2 * Copyright (c) 2003 Peter Wemm.
3 * Copyright (c) 1993 The Regents of the University of California.
4 * Copyright (c) 2008 The DragonFly Project.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * $FreeBSD: src/sys/amd64/include/cpufunc.h,v 1.139 2004/01/28 23:53:04 peter Exp $
35 * Functions to provide access to special i386 instructions.
36 * This in included in sys/systm.h, and that file should be
37 * used in preference to this.
40 #ifndef _CPU_CPUFUNC_H_
41 #define _CPU_CPUFUNC_H_
43 #include <sys/cdefs.h>
44 #include <sys/thread.h>
45 #include <machine/psl.h>
46 #include <machine/smp.h>
49 struct region_descriptor;
53 #define readb(va) (*(volatile u_int8_t *) (va))
54 #define readw(va) (*(volatile u_int16_t *) (va))
55 #define readl(va) (*(volatile u_int32_t *) (va))
56 #define readq(va) (*(volatile u_int64_t *) (va))
58 #define writeb(va, d) (*(volatile u_int8_t *) (va) = (d))
59 #define writew(va, d) (*(volatile u_int16_t *) (va) = (d))
60 #define writel(va, d) (*(volatile u_int32_t *) (va) = (d))
61 #define writeq(va, d) (*(volatile u_int64_t *) (va) = (d))
65 #include <machine/lock.h> /* XXX */
70 __asm __volatile("int $3");
76 __asm __volatile("pause":::"memory");
84 __asm __volatile("bsfl %1,%0" : "=r" (result) : "rm" (mask));
88 static __inline u_long
93 __asm __volatile("bsfq %1,%0" : "=r" (result) : "rm" (mask));
97 static __inline u_long
102 __asm __volatile("bsfq %1,%0" : "=r" (result) : "rm" (mask));
106 static __inline u_int
111 __asm __volatile("bsrl %1,%0" : "=r" (result) : "rm" (mask));
115 static __inline u_long
120 __asm __volatile("bsrq %1,%0" : "=r" (result) : "rm" (mask));
127 __asm __volatile("clflush %0" : : "m" (*(char *) addr));
131 do_cpuid(u_int ax, u_int *p)
133 __asm __volatile("cpuid"
134 : "=a" (p[0]), "=b" (p[1]), "=c" (p[2]), "=d" (p[3])
139 cpuid_count(u_int ax, u_int cx, u_int *p)
141 __asm __volatile("cpuid"
142 : "=a" (p[0]), "=b" (p[1]), "=c" (p[2]), "=d" (p[3])
143 : "0" (ax), "c" (cx));
146 #ifndef _CPU_DISABLE_INTR_DEFINED
149 cpu_disable_intr(void)
151 __asm __volatile("cli" : : : "memory");
156 #ifndef _CPU_ENABLE_INTR_DEFINED
159 cpu_enable_intr(void)
161 __asm __volatile("sti");
167 * Cpu and compiler memory ordering fence. mfence ensures strong read and
170 * A serializing or fence instruction is required here. A locked bus
171 * cycle on data for which we already own cache mastership is the most
177 __asm __volatile("mfence" : : : "memory");
181 * cpu_lfence() ensures strong read ordering for reads issued prior
182 * to the instruction verses reads issued afterwords.
184 * A serializing or fence instruction is required here. A locked bus
185 * cycle on data for which we already own cache mastership is the most
191 __asm __volatile("lfence" : : : "memory");
195 * cpu_sfence() ensures strong write ordering for writes issued prior
196 * to the instruction verses writes issued afterwords. Writes are
197 * ordered on intel cpus so we do not actually have to do anything.
204 * Don't use 'sfence' here, as it will create a lot of
205 * unnecessary stalls.
207 __asm __volatile("" : : : "memory");
211 * cpu_ccfence() prevents the compiler from reordering instructions, in
212 * particular stores, relative to the current cpu. Use cpu_sfence() if
213 * you need to guarentee ordering by both the compiler and by the cpu.
215 * This also prevents the compiler from caching memory loads into local
216 * variables across the routine.
