xref: /linux/arch/x86/include/asm/tlbflush.h (revision 906fd46a65383cd639e5eec72a047efc33045d86)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_X86_TLBFLUSH_H
3 #define _ASM_X86_TLBFLUSH_H
4 
5 #include <linux/mm_types.h>
6 #include <linux/mmu_notifier.h>
7 #include <linux/sched.h>
8 
9 #include <asm/processor.h>
10 #include <asm/cpufeature.h>
11 #include <asm/special_insns.h>
12 #include <asm/smp.h>
13 #include <asm/invpcid.h>
14 #include <asm/pti.h>
15 #include <asm/processor-flags.h>
16 #include <asm/pgtable.h>
17 
18 DECLARE_PER_CPU(u64, tlbstate_untag_mask);
19 
20 void __flush_tlb_all(void);
21 
22 #define TLB_FLUSH_ALL	-1UL
23 #define TLB_GENERATION_INVALID	0
24 
25 void cr4_update_irqsoff(unsigned long set, unsigned long clear);
26 unsigned long cr4_read_shadow(void);
27 
28 /* Set in this cpu's CR4. */
29 static inline void cr4_set_bits_irqsoff(unsigned long mask)
30 {
31 	cr4_update_irqsoff(mask, 0);
32 }
33 
34 /* Clear in this cpu's CR4. */
35 static inline void cr4_clear_bits_irqsoff(unsigned long mask)
36 {
37 	cr4_update_irqsoff(0, mask);
38 }
39 
40 /* Set in this cpu's CR4. */
41 static inline void cr4_set_bits(unsigned long mask)
42 {
43 	unsigned long flags;
44 
45 	local_irq_save(flags);
46 	cr4_set_bits_irqsoff(mask);
47 	local_irq_restore(flags);
48 }
49 
50 /* Clear in this cpu's CR4. */
51 static inline void cr4_clear_bits(unsigned long mask)
52 {
53 	unsigned long flags;
54 
55 	local_irq_save(flags);
56 	cr4_clear_bits_irqsoff(mask);
57 	local_irq_restore(flags);
58 }
59 
60 #ifndef MODULE
61 /*
62  * 6 because 6 should be plenty and struct tlb_state will fit in two cache
63  * lines.
64  */
65 #define TLB_NR_DYN_ASIDS	6
66 
67 struct tlb_context {
68 	u64 ctx_id;
69 	u64 tlb_gen;
70 };
71 
72 struct tlb_state {
73 	/*
74 	 * cpu_tlbstate.loaded_mm should match CR3 whenever interrupts
75 	 * are on.  This means that it may not match current->active_mm,
76 	 * which will contain the previous user mm when we're in lazy TLB
77 	 * mode even if we've already switched back to swapper_pg_dir.
78 	 *
79 	 * During switch_mm_irqs_off(), loaded_mm will be set to
80 	 * LOADED_MM_SWITCHING during the brief interrupts-off window
81 	 * when CR3 and loaded_mm would otherwise be inconsistent.  This
82 	 * is for nmi_uaccess_okay()'s benefit.
83 	 */
84 	struct mm_struct *loaded_mm;
85 
86 #define LOADED_MM_SWITCHING ((struct mm_struct *)1UL)
87 
88 	/* Last user mm for optimizing IBPB */
89 	union {
90 		struct mm_struct	*last_user_mm;
91 		unsigned long		last_user_mm_spec;
92 	};
93 
94 	u16 loaded_mm_asid;
95 	u16 next_asid;
96 
97 	/*
98 	 * If set we changed the page tables in such a way that we
99 	 * needed an invalidation of all contexts (aka. PCIDs / ASIDs).
100 	 * This tells us to go invalidate all the non-loaded ctxs[]
101 	 * on the next context switch.
102 	 *
103 	 * The current ctx was kept up-to-date as it ran and does not
104 	 * need to be invalidated.
105 	 */
106 	bool invalidate_other;
107 
108 #ifdef CONFIG_ADDRESS_MASKING
109 	/*
110 	 * Active LAM mode.
111 	 *
112 	 * X86_CR3_LAM_U57/U48 shifted right by X86_CR3_LAM_U57_BIT or 0 if LAM
113 	 * disabled.
