xref: /linux/arch/arm64/include/asm/tlbflush.h (revision 55d0969c451159cff86949b38c39171cab962069)
1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3  * Based on arch/arm/include/asm/tlbflush.h
4  *
5  * Copyright (C) 1999-2003 Russell King
6  * Copyright (C) 2012 ARM Ltd.
7  */
8 #ifndef __ASM_TLBFLUSH_H
9 #define __ASM_TLBFLUSH_H
10 
11 #ifndef __ASSEMBLY__
12 
13 #include <linux/bitfield.h>
14 #include <linux/mm_types.h>
15 #include <linux/sched.h>
16 #include <linux/mmu_notifier.h>
17 #include <asm/cputype.h>
18 #include <asm/mmu.h>
19 
20 /*
21  * Raw TLBI operations.
22  *
23  * Where necessary, use the __tlbi() macro to avoid asm()
24  * boilerplate. Drivers and most kernel code should use the TLB
25  * management routines in preference to the macro below.
26  *
27  * The macro can be used as __tlbi(op) or __tlbi(op, arg), depending
28  * on whether a particular TLBI operation takes an argument or
29  * not. The macros handles invoking the asm with or without the
30  * register argument as appropriate.
31  */
32 #define __TLBI_0(op, arg) asm (ARM64_ASM_PREAMBLE			       \
33 			       "tlbi " #op "\n"				       \
34 		   ALTERNATIVE("nop\n			nop",		       \
35 			       "dsb ish\n		tlbi " #op,	       \
36 			       ARM64_WORKAROUND_REPEAT_TLBI,		       \
37 			       CONFIG_ARM64_WORKAROUND_REPEAT_TLBI)	       \
38 			    : : )
39 
40 #define __TLBI_1(op, arg) asm (ARM64_ASM_PREAMBLE			       \
41 			       "tlbi " #op ", %0\n"			       \
42 		   ALTERNATIVE("nop\n			nop",		       \
43 			       "dsb ish\n		tlbi " #op ", %0",     \
44 			       ARM64_WORKAROUND_REPEAT_TLBI,		       \
45 			       CONFIG_ARM64_WORKAROUND_REPEAT_TLBI)	       \
46 			    : : "r" (arg))
47 
48 #define __TLBI_N(op, arg, n, ...) __TLBI_##n(op, arg)
49 
50 #define __tlbi(op, ...)		__TLBI_N(op, ##__VA_ARGS__, 1, 0)
51 
52 #define __tlbi_user(op, arg) do {						\
53 	if (arm64_kernel_unmapped_at_el0())					\
54 		__tlbi(op, (arg) | USER_ASID_FLAG);				\
55 } while (0)
56 
57 /* This macro creates a properly formatted VA operand for the TLBI */
58 #define __TLBI_VADDR(addr, asid)				\
59 	({							\
60 		unsigned long __ta = (addr) >> 12;		\
61 		__ta &= GENMASK_ULL(43, 0);			\
62 		__ta |= (unsigned long)(asid) << 48;		\
63 		__ta;						\
64 	})
65 
66 /*
67  * Get translation granule of the system, which is decided by
68  * PAGE_SIZE.  Used by TTL.
69  *  - 4KB	: 1
70  *  - 16KB	: 2
71  *  - 64KB	: 3
72  */
73 #define TLBI_TTL_TG_4K		1
74 #define TLBI_TTL_TG_16K		2
75 #define TLBI_TTL_TG_64K		3
76 
77 static inline unsigned long get_trans_granule(void)
78 {
79 	switch (PAGE_SIZE) {
80 	case SZ_4K:
81 		return TLBI_TTL_TG_4K;
82 	case SZ_16K:
83 		return TLBI_TTL_TG_16K;
84 	case SZ_64K:
85 		return TLBI_TTL_TG_64K;
86 	default:
87 		return 0;
88 	}
89 }
90 
91 /*
92  * Level-based TLBI operations.
93  *
94  * When ARMv8.4-TTL exists, TLBI operations take an additional hint for
95  * the level at which the invalidation must take place. If the level is
96  * wrong, no invalidation may take place. In the case where the level
97  * cannot be easily determined, the value TLBI_TTL_UNKNOWN will perform
98  * a non-hinted invalidation. Any provided level outside the hint range
99  * will also cause fall-back to non-hinted invalidation.
100  *
101  * For Stage-2 invalidation, use the level values provided to that effect
102  * in asm/stage2_pgtable.h.
