xref: /linux/arch/x86/include/asm/mmu_context.h (revision fa84cf094ef9667e2b91c104b0a788fd1896f482)
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 
28 DECLARE_STATIC_KEY_FALSE(rdpmc_always_available_key);
29 
30 static inline void load_mm_cr4(struct mm_struct *mm)
31 {
32 	if (static_branch_unlikely(&rdpmc_always_available_key) ||
33 	    atomic_read(&mm->context.perf_rdpmc_allowed))
34 		cr4_set_bits(X86_CR4_PCE);
35 	else
36 		cr4_clear_bits(X86_CR4_PCE);
37 }
38 #else
39 static inline void load_mm_cr4(struct mm_struct *mm) {}
40 #endif
41 
42 #ifdef CONFIG_MODIFY_LDT_SYSCALL
43 /*
44  * ldt_structs can be allocated, used, and freed, but they are never
45  * modified while live.
46  */
47 struct ldt_struct {
48 	/*
49 	 * Xen requires page-aligned LDTs with special permissions.  This is
50 	 * needed to prevent us from installing evil descriptors such as
51 	 * call gates.  On native, we could merge the ldt_struct and LDT
52 	 * allocations, but it's not worth trying to optimize.
53 	 */
54 	struct desc_struct	*entries;
55 	unsigned int		nr_entries;
56 
57 	/*
58 	 * If PTI is in use, then the entries array is not mapped while we're
59 	 * in user mode.  The whole array will be aliased at the addressed
60 	 * given by ldt_slot_va(slot).  We use two slots so that we can allocate
61 	 * and map, and enable a new LDT without invalidating the mapping
62 	 * of an older, still-in-use LDT.
63 	 *
64 	 * slot will be -1 if this LDT doesn't have an alias mapping.
65 	 */
66 	int			slot;
67 };
68 
69 /* This is a multiple of PAGE_SIZE. */
70 #define LDT_SLOT_STRIDE (LDT_ENTRIES * LDT_ENTRY_SIZE)
71 
72 static inline void *ldt_slot_va(int slot)
73 {
74 #ifdef CONFIG_X86_64
75 	return (void *)(LDT_BASE_ADDR + LDT_SLOT_STRIDE * slot);
76 #else
77 	BUG();
78 	return (void *)fix_to_virt(FIX_HOLE);
79 #endif
80 }
81 
82 /*
83  * Used for LDT copy/destruction.
84  */
85 static inline void init_new_context_ldt(struct mm_struct *mm)
86 {
87 	mm->context.ldt = NULL;
88 	init_rwsem(&mm->context.ldt_usr_sem);
89 }
90 int ldt_dup_context(struct mm_struct *oldmm, struct mm_struct *mm);
91 void destroy_context_ldt(struct mm_struct *mm);
92 void ldt_arch_exit_mmap(struct mm_struct *mm);
93 #else	/* CONFIG_MODIFY_LDT_SYSCALL */
94 static inline void init_new_context_ldt(struct mm_struct *mm) { }
95 static inline int ldt_dup_context(struct mm_struct *oldmm,
96 				  struct mm_struct *mm)
97 {
98 	return 0;
99 }
100 static inline void destroy_context_ldt(struct mm_struct *mm) { }
101 static inline void ldt_arch_exit_mmap(struct mm_struct *mm) { }
102 #endif
103 
104 static inline void load_mm_ldt(struct mm_struct *mm)
105 {
106 #ifdef CONFIG_MODIFY_LDT_SYSCALL
107 	struct ldt_struct *ldt;
108 
109 	/* READ_ONCE synchronizes with smp_store_release */
110 	ldt = READ_ONCE(mm->context.ldt);
111 
112 	/*
113 	 * Any change to mm->context.ldt is followed by an IPI to all
114 	 * CPUs with the mm active.  The LDT will not be freed until
115 	 * after the IPI is handled by all such CPUs.  This means that,
116 	 * if the ldt_struct changes before we return, the values we see
117 	 * will be safe, and the new values will be loaded before we run
118 	 * any user code.
119 	 *
120 	 * NB: don't try to convert this to use RCU without extreme care.
