xref: /openbmc/linux/arch/x86/include/asm/mmu_context.h (revision 4f6cce39)
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 #include <linux/pkeys.h>
8 
9 #include <trace/events/tlb.h>
10 
11 #include <asm/pgalloc.h>
12 #include <asm/tlbflush.h>
13 #include <asm/paravirt.h>
14 #include <asm/mpx.h>
15 #ifndef CONFIG_PARAVIRT
16 static inline void paravirt_activate_mm(struct mm_struct *prev,
17 					struct mm_struct *next)
18 {
19 }
20 #endif	/* !CONFIG_PARAVIRT */
21 
22 #ifdef CONFIG_PERF_EVENTS
23 extern struct static_key rdpmc_always_available;
24 
25 static inline void load_mm_cr4(struct mm_struct *mm)
26 {
27 	if (static_key_false(&rdpmc_always_available) ||
28 	    atomic_read(&mm->context.perf_rdpmc_allowed))
29 		cr4_set_bits(X86_CR4_PCE);
30 	else
31 		cr4_clear_bits(X86_CR4_PCE);
32 }
33 #else
34 static inline void load_mm_cr4(struct mm_struct *mm) {}
35 #endif
36 
37 #ifdef CONFIG_MODIFY_LDT_SYSCALL
38 /*
39  * ldt_structs can be allocated, used, and freed, but they are never
40  * modified while live.
41  */
42 struct ldt_struct {
43 	/*
44 	 * Xen requires page-aligned LDTs with special permissions.  This is
45 	 * needed to prevent us from installing evil descriptors such as
46 	 * call gates.  On native, we could merge the ldt_struct and LDT
47 	 * allocations, but it's not worth trying to optimize.
48 	 */
49 	struct desc_struct *entries;
50 	unsigned int size;
51 };
52 
53 /*
54  * Used for LDT copy/destruction.
55  */
56 int init_new_context_ldt(struct task_struct *tsk, struct mm_struct *mm);
57 void destroy_context_ldt(struct mm_struct *mm);
58 #else	/* CONFIG_MODIFY_LDT_SYSCALL */
59 static inline int init_new_context_ldt(struct task_struct *tsk,
60 				       struct mm_struct *mm)
61 {
62 	return 0;
63 }
64 static inline void destroy_context_ldt(struct mm_struct *mm) {}
65 #endif
66 
67 static inline void load_mm_ldt(struct mm_struct *mm)
68 {
69 #ifdef CONFIG_MODIFY_LDT_SYSCALL
70 	struct ldt_struct *ldt;
71 
72 	/* lockless_dereference synchronizes with smp_store_release */
73 	ldt = lockless_dereference(mm->context.ldt);
74 
75 	/*
76 	 * Any change to mm->context.ldt is followed by an IPI to all
77 	 * CPUs with the mm active.  The LDT will not be freed until
78 	 * after the IPI is handled by all such CPUs.  This means that,
79 	 * if the ldt_struct changes before we return, the values we see
80 	 * will be safe, and the new values will be loaded before we run
81 	 * any user code.
82 	 *
83 	 * NB: don't try to convert this to use RCU without extreme care.
84 	 * We would still need IRQs off, because we don't want to change
85 	 * the local LDT after an IPI loaded a newer value than the one
86 	 * that we can see.
87 	 */
88 
89 	if (unlikely(ldt))
90 		set_ldt(ldt->entries, ldt->size);
91 	else
92 		clear_LDT();
93 #else
94 	clear_LDT();
95 #endif
96 
97 	DEBUG_LOCKS_WARN_ON(preemptible());
98 }
99 
100 static inline void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
101 {
102 #ifdef CONFIG_SMP
103 	if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
104 		this_cpu_write(cpu_tlbstate.state, TLBSTATE_LAZY);
105 #endif
106 }
107 
108 static inline int init_new_context(struct task_struct *tsk,
109 				   struct mm_struct *mm)
110 {
111 	#ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
112 	if (cpu_feature_enabled(X86_FEATURE_OSPKE)) {
113 		/* pkey 0 is the default and always allocated */
114 		mm->context.pkey_allocation_map = 0x1;
115 		/* -1 means unallocated or invalid */
116 		mm->context.execute_only_pkey = -1;
117 	}
118 	#endif
119 	init_new_context_ldt(tsk, mm);
120 
121 	return 0;
122 }
123 static inline void destroy_context(struct mm_struct *mm)
124 {
125 	destroy_context_ldt(mm);
126 }
127 
128 extern void switch_mm(struct mm_struct *prev, struct mm_struct *next,
129 		      struct task_struct *tsk);
130 
131 extern void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
132 			       struct task_struct *tsk);
133 #define switch_mm_irqs_off switch_mm_irqs_off
134 
135 #define activate_mm(prev, next)			\
136 do {						\
