xref: /linux/arch/x86/include/asm/mmu_context.h (revision 06a130e42a5bfc84795464bff023bff4c16f58c5)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_X86_MMU_CONTEXT_H
3 #define _ASM_X86_MMU_CONTEXT_H
4 
5 #include <asm/desc.h>
6 #include <linux/atomic.h>
7 #include <linux/mm_types.h>
8 #include <linux/pkeys.h>
9 
10 #include <trace/events/tlb.h>
11 
12 #include <asm/tlbflush.h>
13 #include <asm/paravirt.h>
14 #include <asm/debugreg.h>
15 #include <asm/gsseg.h>
16 
17 extern atomic64_t last_mm_ctx_id;
18 
19 #ifdef CONFIG_PERF_EVENTS
20 DECLARE_STATIC_KEY_FALSE(rdpmc_never_available_key);
21 DECLARE_STATIC_KEY_FALSE(rdpmc_always_available_key);
22 void cr4_update_pce(void *ignored);
23 #endif
24 
25 #ifdef CONFIG_MODIFY_LDT_SYSCALL
26 /*
27  * ldt_structs can be allocated, used, and freed, but they are never
28  * modified while live.
29  */
30 struct ldt_struct {
31 	/*
32 	 * Xen requires page-aligned LDTs with special permissions.  This is
33 	 * needed to prevent us from installing evil descriptors such as
34 	 * call gates.  On native, we could merge the ldt_struct and LDT
35 	 * allocations, but it's not worth trying to optimize.
36 	 */
37 	struct desc_struct	*entries;
38 	unsigned int		nr_entries;
39 
40 	/*
41 	 * If PTI is in use, then the entries array is not mapped while we're
42 	 * in user mode.  The whole array will be aliased at the addressed
43 	 * given by ldt_slot_va(slot).  We use two slots so that we can allocate
44 	 * and map, and enable a new LDT without invalidating the mapping
45 	 * of an older, still-in-use LDT.
46 	 *
47 	 * slot will be -1 if this LDT doesn't have an alias mapping.
48 	 */
49 	int			slot;
50 };
51 
52 /*
53  * Used for LDT copy/destruction.
54  */
55 static inline void init_new_context_ldt(struct mm_struct *mm)
56 {
57 	mm->context.ldt = NULL;
58 	init_rwsem(&mm->context.ldt_usr_sem);
59 }
60 int ldt_dup_context(struct mm_struct *oldmm, struct mm_struct *mm);
61 void destroy_context_ldt(struct mm_struct *mm);
62 void ldt_arch_exit_mmap(struct mm_struct *mm);
63 #else	/* CONFIG_MODIFY_LDT_SYSCALL */
64 static inline void init_new_context_ldt(struct mm_struct *mm) { }
65 static inline int ldt_dup_context(struct mm_struct *oldmm,
66 				  struct mm_struct *mm)
67 {
68 	return 0;
69 }
70 static inline void destroy_context_ldt(struct mm_struct *mm) { }
71 static inline void ldt_arch_exit_mmap(struct mm_struct *mm) { }
72 #endif
73 
74 #ifdef CONFIG_MODIFY_LDT_SYSCALL
75 extern void load_mm_ldt(struct mm_struct *mm);
76 extern void switch_ldt(struct mm_struct *prev, struct mm_struct *next);
77 #else
78 static inline void load_mm_ldt(struct mm_struct *mm)
79 {
80 	clear_LDT();
81 }
82 static inline void switch_ldt(struct mm_struct *prev, struct mm_struct *next)
83 {
84 	DEBUG_LOCKS_WARN_ON(preemptible());
85 }
86 #endif
87 
88 #ifdef CONFIG_ADDRESS_MASKING
89 static inline unsigned long mm_lam_cr3_mask(struct mm_struct *mm)
90 {
91 	/*
92 	 * When switch_mm_irqs_off() is called for a kthread, it may race with
93 	 * LAM enablement. switch_mm_irqs_off() uses the LAM mask to do two
94 	 * things: populate CR3 and populate 'cpu_tlbstate.lam'. Make sure it
95 	 * reads a single value for both.
96 	 */
97 	return READ_ONCE(mm->context.lam_cr3_mask);
98 }
99 
100 static inline void dup_lam(struct mm_struct *oldmm, struct mm_struct *mm)
101 {
102 	mm->context.lam_cr3_mask = oldmm->context.lam_cr3_mask;
103 	mm->context.untag_mask = oldmm->context.untag_mask;
104 }
105 
106 #define mm_untag_mask mm_untag_mask
107 static inline unsigned long mm_untag_mask(struct mm_struct *mm)
108 {
109 	return mm->context.untag_mask;
110 }
111 
112 static inline void mm_reset_untag_mask(struct mm_struct *mm)
113 {
114 	mm->context.untag_mask = -1UL;
115 }
116 
117 #define arch_pgtable_dma_compat arch_pgtable_dma_compat
118 static inline bool arch_pgtable_dma_compat(struct mm_struct *mm)
119 {
120 	return !mm_lam_cr3_mask(mm) ||
121 		test_bit(MM_CONTEXT_FORCE_TAGGED_SVA, &mm->context.flags);
122 }
123 #else
124 
125 static inline unsigned long mm_lam_cr3_mask(struct mm_struct *mm)
126 {
127 	return 0;
128 }
129 
130 static inline void dup_lam(struct mm_struct *oldmm, struct mm_struct *mm)
131 {
132 }
133 
134 static inline void mm_reset_untag_mask(struct mm_struct *mm)
135 {
136 }
137 #endif
138 
139 #define enter_lazy_tlb enter_lazy_tlb
140 extern void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk);
141 
142 /*
143  * Init a new mm.  Used on mm copies, like at fork()
144  * and on mm's that are brand-new, like at execve().
