1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_X86_PROCESSOR_H
3 #define _ASM_X86_PROCESSOR_H
5 #include <asm/processor-flags.h>
7 /* Forward declaration, a strange C thing */
13 #include <asm/math_emu.h>
14 #include <asm/segment.h>
15 #include <asm/types.h>
16 #include <uapi/asm/sigcontext.h>
17 #include <asm/current.h>
18 #include <asm/cpufeatures.h>
19 #include <asm/cpuid.h>
21 #include <asm/pgtable_types.h>
22 #include <asm/percpu.h>
23 #include <asm/desc_defs.h>
25 #include <asm/special_insns.h>
26 #include <asm/fpu/types.h>
27 #include <asm/unwind_hints.h>
28 #include <asm/vmxfeatures.h>
29 #include <asm/vdso/processor.h>
30 #include <asm/shstk.h>
32 #include <linux/personality.h>
33 #include <linux/cache.h>
34 #include <linux/threads.h>
35 #include <linux/math64.h>
36 #include <linux/err.h>
37 #include <linux/irqflags.h>
38 #include <linux/mem_encrypt.h>
41 * We handle most unaligned accesses in hardware. On the other hand
42 * unaligned DMA can be quite expensive on some Nehalem processors.
44 * Based on this we disable the IP header alignment in network drivers.
46 #define NET_IP_ALIGN 0
51 * These alignment constraints are for performance in the vSMP case,
52 * but in the task_struct case we must also meet hardware imposed
53 * alignment requirements of the FPU state:
55 #ifdef CONFIG_X86_VSMP
56 # define ARCH_MIN_TASKALIGN (1 << INTERNODE_CACHE_SHIFT)
57 # define ARCH_MIN_MMSTRUCT_ALIGN (1 << INTERNODE_CACHE_SHIFT)
59 # define ARCH_MIN_TASKALIGN __alignof__(union fpregs_state)
60 # define ARCH_MIN_MMSTRUCT_ALIGN 0
68 extern u16 __read_mostly tlb_lli_4k[NR_INFO];
69 extern u16 __read_mostly tlb_lli_2m[NR_INFO];
70 extern u16 __read_mostly tlb_lli_4m[NR_INFO];
71 extern u16 __read_mostly tlb_lld_4k[NR_INFO];
72 extern u16 __read_mostly tlb_lld_2m[NR_INFO];
73 extern u16 __read_mostly tlb_lld_4m[NR_INFO];
74 extern u16 __read_mostly tlb_lld_1g[NR_INFO];
77 * CPU type and hardware bug flags. Kept separately for each CPU.
80 struct cpuinfo_topology {
81 // Real APIC ID read from the local APIC
83 // The initial APIC ID provided by CPUID
86 // Physical package ID
89 // Physical die ID on AMD, Relative on Intel
92 // Compute unit ID - AMD specific
95 // Core ID relative to the package
98 // Logical ID mappings
102 // AMD Node ID and Nodes per Package info
105 // Cache level topology IDs
113 * The particular ordering (low-to-high) of (vendor,
114 * family, model) is done in case range of models, like
115 * it is usually done on AMD, need to be compared.
125 /* combined vendor, family, model */
130 /* Number of 4K pages in DTLB/ITLB combined(in pages): */
133 #ifdef CONFIG_X86_VMX_FEATURE_NAMES
134 __u32 vmx_capability[NVMXINTS];
138 /* Max extended CPUID function supported: */
139 __u32 extended_cpuid_level;
140 /* Maximum supported CPUID level, -1=no CPUID: */
143 * Align to size of unsigned long because the x86_capability array
144 * is passed to bitops which require the alignment. Use unnamed
145 * union to enforce the array is aligned to size of unsigned long.
