x86: move switch_to macro to system.h
[sfrench/cifs-2.6.git] / include / asm-x86 / system.h
index 692562b48f2a6534074df384a7dff97217ea06a8..ba3403f1d020ec9c92db925deb5a795e864713ff 100644 (file)
@@ -1,5 +1,371 @@
+#ifndef _ASM_X86_SYSTEM_H_
+#define _ASM_X86_SYSTEM_H_
+
+#include <asm/asm.h>
+
+#include <linux/kernel.h>
+
 #ifdef CONFIG_X86_32
+#define AT_VECTOR_SIZE_ARCH 2 /* entries in ARCH_DLINFO */
+
+struct task_struct; /* one of the stranger aspects of C forward declarations */
+extern struct task_struct *FASTCALL(__switch_to(struct task_struct *prev,
+                                               struct task_struct *next));
+
+/*
+ * Saving eflags is important. It switches not only IOPL between tasks,
+ * it also protects other tasks from NT leaking through sysenter etc.
+ */
+#define switch_to(prev, next, last) do {                               \
+       unsigned long esi, edi;                                         \
+       asm volatile("pushfl\n\t"               /* Save flags */        \
+                    "pushl %%ebp\n\t"                                  \
+                    "movl %%esp,%0\n\t"        /* save ESP */          \
+                    "movl %5,%%esp\n\t"        /* restore ESP */       \
+                    "movl $1f,%1\n\t"          /* save EIP */          \
+                    "pushl %6\n\t"             /* restore EIP */       \
+                    "jmp __switch_to\n"                                \
+                    "1:\t"                                             \
+                    "popl %%ebp\n\t"                                   \
+                    "popfl"                                            \
+                    :"=m" (prev->thread.sp), "=m" (prev->thread.ip),   \
+                     "=a" (last), "=S" (esi), "=D" (edi)               \
+                    :"m" (next->thread.sp), "m" (next->thread.ip),     \
+                     "2" (prev), "d" (next));                          \
+} while (0)
+
 # include "system_32.h"
 #else
+#define __SAVE(reg, offset) "movq %%" #reg ",(14-" #offset ")*8(%%rsp)\n\t"
+#define __RESTORE(reg, offset) "movq (14-" #offset ")*8(%%rsp),%%" #reg "\n\t"
+
+/* frame pointer must be last for get_wchan */
+#define SAVE_CONTEXT    "pushf ; pushq %%rbp ; movq %%rsi,%%rbp\n\t"
+#define RESTORE_CONTEXT "movq %%rbp,%%rsi ; popq %%rbp ; popf\t"
+
+#define __EXTRA_CLOBBER  \
+       , "rcx", "rbx", "rdx", "r8", "r9", "r10", "r11", \
+         "r12", "r13", "r14", "r15"
+
+/* Save restore flags to clear handle leaking NT */
+#define switch_to(prev, next, last) \
+       asm volatile(SAVE_CONTEXT                                         \
+            "movq %%rsp,%P[threadrsp](%[prev])\n\t" /* save RSP */       \
+            "movq %P[threadrsp](%[next]),%%rsp\n\t" /* restore RSP */    \
+            "call __switch_to\n\t"                                       \
+            ".globl thread_return\n"                                     \
+            "thread_return:\n\t"                                         \
+            "movq %%gs:%P[pda_pcurrent],%%rsi\n\t"                       \
+            "movq %P[thread_info](%%rsi),%%r8\n\t"                       \
+            LOCK_PREFIX "btr  %[tif_fork],%P[ti_flags](%%r8)\n\t"        \
+            "movq %%rax,%%rdi\n\t"                                       \
+            "jc   ret_from_fork\n\t"                                     \
+            RESTORE_CONTEXT                                              \
+            : "=a" (last)                                                \
+            : [next] "S" (next), [prev] "D" (prev),                      \
+              [threadrsp] "i" (offsetof(struct task_struct, thread.sp)), \
+              [ti_flags] "i" (offsetof(struct thread_info, flags)),      \
+              [tif_fork] "i" (TIF_FORK),                                 \
+              [thread_info] "i" (offsetof(struct task_struct, stack)),   \
+              [pda_pcurrent] "i" (offsetof(struct x8664_pda, pcurrent))  \
+            : "memory", "cc" __EXTRA_CLOBBER)
 # include "system_64.h"
 #endif
+
+#ifdef __KERNEL__
+#define _set_base(addr, base) do { unsigned long __pr; \
+__asm__ __volatile__ ("movw %%dx,%1\n\t" \
+       "rorl $16,%%edx\n\t" \
+       "movb %%dl,%2\n\t" \
+       "movb %%dh,%3" \
+       :"=&d" (__pr) \
+       :"m" (*((addr)+2)), \
+        "m" (*((addr)+4)), \
+        "m" (*((addr)+7)), \
+        "0" (base) \
+       ); } while (0)
+
+#define _set_limit(addr, limit) do { unsigned long __lr; \
+__asm__ __volatile__ ("movw %%dx,%1\n\t" \
+       "rorl $16,%%edx\n\t" \
+       "movb %2,%%dh\n\t" \
+       "andb $0xf0,%%dh\n\t" \
+       "orb %%dh,%%dl\n\t" \
+       "movb %%dl,%2" \
+       :"=&d" (__lr) \
+       :"m" (*(addr)), \
+        "m" (*((addr)+6)), \
+        "0" (limit) \
+       ); } while (0)
+
+#define set_base(ldt, base) _set_base(((char *)&(ldt)) , (base))
+#define set_limit(ldt, limit) _set_limit(((char *)&(ldt)) , ((limit)-1))
+
+extern void load_gs_index(unsigned);
+
+/*
+ * Load a segment. Fall back on loading the zero
+ * segment if something goes wrong..