221 __asm __volatile("" : : : "memory");
225 * This is a horrible, horrible hack that might have to be put at the
226 * end of certain procedures (on a case by case basis), just before it
227 * returns to avoid what we believe to be an unreported AMD cpu bug.
228 * Found to occur on both a Phenom II X4 820 (two of them), as well
229 * as a 48-core built around an Opteron 6168 (Id = 0x100f91 Stepping = 1).
230 * The problem does not appear to occur w/Intel cpus.
232 * The bug is likely related to either a write combining issue or the
233 * Return Address Stack (RAS) hardware cache.
235 * In particular, we had to do this for GCC's fill_sons_in_loop() routine
236 * which due to its deep recursion and stack flow appears to be able to
237 * tickle the amd cpu bug (w/ gcc-4.4.7). Adding a single 'nop' to the
238 * end of the routine just before it returns works around the bug.
240 * The bug appears to be extremely sensitive to %rip and %rsp values, to
241 * the point where even just inserting an instruction in an unrelated
242 * procedure (shifting the entire code base being run) effects the outcome.
243 * DragonFly is probably able to more readily reproduce the bug due to
244 * the stackgap randomization code. We would expect OpenBSD (where we got
245 * the stackgap randomization code from) to also be able to reproduce the
246 * issue. To date we have only reproduced the issue in DragonFly.
248 #define __AMDCPUBUG_DFLY01_AVAILABLE__
251 cpu_amdcpubug_dfly01(void)
253 __asm __volatile("nop" : : : "memory");
258 #define HAVE_INLINE_FFS
265 * Note that gcc-2's builtin ffs would be used if we didn't declare
266 * this inline or turn off the builtin. The builtin is faster but
267 * broken in gcc-2.4.5 and slower but working in gcc-2.5 and later
270 return (mask == 0 ? mask : (int)bsfl((u_int)mask) + 1);
272 /* Actually, the above is way out of date. The builtins use cmov etc */
273 return (__builtin_ffs(mask));
277 #define HAVE_INLINE_FFSL
282 return (mask == 0 ? mask : (int)bsfq((u_long)mask) + 1);
285 #define HAVE_INLINE_FLS
290 return (mask == 0 ? mask : (int)bsrl((u_int)mask) + 1);
293 #define HAVE_INLINE_FLSL
298 return (mask == 0 ? mask : (int)bsrq((u_long)mask) + 1);
301 #define HAVE_INLINE_FLSLL
304 flsll(long long mask)
306 return (flsl((long)mask));
314 __asm __volatile("hlt");
318 * The following complications are to get around gcc not having a
319 * constraint letter for the range 0..255. We still put "d" in the
320 * constraint because "i" isn't a valid constraint when the port
321 * isn't constant. This only matters for -O0 because otherwise
322 * the non-working version gets optimized away.
324 * Use an expression-statement instead of a conditional expression
325 * because gcc-2.6.0 would promote the operands of the conditional
326 * and produce poor code for "if ((inb(var) & const1) == const2)".
328 * The unnecessary test `(port) < 0x10000' is to generate a warning if
329 * the `port' has type u_short or smaller. Such types are pessimal.
330 * This actually only works for signed types. The range check is
331 * careful to avoid generating warnings.
333 #define inb(port) __extension__ ({ \
335 if (__builtin_constant_p(port) && ((port) & 0xffff) < 0x100 \
336 && (port) < 0x10000) \
337 _data = inbc(port); \
339 _data = inbv(port); \
342 #define outb(port, data) ( \
343 __builtin_constant_p(port) && ((port) & 0xffff) < 0x100 \
344 && (port) < 0x10000 \
345 ? outbc(port, data) : outbv(port, data))
347 static __inline u_char
352 __asm __volatile("inb %1,%0" : "=a" (data) : "id" ((u_short)(port)));
357 outbc(u_int port, u_char data)
359 __asm __volatile("outb %0,%1" : : "a" (data), "id" ((u_short)(port)));
362 static __inline u_char
367 * We use %%dx and not %1 here because i/o is done at %dx and not at
368 * %edx, while gcc generates inferior code (movw instead of movl)
369 * if we tell it to load (u_short) port.
