xref: /linux/arch/arm64/include/asm/tlbflush.h (revision 173b0b5b0e865348684c02bd9cb1d22b5d46e458)
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 __TLBI_VADDR_RANGE(baddr, asid, scale, num, ttl)			\
146 	({									\
147 		unsigned long __ta = (baddr);					\
148 		unsigned long __ttl = (ttl >= 1 && ttl <= 3) ? ttl : 0;		\
149 		__ta &= GENMASK_ULL(36, 0);					\
150 		__ta |= __ttl << 37;						\
151 		__ta |= (unsigned long)(num) << 39;				\
152 		__ta |= (unsigned long)(scale) << 44;				\
153 		__ta |= get_trans_granule() << 46;				\
154 		__ta |= (unsigned long)(asid) << 48;				\
155 		__ta;								\
156 	})
157 
158 /* These macros are used by the TLBI RANGE feature. */
159 #define __TLBI_RANGE_PAGES(num, scale)	\
160 	((unsigned long)((num) + 1) << (5 * (scale) + 1))
161 #define MAX_TLBI_RANGE_PAGES		__TLBI_RANGE_PAGES(31, 3)
162 
163 /*
164  * Generate 'num' values from -1 to 31 with -1 rejected by the
165  * __flush_tlb_range() loop below. Its return value is only
166  * significant for a maximum of MAX_TLBI_RANGE_PAGES pages. If
167  * 'pages' is more than that, you must iterate over the overall
168  * range.
169  */
170 #define __TLBI_RANGE_NUM(pages, scale)					\
171 	({								\
172 		int __pages = min((pages),				\
173 				  __TLBI_RANGE_PAGES(31, (scale)));	\
174 		(__pages >> (5 * (scale) + 1)) - 1;			\
175 	})
176 
177 /*
178  *	TLB Invalidation
179  *	================
180  *
181  * 	This header file implements the low-level TLB invalidation routines
182  *	(sometimes referred to as "flushing" in the kernel) for arm64.
183  *
184  *	Every invalidation operation uses the following template:
185  *
186  *	DSB ISHST	// Ensure prior page-table updates have completed
187  *	TLBI ...	// Invalidate the TLB
188  *	DSB ISH		// Ensure the TLB invalidation has completed
189  *      if (invalidated kernel mappings)
190  *		ISB	// Discard any instructions fetched from the old mapping
191  *
192  *
193  *	The following functions form part of the "core" TLB invalidation API,
194  *	as documented in Documentation/core-api/cachetlb.rst:
195  *
196  *	flush_tlb_all()
197  *		Invalidate the entire TLB (kernel + user) on all CPUs
198  *
199  *	flush_tlb_mm(mm)
200  *		Invalidate an entire user address space on all CPUs.
201  *		The 'mm' argument identifies the ASID to invalidate.
202  *
203  *	flush_tlb_range(vma, start, end)
204  *		Invalidate the virtual-address range '[start, end)' on all
205  *		CPUs for the user address space corresponding to 'vma->mm'.
206  *		Note that this operation also invalidates any walk-cache
207  *		entries associated with translations for the specified address
208  *		range.
209  *
210  *	flush_tlb_kernel_range(start, end)
211  *		Same as flush_tlb_range(..., start, end), but applies to
212  * 		kernel mappings rather than a particular user address space.
213  *		Whilst not explicitly documented, this function is used when
214  *		unmapping pages from vmalloc/io space.
215  *
216  *	flush_tlb_page(vma, addr)
217  *		Invalidate a single user mapping for address 'addr' in the
218  *		address space corresponding to 'vma->mm'.  Note that this
219  *		operation only invalidates a single, last-level page-table
220  *		entry and therefore does not affect any walk-caches.
221  *
222  *
223  *	Next, we have some undocumented invalidation routines that you probably
224  *	don't want to call unless you know what you're doing:
225  *
226  *	local_flush_tlb_all()
227  *		Same as flush_tlb_all(), but only applies to the calling CPU.
