x86, fpu: lazy allocation of FPU area - v5
[linux-2.6.git] / arch / x86 / kernel / i387.c
index d2e39e6..db6839b 100644 (file)
@@ -5,51 +5,48 @@
  *  General FPU state handling cleanups
  *     Gareth Hughes <gareth@valinux.com>, May 2000
  */
-
-#include <linux/sched.h>
 #include <linux/module.h>
 #include <linux/regset.h>
+#include <linux/sched.h>
+
+#include <asm/sigcontext.h>
 #include <asm/processor.h>
-#include <asm/i387.h>
 #include <asm/math_emu.h>
-#include <asm/sigcontext.h>
-#include <asm/user.h>
-#include <asm/ptrace.h>
 #include <asm/uaccess.h>
+#include <asm/ptrace.h>
+#include <asm/i387.h>
+#include <asm/user.h>
 
 #ifdef CONFIG_X86_64
-
-#include <asm/sigcontext32.h>
-#include <asm/user32.h>
-
+# include <asm/sigcontext32.h>
+# include <asm/user32.h>
 #else
-
-#define        save_i387_ia32          save_i387
-#define        restore_i387_ia32       restore_i387
-
-#define _fpstate_ia32          _fpstate
-#define user_i387_ia32_struct  user_i387_struct
-#define user32_fxsr_struct     user_fxsr_struct
-
+# define save_i387_ia32                save_i387
+# define restore_i387_ia32     restore_i387
+# define _fpstate_ia32         _fpstate
+# define user_i387_ia32_struct user_i387_struct
+# define user32_fxsr_struct    user_fxsr_struct
 #endif
 
 #ifdef CONFIG_MATH_EMULATION
-#define HAVE_HWFP (boot_cpu_data.hard_math)
+# define HAVE_HWFP             (boot_cpu_data.hard_math)
 #else
-#define HAVE_HWFP 1
+# define HAVE_HWFP             1
 #endif
 
-static unsigned int mxcsr_feature_mask __read_mostly = 0xffffffffu;
+static unsigned int            mxcsr_feature_mask __read_mostly = 0xffffffffu;
+unsigned int xstate_size;
+static struct i387_fxsave_struct fx_scratch __cpuinitdata;
 
-void mxcsr_feature_mask_init(void)
+void __cpuinit mxcsr_feature_mask_init(void)
 {
        unsigned long mask = 0;
+
        clts();
        if (cpu_has_fxsr) {
-               memset(&current->thread.i387.fxsave, 0,
-                      sizeof(struct i387_fxsave_struct));
-               asm volatile("fxsave %0" : : "m" (current->thread.i387.fxsave));
-               mask = current->thread.i387.fxsave.mxcsr_mask;
+               memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
+               asm volatile("fxsave %0" : : "m" (fx_scratch));
+               mask = fx_scratch.mxcsr_mask;
                if (mask == 0)
                        mask = 0x0000ffbf;
        }
@@ -57,6 +54,16 @@ void mxcsr_feature_mask_init(void)
        stts();
 }
 
+void __init init_thread_xstate(void)
+{
+       if (cpu_has_fxsr)
+               xstate_size = sizeof(struct i387_fxsave_struct);
+#ifdef CONFIG_X86_32
+       else
+               xstate_size = sizeof(struct i387_fsave_struct);
+#endif
+}
+
 #ifdef CONFIG_X86_64
 /*
  * Called at bootup to set up the initial FPU state that is later cloned
@@ -65,14 +72,11 @@ void mxcsr_feature_mask_init(void)
 void __cpuinit fpu_init(void)
 {
        unsigned long oldcr0 = read_cr0();
-       extern void __bad_fxsave_alignment(void);
 
-       if (offsetof(struct task_struct, thread.i387.fxsave) & 15)
-               __bad_fxsave_alignment();
        set_in_cr4(X86_CR4_OSFXSR);
        set_in_cr4(X86_CR4_OSXMMEXCPT);
 
-       write_cr0(oldcr0 & ~((1UL<<3)|(1UL<<2))); /* clear TS and EM */
+       write_cr0(oldcr0 & ~(X86_CR0_TS|X86_CR0_EM)); /* clear TS and EM */
 
        mxcsr_feature_mask_init();
        /* clean state in init */
@@ -87,32 +91,44 @@ void __cpuinit fpu_init(void)
  * value at reset if we support XMM instructions and then
  * remeber the current task has used the FPU.
  */
-void init_fpu(struct task_struct *tsk)
+int init_fpu(struct task_struct *tsk)
 {
        if (tsk_used_math(tsk)) {
                if (tsk == current)
                        unlazy_fpu(tsk);
-               return;
+               return 0;
+       }
+
+       /*
+        * Memory allocation at the first usage of the FPU and other state.
+        */
+       if (!tsk->thread.xstate) {
+               tsk->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
+                                                     GFP_KERNEL);
+               if (!tsk->thread.xstate)
+                       return -ENOMEM;
        }
 