145  */
146 #define init_new_context init_new_context
147 static inline int init_new_context(struct task_struct *tsk,
148 				   struct mm_struct *mm)
149 {
150 	mutex_init(&mm->context.lock);
151 
152 	mm->context.ctx_id = atomic64_inc_return(&last_mm_ctx_id);
153 	atomic64_set(&mm->context.tlb_gen, 0);
154 
155 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
156 	if (cpu_feature_enabled(X86_FEATURE_OSPKE)) {
157 		/* pkey 0 is the default and allocated implicitly */
158 		mm->context.pkey_allocation_map = 0x1;
159 		/* -1 means unallocated or invalid */
160 		mm->context.execute_only_pkey = -1;
161 	}
162 #endif
163 	mm_reset_untag_mask(mm);
164 	init_new_context_ldt(mm);
165 	return 0;
166 }
167 
168 #define destroy_context destroy_context
169 static inline void destroy_context(struct mm_struct *mm)
170 {
171 	destroy_context_ldt(mm);
172 }
173 
174 extern void switch_mm(struct mm_struct *prev, struct mm_struct *next,
175 		      struct task_struct *tsk);
176 
177 extern void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
178 			       struct task_struct *tsk);
179 #define switch_mm_irqs_off switch_mm_irqs_off
180 
181 #define activate_mm(prev, next)			\
182 do {						\
183 	paravirt_enter_mmap(next);		\
184 	switch_mm((prev), (next), NULL);	\
185 } while (0);
186 
187 #ifdef CONFIG_X86_32
188 #define deactivate_mm(tsk, mm)			\
189 do {						\
190 	loadsegment(gs, 0);			\
191 } while (0)
192 #else
193 #define deactivate_mm(tsk, mm)			\
194 do {						\
195 	shstk_free(tsk);			\
196 	load_gs_index(0);			\
197 	loadsegment(fs, 0);			\
198 } while (0)
199 #endif
200 
201 static inline void arch_dup_pkeys(struct mm_struct *oldmm,
202 				  struct mm_struct *mm)
203 {
204 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
205 	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
206 		return;
207 
208 	/* Duplicate the oldmm pkey state in mm: */
209 	mm->context.pkey_allocation_map = oldmm->context.pkey_allocation_map;
210 	mm->context.execute_only_pkey   = oldmm->context.execute_only_pkey;
211 #endif
212 }
213 
214 static inline int arch_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
215 {
216 	arch_dup_pkeys(oldmm, mm);
217 	paravirt_enter_mmap(mm);
218 	dup_lam(oldmm, mm);
219 	return ldt_dup_context(oldmm, mm);
220 }
221 
222 static inline void arch_exit_mmap(struct mm_struct *mm)
223 {
224 	paravirt_arch_exit_mmap(mm);
225 	ldt_arch_exit_mmap(mm);
226 }
227 
228 #ifdef CONFIG_X86_64
229 static inline bool is_64bit_mm(struct mm_struct *mm)
230 {
231 	return	!IS_ENABLED(CONFIG_IA32_EMULATION) ||
232 		!test_bit(MM_CONTEXT_UPROBE_IA32, &mm->context.flags);
233 }
234 #else
235 static inline bool is_64bit_mm(struct mm_struct *mm)
236 {
237 	return false;
238 }
239 #endif
240 
241 /*
242  * We only want to enforce protection keys on the current process
243  * because we effectively have no access to PKRU for other
244  * processes or any way to tell *which * PKRU in a threaded
245  * process we could use.
246  *
247  * So do not enforce things if the VMA is not from the current
248  * mm, or if we are in a kernel thread.
249  */
250 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma,
251 		bool write, bool execute, bool foreign)
252 {
253 	/* pkeys never affect instruction fetches */
254 	if (execute)
255 		return true;
256 	/* allow access if the VMA is not one from this process */
257 	if (foreign || vma_is_foreign(vma))
258 		return true;
259 	return __pkru_allows_pkey(vma_pkey(vma), write);
260 }
261 
262 unsigned long __get_current_cr3_fast(void);
263 
264 #include <asm-generic/mmu_context.h>
265 
266 #endif /* _ASM_X86_MMU_CONTEXT_H */
267