xref: /linux/arch/powerpc/mm/book3s64/hash_native.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * native hashtable management.
4  *
5  * SMP scalability work:
6  *    Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
7  */
8 
9 #undef DEBUG_LOW
10 
11 #include <linux/spinlock.h>
12 #include <linux/bitops.h>
13 #include <linux/of.h>
14 #include <linux/processor.h>
15 #include <linux/threads.h>
16 #include <linux/smp.h>
17 #include <linux/pgtable.h>
18 
19 #include <asm/machdep.h>
20 #include <asm/mmu.h>
21 #include <asm/mmu_context.h>
22 #include <asm/trace.h>
23 #include <asm/tlb.h>
24 #include <asm/cputable.h>
25 #include <asm/udbg.h>
26 #include <asm/kexec.h>
27 #include <asm/ppc-opcode.h>
28 #include <asm/feature-fixups.h>
29 
30 #include <misc/cxl-base.h>
31 
32 #ifdef DEBUG_LOW
33 #define DBG_LOW(fmt...) udbg_printf(fmt)
34 #else
35 #define DBG_LOW(fmt...)
36 #endif
37 
38 #ifdef __BIG_ENDIAN__
39 #define HPTE_LOCK_BIT 3
40 #else
41 #define HPTE_LOCK_BIT (56+3)
42 #endif
43 
44 static DEFINE_RAW_SPINLOCK(native_tlbie_lock);
45 
46 #ifdef CONFIG_LOCKDEP
47 static struct lockdep_map hpte_lock_map =
48 	STATIC_LOCKDEP_MAP_INIT("hpte_lock", &hpte_lock_map);
49 
50 static void acquire_hpte_lock(void)
51 {
52 	lock_map_acquire(&hpte_lock_map);
53 }
54 
55 static void release_hpte_lock(void)
56 {
57 	lock_map_release(&hpte_lock_map);
58 }
59 #else
60 static void acquire_hpte_lock(void)
61 {
62 }
63 
64 static void release_hpte_lock(void)
65 {
66 }
67 #endif
68 
69 static inline unsigned long  ___tlbie(unsigned long vpn, int psize,
70 						int apsize, int ssize)
71 {
72 	unsigned long va;
73 	unsigned int penc;
74 	unsigned long sllp;
75 
76 	/*
77 	 * We need 14 to 65 bits of va for a tlibe of 4K page
78 	 * With vpn we ignore the lower VPN_SHIFT bits already.
79 	 * And top two bits are already ignored because we can
80 	 * only accomodate 76 bits in a 64 bit vpn with a VPN_SHIFT
81 	 * of 12.
82 	 */
83 	va = vpn << VPN_SHIFT;
84 	/*
85 	 * clear top 16 bits of 64bit va, non SLS segment
86 	 * Older versions of the architecture (2.02 and earler) require the
87 	 * masking of the top 16 bits.
88 	 */
89 	if (mmu_has_feature(MMU_FTR_TLBIE_CROP_VA))
90 		va &= ~(0xffffULL << 48);
91 
92 	switch (psize) {
93 	case MMU_PAGE_4K:
94 		/* clear out bits after (52) [0....52.....63] */
95 		va &= ~((1ul << (64 - 52)) - 1);
96 		va |= ssize << 8;
97 		sllp = get_sllp_encoding(apsize);
98 		va |= sllp << 5;
99 		asm volatile(ASM_FTR_IFCLR("tlbie %0,0", PPC_TLBIE(%1,%0), %2)
100 			     : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206)
101 			     : "memory");
102 		break;
103 	default:
104 		/* We need 14 to 14 + i bits of va */
105 		penc = mmu_psize_defs[psize].penc[apsize];
106 		va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1);
107 		va |= penc << 12;
108 		va |= ssize << 8;
109 		/*
110 		 * AVAL bits:
111 		 * We don't need all the bits, but rest of the bits
112 		 * must be ignored by the processor.
113 		 * vpn cover upto 65 bits of va. (0...65) and we need
114 		 * 58..64 bits of va.