        if (cpu_has_fxsr) {
-               memset(&tsk->thread.i387.fxsave, 0,
-                      sizeof(struct i387_fxsave_struct));
-               tsk->thread.i387.fxsave.cwd = 0x37f;
+               struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
+
+               memset(fx, 0, xstate_size);
+               fx->cwd = 0x37f;
                if (cpu_has_xmm)
-                       tsk->thread.i387.fxsave.mxcsr = MXCSR_DEFAULT;
+                       fx->mxcsr = MXCSR_DEFAULT;
        } else {
-               memset(&tsk->thread.i387.fsave, 0,
-                      sizeof(struct i387_fsave_struct));
-               tsk->thread.i387.fsave.cwd = 0xffff037fu;
-               tsk->thread.i387.fsave.swd = 0xffff0000u;
-               tsk->thread.i387.fsave.twd = 0xffffffffu;
-               tsk->thread.i387.fsave.fos = 0xffff0000u;
+               struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
+               memset(fp, 0, xstate_size);
+               fp->cwd = 0xffff037fu;
+               fp->swd = 0xffff0000u;
+               fp->twd = 0xffffffffu;
+               fp->fos = 0xffff0000u;
        }
        /*
         * Only the device not available exception or ptrace can call init_fpu.
         */
        set_stopped_child_used_math(tsk);
+       return 0;
 }
 
 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
@@ -129,13 +145,17 @@ int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
                unsigned int pos, unsigned int count,
                void *kbuf, void __user *ubuf)
 {
+       int ret;
+
        if (!cpu_has_fxsr)
                return -ENODEV;
 
-       init_fpu(target);
+       ret = init_fpu(target);
+       if (ret)
+               return ret;
 
        return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
-                                  &target->thread.i387.fxsave, 0, -1);
+                                  &target->thread.xstate->fxsave, 0, -1);
 }
 
 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
@@ -147,16 +167,19 @@ int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
        if (!cpu_has_fxsr)
                return -ENODEV;
 
-       init_fpu(target);
+       ret = init_fpu(target);
+       if (ret)
+               return ret;
+
        set_stopped_child_used_math(target);
 
        ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
-                                &target->thread.i387.fxsave, 0, -1);
+                                &target->thread.xstate->fxsave, 0, -1);
 
        /*
         * mxcsr reserved bits must be masked to zero for security reasons.
         */
-       target->thread.i387.fxsave.mxcsr &= mxcsr_feature_mask;
+       target->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 
        return ret;
 }
@@ -178,6 +201,7 @@ static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
        tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
        tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
        tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
+
        return tmp;
 }
 
@@ -232,10 +256,10 @@ static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
  * FXSR floating point environment conversions.
  */
 
-static void convert_from_fxsr(struct user_i387_ia32_struct *env,
-                             struct task_struct *tsk)
+static void
+convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
 {
-       struct i387_fxsave_struct *fxsave = &tsk->thread.i387.fxsave;
+       struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
        struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
        struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
        int i;
@@ -252,10 +276,11 @@ static void convert_from_fxsr(struct user_i387_ia32_struct *env,
                 * should be actually ds/cs at fpu exception time, but
                 * that information is not available in 64bit mode.
                 */
-               asm("mov %%ds,%0" : "=r" (env->fos));
-               asm("mov %%cs,%0" : "=r" (env->fcs));
+               asm("mov %%ds, %[fos]" : [fos] "=r" (env->fos));
+               asm("mov %%cs, %[fcs]" : [fcs] "=r" (env->fcs));
        } else {
                struct pt_regs *regs = task_pt_regs(tsk);
+
                env->fos = 0xffff0000 | tsk->thread.ds;
                env->fcs = regs->cs;
        }
@@ -274,7 +299,7 @@ static void convert_to_fxsr(struct task_struct *tsk,
                            const struct user_i387_ia32_struct *env)
 
 {
-       struct i387_fxsave_struct *fxsave = &tsk->thread.i387.fxsave;
+       struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
        struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
        struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
        int i;
@@ -303,15 +328,20 @@ int fpregs_get(struct task_struct *target, const struct user_regset *regset,
               void *kbuf, void __user *ubuf)
 {
        struct user_i387_ia32_struct env;
+       int ret;
 
        if (!HAVE_HWFP)
                return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
 
-       init_fpu(target);
+       ret = init_fpu(target);
+       if (ret)
+               return ret;
 
-       if (!cpu_has_fxsr)
+       if (!cpu_has_fxsr) {
                return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
-                                          &target->thread.i387.fsave, 0, -1);
+                                          &target->thread.xstate->fsave, 0,
+                                          -1);
+       }
 
        if (kbuf && pos == 0 && count == sizeof(env)) {
                convert_from_fxsr(kbuf, target);
@@ -319,6 +349,7 @@ int fpregs_get(struct task_struct *target, const struct user_regset *regset,
        }
 
        convert_from_fxsr(&env, target);
+
        return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 }
 
