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 #include <asm/debugreg.h>
17
18 extern atomic64_t last_mm_ctx_id;
19
20 #ifndef CONFIG_PARAVIRT_XXL
21 static inline void paravirt_activate_mm(struct mm_struct *prev,
22 struct mm_struct *next)
23 {
24 }
25 #endif /* !CONFIG_PARAVIRT_XXL */
26
27 #ifdef CONFIG_PERF_EVENTS
28
29 DECLARE_STATIC_KEY_FALSE(rdpmc_always_available_key);
30
31 static inline void load_mm_cr4_irqsoff(struct mm_struct *mm)
32 {
33 if (static_branch_unlikely(&rdpmc_always_available_key) ||
34 atomic_read(&mm->context.perf_rdpmc_allowed))
35 cr4_set_bits_irqsoff(X86_CR4_PCE);
36 else
37 cr4_clear_bits_irqsoff(X86_CR4_PCE);
38 }
39 #else
40 static inline void load_mm_cr4_irqsoff(struct mm_struct *mm) {}
41 #endif
42
43 #ifdef CONFIG_MODIFY_LDT_SYSCALL
44 /*
45 * ldt_structs can be allocated, used, and freed, but they are never
46 * modified while live.
47 */
48 struct ldt_struct {
49 /*
50 * Xen requires page-aligned LDTs with special permissions. This is
51 * needed to prevent us from installing evil descriptors such as
52 * call gates. On native, we could merge the ldt_struct and LDT
53 * allocations, but it's not worth trying to optimize.
54 */
55 struct desc_struct *entries;
56 unsigned int nr_entries;
57
58 /*
59 * If PTI is in use, then the entries array is not mapped while we're
60 * in user mode. The whole array will be aliased at the addressed
61 * given by ldt_slot_va(slot). We use two slots so that we can allocate
62 * and map, and enable a new LDT without invalidating the mapping
63 * of an older, still-in-use LDT.
64 *
65 * slot will be -1 if this LDT doesn't have an alias mapping.
66 */
67 int slot;
68 };
69
70 /* This is a multiple of PAGE_SIZE. */
71 #define LDT_SLOT_STRIDE (LDT_ENTRIES * LDT_ENTRY_SIZE)
72
73 static inline void *ldt_slot_va(int slot)
74 {
75 return (void *)(LDT_BASE_ADDR + LDT_SLOT_STRIDE * slot);
76 }
77
78 /*
79 * Used for LDT copy/destruction.
80 */
81 static inline void init_new_context_ldt(struct mm_struct *mm)
82 {
83 mm->context.ldt = NULL;
84 init_rwsem(&mm->context.ldt_usr_sem);
85 }
86 int ldt_dup_context(struct mm_struct *oldmm, struct mm_struct *mm);
87 void destroy_context_ldt(struct mm_struct *mm);
88 void ldt_arch_exit_mmap(struct mm_struct *mm);
89 #else /* CONFIG_MODIFY_LDT_SYSCALL */
90 static inline void init_new_context_ldt(struct mm_struct *mm) { }
91 static inline int ldt_dup_context(struct mm_struct *oldmm,
92 struct mm_struct *mm)
93 {
94 return 0;
95 }
96 static inline void destroy_context_ldt(struct mm_struct *mm) { }
97 static inline void ldt_arch_exit_mmap(struct mm_struct *mm) { }
98 #endif
99
100 static inline void load_mm_ldt(struct mm_struct *mm)
101 {
102 #ifdef CONFIG_MODIFY_LDT_SYSCALL
103 struct ldt_struct *ldt;
104
105 /* READ_ONCE synchronizes with smp_store_release */
106 ldt = READ_ONCE(mm->context.ldt);
107
108 /*
109 * Any change to mm->context.ldt is followed by an IPI to all
110 * CPUs with the mm active. The LDT will not be freed until
111 * after the IPI is handled by all such CPUs. This means that,
112 * if the ldt_struct changes before we return, the values we see
113 * will be safe, and the new values will be loaded before we run
114 * any user code.
115 *
116 * NB: don't try to convert this to use RCU without extreme care.
117 * We would still need IRQs off, because we don't want to change
118 * the local LDT after an IPI loaded a newer value than the one
119 * that we can see.
