executed by kernel threads. Add a
call to a new arch-specific function to flush_old_exec(), which will in
turn call fpu_alloc() to allocate a xsave area if necessary.

Signed-off-by: Hans Rosenfeld <[email protected]>
---
arch/x86/include/asm/i387.h | 6 ------
arch/x86/include/asm/thread_info.h | 2 ++
arch/x86/kernel/process.c | 7 +++++++
fs/exec.c | 8 ++++++++
4 files changed, 17 insertions(+), 6 deletions(-)

diff --git a/arch/x86/include/asm/i387.h b/arch/x86/include/asm/i387.h
index 989c0ac..0448f45 100644
--- a/arch/x86/include/asm/i387.h
+++ b/arch/x86/include/asm/i387.h
@@ -329,15 +329,9 @@ static inline void fpu_copy(struct fpu *dst, struct fpu *src)
}

extern void fpu_finit(struct fpu *fpu);
-static union thread_xstate __init_xstate, *init_xstate = &__init_xstate;

static inline void fpu_clear(struct fpu *fpu)
{
- if (!fpu_allocated(fpu)) {
- BUG_ON(init_xstate == NULL);
- fpu->state = init_xstate;
- init_xstate = NULL;
- }
memset(fpu->state, 0, xstate_size);
fpu_finit(fpu);
set_used_math();
diff --git a/arch/x86/include/asm/thread_info.h b/arch/x86/include/asm/thread_info.h
index 0e691c6..07c8cfe 100644
--- a/arch/x86/include/asm/thread_info.h
+++ b/arch/x86/include/asm/thread_info.h
@@ -265,6 +265,8 @@ static inline void set_restore_sigmask(void)
extern void arch_task_cache_init(void);
extern void free_thread_info(struct thread_info *ti);
extern int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src);
+extern int arch_prealloc_fpu(struct task_struct *tsk);
+#define arch_prealloc_fpu arch_prealloc_fpu
#define arch_task_cache_init arch_task_cache_init
#endif
#endif /* _ASM_X86_THREAD_INFO_H */
diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
index 0382f98..3edfbf2 100644
--- a/arch/x86/kernel/process.c
+++ b/arch/x86/kernel/process.c
@@ -26,6 +26,13 @@
struct kmem_cache *task_xstate_cachep;
EXPORT_SYMBOL_GPL(task_xstate_cachep);

+int arch_prealloc_fpu(struct task_struct *tsk)
+{
+ if (!fpu_allocated(&tsk->thread.fpu))
+ return fpu_alloc(&tsk->thread.fpu);
+ return 0;
+}
+
int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
{
int ret;
diff --git a/fs/exec.c b/fs/exec.c
index 5e62d26..932b9ad 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1022,10 +1022,18 @@ void set_task_comm(struct task_struct *tsk, char *buf)
perf_event_comm(tsk);
}

+#if !defined(arch_prealloc_fpu)
+#define arch_prealloc_fpu(tsk) (0)
+#endif
+
int flush_old_exec(struct linux_binprm * bprm)
{
int retval;

+ retval = arch_prealloc_fpu(current);
+ if (retval)
+ goto out;
+
/*
* Make sure we have a private signal table and that
* we are unassociated from the previous thread group.
--
1.5.6.5


2011-04-28 05:55:27

by Hans Rosenfeld

[permalink] [raw]
Subject: [tip:x86/xsave] x86, xsave: fix non-lazy allocation of the xsave area

Commit-ID: 300c6120b4653043e8a80b24e0483c0d223c5aac
Gitweb: http://git.kernel.org/tip/300c6120b4653043e8a80b24e0483c0d223c5aac
Author: Hans Rosenfeld <[email protected]>
AuthorDate: Wed, 27 Apr 2011 16:26:34 +0200
Committer: H. Peter Anvin <[email protected]>
CommitDate: Wed, 27 Apr 2011 18:03:45 -0700

x86, xsave: fix non-lazy allocation of the xsave area

A single static xsave area just for init is not enough, since there are
more user processes that are directly executed by kernel threads. Add a
call to a new arch-specific function to flush_old_exec(), which will in
turn call fpu_alloc() to allocate a xsave area if necessary.

Signed-off-by: Hans Rosenfeld <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: H. Peter Anvin <[email protected]>
---
arch/x86/include/asm/i387.h | 6 ------
arch/x86/include/asm/thread_info.h | 2 ++
arch/x86/kernel/process.c | 7 +++++++
fs/exec.c | 8 ++++++++
4 files changed, 17 insertions(+), 6 deletions(-)

diff --git a/arch/x86/include/asm/i387.h b/arch/x86/include/asm/i387.h
index f410d16..41f1d7e 100644
--- a/arch/x86/include/asm/i387.h
+++ b/arch/x86/include/asm/i387.h
@@ -327,15 +327,9 @@ static inline void fpu_copy(struct fpu *dst, struct fpu *src)
}

extern void fpu_finit(struct fpu *fpu);
-static union thread_xstate __init_xstate, *init_xstate = &__init_xstate;

static inline void fpu_clear(struct fpu *fpu)
{
- if (!fpu_allocated(fpu)) {
- BUG_ON(init_xstate == NULL);
- fpu->state = init_xstate;
- init_xstate = NULL;
- }
memset(fpu->state, 0, xstate_size);
fpu_finit(fpu);
set_used_math();
diff --git a/arch/x86/include/asm/thread_info.h b/arch/x86/include/asm/thread_info.h
index 0e691c6..07c8cfe 100644
--- a/arch/x86/include/asm/thread_info.h
+++ b/arch/x86/include/asm/thread_info.h
@@ -265,6 +265,8 @@ static inline void set_restore_sigmask(void)
extern void arch_task_cache_init(void);
extern void free_thread_info(struct thread_info *ti);
extern int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src);
+extern int arch_prealloc_fpu(struct task_struct *tsk);
+#define arch_prealloc_fpu arch_prealloc_fpu
#define arch_task_cache_init arch_task_cache_init
#endif
#endif /* _ASM_X86_THREAD_INFO_H */
diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
index d46cbe4..a70f620 100644
--- a/arch/x86/kernel/process.c
+++ b/arch/x86/kernel/process.c
@@ -26,6 +26,13 @@
struct kmem_cache *task_xstate_cachep;
EXPORT_SYMBOL_GPL(task_xstate_cachep);

+int arch_prealloc_fpu(struct task_struct *tsk)
+{
+ if (!fpu_allocated(&tsk->thread.fpu))
+ return fpu_alloc(&tsk->thread.fpu);
+ return 0;
+}
+
int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
{
int ret;
diff --git a/fs/exec.c b/fs/exec.c
index 5e62d26..932b9ad 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1022,10 +1022,18 @@ void set_task_comm(struct task_struct *tsk, char *buf)
perf_event_comm(tsk);
}

+#if !defined(arch_prealloc_fpu)
+#define arch_prealloc_fpu(tsk) (0)
+#endif
+
int flush_old_exec(struct linux_binprm * bprm)
{
int retval;

+ retval = arch_prealloc_fpu(current);
+ if (retval)
+ goto out;
+
/*
* Make sure we have a private signal table and that
* we are unassociated from the previous thread group.