103  */
104 #define TLBI_TTL_MASK		GENMASK_ULL(47, 44)
105 
106 #define TLBI_TTL_UNKNOWN	INT_MAX
107 
108 #define __tlbi_level(op, addr, level) do {				\
109 	u64 arg = addr;							\
110 									\
111 	if (alternative_has_cap_unlikely(ARM64_HAS_ARMv8_4_TTL) &&	\
112 	    level >= 0 && level <= 3) {					\
113 		u64 ttl = level & 3;					\
114 		ttl |= get_trans_granule() << 2;			\
115 		arg &= ~TLBI_TTL_MASK;					\
116 		arg |= FIELD_PREP(TLBI_TTL_MASK, ttl);			\
117 	}								\
118 									\
119 	__tlbi(op, arg);						\
120 } while(0)
121 
122 #define __tlbi_user_level(op, arg, level) do {				\
123 	if (arm64_kernel_unmapped_at_el0())				\
124 		__tlbi_level(op, (arg | USER_ASID_FLAG), level);	\
125 } while (0)
126 
127 /*
128  * This macro creates a properly formatted VA operand for the TLB RANGE. The
129  * value bit assignments are:
130  *
131  * +----------+------+-------+-------+-------+----------------------+
132  * |   ASID   |  TG  | SCALE |  NUM  |  TTL  |        BADDR         |
133  * +-----------------+-------+-------+-------+----------------------+
134  * |63      48|47  46|45   44|43   39|38   37|36                   0|
135  *
136  * The address range is determined by below formula: [BADDR, BADDR + (NUM + 1) *
137  * 2^(5*SCALE + 1) * PAGESIZE)
138  *
139  * Note that the first argument, baddr, is pre-shifted; If LPA2 is in use, BADDR
140  * holds addr[52:16]. Else BADDR holds page number. See for example ARM DDI
141  * 0487J.a section C5.5.60 "TLBI VAE1IS, TLBI VAE1ISNXS, TLB Invalidate by VA,
142  * EL1, Inner Shareable".
143  *
144  */
145 #define TLBIR_ASID_MASK		GENMASK_ULL(63, 48)
146 #define TLBIR_TG_MASK		GENMASK_ULL(47, 46)
147 #define TLBIR_SCALE_MASK	GENMASK_ULL(45, 44)
148 #define TLBIR_NUM_MASK		GENMASK_ULL(43, 39)
149 #define TLBIR_TTL_MASK		GENMASK_ULL(38, 37)
150 #define TLBIR_BADDR_MASK	GENMASK_ULL(36,  0)
151 
152 #define __TLBI_VADDR_RANGE(baddr, asid, scale, num, ttl)		\
153 	({								\
154 		unsigned long __ta = 0;					\
155 		unsigned long __ttl = (ttl >= 1 && ttl <= 3) ? ttl : 0;	\
156 		__ta |= FIELD_PREP(TLBIR_BADDR_MASK, baddr);		\
157 		__ta |= FIELD_PREP(TLBIR_TTL_MASK, __ttl);		\
158 		__ta |= FIELD_PREP(TLBIR_NUM_MASK, num);		\
159 		__ta |= FIELD_PREP(TLBIR_SCALE_MASK, scale);		\
160 		__ta |= FIELD_PREP(TLBIR_TG_MASK, get_trans_granule());	\
161 		__ta |= FIELD_PREP(TLBIR_ASID_MASK, asid);		\
162 		__ta;							\
163 	})
164 
165 /* These macros are used by the TLBI RANGE feature. */
166 #define __TLBI_RANGE_PAGES(num, scale)	\
167 	((unsigned long)((num) + 1) << (5 * (scale) + 1))
168 #define MAX_TLBI_RANGE_PAGES		__TLBI_RANGE_PAGES(31, 3)
169 
170 /*
171  * Generate 'num' values from -1 to 31 with -1 rejected by the
172  * __flush_tlb_range() loop below. Its return value is only
173  * significant for a maximum of MAX_TLBI_RANGE_PAGES pages. If
174  * 'pages' is more than that, you must iterate over the overall
175  * range.
176  */
177 #define __TLBI_RANGE_NUM(pages, scale)					\
178 	({								\
179 		int __pages = min((pages),				\
180 				  __TLBI_RANGE_PAGES(31, (scale)));	\
181 		(__pages >> (5 * (scale) + 1)) - 1;			\
182 	})
183 
184 /*
185  *	TLB Invalidation
186  *	================
187  *
188  * 	This header file implements the low-level TLB invalidation routines
189  *	(sometimes referred to as "flushing" in the kernel) for arm64.