228  *
229  *	__flush_tlb_kernel_pgtable(addr)
230  *		Invalidate a single kernel mapping for address 'addr' on all
231  *		CPUs, ensuring that any walk-cache entries associated with the
232  *		translation are also invalidated.
233  *
234  *	__flush_tlb_range(vma, start, end, stride, last_level, tlb_level)
235  *		Invalidate the virtual-address range '[start, end)' on all
236  *		CPUs for the user address space corresponding to 'vma->mm'.
237  *		The invalidation operations are issued at a granularity
238  *		determined by 'stride' and only affect any walk-cache entries
239  *		if 'last_level' is equal to false. tlb_level is the level at
240  *		which the invalidation must take place. If the level is wrong,
241  *		no invalidation may take place. In the case where the level
242  *		cannot be easily determined, the value TLBI_TTL_UNKNOWN will
243  *		perform a non-hinted invalidation.
244  *
245  *
246  *	Finally, take a look at asm/tlb.h to see how tlb_flush() is implemented
247  *	on top of these routines, since that is our interface to the mmu_gather
248  *	API as used by munmap() and friends.
249  */
250 static inline void local_flush_tlb_all(void)
251 {
252 	dsb(nshst);
253 	__tlbi(vmalle1);
254 	dsb(nsh);
255 	isb();
256 }
257 
258 static inline void flush_tlb_all(void)
259 {
260 	dsb(ishst);
261 	__tlbi(vmalle1is);
262 	dsb(ish);
263 	isb();
264 }
265 
266 static inline void flush_tlb_mm(struct mm_struct *mm)
267 {
268 	unsigned long asid;
269 
270 	dsb(ishst);
271 	asid = __TLBI_VADDR(0, ASID(mm));
272 	__tlbi(aside1is, asid);
273 	__tlbi_user(aside1is, asid);
274 	dsb(ish);
275 	mmu_notifier_arch_invalidate_secondary_tlbs(mm, 0, -1UL);
276 }
277 
278 static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
279 					   unsigned long uaddr)
280 {
281 	unsigned long addr;
282 
283 	dsb(ishst);
284 	addr = __TLBI_VADDR(uaddr, ASID(mm));
285 	__tlbi(vale1is, addr);
286 	__tlbi_user(vale1is, addr);
287 	mmu_notifier_arch_invalidate_secondary_tlbs(mm, uaddr & PAGE_MASK,
288 						(uaddr & PAGE_MASK) + PAGE_SIZE);
289 }
290 
291 static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
292 					 unsigned long uaddr)
293 {
294 	return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
295 }
296 
297 static inline void flush_tlb_page(struct vm_area_struct *vma,
298 				  unsigned long uaddr)
299 {
300 	flush_tlb_page_nosync(vma, uaddr);
301 	dsb(ish);
302 }
303 
304 static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
305 {
306 	/*
307 	 * TLB flush deferral is not required on systems which are affected by
308 	 * ARM64_WORKAROUND_REPEAT_TLBI, as __tlbi()/__tlbi_user() implementation
309 	 * will have two consecutive TLBI instructions with a dsb(ish) in between
310 	 * defeating the purpose (i.e save overall 'dsb ish' cost).
311 	 */
312 	if (alternative_has_cap_unlikely(ARM64_WORKAROUND_REPEAT_TLBI))
313 		return false;
314 
315 	return true;
316 }
317 
318 static inline void arch_tlbbatch_add_pending(struct arch_tlbflush_unmap_batch *batch,
319 					     struct mm_struct *mm,
320 					     unsigned long uaddr)
321 {
322 	__flush_tlb_page_nosync(mm, uaddr);
323 }
324 
325 /*
326  * If mprotect/munmap/etc occurs during TLB batched flushing, we need to
327  * synchronise all the TLBI issued with a DSB to avoid the race mentioned in
328  * flush_tlb_batched_pending().