114 	 */
115 	u8 lam;
116 #endif
117 
118 	/*
119 	 * Mask that contains TLB_NR_DYN_ASIDS+1 bits to indicate
120 	 * the corresponding user PCID needs a flush next time we
121 	 * switch to it; see SWITCH_TO_USER_CR3.
122 	 */
123 	unsigned short user_pcid_flush_mask;
124 
125 	/*
126 	 * Access to this CR4 shadow and to H/W CR4 is protected by
127 	 * disabling interrupts when modifying either one.
128 	 */
129 	unsigned long cr4;
130 
131 	/*
132 	 * This is a list of all contexts that might exist in the TLB.
133 	 * There is one per ASID that we use, and the ASID (what the
134 	 * CPU calls PCID) is the index into ctxts.
135 	 *
136 	 * For each context, ctx_id indicates which mm the TLB's user
137 	 * entries came from.  As an invariant, the TLB will never
138 	 * contain entries that are out-of-date as when that mm reached
139 	 * the tlb_gen in the list.
140 	 *
141 	 * To be clear, this means that it's legal for the TLB code to
142 	 * flush the TLB without updating tlb_gen.  This can happen
143 	 * (for now, at least) due to paravirt remote flushes.
144 	 *
145 	 * NB: context 0 is a bit special, since it's also used by
146 	 * various bits of init code.  This is fine -- code that
147 	 * isn't aware of PCID will end up harmlessly flushing
148 	 * context 0.
149 	 */
150 	struct tlb_context ctxs[TLB_NR_DYN_ASIDS];
151 };
152 DECLARE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate);
153 
154 struct tlb_state_shared {
155 	/*
156 	 * We can be in one of several states:
157 	 *
158 	 *  - Actively using an mm.  Our CPU's bit will be set in
159 	 *    mm_cpumask(loaded_mm) and is_lazy == false;
160 	 *
161 	 *  - Not using a real mm.  loaded_mm == &init_mm.  Our CPU's bit
162 	 *    will not be set in mm_cpumask(&init_mm) and is_lazy == false.
163 	 *
164 	 *  - Lazily using a real mm.  loaded_mm != &init_mm, our bit
165 	 *    is set in mm_cpumask(loaded_mm), but is_lazy == true.
166 	 *    We're heuristically guessing that the CR3 load we
167 	 *    skipped more than makes up for the overhead added by
168 	 *    lazy mode.
169 	 */
170 	bool is_lazy;
171 };
172 DECLARE_PER_CPU_SHARED_ALIGNED(struct tlb_state_shared, cpu_tlbstate_shared);
173 
174 bool nmi_uaccess_okay(void);
175 #define nmi_uaccess_okay nmi_uaccess_okay
176 
177 /* Initialize cr4 shadow for this CPU. */
178 static inline void cr4_init_shadow(void)
179 {
180 	this_cpu_write(cpu_tlbstate.cr4, __read_cr4());
181 }
182 
183 extern unsigned long mmu_cr4_features;
184 extern u32 *trampoline_cr4_features;
185 
186 extern void initialize_tlbstate_and_flush(void);
187 
188 /*
189  * TLB flushing:
190  *
191  *  - flush_tlb_all() flushes all processes TLBs
192  *  - flush_tlb_mm(mm) flushes the specified mm context TLB's
193  *  - flush_tlb_page(vma, vmaddr) flushes one page
194  *  - flush_tlb_range(vma, start, end) flushes a range of pages
195  *  - flush_tlb_kernel_range(start, end) flushes a range of kernel pages
196  *  - flush_tlb_multi(cpumask, info) flushes TLBs on multiple cpus
197  *
198  * ..but the i386 has somewhat limited tlb flushing capabilities,
199  * and page-granular flushes are available only on i486 and up.
200  */
201 struct flush_tlb_info {
202 	/*
203 	 * We support several kinds of flushes.
204 	 *
205 	 * - Fully flush a single mm.  .mm will be set, .end will be
206 	 *   TLB_FLUSH_ALL, and .new_tlb_gen will be the tlb_gen to
207 	 *   which the IPI sender is trying to catch us up.
208 	 *
209 	 * - Partially flush a single mm.  .mm will be set, .start and
210 	 *   .end will indicate the range, and .new_tlb_gen will be set
211 	 *   such that the changes between generation .new_tlb_gen-1 and
212 	 *   .new_tlb_gen are entirely contained in the indicated range.