120 */
121
122 if (unlikely(ldt)) {
123 if (static_cpu_has(X86_FEATURE_PTI)) {
124 if (WARN_ON_ONCE((unsigned long)ldt->slot > 1)) {
125 /*
126 * Whoops -- either the new LDT isn't mapped
127 * (if slot == -1) or is mapped into a bogus
128 * slot (if slot > 1).
129 */
130 clear_LDT();
131 return;
132 }
133
134 /*
135 * If page table isolation is enabled, ldt->entries
136 * will not be mapped in the userspace pagetables.
137 * Tell the CPU to access the LDT through the alias
138 * at ldt_slot_va(ldt->slot).
139 */
140 set_ldt(ldt_slot_va(ldt->slot), ldt->nr_entries);
141 } else {
142 set_ldt(ldt->entries, ldt->nr_entries);
143 }
144 } else {
145 clear_LDT();
146 }
147 #else
148 clear_LDT();
149 #endif
150 }
151
152 static inline void switch_ldt(struct mm_struct *prev, struct mm_struct *next)
153 {
154 #ifdef CONFIG_MODIFY_LDT_SYSCALL
155 /*
156 * Load the LDT if either the old or new mm had an LDT.
157 *
158 * An mm will never go from having an LDT to not having an LDT. Two
159 * mms never share an LDT, so we don't gain anything by checking to
160 * see whether the LDT changed. There's also no guarantee that
161 * prev->context.ldt actually matches LDTR, but, if LDTR is non-NULL,
162 * then prev->context.ldt will also be non-NULL.
163 *
164 * If we really cared, we could optimize the case where prev == next
165 * and we're exiting lazy mode. Most of the time, if this happens,
166 * we don't actually need to reload LDTR, but modify_ldt() is mostly
167 * used by legacy code and emulators where we don't need this level of
168 * performance.
169 *
170 * This uses | instead of || because it generates better code.
171 */
172 if (unlikely((unsigned long)prev->context.ldt |
173 (unsigned long)next->context.ldt))
174 load_mm_ldt(next);
175 #endif
176
177 DEBUG_LOCKS_WARN_ON(preemptible());
178 }
179
180 void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk);
181
182 /*
183 * Init a new mm. Used on mm copies, like at fork()
184 * and on mm's that are brand-new, like at execve().
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 void arch_dup_pkeys(struct mm_struct *oldmm,
237 struct mm_struct *mm)
238 {
239 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
240 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
241 return;
242
243 /* Duplicate the oldmm pkey state in mm: */
244 mm->context.pkey_allocation_map = oldmm->context.pkey_allocation_map;
245 mm->context.execute_only_pkey = oldmm->context.execute_only_pkey;
246 #endif
247 }
248
249 static inline int arch_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
250 {
251 arch_dup_pkeys(oldmm, mm);
252 paravirt_arch_dup_mmap(oldmm, mm);
253 return ldt_dup_context(oldmm, mm);
254 }
255
256 static inline void arch_exit_mmap(struct mm_struct *mm)
257 {
258 paravirt_arch_exit_mmap(mm);
259 ldt_arch_exit_mmap(mm);
260 }
261
262 #ifdef CONFIG_X86_64
263 static inline bool is_64bit_mm(struct mm_struct *mm)
264 {
265 return !IS_ENABLED(CONFIG_IA32_EMULATION) ||
266 !(mm->context.ia32_compat == TIF_IA32);
267 }
268 #else
269 static inline bool is_64bit_mm(struct mm_struct *mm)
270 {
271 return false;
272 }
273 #endif
274
275 static inline void arch_bprm_mm_init(struct mm_struct *mm,
276 struct vm_area_struct *vma)
277 {
278 mpx_mm_init(mm);
279 }
280
281 static inline void arch_unmap(struct mm_struct *mm, unsigned long start,
282 unsigned long end)
283 {
284 /*
285 * mpx_notify_unmap() goes and reads a rarely-hot
286 * cacheline in the mm_struct. That can be expensive
287 * enough to be seen in profiles.
288 *
289 * The mpx_notify_unmap() call and its contents have been
290 * observed to affect munmap() performance on hardware
291 * where MPX is not present.