371 __asm __volatile("inb %%dx,%0" : "=a" (data) : "d" (port));
375 static __inline u_int
380 __asm __volatile("inl %%dx,%0" : "=a" (data) : "d" (port));
385 insb(u_int port, void *addr, size_t cnt)
387 __asm __volatile("cld; rep; insb"
388 : "+D" (addr), "+c" (cnt)
394 insw(u_int port, void *addr, size_t cnt)
396 __asm __volatile("cld; rep; insw"
397 : "+D" (addr), "+c" (cnt)
403 insl(u_int port, void *addr, size_t cnt)
405 __asm __volatile("cld; rep; insl"
406 : "+D" (addr), "+c" (cnt)
414 __asm __volatile("invd");
419 #ifndef _CPU_INVLPG_DEFINED
422 * Invalidate a particular VA on this cpu only
424 * TLB flush for an individual page (even if it has PG_G).
425 * Only works on 486+ CPUs (i386 does not have PG_G).
428 cpu_invlpg(void *addr)
430 __asm __volatile("invlpg %0" : : "m" (*(char *)addr) : "memory");
438 __asm __volatile("rep; nop");
443 static __inline u_short
448 __asm __volatile("inw %%dx,%0" : "=a" (data) : "d" (port));
452 static __inline u_int
453 loadandclear(volatile u_int *addr)
457 __asm __volatile("xorl %0,%0; xchgl %1,%0"
458 : "=&r" (result) : "m" (*addr));
463 outbv(u_int port, u_char data)
467 * Use an unnecessary assignment to help gcc's register allocator.
468 * This make a large difference for gcc-1.40 and a tiny difference
469 * for gcc-2.6.0. For gcc-1.40, al had to be ``asm("ax")'' for
470 * best results. gcc-2.6.0 can't handle this.
473 __asm __volatile("outb %0,%%dx" : : "a" (al), "d" (port));
477 outl(u_int port, u_int data)
480 * outl() and outw() aren't used much so we haven't looked at
481 * possible micro-optimizations such as the unnecessary
482 * assignment for them.
484 __asm __volatile("outl %0,%%dx" : : "a" (data), "d" (port));
488 outsb(u_int port, const void *addr, size_t cnt)
490 __asm __volatile("cld; rep; outsb"
491 : "+S" (addr), "+c" (cnt)
496 outsw(u_int port, const void *addr, size_t cnt)
498 __asm __volatile("cld; rep; outsw"
499 : "+S" (addr), "+c" (cnt)
504 outsl(u_int port, const void *addr, size_t cnt)
506 __asm __volatile("cld; rep; outsl"
507 : "+S" (addr), "+c" (cnt)
512 outw(u_int port, u_short data)
514 __asm __volatile("outw %0,%%dx" : : "a" (data), "d" (port));
520 __asm __volatile("pause");
523 static __inline u_long
528 __asm __volatile("pushfq; popq %0" : "=r" (rf));
532 static __inline u_int64_t
537 __asm __volatile("rdmsr" : "=a" (low), "=d" (high) : "c" (msr));
538 return (low | ((u_int64_t)high << 32));
541 static __inline u_int64_t
546 __asm __volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (pmc));
547 return (low | ((u_int64_t)high << 32));
550 #define _RDTSC_SUPPORTED_
552 static __inline u_int64_t
557 __asm __volatile("rdtsc" : "=a" (low), "=d" (high));
558 return (low | ((u_int64_t)high << 32));
562 #include <machine/cputypes.h>
563 #include <machine/md_var.h>
565 static __inline u_int64_t
568 if (cpu_vendor_id == CPU_VENDOR_INTEL)
579 __asm __volatile("wbinvd");
583 void cpu_wbinvd_on_all_cpus_callback(void *arg);
586 cpu_wbinvd_on_all_cpus(void)