329  */
330 static inline void arch_flush_tlb_batched_pending(struct mm_struct *mm)
331 {
332 	dsb(ish);
333 }
334 
335 /*
336  * To support TLB batched flush for multiple pages unmapping, we only send
337  * the TLBI for each page in arch_tlbbatch_add_pending() and wait for the
338  * completion at the end in arch_tlbbatch_flush(). Since we've already issued
339  * TLBI for each page so only a DSB is needed to synchronise its effect on the
340  * other CPUs.
341  *
342  * This will save the time waiting on DSB comparing issuing a TLBI;DSB sequence
343  * for each page.
344  */
345 static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
346 {
347 	dsb(ish);
348 }
349 
350 /*
351  * This is meant to avoid soft lock-ups on large TLB flushing ranges and not
352  * necessarily a performance improvement.
353  */
354 #define MAX_DVM_OPS	PTRS_PER_PTE
355 
356 /*
357  * __flush_tlb_range_op - Perform TLBI operation upon a range
358  *
359  * @op:	TLBI instruction that operates on a range (has 'r' prefix)
360  * @start:	The start address of the range
361  * @pages:	Range as the number of pages from 'start'
362  * @stride:	Flush granularity
363  * @asid:	The ASID of the task (0 for IPA instructions)
364  * @tlb_level:	Translation Table level hint, if known
365  * @tlbi_user:	If 'true', call an additional __tlbi_user()
366  *              (typically for user ASIDs). 'flase' for IPA instructions
367  * @lpa2:	If 'true', the lpa2 scheme is used as set out below
368  *
369  * When the CPU does not support TLB range operations, flush the TLB
370  * entries one by one at the granularity of 'stride'. If the TLB
371  * range ops are supported, then:
372  *
373  * 1. If FEAT_LPA2 is in use, the start address of a range operation must be
374  *    64KB aligned, so flush pages one by one until the alignment is reached
375  *    using the non-range operations. This step is skipped if LPA2 is not in
376  *    use.
377  *
378  * 2. The minimum range granularity is decided by 'scale', so multiple range
379  *    TLBI operations may be required. Start from scale = 3, flush the largest
380  *    possible number of pages ((num+1)*2^(5*scale+1)) that fit into the
381  *    requested range, then decrement scale and continue until one or zero pages
382  *    are left. We must start from highest scale to ensure 64KB start alignment
383  *    is maintained in the LPA2 case.
384  *
385  * 3. If there is 1 page remaining, flush it through non-range operations. Range
386  *    operations can only span an even number of pages. We save this for last to
387  *    ensure 64KB start alignment is maintained for the LPA2 case.
388  */
389 #define __flush_tlb_range_op(op, start, pages, stride,			\
390 				asid, tlb_level, tlbi_user, lpa2)	\
391 do {									\
392 	int num = 0;							\
393 	int scale = 3;							\
394 	int shift = lpa2 ? 16 : PAGE_SHIFT;				\
395 	unsigned long addr;						\
396 									\
397 	while (pages > 0) {						\
398 		if (!system_supports_tlb_range() ||			\
399 		    pages == 1 ||					\
400 		    (lpa2 && start != ALIGN(start, SZ_64K))) {		\
401 			addr = __TLBI_VADDR(start, asid);		\
402 			__tlbi_level(op, addr, tlb_level);		\
403 			if (tlbi_user)					\
404 				__tlbi_user_level(op, addr, tlb_level);	\
405 			start += stride;				\
406 			pages -= stride >> PAGE_SHIFT;			\
407 			continue;					\
408 		}							\
409 									\
410 		num = __TLBI_RANGE_NUM(pages, scale);			\
411 		if (num >= 0) {						\
412 			addr = __TLBI_VADDR_RANGE(start >> shift, asid, \
413 						scale, num, tlb_level);	\
414 			__tlbi(r##op, addr);				\
415 			if (tlbi_user)					\
416 				__tlbi_user(r##op, addr);		\
417 			start += __TLBI_RANGE_PAGES(num, scale) << PAGE_SHIFT; \
418 			pages -= __TLBI_RANGE_PAGES(num, scale);	\
419 		}							\
420 		scale--;						\
421 	}								\
422 } while (0)
423 
424 #define __flush_s2_tlb_range_op(op, start, pages, stride, tlb_level) \
425 	__flush_tlb_range_op(op, start, pages, stride, 0, tlb_level, false, kvm_lpa2_is_enabled());
426 
427 static inline void __flush_tlb_range_nosync(struct vm_area_struct *vma,
428 				     unsigned long start, unsigned long end,
429 				     unsigned long stride, bool last_level,
430 				     int tlb_level)
431 {
432 	unsigned long asid, pages;
433 
434 	start = round_down(start, stride);
435 	end = round_up(end, stride);
436 	pages = (end - start) >> PAGE_SHIFT;
437 
438 	/*
439 	 * When not uses TLB range ops, we can handle up to
440 	 * (MAX_DVM_OPS - 1) pages;
441 	 * When uses TLB range ops, we can handle up to
442 	 * (MAX_TLBI_RANGE_PAGES - 1) pages.