+ */
+#define loadsegment(seg, value)                        \
+       asm volatile("\n"                       \
+               "1:\t"                          \
+               "movl %k0,%%" #seg "\n"         \
+               "2:\n"                          \
+               ".section .fixup,\"ax\"\n"      \
+               "3:\t"                          \
+               "movl %k1, %%" #seg "\n\t"      \
+               "jmp 2b\n"                      \
+               ".previous\n"                   \
+               ".section __ex_table,\"a\"\n\t" \
+               _ASM_ALIGN "\n\t"               \
+               _ASM_PTR " 1b,3b\n"             \
+               ".previous"                     \
+               : :"r" (value), "r" (0))
+
+
+/*
+ * Save a segment register away
+ */
+#define savesegment(seg, value) \
+       asm volatile("mov %%" #seg ",%0":"=rm" (value))
+
+static inline unsigned long get_limit(unsigned long segment)
+{
+       unsigned long __limit;
+       __asm__("lsll %1,%0"
+               :"=r" (__limit):"r" (segment));
+       return __limit+1;
+}
+
+static inline void native_clts(void)
+{
+       asm volatile ("clts");
+}
+
+/*
+ * Volatile isn't enough to prevent the compiler from reordering the
+ * read/write functions for the control registers and messing everything up.
+ * A memory clobber would solve the problem, but would prevent reordering of
+ * all loads stores around it, which can hurt performance. Solution is to
+ * use a variable and mimic reads and writes to it to enforce serialization
+ */
+static unsigned long __force_order;
+
+static inline unsigned long native_read_cr0(void)
+{
+       unsigned long val;
+       asm volatile("mov %%cr0,%0\n\t" :"=r" (val), "=m" (__force_order));
+       return val;
+}
+
+static inline void native_write_cr0(unsigned long val)
+{
+       asm volatile("mov %0,%%cr0": :"r" (val), "m" (__force_order));
+}
+
+static inline unsigned long native_read_cr2(void)
+{
+       unsigned long val;
+       asm volatile("mov %%cr2,%0\n\t" :"=r" (val), "=m" (__force_order));
+       return val;
+}
+
+static inline void native_write_cr2(unsigned long val)
+{
+       asm volatile("mov %0,%%cr2": :"r" (val), "m" (__force_order));
+}
+
+static inline unsigned long native_read_cr3(void)
+{
+       unsigned long val;
+       asm volatile("mov %%cr3,%0\n\t" :"=r" (val), "=m" (__force_order));
+       return val;
+}
+
+static inline void native_write_cr3(unsigned long val)
+{
+       asm volatile("mov %0,%%cr3": :"r" (val), "m" (__force_order));
+}
+
+static inline unsigned long native_read_cr4(void)
+{
+       unsigned long val;
+       asm volatile("mov %%cr4,%0\n\t" :"=r" (val), "=m" (__force_order));
+       return val;
+}
+
+static inline unsigned long native_read_cr4_safe(void)
+{
+       unsigned long val;
+       /* This could fault if %cr4 does not exist. In x86_64, a cr4 always
+        * exists, so it will never fail. */
+#ifdef CONFIG_X86_32
+       asm volatile("1: mov %%cr4, %0          \n"
+               "2:                             \n"
+               ".section __ex_table,\"a\"      \n"
+               ".long 1b,2b                    \n"
+               ".previous                      \n"
+               : "=r" (val), "=m" (__force_order) : "0" (0));
+#else
+       val = native_read_cr4();
+#endif
+       return val;
+}
+
+static inline void native_write_cr4(unsigned long val)
+{
+       asm volatile("mov %0,%%cr4": :"r" (val), "m" (__force_order));
+}
+
+static inline void native_wbinvd(void)
+{
+       asm volatile("wbinvd": : :"memory");
+}
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#else
+#define read_cr0()     (native_read_cr0())
+#define write_cr0(x)   (native_write_cr0(x))
+#define read_cr2()     (native_read_cr2())
+#define write_cr2(x)   (native_write_cr2(x))
+#define read_cr3()     (native_read_cr3())
+#define write_cr3(x)   (native_write_cr3(x))
+#define read_cr4()     (native_read_cr4())
+#define read_cr4_safe()        (native_read_cr4_safe())
+#define write_cr4(x)   (native_write_cr4(x))
+#define wbinvd()       (native_wbinvd())
+
+/* Clear the 'TS' bit */
+#define clts()         (native_clts())
+
+#endif/* CONFIG_PARAVIRT */
+
+#define stts() write_cr0(8 | read_cr0())
+
+#endif /* __KERNEL__ */
+
+static inline void clflush(void *__p)
+{
+       asm volatile("clflush %0" : "+m" (*(char __force *)__p));
+}
+
+#define nop() __asm__ __volatile__ ("nop")
+
+void disable_hlt(void);
+void enable_hlt(void);
+
+extern int es7000_plat;
+void cpu_idle_wait(void);
+
+extern unsigned long arch_align_stack(unsigned long sp);
+extern void free_init_pages(char *what, unsigned long begin, unsigned long end);
+
+void default_idle(void);
+
+/*
+ * Force strict CPU ordering.