115 		 */
116 		va |= (vpn & 0xfe); /* AVAL */
117 		va |= 1; /* L */
118 		asm volatile(ASM_FTR_IFCLR("tlbie %0,1", PPC_TLBIE(%1,%0), %2)
119 			     : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206)
120 			     : "memory");
121 		break;
122 	}
123 	return va;
124 }
125 
126 static inline void fixup_tlbie_vpn(unsigned long vpn, int psize,
127 				   int apsize, int ssize)
128 {
129 	if (cpu_has_feature(CPU_FTR_P9_TLBIE_ERAT_BUG)) {
130 		/* Radix flush for a hash guest */
131 
132 		unsigned long rb,rs,prs,r,ric;
133 
134 		rb = PPC_BIT(52); /* IS = 2 */
135 		rs = 0;  /* lpid = 0 */
136 		prs = 0; /* partition scoped */
137 		r = 1;   /* radix format */
138 		ric = 0; /* RIC_FLSUH_TLB */
139 
140 		/*
141 		 * Need the extra ptesync to make sure we don't
142 		 * re-order the tlbie
143 		 */
144 		asm volatile("ptesync": : :"memory");
145 		asm volatile(PPC_TLBIE_5(%0, %4, %3, %2, %1)
146 			     : : "r"(rb), "i"(r), "i"(prs),
147 			       "i"(ric), "r"(rs) : "memory");
148 	}
149 
150 
151 	if (cpu_has_feature(CPU_FTR_P9_TLBIE_STQ_BUG)) {
152 		/* Need the extra ptesync to ensure we don't reorder tlbie*/
153 		asm volatile("ptesync": : :"memory");
154 		___tlbie(vpn, psize, apsize, ssize);
155 	}
156 }
157 
158 static inline void __tlbie(unsigned long vpn, int psize, int apsize, int ssize)
159 {
160 	unsigned long rb;
161 
162 	rb = ___tlbie(vpn, psize, apsize, ssize);
163 	trace_tlbie(0, 0, rb, 0, 0, 0, 0);
164 }
165 
166 static inline void __tlbiel(unsigned long vpn, int psize, int apsize, int ssize)
167 {
168 	unsigned long va;
169 	unsigned int penc;
170 	unsigned long sllp;
171 
172 	/* VPN_SHIFT can be atmost 12 */
173 	va = vpn << VPN_SHIFT;
174 	/*
175 	 * clear top 16 bits of 64 bit va, non SLS segment
176 	 * Older versions of the architecture (2.02 and earler) require the
177 	 * masking of the top 16 bits.
178 	 */
179 	if (mmu_has_feature(MMU_FTR_TLBIE_CROP_VA))
180 		va &= ~(0xffffULL << 48);
181 
182 	switch (psize) {
183 	case MMU_PAGE_4K:
184 		/* clear out bits after(52) [0....52.....63] */
185 		va &= ~((1ul << (64 - 52)) - 1);
186 		va |= ssize << 8;
187 		sllp = get_sllp_encoding(apsize);
188 		va |= sllp << 5;
189 		asm volatile(ASM_FTR_IFSET("tlbiel %0", PPC_TLBIEL_v205(%0, 0), %1)
190 			     : : "r" (va), "i" (CPU_FTR_ARCH_206)
191 			     : "memory");
192 		break;
193 	default:
194 		/* We need 14 to 14 + i bits of va */
195 		penc = mmu_psize_defs[psize].penc[apsize];
196 		va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1);
197 		va |= penc << 12;
198 		va |= ssize << 8;
199 		/*
200 		 * AVAL bits:
201 		 * We don't need all the bits, but rest of the bits
202 		 * must be ignored by the processor.
203 		 * vpn cover upto 65 bits of va. (0...65) and we need
204 		 * 58..64 bits of va.
