x86/mm, sched/core: Turn off IRQs in switch_mm()
[deliverable/linux.git] / 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_false(&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 #ifdef CONFIG_MODIFY_LDT_SYSCALL
37 /*
38 * ldt_structs can be allocated, used, and freed, but they are never
39 * modified while live.
40 */
41 struct ldt_struct {
42 /*
43 * Xen requires page-aligned LDTs with special permissions. This is
44 * needed to prevent us from installing evil descriptors such as
45 * call gates. On native, we could merge the ldt_struct and LDT
46 * allocations, but it's not worth trying to optimize.
47 */
48 struct desc_struct *entries;
49 int size;
50 };
51
52 /*
53 * Used for LDT copy/destruction.
54 */
55 int init_new_context_ldt(struct task_struct *tsk, struct mm_struct *mm);
56 void destroy_context_ldt(struct mm_struct *mm);
57 #else /* CONFIG_MODIFY_LDT_SYSCALL */
58 static inline int init_new_context_ldt(struct task_struct *tsk,
59 struct mm_struct *mm)
60 {
61 return 0;
62 }
63 static inline void destroy_context_ldt(struct mm_struct *mm) {}
64 #endif
65
66 static inline void load_mm_ldt(struct mm_struct *mm)
67 {
68 #ifdef CONFIG_MODIFY_LDT_SYSCALL
69 struct ldt_struct *ldt;
70
71 /* lockless_dereference synchronizes with smp_store_release */
72 ldt = lockless_dereference(mm->context.ldt);
73
74 /*
75 * Any change to mm->context.ldt is followed by an IPI to all
76 * CPUs with the mm active. The LDT will not be freed until
77 * after the IPI is handled by all such CPUs. This means that,
78 * if the ldt_struct changes before we return, the values we see
79 * will be safe, and the new values will be loaded before we run
80 * any user code.
81 *
82 * NB: don't try to convert this to use RCU without extreme care.
83 * We would still need IRQs off, because we don't want to change
84 * the local LDT after an IPI loaded a newer value than the one
85 * that we can see.
86 */
87
88 if (unlikely(ldt))
89 set_ldt(ldt->entries, ldt->size);
90 else
91 clear_LDT();
92 #else
93 clear_LDT();
94 #endif
95
96 DEBUG_LOCKS_WARN_ON(preemptible());
97 }
98
99 static inline void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
100 {
101 #ifdef CONFIG_SMP
102 if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
103 this_cpu_write(cpu_tlbstate.state, TLBSTATE_LAZY);
104 #endif
105 }
106
107 static inline int init_new_context(struct task_struct *tsk,
108 struct mm_struct *mm)
109 {
110 init_new_context_ldt(tsk, mm);
111 return 0;
112 }
113 static inline void destroy_context(struct mm_struct *mm)
114 {
115 destroy_context_ldt(mm);
116 }
117
118 extern void switch_mm(struct mm_struct *prev, struct mm_struct *next,
119 struct task_struct *tsk);
120
121 extern void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
122 struct task_struct *tsk);
123 #define switch_mm_irqs_off switch_mm_irqs_off
124
125 #define activate_mm(prev, next) \
126 do { \
127 paravirt_activate_mm((prev), (next)); \
128 switch_mm((prev), (next), NULL); \
129 } while (0);
130
131 #ifdef CONFIG_X86_32
132 #define deactivate_mm(tsk, mm) \
133 do { \
134 lazy_load_gs(0); \
135 } while (0)
136 #else
137 #define deactivate_mm(tsk, mm) \
138 do { \
139 load_gs_index(0); \
140 loadsegment(fs, 0); \
141 } while (0)
142 #endif
143
144 static inline void arch_dup_mmap(struct mm_struct *oldmm,
145 struct mm_struct *mm)
146 {
147 paravirt_arch_dup_mmap(oldmm, mm);
148 }
149
150 static inline void arch_exit_mmap(struct mm_struct *mm)
151 {
152 paravirt_arch_exit_mmap(mm);
153 }
154
155 #ifdef CONFIG_X86_64
156 static inline bool is_64bit_mm(struct mm_struct *mm)
157 {
158 return !config_enabled(CONFIG_IA32_EMULATION) ||
159 !(mm->context.ia32_compat == TIF_IA32);
160 }
161 #else
162 static inline bool is_64bit_mm(struct mm_struct *mm)
163 {
164 return false;
165 }
166 #endif
167
168 static inline void arch_bprm_mm_init(struct mm_struct *mm,
169 struct vm_area_struct *vma)
170 {
171 mpx_mm_init(mm);
172 }
173
174 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma,
175 unsigned long start, unsigned long end)
176 {
177 /*
178 * mpx_notify_unmap() goes and reads a rarely-hot
179 * cacheline in the mm_struct. That can be expensive
180 * enough to be seen in profiles.
181 *
182 * The mpx_notify_unmap() call and its contents have been
183 * observed to affect munmap() performance on hardware
184 * where MPX is not present.
185 *
186 * The unlikely() optimizes for the fast case: no MPX
187 * in the CPU, or no MPX use in the process. Even if
188 * we get this wrong (in the unlikely event that MPX
189 * is widely enabled on some system) the overhead of
190 * MPX itself (reading bounds tables) is expected to
191 * overwhelm the overhead of getting this unlikely()
192 * consistently wrong.
193 */
194 if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX)))
195 mpx_notify_unmap(mm, vma, start, end);
196 }
197
198 static inline int vma_pkey(struct vm_area_struct *vma)
199 {
200 u16 pkey = 0;
201 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
202 unsigned long vma_pkey_mask = VM_PKEY_BIT0 | VM_PKEY_BIT1 |
203 VM_PKEY_BIT2 | VM_PKEY_BIT3;
204 pkey = (vma->vm_flags & vma_pkey_mask) >> VM_PKEY_SHIFT;
205 #endif
206 return pkey;
207 }
208
209 static inline bool __pkru_allows_pkey(u16 pkey, bool write)
210 {
211 u32 pkru = read_pkru();
212
213 if (!__pkru_allows_read(pkru, pkey))
214 return false;
215 if (write && !__pkru_allows_write(pkru, pkey))
216 return false;
217
218 return true;
219 }
220
221 /*
222 * We only want to enforce protection keys on the current process
223 * because we effectively have no access to PKRU for other
224 * processes or any way to tell *which * PKRU in a threaded
225 * process we could use.
226 *
227 * So do not enforce things if the VMA is not from the current
228 * mm, or if we are in a kernel thread.
229 */
230 static inline bool vma_is_foreign(struct vm_area_struct *vma)
231 {
232 if (!current->mm)
233 return true;
234 /*
235 * Should PKRU be enforced on the access to this VMA? If
236 * the VMA is from another process, then PKRU has no
237 * relevance and should not be enforced.
238 */
239 if (current->mm != vma->vm_mm)
240 return true;
241
242 return false;
243 }
244
245 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma,
246 bool write, bool execute, bool foreign)
247 {
248 /* pkeys never affect instruction fetches */
249 if (execute)
250 return true;
251 /* allow access if the VMA is not one from this process */
252 if (foreign || vma_is_foreign(vma))
253 return true;
254 return __pkru_allows_pkey(vma_pkey(vma), write);
255 }
256
257 static inline bool arch_pte_access_permitted(pte_t pte, bool write)
258 {
259 return __pkru_allows_pkey(pte_flags_pkey(pte_flags(pte)), write);
260 }
261
262 #endif /* _ASM_X86_MMU_CONTEXT_H */
This page took 0.039122 seconds and 5 git commands to generate.