588 lwkt_cpusync_simple(smp_active_mask, cpu_wbinvd_on_all_cpus_callback, NULL);
593 write_rflags(u_long rf)
595 __asm __volatile("pushq %0; popfq" : : "r" (rf));
599 wrmsr(u_int msr, u_int64_t newval)
605 __asm __volatile("wrmsr" : : "a" (low), "d" (high), "c" (msr));
609 xsetbv(u_int ecx, u_int eax, u_int edx)
611 __asm __volatile(".byte 0x0f,0x01,0xd1"
613 : "a" (eax), "c" (ecx), "d" (edx));
617 load_cr0(u_long data)
620 __asm __volatile("movq %0,%%cr0" : : "r" (data));
623 static __inline u_long
628 __asm __volatile("movq %%cr0,%0" : "=r" (data));
632 static __inline u_long
637 __asm __volatile("movq %%cr2,%0" : "=r" (data));
642 load_cr3(u_long data)
645 __asm __volatile("movq %0,%%cr3" : : "r" (data) : "memory");
648 static __inline u_long
653 __asm __volatile("movq %%cr3,%0" : "=r" (data));
658 load_cr4(u_long data)
660 __asm __volatile("movq %0,%%cr4" : : "r" (data));
663 static __inline u_long
668 __asm __volatile("movq %%cr4,%0" : "=r" (data));
672 #ifndef _CPU_INVLTLB_DEFINED
675 * Invalidate the TLB on this cpu only
681 #if defined(SWTCH_OPTIM_STATS)
688 extern void smp_invltlb(void);
689 extern void smp_sniff(void);
691 static __inline u_short
695 __asm __volatile("movw %%fs,%0" : "=rm" (sel));
699 static __inline u_short
703 __asm __volatile("movw %%gs,%0" : "=rm" (sel));
710 __asm __volatile("movw %0,%%ds" : : "rm" (sel));
716 __asm __volatile("movw %0,%%es" : : "rm" (sel));
720 /* This is defined in <machine/specialreg.h> but is too painful to get to */
722 #define MSR_FSBASE 0xc0000100
727 /* Preserve the fsbase value across the selector load */
728 __asm __volatile("rdmsr; movw %0,%%fs; wrmsr"
729 : : "rm" (sel), "c" (MSR_FSBASE) : "eax", "edx");
733 #define MSR_GSBASE 0xc0000101
739 * Preserve the gsbase value across the selector load.
740 * Note that we have to disable interrupts because the gsbase
741 * being trashed happens to be the kernel gsbase at the time.
743 __asm __volatile("pushfq; cli; rdmsr; movw %0,%%gs; wrmsr; popfq"
744 : : "rm" (sel), "c" (MSR_GSBASE) : "eax", "edx");
747 /* Usable by userland */
751 __asm __volatile("movw %0,%%fs" : : "rm" (sel));
757 __asm __volatile("movw %0,%%gs" : : "rm" (sel));
761 /* void lidt(struct region_descriptor *addr); */
763 lidt(struct region_descriptor *addr)
765 __asm __volatile("lidt (%0)" : : "r" (addr));
768 /* void lldt(u_short sel); */
772 __asm __volatile("lldt %0" : : "r" (sel));
775 /* void ltr(u_short sel); */
779 __asm __volatile("ltr %0" : : "r" (sel));
782 static __inline u_int64_t
786 __asm __volatile("movq %%dr0,%0" : "=r" (data));
791 load_dr0(u_int64_t dr0)
793 __asm __volatile("movq %0,%%dr0" : : "r" (dr0));
796 static __inline u_int64_t
800 __asm __volatile("movq %%dr1,%0" : "=r" (data));
805 load_dr1(u_int64_t dr1)
807 __asm __volatile("movq %0,%%dr1" : : "r" (dr1));
810 static __inline u_int64_t
814 __asm __volatile("movq %%dr2,%0" : "=r" (data));
819 load_dr2(u_int64_t dr2)
821 __asm __volatile("movq %0,%%dr2" : : "r" (dr2));
824 static __inline u_int64_t
828 __asm __volatile("movq %%dr3,%0" : "=r" (data));