190  *
191  *	Every invalidation operation uses the following template:
192  *
193  *	DSB ISHST	// Ensure prior page-table updates have completed
194  *	TLBI ...	// Invalidate the TLB
195  *	DSB ISH		// Ensure the TLB invalidation has completed
196  *      if (invalidated kernel mappings)
197  *		ISB	// Discard any instructions fetched from the old mapping
198  *
199  *
200  *	The following functions form part of the "core" TLB invalidation API,
201  *	as documented in Documentation/core-api/cachetlb.rst:
202  *
203  *	flush_tlb_all()
204  *		Invalidate the entire TLB (kernel + user) on all CPUs
205  *
206  *	flush_tlb_mm(mm)
207  *		Invalidate an entire user address space on all CPUs.
208  *		The 'mm' argument identifies the ASID to invalidate.
209  *
210  *	flush_tlb_range(vma, start, end)
211  *		Invalidate the virtual-address range '[start, end)' on all
212  *		CPUs for the user address space corresponding to 'vma->mm'.
213  *		Note that this operation also invalidates any walk-cache
214  *		entries associated with translations for the specified address
215  *		range.
216  *
217  *	flush_tlb_kernel_range(start, end)
218  *		Same as flush_tlb_range(..., start, end), but applies to
219  * 		kernel mappings rather than a particular user address space.
220  *		Whilst not explicitly documented, this function is used when
221  *		unmapping pages from vmalloc/io space.
222  *
223  *	flush_tlb_page(vma, addr)
224  *		Invalidate a single user mapping for address 'addr' in the
225  *		address space corresponding to 'vma->mm'.  Note that this
226  *		operation only invalidates a single, last-level page-table
227  *		entry and therefore does not affect any walk-caches.
228  *
229  *
230  *	Next, we have some undocumented invalidation routines that you probably
231  *	don't want to call unless you know what you're doing:
232  *
233  *	local_flush_tlb_all()
234  *		Same as flush_tlb_all(), but only applies to the calling CPU.
235  *
236  *	__flush_tlb_kernel_pgtable(addr)
237  *		Invalidate a single kernel mapping for address 'addr' on all
238  *		CPUs, ensuring that any walk-cache entries associated with the
239  *		translation are also invalidated.
240  *
241  *	__flush_tlb_range(vma, start, end, stride, last_level, tlb_level)
242  *		Invalidate the virtual-address range '[start, end)' on all
243  *		CPUs for the user address space corresponding to 'vma->mm'.
244  *		The invalidation operations are issued at a granularity
245  *		determined by 'stride' and only affect any walk-cache entries
246  *		if 'last_level' is equal to false. tlb_level is the level at
247  *		which the invalidation must take place. If the level is wrong,
248  *		no invalidation may take place. In the case where the level
249  *		cannot be easily determined, the value TLBI_TTL_UNKNOWN will
250  *		perform a non-hinted invalidation.
251  *
252  *
253  *	Finally, take a look at asm/tlb.h to see how tlb_flush() is implemented
254  *	on top of these routines, since that is our interface to the mmu_gather
255  *	API as used by munmap() and friends.
256  */
257 static inline void local_flush_tlb_all(void)
258 {
259 	dsb(nshst);
260 	__tlbi(vmalle1);
261 	dsb(nsh);
262 	isb();
263 }
264 
265 static inline void flush_tlb_all(void)
266 {
267 	dsb(ishst);
268 	__tlbi(vmalle1is);
269 	dsb(ish);
270 	isb();
271 }
272 
273 static inline void flush_tlb_mm(struct mm_struct *mm)
274 {
275 	unsigned long asid;
276 
277 	dsb(ishst);
278 	asid = __TLBI_VADDR(0, ASID(mm));
279 	__tlbi(aside1is, asid);
280 	__tlbi_user(aside1is, asid);
281 	dsb(ish);
282 	mmu_notifier_arch_invalidate_secondary_tlbs(mm, 0, -1UL);
283 }
284 
285 static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
286 					   unsigned long uaddr)
287 {
288 	unsigned long addr;
289 
290 	dsb(ishst);
291 	addr = __TLBI_VADDR(uaddr, ASID(mm));
292 	__tlbi(vale1is, addr);
293 	__tlbi_user(vale1is, addr);
294 	mmu_notifier_arch_invalidate_secondary_tlbs(mm, uaddr & PAGE_MASK,
295 						(uaddr & PAGE_MASK) + PAGE_SIZE);
296 }
297 
298 static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
299 					 unsigned long uaddr)
300 {
301 	return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
302 }
303 
304 static inline void flush_tlb_page(struct vm_area_struct *vma,
305 				  unsigned long uaddr)
306 {
307 	flush_tlb_page_nosync(vma, uaddr);
308 	dsb(ish);
309 }
310 
311 static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
312 {
313 	/*
314 	 * TLB flush deferral is not required on systems which are affected by
315 	 * ARM64_WORKAROUND_REPEAT_TLBI, as __tlbi()/__tlbi_user() implementation
316 	 * will have two consecutive TLBI instructions with a dsb(ish) in between
317 	 * defeating the purpose (i.e save overall 'dsb ish' cost).