121 	 * We would still need IRQs off, because we don't want to change
122 	 * the local LDT after an IPI loaded a newer value than the one
123 	 * that we can see.
124 	 */
125 
126 	if (unlikely(ldt)) {
127 		if (static_cpu_has(X86_FEATURE_PTI)) {
128 			if (WARN_ON_ONCE((unsigned long)ldt->slot > 1)) {
129 				/*
130 				 * Whoops -- either the new LDT isn't mapped
131 				 * (if slot == -1) or is mapped into a bogus
132 				 * slot (if slot > 1).
133 				 */
134 				clear_LDT();
135 				return;
136 			}
137 
138 			/*
139 			 * If page table isolation is enabled, ldt->entries
140 			 * will not be mapped in the userspace pagetables.
141 			 * Tell the CPU to access the LDT through the alias
142 			 * at ldt_slot_va(ldt->slot).
143 			 */
144 			set_ldt(ldt_slot_va(ldt->slot), ldt->nr_entries);
145 		} else {
146 			set_ldt(ldt->entries, ldt->nr_entries);
147 		}
148 	} else {
149 		clear_LDT();
150 	}
151 #else
152 	clear_LDT();
153 #endif
154 }
155 
156 static inline void switch_ldt(struct mm_struct *prev, struct mm_struct *next)
157 {
158 #ifdef CONFIG_MODIFY_LDT_SYSCALL
159 	/*
160 	 * Load the LDT if either the old or new mm had an LDT.
161 	 *
162 	 * An mm will never go from having an LDT to not having an LDT.  Two
163 	 * mms never share an LDT, so we don't gain anything by checking to
164 	 * see whether the LDT changed.  There's also no guarantee that
165 	 * prev->context.ldt actually matches LDTR, but, if LDTR is non-NULL,
166 	 * then prev->context.ldt will also be non-NULL.
167 	 *
168 	 * If we really cared, we could optimize the case where prev == next
169 	 * and we're exiting lazy mode.  Most of the time, if this happens,
170 	 * we don't actually need to reload LDTR, but modify_ldt() is mostly
171 	 * used by legacy code and emulators where we don't need this level of
172 	 * performance.
173 	 *
174 	 * This uses | instead of || because it generates better code.
175 	 */
176 	if (unlikely((unsigned long)prev->context.ldt |
177 		     (unsigned long)next->context.ldt))
178 		load_mm_ldt(next);
179 #endif
180 
181 	DEBUG_LOCKS_WARN_ON(preemptible());
182 }
183 
184 void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk);
185 
186 static inline int init_new_context(struct task_struct *tsk,
187 				   struct mm_struct *mm)
188 {
189 	mutex_init(&mm->context.lock);
190 
191 	mm->context.ctx_id = atomic64_inc_return(&last_mm_ctx_id);
192 	atomic64_set(&mm->context.tlb_gen, 0);
193 
194 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
195 	if (cpu_feature_enabled(X86_FEATURE_OSPKE)) {
196 		/* pkey 0 is the default and allocated implicitly */
197 		mm->context.pkey_allocation_map = 0x1;
198 		/* -1 means unallocated or invalid */
199 		mm->context.execute_only_pkey = -1;
200 	}
201 #endif
202 	init_new_context_ldt(mm);
203 	return 0;
204 }
205 static inline void destroy_context(struct mm_struct *mm)
206 {
207 	destroy_context_ldt(mm);
208 }
209 
210 extern void switch_mm(struct mm_struct *prev, struct mm_struct *next,
211 		      struct task_struct *tsk);
212 
213 extern void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
214 			       struct task_struct *tsk);
215 #define switch_mm_irqs_off switch_mm_irqs_off
216 
217 #define activate_mm(prev, next)			\
218 do {						\
219 	paravirt_activate_mm((prev), (next));	\
220 	switch_mm((prev), (next), NULL);	\
221 } while (0);
222 
223 #ifdef CONFIG_X86_32
224 #define deactivate_mm(tsk, mm)			\
225 do {						\
226 	lazy_load_gs(0);			\
227 } while (0)
228 #else
229 #define deactivate_mm(tsk, mm)			\
230 do {						\
231 	load_gs_index(0);			\
232 	loadsegment(fs, 0);			\
233 } while (0)