148 __u32 x86_capability[NCAPINTS + NBUGINTS];
149 unsigned long x86_capability_alignment;
151 char x86_vendor_id[16];
152 char x86_model_id[64];
153 struct cpuinfo_topology topo;
154 /* in KB - valid for CPUS which support this call: */
155 unsigned int x86_cache_size;
156 int x86_cache_alignment; /* In bytes */
157 /* Cache QoS architectural values, valid only on the BSP: */
158 int x86_cache_max_rmid; /* max index */
159 int x86_cache_occ_scale; /* scale to bytes */
160 int x86_cache_mbm_width_offset;
162 unsigned long loops_per_jiffy;
163 /* protected processor identification number */
165 u16 x86_clflush_size;
166 /* number of cores as seen by the OS: */
168 /* Index into per_cpu list: */
170 /* Is SMT active on this core? */
173 /* Address space bits used by the cache internally */
175 unsigned initialized : 1;
176 } __randomize_layout;
178 #define X86_VENDOR_INTEL 0
179 #define X86_VENDOR_CYRIX 1
180 #define X86_VENDOR_AMD 2
181 #define X86_VENDOR_UMC 3
182 #define X86_VENDOR_CENTAUR 5
183 #define X86_VENDOR_TRANSMETA 7
184 #define X86_VENDOR_NSC 8
185 #define X86_VENDOR_HYGON 9
186 #define X86_VENDOR_ZHAOXIN 10
187 #define X86_VENDOR_VORTEX 11
188 #define X86_VENDOR_NUM 12
190 #define X86_VENDOR_UNKNOWN 0xff
193 * capabilities of CPUs
195 extern struct cpuinfo_x86 boot_cpu_data;
196 extern struct cpuinfo_x86 new_cpu_data;
198 extern __u32 cpu_caps_cleared[NCAPINTS + NBUGINTS];
199 extern __u32 cpu_caps_set[NCAPINTS + NBUGINTS];
201 DECLARE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
202 #define cpu_data(cpu) per_cpu(cpu_info, cpu)
204 extern const struct seq_operations cpuinfo_op;
206 #define cache_line_size() (boot_cpu_data.x86_cache_alignment)
208 extern void cpu_detect(struct cpuinfo_x86 *c);
210 static inline unsigned long long l1tf_pfn_limit(void)
212 return BIT_ULL(boot_cpu_data.x86_cache_bits - 1 - PAGE_SHIFT);
215 extern void early_cpu_init(void);
216 extern void identify_secondary_cpu(struct cpuinfo_x86 *);
217 extern void print_cpu_info(struct cpuinfo_x86 *);
218 void print_cpu_msr(struct cpuinfo_x86 *);
221 * Friendlier CR3 helpers.
223 static inline unsigned long read_cr3_pa(void)
225 return __read_cr3() & CR3_ADDR_MASK;
228 static inline unsigned long native_read_cr3_pa(void)
230 return __native_read_cr3() & CR3_ADDR_MASK;
233 static inline void load_cr3(pgd_t *pgdir)
235 write_cr3(__sme_pa(pgdir));
239 * Note that while the legacy 'TSS' name comes from 'Task State Segment',
240 * on modern x86 CPUs the TSS also holds information important to 64-bit mode,
241 * unrelated to the task-switch mechanism:
244 /* This is the TSS defined by the hardware. */
246 unsigned short back_link, __blh;
248 unsigned short ss0, __ss0h;
252 * We don't use ring 1, so ss1 is a convenient scratch space in
253 * the same cacheline as sp0. We use ss1 to cache the value in
254 * MSR_IA32_SYSENTER_CS. When we context switch
255 * MSR_IA32_SYSENTER_CS, we first check if the new value being
256 * written matches ss1, and, if it's not, then we wrmsr the new
257 * value and update ss1.
259 * The only reason we context switch MSR_IA32_SYSENTER_CS is
260 * that we set it to zero in vm86 tasks to avoid corrupting the
261 * stack if we were to go through the sysenter path from vm86
264 unsigned short ss1; /* MSR_IA32_SYSENTER_CS */
266 unsigned short __ss1h;
268 unsigned short ss2, __ss2h;
280 unsigned short es, __esh;
281 unsigned short cs, __csh;
282 unsigned short ss, __ssh;
283 unsigned short ds, __dsh;
284 unsigned short fs, __fsh;
285 unsigned short gs, __gsh;
286 unsigned short ldt, __ldth;
287 unsigned short trace;
288 unsigned short io_bitmap_base;
290 } __attribute__((packed));
298 * Since Linux does not use ring 2, the 'sp2' slot is unused by
299 * hardware. entry_SYSCALL_64 uses it as scratch space to stash
300 * the user RSP value.