213 	 *
214 	 * - Fully flush all mms whose tlb_gens have been updated.  .mm
215 	 *   will be NULL, .end will be TLB_FLUSH_ALL, and .new_tlb_gen
216 	 *   will be zero.
217 	 */
218 	struct mm_struct	*mm;
219 	unsigned long		start;
220 	unsigned long		end;
221 	u64			new_tlb_gen;
222 	unsigned int		initiating_cpu;
223 	u8			stride_shift;
224 	u8			freed_tables;
225 };
226 
227 void flush_tlb_local(void);
228 void flush_tlb_one_user(unsigned long addr);
229 void flush_tlb_one_kernel(unsigned long addr);
230 void flush_tlb_multi(const struct cpumask *cpumask,
231 		      const struct flush_tlb_info *info);
232 
233 #ifdef CONFIG_PARAVIRT
234 #include <asm/paravirt.h>
235 #endif
236 
237 #define flush_tlb_mm(mm)						\
238 		flush_tlb_mm_range(mm, 0UL, TLB_FLUSH_ALL, 0UL, true)
239 
240 #define flush_tlb_range(vma, start, end)				\
241 	flush_tlb_mm_range((vma)->vm_mm, start, end,			\
242 			   ((vma)->vm_flags & VM_HUGETLB)		\
243 				? huge_page_shift(hstate_vma(vma))	\
244 				: PAGE_SHIFT, false)
245 
246 extern void flush_tlb_all(void);
247 extern void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
248 				unsigned long end, unsigned int stride_shift,
249 				bool freed_tables);
250 extern void flush_tlb_kernel_range(unsigned long start, unsigned long end);
251 
252 static inline void flush_tlb_page(struct vm_area_struct *vma, unsigned long a)
253 {
254 	flush_tlb_mm_range(vma->vm_mm, a, a + PAGE_SIZE, PAGE_SHIFT, false);
255 }
256 
257 static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
258 {
259 	bool should_defer = false;
260 
261 	/* If remote CPUs need to be flushed then defer batch the flush */
262 	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
263 		should_defer = true;
264 	put_cpu();
265 
266 	return should_defer;
267 }
268 
269 static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
270 {
271 	/*
272 	 * Bump the generation count.  This also serves as a full barrier
273 	 * that synchronizes with switch_mm(): callers are required to order
274 	 * their read of mm_cpumask after their writes to the paging
275 	 * structures.
276 	 */
277 	return atomic64_inc_return(&mm->context.tlb_gen);
278 }
279 
280 static inline void arch_tlbbatch_add_pending(struct arch_tlbflush_unmap_batch *batch,
281 					     struct mm_struct *mm,
282 					     unsigned long uaddr)
283 {
284 	inc_mm_tlb_gen(mm);
285 	cpumask_or(&batch->cpumask, &batch->cpumask, mm_cpumask(mm));
286 	mmu_notifier_arch_invalidate_secondary_tlbs(mm, 0, -1UL);
287 }
288 
289 static inline void arch_flush_tlb_batched_pending(struct mm_struct *mm)
290 {
291 	flush_tlb_mm(mm);
292 }
293 
294 extern void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch);
295 
296 static inline bool pte_flags_need_flush(unsigned long oldflags,
297 					unsigned long newflags,
298 					bool ignore_access)
299 {
300 	/*
301 	 * Flags that require a flush when cleared but not when they are set.
302 	 * Only include flags that would not trigger spurious page-faults.
303 	 * Non-present entries are not cached. Hardware would set the
304 	 * dirty/access bit if needed without a fault.
305 	 */
306 	const pteval_t flush_on_clear = _PAGE_DIRTY | _PAGE_PRESENT |
307 					_PAGE_ACCESSED;
308 	const pteval_t software_flags = _PAGE_SOFTW1 | _PAGE_SOFTW2 |
309 					_PAGE_SOFTW3 | _PAGE_SOFTW4 |
310 					_PAGE_SAVED_DIRTY;
311 	const pteval_t flush_on_change = _PAGE_RW | _PAGE_USER | _PAGE_PWT |
312 			  _PAGE_PCD | _PAGE_PSE | _PAGE_GLOBAL | _PAGE_PAT |
313 			  _PAGE_PAT_LARGE | _PAGE_PKEY_BIT0 | _PAGE_PKEY_BIT1 |
314 			  _PAGE_PKEY_BIT2 | _PAGE_PKEY_BIT3 | _PAGE_NX;
315 	unsigned long diff = oldflags ^ newflags;
316 
317 	BUILD_BUG_ON(flush_on_clear & software_flags);
318 	BUILD_BUG_ON(flush_on_clear & flush_on_change);
319 	BUILD_BUG_ON(flush_on_change & software_flags);
320 
321 	/* Ignore software flags */
322 	diff &= ~software_flags;
323 
324 	if (ignore_access)
325 		diff &= ~_PAGE_ACCESSED;
326 
327 	/*
328 	 * Did any of the 'flush_on_clear' flags was clleared set from between
329 	 * 'oldflags' and 'newflags'?
