883f6b933fa4b6501af7a050fc161eafdb8f8d91
[kvmfornfv.git] / kernel / arch / x86 / include / asm / mmu_context.h
1 #ifndef _ASM_X86_MMU_CONTEXT_H
2 #define _ASM_X86_MMU_CONTEXT_H
3
4 #include <asm/desc.h>
5 #include <linux/atomic.h>
6 #include <linux/mm_types.h>
7
8 #include <trace/events/tlb.h>
9
10 #include <asm/pgalloc.h>
11 #include <asm/tlbflush.h>
12 #include <asm/paravirt.h>
13 #include <asm/mpx.h>
14 #ifndef CONFIG_PARAVIRT
15 static inline void paravirt_activate_mm(struct mm_struct *prev,
16                                         struct mm_struct *next)
17 {
18 }
19 #endif  /* !CONFIG_PARAVIRT */
20
21 #ifdef CONFIG_PERF_EVENTS
22 extern struct static_key rdpmc_always_available;
23
24 static inline void load_mm_cr4(struct mm_struct *mm)
25 {
26         if (static_key_true(&rdpmc_always_available) ||
27             atomic_read(&mm->context.perf_rdpmc_allowed))
28                 cr4_set_bits(X86_CR4_PCE);
29         else
30                 cr4_clear_bits(X86_CR4_PCE);
31 }
32 #else
33 static inline void load_mm_cr4(struct mm_struct *mm) {}
34 #endif
35
36 /*
37  * Used for LDT copy/destruction.
38  */
39 int init_new_context(struct task_struct *tsk, struct mm_struct *mm);
40 void destroy_context(struct mm_struct *mm);
41
42
43 static inline void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
44 {
45 #ifdef CONFIG_SMP
46         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
47                 this_cpu_write(cpu_tlbstate.state, TLBSTATE_LAZY);
48 #endif
49 }
50
51 static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
52                              struct task_struct *tsk)
53 {
54         unsigned cpu = smp_processor_id();
55
56         if (likely(prev != next)) {
57 #ifdef CONFIG_SMP
58                 this_cpu_write(cpu_tlbstate.state, TLBSTATE_OK);
59                 this_cpu_write(cpu_tlbstate.active_mm, next);
60 #endif
61                 cpumask_set_cpu(cpu, mm_cpumask(next));
62
63                 /* Re-load page tables */
64                 load_cr3(next->pgd);
65                 trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
66
67                 /* Stop flush ipis for the previous mm */
68                 cpumask_clear_cpu(cpu, mm_cpumask(prev));
69
70                 /* Load per-mm CR4 state */
71                 load_mm_cr4(next);
72
73                 /*
74                  * Load the LDT, if the LDT is different.
75                  *
76                  * It's possible that prev->context.ldt doesn't match
77                  * the LDT register.  This can happen if leave_mm(prev)
78                  * was called and then modify_ldt changed
79                  * prev->context.ldt but suppressed an IPI to this CPU.
80                  * In this case, prev->context.ldt != NULL, because we
81                  * never free an LDT while the mm still exists.  That
82                  * means that next->context.ldt != prev->context.ldt,
83                  * because mms never share an LDT.
84                  */
85                 if (unlikely(prev->context.ldt != next->context.ldt))
86                         load_LDT_nolock(&next->context);
87         }
88 #ifdef CONFIG_SMP
89           else {
90                 this_cpu_write(cpu_tlbstate.state, TLBSTATE_OK);
91                 BUG_ON(this_cpu_read(cpu_tlbstate.active_mm) != next);
92
93                 if (!cpumask_test_cpu(cpu, mm_cpumask(next))) {
94                         /*
95                          * On established mms, the mm_cpumask is only changed
96                          * from irq context, from ptep_clear_flush() while in
97                          * lazy tlb mode, and here. Irqs are blocked during
98                          * schedule, protecting us from simultaneous changes.
99                          */
100                         cpumask_set_cpu(cpu, mm_cpumask(next));
101                         /*
102                          * We were in lazy tlb mode and leave_mm disabled
103                          * tlb flush IPI delivery. We must reload CR3
104                          * to make sure to use no freed page tables.
105                          */
106                         load_cr3(next->pgd);
107                         trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
108                         load_mm_cr4(next);
109                         load_LDT_nolock(&next->context);
110                 }
111         }
112 #endif
113 }
114
115 #define activate_mm(prev, next)                 \
116 do {                                            \
117         paravirt_activate_mm((prev), (next));   \
118         switch_mm((prev), (next), NULL);        \
119 } while (0);
120
121 #ifdef CONFIG_X86_32
122 #define deactivate_mm(tsk, mm)                  \
123 do {                                            \
124         lazy_load_gs(0);                        \
125 } while (0)
126 #else
127 #define deactivate_mm(tsk, mm)                  \
128 do {                                            \
129         load_gs_index(0);                       \
130         loadsegment(fs, 0);                     \
131 } while (0)
132 #endif
133
134 static inline void arch_dup_mmap(struct mm_struct *oldmm,
135                                  struct mm_struct *mm)
136 {
137         paravirt_arch_dup_mmap(oldmm, mm);
138 }
139
140 static inline void arch_exit_mmap(struct mm_struct *mm)
141 {
142         paravirt_arch_exit_mmap(mm);
143 }
144
145 static inline void arch_bprm_mm_init(struct mm_struct *mm,
146                 struct vm_area_struct *vma)
147 {
148         mpx_mm_init(mm);
149 }
150
151 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma,
152                               unsigned long start, unsigned long end)
153 {
154         /*
155          * mpx_notify_unmap() goes and reads a rarely-hot
156          * cacheline in the mm_struct.  That can be expensive
157          * enough to be seen in profiles.
158          *
159          * The mpx_notify_unmap() call and its contents have been
160          * observed to affect munmap() performance on hardware
161          * where MPX is not present.
162          *
163          * The unlikely() optimizes for the fast case: no MPX
164          * in the CPU, or no MPX use in the process.  Even if
165          * we get this wrong (in the unlikely event that MPX
166          * is widely enabled on some system) the overhead of
167          * MPX itself (reading bounds tables) is expected to
168          * overwhelm the overhead of getting this unlikely()
169          * consistently wrong.
170          */
171         if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX)))
172                 mpx_notify_unmap(mm, vma, start, end);
173 }
174
175 #endif /* _ASM_X86_MMU_CONTEXT_H */