311 } __attribute__((packed));
317 #define IO_BITMAP_BITS 65536
318 #define IO_BITMAP_BYTES (IO_BITMAP_BITS / BITS_PER_BYTE)
319 #define IO_BITMAP_LONGS (IO_BITMAP_BYTES / sizeof(long))
321 #define IO_BITMAP_OFFSET_VALID_MAP \
322 (offsetof(struct tss_struct, io_bitmap.bitmap) - \
323 offsetof(struct tss_struct, x86_tss))
325 #define IO_BITMAP_OFFSET_VALID_ALL \
326 (offsetof(struct tss_struct, io_bitmap.mapall) - \
327 offsetof(struct tss_struct, x86_tss))
329 #ifdef CONFIG_X86_IOPL_IOPERM
331 * sizeof(unsigned long) coming from an extra "long" at the end of the
332 * iobitmap. The limit is inclusive, i.e. the last valid byte.
334 # define __KERNEL_TSS_LIMIT \
335 (IO_BITMAP_OFFSET_VALID_ALL + IO_BITMAP_BYTES + \
336 sizeof(unsigned long) - 1)
338 # define __KERNEL_TSS_LIMIT \
339 (offsetof(struct tss_struct, x86_tss) + sizeof(struct x86_hw_tss) - 1)
342 /* Base offset outside of TSS_LIMIT so unpriviledged IO causes #GP */
343 #define IO_BITMAP_OFFSET_INVALID (__KERNEL_TSS_LIMIT + 1)
346 char stack[PAGE_SIZE];
349 struct entry_stack_page {
350 struct entry_stack stack;
351 } __aligned(PAGE_SIZE);
354 * All IO bitmap related data stored in the TSS:
356 struct x86_io_bitmap {
357 /* The sequence number of the last active bitmap. */
361 * Store the dirty size of the last io bitmap offender. The next
362 * one will have to do the cleanup as the switch out to a non io
363 * bitmap user will just set x86_tss.io_bitmap_base to a value
364 * outside of the TSS limit. So for sane tasks there is no need to
365 * actually touch the io_bitmap at all.
367 unsigned int prev_max;
370 * The extra 1 is there because the CPU will access an
371 * additional byte beyond the end of the IO permission
372 * bitmap. The extra byte must be all 1 bits, and must
373 * be within the limit.
375 unsigned long bitmap[IO_BITMAP_LONGS + 1];
378 * Special I/O bitmap to emulate IOPL(3). All bytes zero,
379 * except the additional byte at the end.
381 unsigned long mapall[IO_BITMAP_LONGS + 1];
386 * The fixed hardware portion. This must not cross a page boundary
387 * at risk of violating the SDM's advice and potentially triggering
390 struct x86_hw_tss x86_tss;
392 struct x86_io_bitmap io_bitmap;
393 } __aligned(PAGE_SIZE);
395 DECLARE_PER_CPU_PAGE_ALIGNED(struct tss_struct, cpu_tss_rw);
397 /* Per CPU interrupt stacks */
399 char stack[IRQ_STACK_SIZE];
400 } __aligned(IRQ_STACK_SIZE);
403 struct fixed_percpu_data {
405 * GCC hardcodes the stack canary as %gs:40. Since the
406 * irq_stack is the object at %gs:0, we reserve the bottom
407 * 48 bytes of the irq stack for the canary.
409 * Once we are willing to require -mstack-protector-guard-symbol=
410 * support for x86_64 stackprotector, we can get rid of this.
413 unsigned long stack_canary;
416 DECLARE_PER_CPU_FIRST(struct fixed_percpu_data, fixed_percpu_data) __visible;
417 DECLARE_INIT_PER_CPU(fixed_percpu_data);
419 static inline unsigned long cpu_kernelmode_gs_base(int cpu)
421 return (unsigned long)per_cpu(fixed_percpu_data.gs_base, cpu);
424 extern asmlinkage void entry_SYSCALL32_ignore(void);
426 /* Save actual FS/GS selectors and bases to current->thread */
427 void current_save_fsgs(void);
429 #ifdef CONFIG_STACKPROTECTOR
430 DECLARE_PER_CPU(unsigned long, __stack_chk_guard);
436 struct thread_struct {
437 /* Cached TLS descriptors: */
438 struct desc_struct tls_array[GDT_ENTRY_TLS_ENTRIES];
444 unsigned long sysenter_cs;
448 unsigned short fsindex;
449 unsigned short gsindex;
453 unsigned long fsbase;
454 unsigned long gsbase;
457 * XXX: this could presumably be unsigned short. Alternatively,
458 * 32-bit kernels could be taught to use fsindex instead.