+ * And yes, this is required on UP too when we're talking
+ * to devices.
+ */
+#ifdef CONFIG_X86_32
+/*
+ * For now, "wmb()" doesn't actually do anything, as all
+ * Intel CPU's follow what Intel calls a *Processor Order*,
+ * in which all writes are seen in the program order even
+ * outside the CPU.
+ *
+ * I expect future Intel CPU's to have a weaker ordering,
+ * but I'd also expect them to finally get their act together
+ * and add some real memory barriers if so.
+ *
+ * Some non intel clones support out of order store. wmb() ceases to be a
+ * nop for these.
+ */
+#define mb() alternative("lock; addl $0,0(%%esp)", "mfence", X86_FEATURE_XMM2)
+#define rmb() alternative("lock; addl $0,0(%%esp)", "lfence", X86_FEATURE_XMM2)
+#define wmb() alternative("lock; addl $0,0(%%esp)", "sfence", X86_FEATURE_XMM)
+#else
+#define mb()   asm volatile("mfence":::"memory")
+#define rmb()  asm volatile("lfence":::"memory")
+#define wmb()  asm volatile("sfence" ::: "memory")
+#endif
+
+/**
+ * read_barrier_depends - Flush all pending reads that subsequents reads
+ * depend on.
+ *
+ * No data-dependent reads from memory-like regions are ever reordered
+ * over this barrier.  All reads preceding this primitive are guaranteed
+ * to access memory (but not necessarily other CPUs' caches) before any
+ * reads following this primitive that depend on the data return by
+ * any of the preceding reads.  This primitive is much lighter weight than
+ * rmb() on most CPUs, and is never heavier weight than is
+ * rmb().
+ *
+ * These ordering constraints are respected by both the local CPU
+ * and the compiler.
+ *
+ * Ordering is not guaranteed by anything other than these primitives,
+ * not even by data dependencies.  See the documentation for
+ * memory_barrier() for examples and URLs to more information.
+ *
+ * For example, the following code would force ordering (the initial
+ * value of "a" is zero, "b" is one, and "p" is "&a"):
+ *
+ * <programlisting>
+ *     CPU 0                           CPU 1
+ *
+ *     b = 2;
+ *     memory_barrier();
+ *     p = &b;                         q = p;
+ *                                     read_barrier_depends();
+ *                                     d = *q;
+ * </programlisting>
+ *
+ * because the read of "*q" depends on the read of "p" and these
+ * two reads are separated by a read_barrier_depends().  However,
+ * the following code, with the same initial values for "a" and "b":
+ *
+ * <programlisting>
+ *     CPU 0                           CPU 1
+ *
+ *     a = 2;
+ *     memory_barrier();
+ *     b = 3;                          y = b;
+ *                                     read_barrier_depends();
+ *                                     x = a;
+ * </programlisting>
+ *
+ * does not enforce ordering, since there is no data dependency between
+ * the read of "a" and the read of "b".  Therefore, on some CPUs, such
+ * as Alpha, "y" could be set to 3 and "x" to 0.  Use rmb()
+ * in cases like this where there are no data dependencies.
+ **/
+
+#define read_barrier_depends() do { } while (0)
+
+#ifdef CONFIG_SMP
+#define smp_mb()       mb()
+#ifdef CONFIG_X86_PPRO_FENCE
+# define smp_rmb()     rmb()
+#else
+# define smp_rmb()     barrier()
+#endif
+#ifdef CONFIG_X86_OOSTORE
+# define smp_wmb()     wmb()
+#else
+# define smp_wmb()     barrier()
+#endif
+#define smp_read_barrier_depends()     read_barrier_depends()
+#define set_mb(var, value) do { (void) xchg(&var, value); } while (0)
+#else
+#define smp_mb()       barrier()
+#define smp_rmb()      barrier()
+#define smp_wmb()      barrier()
+#define smp_read_barrier_depends()     do { } while (0)
+#define set_mb(var, value) do { var = value; barrier(); } while (0)
+#endif
+
+
+#endif