205 		 */
206 		va |= (vpn & 0xfe);
207 		va |= 1; /* L */
208 		asm volatile(ASM_FTR_IFSET("tlbiel %0", PPC_TLBIEL_v205(%0, 1), %1)
209 			     : : "r" (va), "i" (CPU_FTR_ARCH_206)
210 			     : "memory");
211 		break;
212 	}
213 	trace_tlbie(0, 1, va, 0, 0, 0, 0);
214 
215 }
216 
217 static inline void tlbie(unsigned long vpn, int psize, int apsize,
218 			 int ssize, int local)
219 {
220 	unsigned int use_local;
221 	int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
222 
223 	use_local = local && mmu_has_feature(MMU_FTR_TLBIEL) && !cxl_ctx_in_use();
224 
225 	if (use_local)
226 		use_local = mmu_psize_defs[psize].tlbiel;
227 	if (lock_tlbie && !use_local)
228 		raw_spin_lock(&native_tlbie_lock);
229 	asm volatile("ptesync": : :"memory");
230 	if (use_local) {
231 		__tlbiel(vpn, psize, apsize, ssize);
232 		ppc_after_tlbiel_barrier();
233 	} else {
234 		__tlbie(vpn, psize, apsize, ssize);
235 		fixup_tlbie_vpn(vpn, psize, apsize, ssize);
236 		asm volatile("eieio; tlbsync; ptesync": : :"memory");
237 	}
238 	if (lock_tlbie && !use_local)
239 		raw_spin_unlock(&native_tlbie_lock);
240 }
241 
242 static inline void native_lock_hpte(struct hash_pte *hptep)
243 {
244 	unsigned long *word = (unsigned long *)&hptep->v;
245 
246 	acquire_hpte_lock();
247 	while (1) {
248 		if (!test_and_set_bit_lock(HPTE_LOCK_BIT, word))
249 			break;
250 		spin_begin();
251 		while(test_bit(HPTE_LOCK_BIT, word))
252 			spin_cpu_relax();
253 		spin_end();
254 	}
255 }
256 
257 static inline void native_unlock_hpte(struct hash_pte *hptep)
258 {
259 	unsigned long *word = (unsigned long *)&hptep->v;
260 
261 	release_hpte_lock();
262 	clear_bit_unlock(HPTE_LOCK_BIT, word);
263 }
264 
265 static long native_hpte_insert(unsigned long hpte_group, unsigned long vpn,
266 			unsigned long pa, unsigned long rflags,
267 			unsigned long vflags, int psize, int apsize, int ssize)
268 {
269 	struct hash_pte *hptep = htab_address + hpte_group;
270 	unsigned long hpte_v, hpte_r;
271 	unsigned long flags;
272 	int i;
273 
274 	local_irq_save(flags);
275 
276 	if (!(vflags & HPTE_V_BOLTED)) {
277 		DBG_LOW("    insert(group=%lx, vpn=%016lx, pa=%016lx,"
278 			" rflags=%lx, vflags=%lx, psize=%d)\n",
279 			hpte_group, vpn, pa, rflags, vflags, psize);
280 	}
281 
282 	for (i = 0; i < HPTES_PER_GROUP; i++) {
283 		if (! (be64_to_cpu(hptep->v) & HPTE_V_VALID)) {
284 			/* retry with lock held */
285 			native_lock_hpte(hptep);
286 			if (! (be64_to_cpu(hptep->v) & HPTE_V_VALID))
287 				break;
288 			native_unlock_hpte(hptep);
289 		}
290 
291 		hptep++;
292 	}
293 
294 	if (i == HPTES_PER_GROUP) {
295 		local_irq_restore(flags);
296 		return -1;
297 	}
298 
299 	hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID;
300 	hpte_r = hpte_encode_r(pa, psize, apsize) | rflags;
301 
302 	if (!(vflags & HPTE_V_BOLTED)) {
303 		DBG_LOW(" i=%x hpte_v=%016lx, hpte_r=%016lx\n",
304 			i, hpte_v, hpte_r);
305 	}
306 
307 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
308 		hpte_r = hpte_old_to_new_r(hpte_v, hpte_r);
309 		hpte_v = hpte_old_to_new_v(hpte_v);
310 	}
311 
312 	hptep->r = cpu_to_be64(hpte_r);
313 	/* Guarantee the second dword is visible before the valid bit */
314 	eieio();
315 	/*
316 	 * Now set the first dword including the valid bit
317 	 * NOTE: this also unlocks the hpte
318 	 */
319 	release_hpte_lock();
320 	hptep->v = cpu_to_be64(hpte_v);
321 
322 	__asm__ __volatile__ ("ptesync" : : : "memory");
323 
324 	local_irq_restore(flags);
325 
326 	return i | (!!(vflags & HPTE_V_SECONDARY) << 3);
327 }
328 
329 static long native_hpte_remove(unsigned long hpte_group)
330 {
331 	unsigned long hpte_v, flags;
332 	struct hash_pte *hptep;
333 	int i;