292 *
293 * The unlikely() optimizes for the fast case: no MPX
294 * in the CPU, or no MPX use in the process. Even if
295 * we get this wrong (in the unlikely event that MPX
296 * is widely enabled on some system) the overhead of
297 * MPX itself (reading bounds tables) is expected to
298 * overwhelm the overhead of getting this unlikely()
299 * consistently wrong.
300 */
301 if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX)))
302 mpx_notify_unmap(mm, start, end);
303 }
304
305 /*
306 * We only want to enforce protection keys on the current process
307 * because we effectively have no access to PKRU for other
308 * processes or any way to tell *which * PKRU in a threaded
309 * process we could use.
310 *
311 * So do not enforce things if the VMA is not from the current
312 * mm, or if we are in a kernel thread.
313 */
314 static inline bool vma_is_foreign(struct vm_area_struct *vma)
315 {
316 if (!current->mm)
317 return true;
318 /*
319 * Should PKRU be enforced on the access to this VMA? If
320 * the VMA is from another process, then PKRU has no
321 * relevance and should not be enforced.
322 */
323 if (current->mm != vma->vm_mm)
324 return true;
325
326 return false;
327 }
328
329 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma,
330 bool write, bool execute, bool foreign)
331 {
332 /* pkeys never affect instruction fetches */
333 if (execute)
334 return true;
335 /* allow access if the VMA is not one from this process */
336 if (foreign || vma_is_foreign(vma))
337 return true;
338 return __pkru_allows_pkey(vma_pkey(vma), write);
339 }
340
341 /*
342 * This can be used from process context to figure out what the value of
343 * CR3 is without needing to do a (slow) __read_cr3().
344 *
345 * It's intended to be used for code like KVM that sneakily changes CR3
346 * and needs to restore it. It needs to be used very carefully.
347 */
348 static inline unsigned long __get_current_cr3_fast(void)
349 {
350 unsigned long cr3 = build_cr3(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd,
351 this_cpu_read(cpu_tlbstate.loaded_mm_asid));
352
353 /* For now, be very restrictive about when this can be called. */
354 VM_WARN_ON(in_nmi() || preemptible());
355
356 VM_BUG_ON(cr3 != __read_cr3());
357 return cr3;
358 }
359
360 typedef struct {
361 struct mm_struct *mm;
362 } temp_mm_state_t;
363
364 /*
365 * Using a temporary mm allows to set temporary mappings that are not accessible
366 * by other CPUs. Such mappings are needed to perform sensitive memory writes
367 * that override the kernel memory protections (e.g., W^X), without exposing the
368 * temporary page-table mappings that are required for these write operations to
369 * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the
370 * mapping is torn down.
371 *
372 * Context: The temporary mm needs to be used exclusively by a single core. To
373 * harden security IRQs must be disabled while the temporary mm is
374 * loaded, thereby preventing interrupt handler bugs from overriding
375 * the kernel memory protection.
376 */
377 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm)
378 {
379 temp_mm_state_t temp_state;
380
381 lockdep_assert_irqs_disabled();
382 temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm);
383 switch_mm_irqs_off(NULL, mm, current);
384
385 /*
386 * If breakpoints are enabled, disable them while the temporary mm is
387 * used. Userspace might set up watchpoints on addresses that are used
388 * in the temporary mm, which would lead to wrong signals being sent or
389 * crashes.
390 *
391 * Note that breakpoints are not disabled selectively, which also causes
392 * kernel breakpoints (e.g., perf's) to be disabled. This might be
393 * undesirable, but still seems reasonable as the code that runs in the
394 * temporary mm should be short.
395 */
396 if (hw_breakpoint_active())
397 hw_breakpoint_disable();
398
399 return temp_state;
400 }
401
402 static inline void unuse_temporary_mm(temp_mm_state_t prev_state)
403 {
404 lockdep_assert_irqs_disabled();
405 switch_mm_irqs_off(NULL, prev_state.mm, current);
406
407 /*
408 * Restore the breakpoints if they were disabled before the temporary mm
409 * was loaded.
410 */
411 if (hw_breakpoint_active())
412 hw_breakpoint_restore();
413 }
414
415 #endif /* _ASM_X86_MMU_CONTEXT_H */