833 load_dr3(u_int64_t dr3)
835 __asm __volatile("movq %0,%%dr3" : : "r" (dr3));
838 static __inline u_int64_t
842 __asm __volatile("movq %%dr4,%0" : "=r" (data));
847 load_dr4(u_int64_t dr4)
849 __asm __volatile("movq %0,%%dr4" : : "r" (dr4));
852 static __inline u_int64_t
856 __asm __volatile("movq %%dr5,%0" : "=r" (data));
861 load_dr5(u_int64_t dr5)
863 __asm __volatile("movq %0,%%dr5" : : "r" (dr5));
866 static __inline u_int64_t
870 __asm __volatile("movq %%dr6,%0" : "=r" (data));
875 load_dr6(u_int64_t dr6)
877 __asm __volatile("movq %0,%%dr6" : : "r" (dr6));
880 static __inline u_int64_t
884 __asm __volatile("movq %%dr7,%0" : "=r" (data));
889 load_dr7(u_int64_t dr7)
891 __asm __volatile("movq %0,%%dr7" : : "r" (dr7));
894 static __inline register_t
899 rflags = read_rflags();
905 intr_restore(register_t rflags)
907 write_rflags(rflags);
910 #else /* !__GNUC__ */
912 int breakpoint(void);
913 void cpu_pause(void);
914 u_int bsfl(u_int mask);
915 u_int bsrl(u_int mask);
916 void cpu_disable_intr(void);
917 void cpu_enable_intr(void);
918 void cpu_invlpg(u_long addr);
919 void cpu_invlpg_range(u_long start, u_long end);
920 void do_cpuid(u_int ax, u_int *p);
922 u_char inb(u_int port);
923 u_int inl(u_int port);
924 void insb(u_int port, void *addr, size_t cnt);
925 void insl(u_int port, void *addr, size_t cnt);
926 void insw(u_int port, void *addr, size_t cnt);
928 void invlpg_range(u_int start, u_int end);
929 void cpu_invltlb(void);
930 u_short inw(u_int port);
931 void load_cr0(u_int cr0);
932 void load_cr3(u_int cr3);
933 void load_cr4(u_int cr4);
934 void load_fs(u_int sel);
935 void load_gs(u_int sel);
936 struct region_descriptor;
937 void lidt(struct region_descriptor *addr);
938 void lldt(u_short sel);
939 void ltr(u_short sel);
940 void outb(u_int port, u_char data);
941 void outl(u_int port, u_int data);
942 void outsb(u_int port, void *addr, size_t cnt);
943 void outsl(u_int port, void *addr, size_t cnt);
944 void outsw(u_int port, void *addr, size_t cnt);
945 void outw(u_int port, u_short data);
946 void ia32_pause(void);
953 u_int64_t rdmsr(u_int msr);
954 u_int64_t rdpmc(u_int pmc);
955 u_int64_t rdtsc(void);
956 u_int read_rflags(void);
958 void write_rflags(u_int rf);
959 void wrmsr(u_int msr, u_int64_t newval);
960 u_int64_t rdr0(void);
961 void load_dr0(u_int64_t dr0);
962 u_int64_t rdr1(void);
963 void load_dr1(u_int64_t dr1);
964 u_int64_t rdr2(void);
965 void load_dr2(u_int64_t dr2);
966 u_int64_t rdr3(void);
967 void load_dr3(u_int64_t dr3);
968 u_int64_t rdr4(void);
969 void load_dr4(u_int64_t dr4);
970 u_int64_t rdr5(void);
971 void load_dr5(u_int64_t dr5);
972 u_int64_t rdr6(void);
973 void load_dr6(u_int64_t dr6);
974 u_int64_t rdr7(void);
975 void load_dr7(u_int64_t dr7);
976 register_t intr_disable(void);
977 void intr_restore(register_t rf);
979 #endif /* __GNUC__ */
981 int rdmsr_safe(u_int msr, uint64_t *val);
982 int wrmsr_safe(u_int msr, uint64_t newval);
983 void reset_dbregs(void);
987 #endif /* !_CPU_CPUFUNC_H_ */