xref: /openbmc/linux/arch/x86/include/asm/mmu_context.h (revision 9dae47aba0a055f761176d9297371d5bb24289ec)
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/pgalloc.h>
13 #include <asm/tlbflush.h>
14 #include <asm/paravirt.h>
15 #include <asm/mpx.h>
16 
17 extern atomic64_t last_mm_ctx_id;
18 
19 #ifndef CONFIG_PARAVIRT
20 static inline void paravirt_activate_mm(struct mm_struct *prev,
21 					struct mm_struct *next)
22 {
23 }
24 #endif	/* !CONFIG_PARAVIRT */
25 
26 #ifdef CONFIG_PERF_EVENTS
27 extern struct static_key rdpmc_always_available;
28 
29 static inline void load_mm_cr4(struct mm_struct *mm)
30 {
31 	if (static_key_false(&rdpmc_always_available) ||
32 	    atomic_read(&mm->context.perf_rdpmc_allowed))
33 		cr4_set_bits(X86_CR4_PCE);
34 	else
35 		cr4_clear_bits(X86_CR4_PCE);
36 }
37 #else
38 static inline void load_mm_cr4(struct mm_struct *mm) {}
39 #endif
40 
41 #ifdef CONFIG_MODIFY_LDT_SYSCALL
42 /*
43  * ldt_structs can be allocated, used, and freed, but they are never
44  * modified while live.
45  */
46 struct ldt_struct {
47 	/*
48 	 * Xen requires page-aligned LDTs with special permissions.  This is
49 	 * needed to prevent us from installing evil descriptors such as
50 	 * call gates.  On native, we could merge the ldt_struct and LDT
51 	 * allocations, but it's not worth trying to optimize.
52 	 */
53 	struct desc_struct *entries;
54 	unsigned int nr_entries;
55 };
56 
57 /*
58  * Used for LDT copy/destruction.
59  */
60 static inline void init_new_context_ldt(struct mm_struct *mm)
61 {
62 	mm->context.ldt = NULL;
63 	init_rwsem(&mm->context.ldt_usr_sem);
64 }
65 int ldt_dup_context(struct mm_struct *oldmm, struct mm_struct *mm);
66 void destroy_context_ldt(struct mm_struct *mm);
67 #else	/* CONFIG_MODIFY_LDT_SYSCALL */
68 static inline void init_new_context_ldt(struct mm_struct *mm) { }
69 static inline int ldt_dup_context(struct mm_struct *oldmm,
70 				  struct mm_struct *mm)
71 {
72 	return 0;
73 }
74 static inline void destroy_context_ldt(struct mm_struct *mm) {}
75 #endif
76 
77 static inline void load_mm_ldt(struct mm_struct *mm)
78 {
79 #ifdef CONFIG_MODIFY_LDT_SYSCALL
80 	struct ldt_struct *ldt;
81 
82 	/* READ_ONCE synchronizes with smp_store_release */
83 	ldt = READ_ONCE(mm->context.ldt);
84 
85 	/*
86 	 * Any change to mm->context.ldt is followed by an IPI to all
87 	 * CPUs with the mm active.  The LDT will not be freed until
88 	 * after the IPI is handled by all such CPUs.  This means that,
89 	 * if the ldt_struct changes before we return, the values we see
90 	 * will be safe, and the new values will be loaded before we run
91 	 * any user code.
92 	 *
93 	 * NB: don't try to convert this to use RCU without extreme care.
94 	 * We would still need IRQs off, because we don't want to change
95 	 * the local LDT after an IPI loaded a newer value than the one
96 	 * that we can see.
97 	 */
98 
99 	if (unlikely(ldt))
100 		set_ldt(ldt->entries, ldt->nr_entries);
101 	else
102 		clear_LDT();
103 #else
104 	clear_LDT();
105 #endif
106 }
107 
108 static inline void switch_ldt(struct mm_struct *prev, struct mm_struct *next)
109 {
110 #ifdef CONFIG_MODIFY_LDT_SYSCALL
111 	/*
112 	 * Load the LDT if either the old or new mm had an LDT.
113 	 *
114 	 * An mm will never go from having an LDT to not having an LDT.  Two
115 	 * mms never share an LDT, so we don't gain anything by checking to
116 	 * see whether the LDT changed.  There's also no guarantee that
117 	 * prev->context.ldt actually matches LDTR, but, if LDTR is non-NULL,
118 	 * then prev->context.ldt will also be non-NULL.
119 	 *
120 	 * If we really cared, we could optimize the case where prev == next
121 	 * and we're exiting lazy mode.  Most of the time, if this happens,
122 	 * we don't actually need to reload LDTR, but modify_ldt() is mostly
123 	 * used by legacy code and emulators where we don't need this level of
124 	 * performance.
125 	 *
126 	 * This uses | instead of || because it generates better code.