464 /* Save middle states of ptrace breakpoints */
465 struct perf_event *ptrace_bps[HBP_NUM];
466 /* Debug status used for traps, single steps, etc... */
467 unsigned long virtual_dr6;
468 /* Keep track of the exact dr7 value set by the user */
469 unsigned long ptrace_dr7;
472 unsigned long trap_nr;
473 unsigned long error_code;
475 /* Virtual 86 mode info */
478 /* IO permissions: */
479 struct io_bitmap *io_bitmap;
482 * IOPL. Privilege level dependent I/O permission which is
483 * emulated via the I/O bitmap to prevent user space from disabling
486 unsigned long iopl_emul;
488 unsigned int iopl_warn:1;
491 * Protection Keys Register for Userspace. Loaded immediately on
492 * context switch. Store it in thread_struct to avoid a lookup in
493 * the tasks's FPU xstate buffer. This value is only valid when a
494 * task is scheduled out. For 'current' the authoritative source of
495 * PKRU is the hardware itself.
499 #ifdef CONFIG_X86_USER_SHADOW_STACK
500 unsigned long features;
501 unsigned long features_locked;
503 struct thread_shstk shstk;
506 /* Floating point and extended processor state */
509 * WARNING: 'fpu' is dynamically-sized. It *MUST* be at
514 extern void fpu_thread_struct_whitelist(unsigned long *offset, unsigned long *size);
516 static inline void arch_thread_struct_whitelist(unsigned long *offset,
519 fpu_thread_struct_whitelist(offset, size);
523 native_load_sp0(unsigned long sp0)
525 this_cpu_write(cpu_tss_rw.x86_tss.sp0, sp0);
528 static __always_inline void native_swapgs(void)
531 asm volatile("swapgs" ::: "memory");
535 static __always_inline unsigned long current_top_of_stack(void)
538 * We can't read directly from tss.sp0: sp0 on x86_32 is special in
539 * and around vm86 mode and sp0 on x86_64 is special because of the
542 if (IS_ENABLED(CONFIG_USE_X86_SEG_SUPPORT))
543 return this_cpu_read_const(const_pcpu_hot.top_of_stack);
545 return this_cpu_read_stable(pcpu_hot.top_of_stack);
548 static __always_inline bool on_thread_stack(void)
550 return (unsigned long)(current_top_of_stack() -
551 current_stack_pointer) < THREAD_SIZE;
554 #ifdef CONFIG_PARAVIRT_XXL
555 #include <asm/paravirt.h>
558 static inline void load_sp0(unsigned long sp0)
560 native_load_sp0(sp0);
563 #endif /* CONFIG_PARAVIRT_XXL */
565 unsigned long __get_wchan(struct task_struct *p);
567 extern void select_idle_routine(void);
568 extern void amd_e400_c1e_apic_setup(void);
570 extern unsigned long boot_option_idle_override;
572 enum idle_boot_override {IDLE_NO_OVERRIDE=0, IDLE_HALT, IDLE_NOMWAIT,
575 extern void enable_sep_cpu(void);
578 /* Defined in head.S */
579 extern struct desc_ptr early_gdt_descr;
581 extern void switch_gdt_and_percpu_base(int);
582 extern void load_direct_gdt(int);
583 extern void load_fixmap_gdt(int);
584 extern void cpu_init(void);
585 extern void cpu_init_exception_handling(void);
586 extern void cr4_init(void);
588 extern void set_task_blockstep(struct task_struct *task, bool on);
590 /* Boot loader type from the setup header: */
591 extern int bootloader_type;
592 extern int bootloader_version;
594 extern char ignore_fpu_irq;
596 #define HAVE_ARCH_PICK_MMAP_LAYOUT 1
597 #define ARCH_HAS_PREFETCHW
600 # define BASE_PREFETCH ""
601 # define ARCH_HAS_PREFETCH
603 # define BASE_PREFETCH "prefetcht0 %1"
607 * Prefetch instructions for Pentium III (+) and AMD Athlon (+)
609 * It's not worth to care about 3dnow prefetches for the K6
610 * because they are microcoded there and very slow.
612 static inline void prefetch(const void *x)
614 alternative_input(BASE_PREFETCH, "prefetchnta %1",
616 "m" (*(const char *)x));
620 * 3dnow prefetch to get an exclusive cache line.