318 	 */
319 	if (alternative_has_cap_unlikely(ARM64_WORKAROUND_REPEAT_TLBI))
320 		return false;
321 
322 	return true;
323 }
324 
325 static inline void arch_tlbbatch_add_pending(struct arch_tlbflush_unmap_batch *batch,
326 					     struct mm_struct *mm,
327 					     unsigned long uaddr)
328 {
329 	__flush_tlb_page_nosync(mm, uaddr);
330 }
331 
332 /*
333  * If mprotect/munmap/etc occurs during TLB batched flushing, we need to
334  * synchronise all the TLBI issued with a DSB to avoid the race mentioned in
335  * flush_tlb_batched_pending().
336  */
337 static inline void arch_flush_tlb_batched_pending(struct mm_struct *mm)
338 {
339 	dsb(ish);
340 }
341 
342 /*
343  * To support TLB batched flush for multiple pages unmapping, we only send
344  * the TLBI for each page in arch_tlbbatch_add_pending() and wait for the
345  * completion at the end in arch_tlbbatch_flush(). Since we've already issued
346  * TLBI for each page so only a DSB is needed to synchronise its effect on the
347  * other CPUs.
348  *
349  * This will save the time waiting on DSB comparing issuing a TLBI;DSB sequence
350  * for each page.
351  */
352 static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
353 {
354 	dsb(ish);
355 }
356 
357 /*
358  * This is meant to avoid soft lock-ups on large TLB flushing ranges and not
359  * necessarily a performance improvement.
360  */
361 #define MAX_DVM_OPS	PTRS_PER_PTE
362 
363 /*
364  * __flush_tlb_range_op - Perform TLBI operation upon a range
365  *
366  * @op:	TLBI instruction that operates on a range (has 'r' prefix)
367  * @start:	The start address of the range
368  * @pages:	Range as the number of pages from 'start'
369  * @stride:	Flush granularity
370  * @asid:	The ASID of the task (0 for IPA instructions)
371  * @tlb_level:	Translation Table level hint, if known
372  * @tlbi_user:	If 'true', call an additional __tlbi_user()
373  *              (typically for user ASIDs). 'flase' for IPA instructions
374  * @lpa2:	If 'true', the lpa2 scheme is used as set out below
375  *
376  * When the CPU does not support TLB range operations, flush the TLB
377  * entries one by one at the granularity of 'stride'. If the TLB
378  * range ops are supported, then:
379  *
380  * 1. If FEAT_LPA2 is in use, the start address of a range operation must be
381  *    64KB aligned, so flush pages one by one until the alignment is reached
382  *    using the non-range operations. This step is skipped if LPA2 is not in
383  *    use.
384  *
385  * 2. The minimum range granularity is decided by 'scale', so multiple range
386  *    TLBI operations may be required. Start from scale = 3, flush the largest
387  *    possible number of pages ((num+1)*2^(5*scale+1)) that fit into the
388  *    requested range, then decrement scale and continue until one or zero pages
389  *    are left. We must start from highest scale to ensure 64KB start alignment
390  *    is maintained in the LPA2 case.
391  *
392  * 3. If there is 1 page remaining, flush it through non-range operations. Range
393  *    operations can only span an even number of pages. We save this for last to
394  *    ensure 64KB start alignment is maintained for the LPA2 case.