334 	int slot_offset;
335 
336 	local_irq_save(flags);
337 
338 	DBG_LOW("    remove(group=%lx)\n", hpte_group);
339 
340 	/* pick a random entry to start at */
341 	slot_offset = mftb() & 0x7;
342 
343 	for (i = 0; i < HPTES_PER_GROUP; i++) {
344 		hptep = htab_address + hpte_group + slot_offset;
345 		hpte_v = be64_to_cpu(hptep->v);
346 
347 		if ((hpte_v & HPTE_V_VALID) && !(hpte_v & HPTE_V_BOLTED)) {
348 			/* retry with lock held */
349 			native_lock_hpte(hptep);
350 			hpte_v = be64_to_cpu(hptep->v);
351 			if ((hpte_v & HPTE_V_VALID)
352 			    && !(hpte_v & HPTE_V_BOLTED))
353 				break;
354 			native_unlock_hpte(hptep);
355 		}
356 
357 		slot_offset++;
358 		slot_offset &= 0x7;
359 	}
360 
361 	if (i == HPTES_PER_GROUP) {
362 		i = -1;
363 		goto out;
364 	}
365 
366 	/* Invalidate the hpte. NOTE: this also unlocks it */
367 	release_hpte_lock();
368 	hptep->v = 0;
369 out:
370 	local_irq_restore(flags);
371 	return i;
372 }
373 
374 static long native_hpte_updatepp(unsigned long slot, unsigned long newpp,
375 				 unsigned long vpn, int bpsize,
376 				 int apsize, int ssize, unsigned long flags)
377 {
378 	struct hash_pte *hptep = htab_address + slot;
379 	unsigned long hpte_v, want_v;
380 	int ret = 0, local = 0;
381 	unsigned long irqflags;
382 
383 	local_irq_save(irqflags);
384 
385 	want_v = hpte_encode_avpn(vpn, bpsize, ssize);
386 
387 	DBG_LOW("    update(vpn=%016lx, avpnv=%016lx, group=%lx, newpp=%lx)",
388 		vpn, want_v & HPTE_V_AVPN, slot, newpp);
389 
390 	hpte_v = hpte_get_old_v(hptep);
391 	/*
392 	 * We need to invalidate the TLB always because hpte_remove doesn't do
393 	 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less
394 	 * random entry from it. When we do that we don't invalidate the TLB
395 	 * (hpte_remove) because we assume the old translation is still
396 	 * technically "valid".
397 	 */
398 	if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID)) {
399 		DBG_LOW(" -> miss\n");
400 		ret = -1;
401 	} else {
402 		native_lock_hpte(hptep);
403 		/* recheck with locks held */
404 		hpte_v = hpte_get_old_v(hptep);
405 		if (unlikely(!HPTE_V_COMPARE(hpte_v, want_v) ||
406 			     !(hpte_v & HPTE_V_VALID))) {
407 			ret = -1;
408 		} else {
409 			DBG_LOW(" -> hit\n");
410 			/* Update the HPTE */
411 			hptep->r = cpu_to_be64((be64_to_cpu(hptep->r) &
412 						~(HPTE_R_PPP | HPTE_R_N)) |
413 					       (newpp & (HPTE_R_PPP | HPTE_R_N |
414 							 HPTE_R_C)));
415 		}
416 		native_unlock_hpte(hptep);
417 	}
418 
419 	if (flags & HPTE_LOCAL_UPDATE)
420 		local = 1;
421 	/*
422 	 * Ensure it is out of the tlb too if it is not a nohpte fault
423 	 */
424 	if (!(flags & HPTE_NOHPTE_UPDATE))
425 		tlbie(vpn, bpsize, apsize, ssize, local);
426 
427 	local_irq_restore(irqflags);
428 
429 	return ret;
430 }
431 
432 static long __native_hpte_find(unsigned long want_v, unsigned long slot)
433 {
434 	struct hash_pte *hptep;
435 	unsigned long hpte_v;
436 	unsigned long i;
437 
438 	for (i = 0; i < HPTES_PER_GROUP; i++) {
439 
440 		hptep = htab_address + slot;
441 		hpte_v = hpte_get_old_v(hptep);
442 		if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID))
443 			/* HPTE matches */
444 			return slot;
445 		++slot;
446 	}
447 
448 	return -1;
449 }
450 
451 static long native_hpte_find(unsigned long vpn, int psize, int ssize)
452 {
453 	unsigned long hpte_group;
454 	unsigned long want_v;
455 	unsigned long hash;
456 	long slot;
457 
458 	hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize);
459 	want_v = hpte_encode_avpn(vpn, psize, ssize);
460 
461 	/*
462 	 * We try to keep bolted entries always in primary hash
463 	 * But in some case we can find them in secondary too.