443 	 */
444 	if ((!system_supports_tlb_range() &&
445 	     (end - start) >= (MAX_DVM_OPS * stride)) ||
446 	    pages >= MAX_TLBI_RANGE_PAGES) {
447 		flush_tlb_mm(vma->vm_mm);
448 		return;
449 	}
450 
451 	dsb(ishst);
452 	asid = ASID(vma->vm_mm);
453 
454 	if (last_level)
455 		__flush_tlb_range_op(vale1is, start, pages, stride, asid,
456 				     tlb_level, true, lpa2_is_enabled());
457 	else
458 		__flush_tlb_range_op(vae1is, start, pages, stride, asid,
459 				     tlb_level, true, lpa2_is_enabled());
460 
461 	mmu_notifier_arch_invalidate_secondary_tlbs(vma->vm_mm, start, end);
462 }
463 
464 static inline void __flush_tlb_range(struct vm_area_struct *vma,
465 				     unsigned long start, unsigned long end,
466 				     unsigned long stride, bool last_level,
467 				     int tlb_level)
468 {
469 	__flush_tlb_range_nosync(vma, start, end, stride,
470 				 last_level, tlb_level);
471 	dsb(ish);
472 }
473 
474 static inline void flush_tlb_range(struct vm_area_struct *vma,
475 				   unsigned long start, unsigned long end)
476 {
477 	/*
478 	 * We cannot use leaf-only invalidation here, since we may be invalidating
479 	 * table entries as part of collapsing hugepages or moving page tables.
480 	 * Set the tlb_level to TLBI_TTL_UNKNOWN because we can not get enough
481 	 * information here.
482 	 */
483 	__flush_tlb_range(vma, start, end, PAGE_SIZE, false, TLBI_TTL_UNKNOWN);
484 }
485 
486 static inline void flush_tlb_kernel_range(unsigned long start, unsigned long end)
487 {
488 	unsigned long addr;
489 
490 	if ((end - start) > (MAX_DVM_OPS * PAGE_SIZE)) {
491 		flush_tlb_all();
492 		return;
493 	}
494 
495 	start = __TLBI_VADDR(start, 0);
496 	end = __TLBI_VADDR(end, 0);
497 
498 	dsb(ishst);
499 	for (addr = start; addr < end; addr += 1 << (PAGE_SHIFT - 12))
500 		__tlbi(vaale1is, addr);
501 	dsb(ish);
502 	isb();
503 }
504 
505 /*
506  * Used to invalidate the TLB (walk caches) corresponding to intermediate page
507  * table levels (pgd/pud/pmd).
508  */
509 static inline void __flush_tlb_kernel_pgtable(unsigned long kaddr)
510 {
511 	unsigned long addr = __TLBI_VADDR(kaddr, 0);
512 
513 	dsb(ishst);
514 	__tlbi(vaae1is, addr);
515 	dsb(ish);
516 	isb();
517 }
518 #endif
519 
520 #endif
521