234 #endif
235 
236 static inline int arch_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
237 {
238 	paravirt_arch_dup_mmap(oldmm, mm);
239 	return ldt_dup_context(oldmm, mm);
240 }
241 
242 static inline void arch_exit_mmap(struct mm_struct *mm)
243 {
244 	paravirt_arch_exit_mmap(mm);
245 	ldt_arch_exit_mmap(mm);
246 }
247 
248 #ifdef CONFIG_X86_64
249 static inline bool is_64bit_mm(struct mm_struct *mm)
250 {
251 	return	!IS_ENABLED(CONFIG_IA32_EMULATION) ||
252 		!(mm->context.ia32_compat == TIF_IA32);
253 }
254 #else
255 static inline bool is_64bit_mm(struct mm_struct *mm)
256 {
257 	return false;
258 }
259 #endif
260 
261 static inline void arch_bprm_mm_init(struct mm_struct *mm,
262 		struct vm_area_struct *vma)
263 {
264 	mpx_mm_init(mm);
265 }
266 
267 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma,
268 			      unsigned long start, unsigned long end)
269 {
270 	/*
271 	 * mpx_notify_unmap() goes and reads a rarely-hot
272 	 * cacheline in the mm_struct.  That can be expensive
273 	 * enough to be seen in profiles.
274 	 *
275 	 * The mpx_notify_unmap() call and its contents have been
276 	 * observed to affect munmap() performance on hardware
277 	 * where MPX is not present.
278 	 *
279 	 * The unlikely() optimizes for the fast case: no MPX
280 	 * in the CPU, or no MPX use in the process.  Even if
281 	 * we get this wrong (in the unlikely event that MPX
282 	 * is widely enabled on some system) the overhead of
283 	 * MPX itself (reading bounds tables) is expected to
284 	 * overwhelm the overhead of getting this unlikely()
285 	 * consistently wrong.
286 	 */
287 	if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX)))
288 		mpx_notify_unmap(mm, vma, start, end);
289 }
290 
291 /*
292  * We only want to enforce protection keys on the current process
293  * because we effectively have no access to PKRU for other
294  * processes or any way to tell *which * PKRU in a threaded
295  * process we could use.
296  *
297  * So do not enforce things if the VMA is not from the current
298  * mm, or if we are in a kernel thread.
299  */
300 static inline bool vma_is_foreign(struct vm_area_struct *vma)
301 {
302 	if (!current->mm)
303 		return true;
304 	/*
305 	 * Should PKRU be enforced on the access to this VMA?  If
306 	 * the VMA is from another process, then PKRU has no
307 	 * relevance and should not be enforced.
308 	 */
309 	if (current->mm != vma->vm_mm)
310 		return true;
311 
312 	return false;
313 }
314 
315 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma,
316 		bool write, bool execute, bool foreign)
317 {
318 	/* pkeys never affect instruction fetches */
319 	if (execute)
320 		return true;
321 	/* allow access if the VMA is not one from this process */
322 	if (foreign || vma_is_foreign(vma))
323 		return true;
324 	return __pkru_allows_pkey(vma_pkey(vma), write);
325 }
326 
327 /*
328  * This can be used from process context to figure out what the value of
329  * CR3 is without needing to do a (slow) __read_cr3().
330  *
331  * It's intended to be used for code like KVM that sneakily changes CR3
332  * and needs to restore it.  It needs to be used very carefully.
333  */
334 static inline unsigned long __get_current_cr3_fast(void)
335 {
336 	unsigned long cr3 = build_cr3(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd,
337 		this_cpu_read(cpu_tlbstate.loaded_mm_asid));
338 
339 	/* For now, be very restrictive about when this can be called. */
340 	VM_WARN_ON(in_nmi() || preemptible());
341 
342 	VM_BUG_ON(cr3 != __read_cr3());
343 	return cr3;
344 }
345 
346 #endif /* _ASM_X86_MMU_CONTEXT_H */
347