464 	 */
465 	hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
466 	slot = __native_hpte_find(want_v, hpte_group);
467 	if (slot < 0) {
468 		/* Try in secondary */
469 		hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
470 		slot = __native_hpte_find(want_v, hpte_group);
471 		if (slot < 0)
472 			return -1;
473 	}
474 
475 	return slot;
476 }
477 
478 /*
479  * Update the page protection bits. Intended to be used to create
480  * guard pages for kernel data structures on pages which are bolted
481  * in the HPT. Assumes pages being operated on will not be stolen.
482  *
483  * No need to lock here because we should be the only user.
484  */
485 static void native_hpte_updateboltedpp(unsigned long newpp, unsigned long ea,
486 				       int psize, int ssize)
487 {
488 	unsigned long vpn;
489 	unsigned long vsid;
490 	long slot;
491 	struct hash_pte *hptep;
492 	unsigned long flags;
493 
494 	local_irq_save(flags);
495 
496 	vsid = get_kernel_vsid(ea, ssize);
497 	vpn = hpt_vpn(ea, vsid, ssize);
498 
499 	slot = native_hpte_find(vpn, psize, ssize);
500 	if (slot == -1)
501 		panic("could not find page to bolt\n");
502 	hptep = htab_address + slot;
503 
504 	/* Update the HPTE */
505 	hptep->r = cpu_to_be64((be64_to_cpu(hptep->r) &
506 				~(HPTE_R_PPP | HPTE_R_N)) |
507 			       (newpp & (HPTE_R_PPP | HPTE_R_N)));
508 	/*
509 	 * Ensure it is out of the tlb too. Bolted entries base and
510 	 * actual page size will be same.
511 	 */
512 	tlbie(vpn, psize, psize, ssize, 0);
513 
514 	local_irq_restore(flags);
515 }
516 
517 /*
518  * Remove a bolted kernel entry. Memory hotplug uses this.
519  *
520  * No need to lock here because we should be the only user.
521  */
522 static int native_hpte_removebolted(unsigned long ea, int psize, int ssize)
523 {
524 	unsigned long vpn;
525 	unsigned long vsid;
526 	long slot;
527 	struct hash_pte *hptep;
528 	unsigned long flags;
529 
530 	local_irq_save(flags);
531 
532 	vsid = get_kernel_vsid(ea, ssize);
533 	vpn = hpt_vpn(ea, vsid, ssize);
534 
535 	slot = native_hpte_find(vpn, psize, ssize);
536 	if (slot == -1)
537 		return -ENOENT;
538 
539 	hptep = htab_address + slot;
540 
541 	VM_WARN_ON(!(be64_to_cpu(hptep->v) & HPTE_V_BOLTED));
542 
543 	/* Invalidate the hpte */
544 	hptep->v = 0;
545 
546 	/* Invalidate the TLB */
547 	tlbie(vpn, psize, psize, ssize, 0);
548 
549 	local_irq_restore(flags);
550 
551 	return 0;
552 }
553 
554 
555 static void native_hpte_invalidate(unsigned long slot, unsigned long vpn,
556 				   int bpsize, int apsize, int ssize, int local)
557 {
558 	struct hash_pte *hptep = htab_address + slot;
559 	unsigned long hpte_v;
560 	unsigned long want_v;
561 	unsigned long flags;
562 
563 	local_irq_save(flags);
564 
565 	DBG_LOW("    invalidate(vpn=%016lx, hash: %lx)\n", vpn, slot);
566 
567 	want_v = hpte_encode_avpn(vpn, bpsize, ssize);
568 	hpte_v = hpte_get_old_v(hptep);
569 
570 	if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
571 		native_lock_hpte(hptep);
572 		/* recheck with locks held */
573 		hpte_v = hpte_get_old_v(hptep);
574 
575 		if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
576 			/* Invalidate the hpte. NOTE: this also unlocks it */
577 			release_hpte_lock();
578 			hptep->v = 0;
579 		} else
580 			native_unlock_hpte(hptep);
581 	}
582 	/*
583 	 * We need to invalidate the TLB always because hpte_remove doesn't do
584 	 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less
585 	 * random entry from it. When we do that we don't invalidate the TLB
586 	 * (hpte_remove) because we assume the old translation is still
587 	 * technically "valid".