395  */
396 #define __flush_tlb_range_op(op, start, pages, stride,			\
397 				asid, tlb_level, tlbi_user, lpa2)	\
398 do {									\
399 	int num = 0;							\
400 	int scale = 3;							\
401 	int shift = lpa2 ? 16 : PAGE_SHIFT;				\
402 	unsigned long addr;						\
403 									\
404 	while (pages > 0) {						\
405 		if (!system_supports_tlb_range() ||			\
406 		    pages == 1 ||					\
407 		    (lpa2 && start != ALIGN(start, SZ_64K))) {		\
408 			addr = __TLBI_VADDR(start, asid);		\
409 			__tlbi_level(op, addr, tlb_level);		\
410 			if (tlbi_user)					\
411 				__tlbi_user_level(op, addr, tlb_level);	\
412 			start += stride;				\
413 			pages -= stride >> PAGE_SHIFT;			\
414 			continue;					\
415 		}							\
416 									\
417 		num = __TLBI_RANGE_NUM(pages, scale);			\
418 		if (num >= 0) {						\
419 			addr = __TLBI_VADDR_RANGE(start >> shift, asid, \
420 						scale, num, tlb_level);	\
421 			__tlbi(r##op, addr);				\
422 			if (tlbi_user)					\
423 				__tlbi_user(r##op, addr);		\
424 			start += __TLBI_RANGE_PAGES(num, scale) << PAGE_SHIFT; \
425 			pages -= __TLBI_RANGE_PAGES(num, scale);	\
426 		}							\
427 		scale--;						\
428 	}								\
429 } while (0)
430 
431 #define __flush_s2_tlb_range_op(op, start, pages, stride, tlb_level) \
432 	__flush_tlb_range_op(op, start, pages, stride, 0, tlb_level, false, kvm_lpa2_is_enabled());
433 
434 static inline void __flush_tlb_range_nosync(struct vm_area_struct *vma,
435 				     unsigned long start, unsigned long end,
436 				     unsigned long stride, bool last_level,
437 				     int tlb_level)
438 {
439 	unsigned long asid, pages;
440 
441 	start = round_down(start, stride);
442 	end = round_up(end, stride);
443 	pages = (end - start) >> PAGE_SHIFT;
444 
445 	/*
446 	 * When not uses TLB range ops, we can handle up to
447 	 * (MAX_DVM_OPS - 1) pages;
448 	 * When uses TLB range ops, we can handle up to
449 	 * MAX_TLBI_RANGE_PAGES pages.
450 	 */
451 	if ((!system_supports_tlb_range() &&
452 	     (end - start) >= (MAX_DVM_OPS * stride)) ||
453 	    pages > MAX_TLBI_RANGE_PAGES) {
454 		flush_tlb_mm(vma->vm_mm);
455 		return;
456 	}
457 
458 	dsb(ishst);
459 	asid = ASID(vma->vm_mm);
460 
461 	if (last_level)
462 		__flush_tlb_range_op(vale1is, start, pages, stride, asid,
463 				     tlb_level, true, lpa2_is_enabled());
464 	else
465 		__flush_tlb_range_op(vae1is, start, pages, stride, asid,
466 				     tlb_level, true, lpa2_is_enabled());
467 
468 	mmu_notifier_arch_invalidate_secondary_tlbs(vma->vm_mm, start, end);
469 }
470 
471 static inline void __flush_tlb_range(struct vm_area_struct *vma,
472 				     unsigned long start, unsigned long end,
473 				     unsigned long stride, bool last_level,
474 				     int tlb_level)
475 {
476 	__flush_tlb_range_nosync(vma, start, end, stride,
477 				 last_level, tlb_level);
478 	dsb(ish);
479 }
480 
481 static inline void flush_tlb_range(struct vm_area_struct *vma,
482 				   unsigned long start, unsigned long end)
483 {
484 	/*
485 	 * We cannot use leaf-only invalidation here, since we may be invalidating
486 	 * table entries as part of collapsing hugepages or moving page tables.
487 	 * Set the tlb_level to TLBI_TTL_UNKNOWN because we can not get enough
488 	 * information here.
489 	 */
490 	__flush_tlb_range(vma, start, end, PAGE_SIZE, false, TLBI_TTL_UNKNOWN);
491 }
492 
493 static inline void flush_tlb_kernel_range(unsigned long start, unsigned long end)
494 {
495 	unsigned long addr;
496 
497 	if ((end - start) > (MAX_DVM_OPS * PAGE_SIZE)) {
498 		flush_tlb_all();
499 		return;
500 	}
501 
502 	start = __TLBI_VADDR(start, 0);
503 	end = __TLBI_VADDR(end, 0);
504 
505 	dsb(ishst);
506 	for (addr = start; addr < end; addr += 1 << (PAGE_SHIFT - 12))
507 		__tlbi(vaale1is, addr);
508 	dsb(ish);
509 	isb();
510 }
511 
512 /*
513  * Used to invalidate the TLB (walk caches) corresponding to intermediate page
514  * table levels (pgd/pud/pmd).
515  */
516 static inline void __flush_tlb_kernel_pgtable(unsigned long kaddr)
517 {
518 	unsigned long addr = __TLBI_VADDR(kaddr, 0);
519 
520 	dsb(ishst);
521 	__tlbi(vaae1is, addr);
522 	dsb(ish);
523 	isb();
524 }
525 #endif
526 
527 #endif
528