330 	 */
331 	if (diff & oldflags & flush_on_clear)
332 		return true;
333 
334 	/* Flush on modified flags. */
335 	if (diff & flush_on_change)
336 		return true;
337 
338 	/* Ensure there are no flags that were left behind */
339 	if (IS_ENABLED(CONFIG_DEBUG_VM) &&
340 	    (diff & ~(flush_on_clear | software_flags | flush_on_change))) {
341 		VM_WARN_ON_ONCE(1);
342 		return true;
343 	}
344 
345 	return false;
346 }
347 
348 /*
349  * pte_needs_flush() checks whether permissions were demoted and require a
350  * flush. It should only be used for userspace PTEs.
351  */
352 static inline bool pte_needs_flush(pte_t oldpte, pte_t newpte)
353 {
354 	/* !PRESENT -> * ; no need for flush */
355 	if (!(pte_flags(oldpte) & _PAGE_PRESENT))
356 		return false;
357 
358 	/* PFN changed ; needs flush */
359 	if (pte_pfn(oldpte) != pte_pfn(newpte))
360 		return true;
361 
362 	/*
363 	 * check PTE flags; ignore access-bit; see comment in
364 	 * ptep_clear_flush_young().
365 	 */
366 	return pte_flags_need_flush(pte_flags(oldpte), pte_flags(newpte),
367 				    true);
368 }
369 #define pte_needs_flush pte_needs_flush
370 
371 /*
372  * huge_pmd_needs_flush() checks whether permissions were demoted and require a
373  * flush. It should only be used for userspace huge PMDs.
374  */
375 static inline bool huge_pmd_needs_flush(pmd_t oldpmd, pmd_t newpmd)
376 {
377 	/* !PRESENT -> * ; no need for flush */
378 	if (!(pmd_flags(oldpmd) & _PAGE_PRESENT))
379 		return false;
380 
381 	/* PFN changed ; needs flush */
382 	if (pmd_pfn(oldpmd) != pmd_pfn(newpmd))
383 		return true;
384 
385 	/*
386 	 * check PMD flags; do not ignore access-bit; see
387 	 * pmdp_clear_flush_young().
388 	 */
389 	return pte_flags_need_flush(pmd_flags(oldpmd), pmd_flags(newpmd),
390 				    false);
391 }
392 #define huge_pmd_needs_flush huge_pmd_needs_flush
393 
394 #ifdef CONFIG_ADDRESS_MASKING
395 static inline  u64 tlbstate_lam_cr3_mask(void)
396 {
397 	u64 lam = this_cpu_read(cpu_tlbstate.lam);
398 
399 	return lam << X86_CR3_LAM_U57_BIT;
400 }
401 
402 static inline void set_tlbstate_lam_mode(struct mm_struct *mm)
403 {
404 	this_cpu_write(cpu_tlbstate.lam,
405 		       mm->context.lam_cr3_mask >> X86_CR3_LAM_U57_BIT);
406 	this_cpu_write(tlbstate_untag_mask, mm->context.untag_mask);
407 }
408 
409 #else
410 
411 static inline u64 tlbstate_lam_cr3_mask(void)
412 {
413 	return 0;
414 }
415 
416 static inline void set_tlbstate_lam_mode(struct mm_struct *mm)
417 {
418 }
419 #endif
420 #endif /* !MODULE */
421 
422 static inline void __native_tlb_flush_global(unsigned long cr4)
423 {
424 	native_write_cr4(cr4 ^ X86_CR4_PGE);
425 	native_write_cr4(cr4);
426 }
427 #endif /* _ASM_X86_TLBFLUSH_H */
428