127 	 */
128 	if (unlikely((unsigned long)prev->context.ldt |
129 		     (unsigned long)next->context.ldt))
130 		load_mm_ldt(next);
131 #endif
132 
133 	DEBUG_LOCKS_WARN_ON(preemptible());
134 }
135 
136 void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk);
137 
138 static inline int init_new_context(struct task_struct *tsk,
139 				   struct mm_struct *mm)
140 {
141 	mutex_init(&mm->context.lock);
142 
143 	mm->context.ctx_id = atomic64_inc_return(&last_mm_ctx_id);
144 	atomic64_set(&mm->context.tlb_gen, 0);
145 
146 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
147 	if (cpu_feature_enabled(X86_FEATURE_OSPKE)) {
148 		/* pkey 0 is the default and always allocated */
149 		mm->context.pkey_allocation_map = 0x1;
150 		/* -1 means unallocated or invalid */
151 		mm->context.execute_only_pkey = -1;
152 	}
153 #endif
154 	init_new_context_ldt(mm);
155 	return 0;
156 }
157 static inline void destroy_context(struct mm_struct *mm)
158 {
159 	destroy_context_ldt(mm);
160 }
161 
162 extern void switch_mm(struct mm_struct *prev, struct mm_struct *next,
163 		      struct task_struct *tsk);
164 
165 extern void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
166 			       struct task_struct *tsk);
167 #define switch_mm_irqs_off switch_mm_irqs_off
168 
169 #define activate_mm(prev, next)			\
170 do {						\
171 	paravirt_activate_mm((prev), (next));	\
172 	switch_mm((prev), (next), NULL);	\
173 } while (0);
174 
175 #ifdef CONFIG_X86_32
176 #define deactivate_mm(tsk, mm)			\
177 do {						\
178 	lazy_load_gs(0);			\
179 } while (0)
180 #else
181 #define deactivate_mm(tsk, mm)			\
182 do {						\
183 	load_gs_index(0);			\
184 	loadsegment(fs, 0);			\
185 } while (0)
186 #endif
187 
188 static inline int arch_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
189 {
190 	paravirt_arch_dup_mmap(oldmm, mm);
191 	return ldt_dup_context(oldmm, mm);
192 }
193 
194 static inline void arch_exit_mmap(struct mm_struct *mm)
195 {
196 	paravirt_arch_exit_mmap(mm);
197 }
198 
199 #ifdef CONFIG_X86_64
200 static inline bool is_64bit_mm(struct mm_struct *mm)
201 {
202 	return	!IS_ENABLED(CONFIG_IA32_EMULATION) ||
203 		!(mm->context.ia32_compat == TIF_IA32);
204 }
205 #else
206 static inline bool is_64bit_mm(struct mm_struct *mm)
207 {
208 	return false;
209 }
210 #endif
211 
212 static inline void arch_bprm_mm_init(struct mm_struct *mm,
213 		struct vm_area_struct *vma)
214 {
215 	mpx_mm_init(mm);
216 }
217 
218 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma,
219 			      unsigned long start, unsigned long end)
220 {
221 	/*
222 	 * mpx_notify_unmap() goes and reads a rarely-hot
223 	 * cacheline in the mm_struct.  That can be expensive
224 	 * enough to be seen in profiles.
225 	 *
226 	 * The mpx_notify_unmap() call and its contents have been
227 	 * observed to affect munmap() performance on hardware
228 	 * where MPX is not present.
229 	 *
230 	 * The unlikely() optimizes for the fast case: no MPX
231 	 * in the CPU, or no MPX use in the process.  Even if
232 	 * we get this wrong (in the unlikely event that MPX
233 	 * is widely enabled on some system) the overhead of
234 	 * MPX itself (reading bounds tables) is expected to
235 	 * overwhelm the overhead of getting this unlikely()
236 	 * consistently wrong.
237 	 */
238 	if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX)))
239 		mpx_notify_unmap(mm, vma, start, end);
240 }
241 
242 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
243 static inline int vma_pkey(struct vm_area_struct *vma)
244 {
245 	unsigned long vma_pkey_mask = VM_PKEY_BIT0 | VM_PKEY_BIT1 |
246 				      VM_PKEY_BIT2 | VM_PKEY_BIT3;
247 
248 	return (vma->vm_flags & vma_pkey_mask) >> VM_PKEY_SHIFT;
249 }
250 #else
251 static inline int vma_pkey(struct vm_area_struct *vma)
252 {
253 	return 0;
254 }
255 #endif
256 
257 /*
258  * We only want to enforce protection keys on the current process
259  * because we effectively have no access to PKRU for other
260  * processes or any way to tell *which * PKRU in a threaded
261  * process we could use.
262  *
263  * So do not enforce things if the VMA is not from the current
264  * mm, or if we are in a kernel thread.
265  */
266 static inline bool vma_is_foreign(struct vm_area_struct *vma)
267 {
268 	if (!current->mm)
269 		return true;
270 	/*
271 	 * Should PKRU be enforced on the access to this VMA?  If
272 	 * the VMA is from another process, then PKRU has no
273 	 * relevance and should not be enforced.
274 	 */
275 	if (current->mm != vma->vm_mm)
276 		return true;
277 
278 	return false;
279 }
280 
281 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma,
282 		bool write, bool execute, bool foreign)
283 {
284 	/* pkeys never affect instruction fetches */
285 	if (execute)
286 		return true;
287 	/* allow access if the VMA is not one from this process */
288 	if (foreign || vma_is_foreign(vma))
289 		return true;
290 	return __pkru_allows_pkey(vma_pkey(vma), write);
291 }
292 
293 /*
294  * This can be used from process context to figure out what the value of
295  * CR3 is without needing to do a (slow) __read_cr3().
296  *
297  * It's intended to be used for code like KVM that sneakily changes CR3
298  * and needs to restore it.  It needs to be used very carefully.
299  */
300 static inline unsigned long __get_current_cr3_fast(void)
301 {
302 	unsigned long cr3 = build_cr3(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd,
303 		this_cpu_read(cpu_tlbstate.loaded_mm_asid));
304 
305 	/* For now, be very restrictive about when this can be called. */
306 	VM_WARN_ON(in_nmi() || preemptible());
307 
308 	VM_BUG_ON(cr3 != __read_cr3());
309 	return cr3;
310 }
311 
312 #endif /* _ASM_X86_MMU_CONTEXT_H */
313