588 	 */
589 	tlbie(vpn, bpsize, apsize, ssize, local);
590 
591 	local_irq_restore(flags);
592 }
593 
594 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
595 static void native_hugepage_invalidate(unsigned long vsid,
596 				       unsigned long addr,
597 				       unsigned char *hpte_slot_array,
598 				       int psize, int ssize, int local)
599 {
600 	int i;
601 	struct hash_pte *hptep;
602 	int actual_psize = MMU_PAGE_16M;
603 	unsigned int max_hpte_count, valid;
604 	unsigned long flags, s_addr = addr;
605 	unsigned long hpte_v, want_v, shift;
606 	unsigned long hidx, vpn = 0, hash, slot;
607 
608 	shift = mmu_psize_defs[psize].shift;
609 	max_hpte_count = 1U << (PMD_SHIFT - shift);
610 
611 	local_irq_save(flags);
612 	for (i = 0; i < max_hpte_count; i++) {
613 		valid = hpte_valid(hpte_slot_array, i);
614 		if (!valid)
615 			continue;
616 		hidx =  hpte_hash_index(hpte_slot_array, i);
617 
618 		/* get the vpn */
619 		addr = s_addr + (i * (1ul << shift));
620 		vpn = hpt_vpn(addr, vsid, ssize);
621 		hash = hpt_hash(vpn, shift, ssize);
622 		if (hidx & _PTEIDX_SECONDARY)
623 			hash = ~hash;
624 
625 		slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
626 		slot += hidx & _PTEIDX_GROUP_IX;
627 
628 		hptep = htab_address + slot;
629 		want_v = hpte_encode_avpn(vpn, psize, ssize);
630 		hpte_v = hpte_get_old_v(hptep);
631 
632 		/* Even if we miss, we need to invalidate the TLB */
633 		if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
634 			/* recheck with locks held */
635 			native_lock_hpte(hptep);
636 			hpte_v = hpte_get_old_v(hptep);
637 
638 			if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
639 				/* Invalidate the hpte. NOTE: this also unlocks it */
640 				release_hpte_lock();
641 				hptep->v = 0;
642 			} else
643 				native_unlock_hpte(hptep);
644 		}
645 		/*
646 		 * We need to do tlb invalidate for all the address, tlbie
647 		 * instruction compares entry_VA in tlb with the VA specified
648 		 * here
649 		 */
650 		tlbie(vpn, psize, actual_psize, ssize, local);
651 	}
652 	local_irq_restore(flags);
653 }
654 #else
655 static void native_hugepage_invalidate(unsigned long vsid,
656 				       unsigned long addr,
657 				       unsigned char *hpte_slot_array,
658 				       int psize, int ssize, int local)
659 {
660 	WARN(1, "%s called without THP support\n", __func__);
661 }
662 #endif
663 
664 static void hpte_decode(struct hash_pte *hpte, unsigned long slot,
665 			int *psize, int *apsize, int *ssize, unsigned long *vpn)
666 {
667 	unsigned long avpn, pteg, vpi;
668 	unsigned long hpte_v = be64_to_cpu(hpte->v);
669 	unsigned long hpte_r = be64_to_cpu(hpte->r);
670 	unsigned long vsid, seg_off;
671 	int size, a_size, shift;
672 	/* Look at the 8 bit LP value */
673 	unsigned int lp = (hpte_r >> LP_SHIFT) & ((1 << LP_BITS) - 1);
674 
675 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
676 		hpte_v = hpte_new_to_old_v(hpte_v, hpte_r);
677 		hpte_r = hpte_new_to_old_r(hpte_r);