@@ -332,12 +363,16 @@ int fpregs_set(struct task_struct *target, const struct user_regset *regset,
        if (!HAVE_HWFP)
                return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
 
-       init_fpu(target);
+       ret = init_fpu(target);
+       if (ret)
+               return ret;
+
        set_stopped_child_used_math(target);
 
-       if (!cpu_has_fxsr)
+       if (!cpu_has_fxsr) {
                return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
-                                         &target->thread.i387.fsave, 0, -1);
+                                         &target->thread.xstate->fsave, 0, -1);
+       }
 
        if (pos > 0 || count < sizeof(env))
                convert_from_fxsr(&env, target);
@@ -356,11 +391,11 @@ int fpregs_set(struct task_struct *target, const struct user_regset *regset,
 static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
 {
        struct task_struct *tsk = current;
+       struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 
        unlazy_fpu(tsk);
-       tsk->thread.i387.fsave.status = tsk->thread.i387.fsave.swd;
-       if (__copy_to_user(buf, &tsk->thread.i387.fsave,
-                          sizeof(struct i387_fsave_struct)))
+       fp->status = fp->swd;
+       if (__copy_to_user(buf, fp, sizeof(struct i387_fsave_struct)))
                return -1;
        return 1;
 }
@@ -368,6 +403,7 @@ static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
 static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
 {
        struct task_struct *tsk = current;
+       struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
        struct user_i387_ia32_struct env;
        int err = 0;
 
@@ -377,12 +413,12 @@ static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
        if (__copy_to_user(buf, &env, sizeof(env)))
                return -1;
 
-       err |= __put_user(tsk->thread.i387.fxsave.swd, &buf->status);
+       err |= __put_user(fx->swd, &buf->status);
        err |= __put_user(X86_FXSR_MAGIC, &buf->magic);
        if (err)
                return -1;
 
-       if (__copy_to_user(&buf->_fxsr_env[0], &tsk->thread.i387.fxsave,
+       if (__copy_to_user(&buf->_fxsr_env[0], fx,
                           sizeof(struct i387_fxsave_struct)))
                return -1;
        return 1;
@@ -392,46 +428,48 @@ int save_i387_ia32(struct _fpstate_ia32 __user *buf)
 {
        if (!used_math())
                return 0;
-
-       /* This will cause a "finit" to be triggered by the next
+       /*
+        * This will cause a "finit" to be triggered by the next
         * attempted FPU operation by the 'current' process.
         */
        clear_used_math();
 
-       if (HAVE_HWFP) {
-               if (cpu_has_fxsr) {
-                       return save_i387_fxsave(buf);
-               } else {
-                       return save_i387_fsave(buf);
-               }
-       } else {
+       if (!HAVE_HWFP) {
                return fpregs_soft_get(current, NULL,
                                       0, sizeof(struct user_i387_ia32_struct),
                                       NULL, buf) ? -1 : 1;
        }
+
+       if (cpu_has_fxsr)
+               return save_i387_fxsave(buf);
+       else
+               return save_i387_fsave(buf);
 }
 
 static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
 {
        struct task_struct *tsk = current;
+
        clear_fpu(tsk);
-       return __copy_from_user(&tsk->thread.i387.fsave, buf,
+       return __copy_from_user(&tsk->thread.xstate->fsave, buf,
                                sizeof(struct i387_fsave_struct));
 }
 
 static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf)
 {
-       int err;
        struct task_struct *tsk = current;
        struct user_i387_ia32_struct env;
+       int err;
+
        clear_fpu(tsk);
-       err = __copy_from_user(&tsk->thread.i387.fxsave, &buf->_fxsr_env[0],
+       err = __copy_from_user(&tsk->thread.xstate->fxsave, &buf->_fxsr_env[0],
                               sizeof(struct i387_fxsave_struct));
        /* mxcsr reserved bits must be masked to zero for security reasons */
-       tsk->thread.i387.fxsave.mxcsr &= mxcsr_feature_mask;
+       tsk->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
        if (err || __copy_from_user(&env, buf, sizeof(env)))
                return 1;
        convert_to_fxsr(tsk, &env);
+
        return 0;
 }
 
@@ -440,17 +478,17 @@ int restore_i387_ia32(struct _fpstate_ia32 __user *buf)
        int err;
 
        if (HAVE_HWFP) {
-               if (cpu_has_fxsr) {
+               if (cpu_has_fxsr)
                        err = restore_i387_fxsave(buf);
-               } else {
+               else
                        err = restore_i387_fsave(buf);
-               }
        } else {
                err = fpregs_soft_set(current, NULL,
                                      0, sizeof(struct user_i387_ia32_struct),
                                      NULL, buf) != 0;
        }
        set_used_math();
+
        return err;
 }
 
@@ -463,8 +501,8 @@ int restore_i387_ia32(struct _fpstate_ia32 __user *buf)
  */
 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
 {
-       int fpvalid;
        struct task_struct *tsk = current;
+       int fpvalid;
 
        fpvalid = !!used_math();
        if (fpvalid)