137 	paravirt_activate_mm((prev), (next));	\
138 	switch_mm((prev), (next), NULL);	\
139 } while (0);
140 
141 #ifdef CONFIG_X86_32
142 #define deactivate_mm(tsk, mm)			\
143 do {						\
144 	lazy_load_gs(0);			\
145 } while (0)
146 #else
147 #define deactivate_mm(tsk, mm)			\
148 do {						\
149 	load_gs_index(0);			\
150 	loadsegment(fs, 0);			\
151 } while (0)
152 #endif
153 
154 static inline void arch_dup_mmap(struct mm_struct *oldmm,
155 				 struct mm_struct *mm)
156 {
157 	paravirt_arch_dup_mmap(oldmm, mm);
158 }
159 
160 static inline void arch_exit_mmap(struct mm_struct *mm)
161 {
162 	paravirt_arch_exit_mmap(mm);
163 }
164 
165 #ifdef CONFIG_X86_64
166 static inline bool is_64bit_mm(struct mm_struct *mm)
167 {
168 	return	!IS_ENABLED(CONFIG_IA32_EMULATION) ||
169 		!(mm->context.ia32_compat == TIF_IA32);
170 }
171 #else
172 static inline bool is_64bit_mm(struct mm_struct *mm)
173 {
174 	return false;
175 }
176 #endif
177 
178 static inline void arch_bprm_mm_init(struct mm_struct *mm,
179 		struct vm_area_struct *vma)
180 {
181 	mpx_mm_init(mm);
182 }
183 
184 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma,
185 			      unsigned long start, unsigned long end)
186 {
187 	/*
188 	 * mpx_notify_unmap() goes and reads a rarely-hot
189 	 * cacheline in the mm_struct.  That can be expensive
190 	 * enough to be seen in profiles.
191 	 *
192 	 * The mpx_notify_unmap() call and its contents have been
193 	 * observed to affect munmap() performance on hardware
194 	 * where MPX is not present.
195 	 *
196 	 * The unlikely() optimizes for the fast case: no MPX
197 	 * in the CPU, or no MPX use in the process.  Even if
198 	 * we get this wrong (in the unlikely event that MPX
199 	 * is widely enabled on some system) the overhead of
200 	 * MPX itself (reading bounds tables) is expected to
201 	 * overwhelm the overhead of getting this unlikely()
202 	 * consistently wrong.
203 	 */
204 	if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX)))
205 		mpx_notify_unmap(mm, vma, start, end);
206 }
207 
208 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
209 static inline int vma_pkey(struct vm_area_struct *vma)
210 {
211 	unsigned long vma_pkey_mask = VM_PKEY_BIT0 | VM_PKEY_BIT1 |
212 				      VM_PKEY_BIT2 | VM_PKEY_BIT3;
213 
214 	return (vma->vm_flags & vma_pkey_mask) >> VM_PKEY_SHIFT;
215 }
216 #else
217 static inline int vma_pkey(struct vm_area_struct *vma)
218 {
219 	return 0;
220 }
221 #endif
222 
223 static inline bool __pkru_allows_pkey(u16 pkey, bool write)
224 {
225 	u32 pkru = read_pkru();
226 
227 	if (!__pkru_allows_read(pkru, pkey))
228 		return false;
229 	if (write && !__pkru_allows_write(pkru, pkey))
230 		return false;
231 
232 	return true;
233 }
234 
235 /*
236  * We only want to enforce protection keys on the current process
237  * because we effectively have no access to PKRU for other
238  * processes or any way to tell *which * PKRU in a threaded
239  * process we could use.
240  *
241  * So do not enforce things if the VMA is not from the current
242  * mm, or if we are in a kernel thread.
243  */
244 static inline bool vma_is_foreign(struct vm_area_struct *vma)
245 {
246 	if (!current->mm)
247 		return true;
248 	/*
249 	 * Should PKRU be enforced on the access to this VMA?  If
250 	 * the VMA is from another process, then PKRU has no
251 	 * relevance and should not be enforced.
252 	 */
253 	if (current->mm != vma->vm_mm)
254 		return true;
255 
256 	return false;
257 }
258 
259 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma,
260 		bool write, bool execute, bool foreign)
261 {
262 	/* pkeys never affect instruction fetches */
263 	if (execute)
264 		return true;
265 	/* allow access if the VMA is not one from this process */
266 	if (foreign || vma_is_foreign(vma))
267 		return true;
268 	return __pkru_allows_pkey(vma_pkey(vma), write);
269 }
270 
271 static inline bool arch_pte_access_permitted(pte_t pte, bool write)
272 {
273 	return __pkru_allows_pkey(pte_flags_pkey(pte_flags(pte)), write);
274 }
275 #endif /* _ASM_X86_MMU_CONTEXT_H */
276