678 	}
679 	if (!(hpte_v & HPTE_V_LARGE)) {
680 		size   = MMU_PAGE_4K;
681 		a_size = MMU_PAGE_4K;
682 	} else {
683 		size = hpte_page_sizes[lp] & 0xf;
684 		a_size = hpte_page_sizes[lp] >> 4;
685 	}
686 	/* This works for all page sizes, and for 256M and 1T segments */
687 	*ssize = hpte_v >> HPTE_V_SSIZE_SHIFT;
688 	shift = mmu_psize_defs[size].shift;
689 
690 	avpn = (HPTE_V_AVPN_VAL(hpte_v) & ~mmu_psize_defs[size].avpnm);
691 	pteg = slot / HPTES_PER_GROUP;
692 	if (hpte_v & HPTE_V_SECONDARY)
693 		pteg = ~pteg;
694 
695 	switch (*ssize) {
696 	case MMU_SEGSIZE_256M:
697 		/* We only have 28 - 23 bits of seg_off in avpn */
698 		seg_off = (avpn & 0x1f) << 23;
699 		vsid    =  avpn >> 5;
700 		/* We can find more bits from the pteg value */
701 		if (shift < 23) {
702 			vpi = (vsid ^ pteg) & htab_hash_mask;
703 			seg_off |= vpi << shift;
704 		}
705 		*vpn = vsid << (SID_SHIFT - VPN_SHIFT) | seg_off >> VPN_SHIFT;
706 		break;
707 	case MMU_SEGSIZE_1T:
708 		/* We only have 40 - 23 bits of seg_off in avpn */
709 		seg_off = (avpn & 0x1ffff) << 23;
710 		vsid    = avpn >> 17;
711 		if (shift < 23) {
712 			vpi = (vsid ^ (vsid << 25) ^ pteg) & htab_hash_mask;
713 			seg_off |= vpi << shift;
714 		}
715 		*vpn = vsid << (SID_SHIFT_1T - VPN_SHIFT) | seg_off >> VPN_SHIFT;
716 		break;
717 	default:
718 		*vpn = size = 0;
719 	}
720 	*psize  = size;
721 	*apsize = a_size;
722 }
723 
724 /*
725  * clear all mappings on kexec.  All cpus are in real mode (or they will
726  * be when they isi), and we are the only one left.  We rely on our kernel
727  * mapping being 0xC0's and the hardware ignoring those two real bits.
728  *
729  * This must be called with interrupts disabled.
730  *
731  * Taking the native_tlbie_lock is unsafe here due to the possibility of
732  * lockdep being on. On pre POWER5 hardware, not taking the lock could
733  * cause deadlock. POWER5 and newer not taking the lock is fine. This only
734  * gets called during boot before secondary CPUs have come up and during
735  * crashdump and all bets are off anyway.
736  *
737  * TODO: add batching support when enabled.  remember, no dynamic memory here,
738  * although there is the control page available...
739  */
740 static notrace void native_hpte_clear(void)
741 {
742 	unsigned long vpn = 0;
743 	unsigned long slot, slots;
744 	struct hash_pte *hptep = htab_address;
745 	unsigned long hpte_v;
746 	unsigned long pteg_count;
747 	int psize, apsize, ssize;
748 
749 	pteg_count = htab_hash_mask + 1;
750 
751 	slots = pteg_count * HPTES_PER_GROUP;
752 
753 	for (slot = 0; slot < slots; slot++, hptep++) {
754 		/*
755 		 * we could lock the pte here, but we are the only cpu
756 		 * running,  right?  and for crash dump, we probably
757 		 * don't want to wait for a maybe bad cpu.
758 		 */
759 		hpte_v = be64_to_cpu(hptep->v);
760 
761 		/*
762 		 * Call __tlbie() here rather than tlbie() since we can't take the
763 		 * native_tlbie_lock.