621 * Useful for spinlocks to avoid one state transition in the
622 * cache coherency protocol:
624 static __always_inline void prefetchw(const void *x)
626 alternative_input(BASE_PREFETCH, "prefetchw %1",
627 X86_FEATURE_3DNOWPREFETCH,
628 "m" (*(const char *)x));
631 #define TOP_OF_INIT_STACK ((unsigned long)&init_stack + sizeof(init_stack) - \
632 TOP_OF_KERNEL_STACK_PADDING)
634 #define task_top_of_stack(task) ((unsigned long)(task_pt_regs(task) + 1))
636 #define task_pt_regs(task) \
638 unsigned long __ptr = (unsigned long)task_stack_page(task); \
639 __ptr += THREAD_SIZE - TOP_OF_KERNEL_STACK_PADDING; \
640 ((struct pt_regs *)__ptr) - 1; \
644 #define INIT_THREAD { \
645 .sp0 = TOP_OF_INIT_STACK, \
646 .sysenter_cs = __KERNEL_CS, \
649 #define KSTK_ESP(task) (task_pt_regs(task)->sp)
652 extern unsigned long __top_init_kernel_stack[];
654 #define INIT_THREAD { \
655 .sp = (unsigned long)&__top_init_kernel_stack, \
658 extern unsigned long KSTK_ESP(struct task_struct *task);
660 #endif /* CONFIG_X86_64 */
662 extern void start_thread(struct pt_regs *regs, unsigned long new_ip,
663 unsigned long new_sp);
666 * This decides where the kernel will search for a free chunk of vm
667 * space during mmap's.
669 #define __TASK_UNMAPPED_BASE(task_size) (PAGE_ALIGN(task_size / 3))
670 #define TASK_UNMAPPED_BASE __TASK_UNMAPPED_BASE(TASK_SIZE_LOW)
672 #define KSTK_EIP(task) (task_pt_regs(task)->ip)
674 /* Get/set a process' ability to use the timestamp counter instruction */
675 #define GET_TSC_CTL(adr) get_tsc_mode((adr))
676 #define SET_TSC_CTL(val) set_tsc_mode((val))
678 extern int get_tsc_mode(unsigned long adr);
679 extern int set_tsc_mode(unsigned int val);
681 DECLARE_PER_CPU(u64, msr_misc_features_shadow);
683 static inline u32 per_cpu_llc_id(unsigned int cpu)
685 return per_cpu(cpu_info.topo.llc_id, cpu);
688 static inline u32 per_cpu_l2c_id(unsigned int cpu)
690 return per_cpu(cpu_info.topo.l2c_id, cpu);
693 #ifdef CONFIG_CPU_SUP_AMD
694 extern u32 amd_get_highest_perf(void);
695 extern void amd_clear_divider(void);
696 extern void amd_check_microcode(void);
698 static inline u32 amd_get_highest_perf(void) { return 0; }
699 static inline void amd_clear_divider(void) { }
700 static inline void amd_check_microcode(void) { }
703 extern unsigned long arch_align_stack(unsigned long sp);
704 void free_init_pages(const char *what, unsigned long begin, unsigned long end);
705 extern void free_kernel_image_pages(const char *what, void *begin, void *end);
707 void default_idle(void);
709 bool xen_set_default_idle(void);
711 #define xen_set_default_idle 0
714 void __noreturn stop_this_cpu(void *dummy);
715 void microcode_check(struct cpuinfo_x86 *prev_info);
716 void store_cpu_caps(struct cpuinfo_x86 *info);
718 enum l1tf_mitigations {
720 L1TF_MITIGATION_FLUSH_NOWARN,
721 L1TF_MITIGATION_FLUSH,
722 L1TF_MITIGATION_FLUSH_NOSMT,
723 L1TF_MITIGATION_FULL,
724 L1TF_MITIGATION_FULL_FORCE
727 extern enum l1tf_mitigations l1tf_mitigation;
729 enum mds_mitigations {
732 MDS_MITIGATION_VMWERV,
735 extern bool gds_ucode_mitigated(void);
738 * Make previous memory operations globally visible before
741 * MFENCE makes writes visible, but only affects load/store
742 * instructions. WRMSR is unfortunately not a load/store
743 * instruction and is unaffected by MFENCE. The LFENCE ensures
744 * that the WRMSR is not reordered.
746 * Most WRMSRs are full serializing instructions themselves and
747 * do not require this barrier. This is only required for the
748 * IA32_TSC_DEADLINE and X2APIC MSRs.
750 static inline void weak_wrmsr_fence(void)
752 alternative("mfence; lfence", "", ALT_NOT(X86_FEATURE_APIC_MSRS_FENCE));
755 #endif /* _ASM_X86_PROCESSOR_H */