764 		 */
765 		if (hpte_v & HPTE_V_VALID) {
766 			hpte_decode(hptep, slot, &psize, &apsize, &ssize, &vpn);
767 			hptep->v = 0;
768 			___tlbie(vpn, psize, apsize, ssize);
769 		}
770 	}
771 
772 	asm volatile("eieio; tlbsync; ptesync":::"memory");
773 }
774 
775 /*
776  * Batched hash table flush, we batch the tlbie's to avoid taking/releasing
777  * the lock all the time
778  */
779 static void native_flush_hash_range(unsigned long number, int local)
780 {
781 	unsigned long vpn = 0;
782 	unsigned long hash, index, hidx, shift, slot;
783 	struct hash_pte *hptep;
784 	unsigned long hpte_v;
785 	unsigned long want_v;
786 	unsigned long flags;
787 	real_pte_t pte;
788 	struct ppc64_tlb_batch *batch = this_cpu_ptr(&ppc64_tlb_batch);
789 	unsigned long psize = batch->psize;
790 	int ssize = batch->ssize;
791 	int i;
792 	unsigned int use_local;
793 
794 	use_local = local && mmu_has_feature(MMU_FTR_TLBIEL) &&
795 		mmu_psize_defs[psize].tlbiel && !cxl_ctx_in_use();
796 
797 	local_irq_save(flags);
798 
799 	for (i = 0; i < number; i++) {
800 		vpn = batch->vpn[i];
801 		pte = batch->pte[i];
802 
803 		pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
804 			hash = hpt_hash(vpn, shift, ssize);
805 			hidx = __rpte_to_hidx(pte, index);
806 			if (hidx & _PTEIDX_SECONDARY)
807 				hash = ~hash;
808 			slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
809 			slot += hidx & _PTEIDX_GROUP_IX;
810 			hptep = htab_address + slot;
811 			want_v = hpte_encode_avpn(vpn, psize, ssize);
812 			hpte_v = hpte_get_old_v(hptep);
813 
814 			if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID))
815 				continue;
816 			/* lock and try again */
817 			native_lock_hpte(hptep);
818 			hpte_v = hpte_get_old_v(hptep);
819 
820 			if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID))
821 				native_unlock_hpte(hptep);
822 			else {
823 				release_hpte_lock();
824 				hptep->v = 0;
825 			}
826 
827 		} pte_iterate_hashed_end();
828 	}
829 
830 	if (use_local) {
831 		asm volatile("ptesync":::"memory");
832 		for (i = 0; i < number; i++) {
833 			vpn = batch->vpn[i];
834 			pte = batch->pte[i];
835 
836 			pte_iterate_hashed_subpages(pte, psize,
837 						    vpn, index, shift) {
838 				__tlbiel(vpn, psize, psize, ssize);
839 			} pte_iterate_hashed_end();
840 		}
841 		ppc_after_tlbiel_barrier();
842 	} else {
843 		int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
844 
845 		if (lock_tlbie)
846 			raw_spin_lock(&native_tlbie_lock);
847 
848 		asm volatile("ptesync":::"memory");
849 		for (i = 0; i < number; i++) {
850 			vpn = batch->vpn[i];
851 			pte = batch->pte[i];
852 
853 			pte_iterate_hashed_subpages(pte, psize,
854 						    vpn, index, shift) {
855 				__tlbie(vpn, psize, psize, ssize);
856 			} pte_iterate_hashed_end();
857 		}
858 		/*
859 		 * Just do one more with the last used values.
860 		 */
861 		fixup_tlbie_vpn(vpn, psize, psize, ssize);
862 		asm volatile("eieio; tlbsync; ptesync":::"memory");
863 
864 		if (lock_tlbie)
865 			raw_spin_unlock(&native_tlbie_lock);
866 	}
867 
868 	local_irq_restore(flags);
869 }
870 
871 void __init hpte_init_native(void)
872 {
873 	mmu_hash_ops.hpte_invalidate	= native_hpte_invalidate;
874 	mmu_hash_ops.hpte_updatepp	= native_hpte_updatepp;
875 	mmu_hash_ops.hpte_updateboltedpp = native_hpte_updateboltedpp;
876 	mmu_hash_ops.hpte_removebolted = native_hpte_removebolted;
877 	mmu_hash_ops.hpte_insert	= native_hpte_insert;
878 	mmu_hash_ops.hpte_remove	= native_hpte_remove;
879 	mmu_hash_ops.hpte_clear_all	= native_hpte_clear;
880 	mmu_hash_ops.flush_hash_range = native_flush_hash_range;
881 	mmu_hash_ops.hugepage_invalidate   = native_hugepage_invalidate;
882 }
883