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 
18 #include <asm/machdep.h>
19 #include <asm/mmu.h>
20 #include <asm/mmu_context.h>
21 #include <asm/pgtable.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 static inline void tlbiel_hash_set_isa206(unsigned int set, unsigned int is)
47 {
48 	unsigned long rb;
49 
50 	rb = (set << PPC_BITLSHIFT(51)) | (is << PPC_BITLSHIFT(53));
51 
52 	asm volatile("tlbiel %0" : : "r" (rb));
53 }
54 
55 /*
56  * tlbiel instruction for hash, set invalidation
57  * i.e., r=1 and is=01 or is=10 or is=11
58  */
59 static __always_inline void tlbiel_hash_set_isa300(unsigned int set, unsigned int is,
60 					unsigned int pid,
61 					unsigned int ric, unsigned int prs)
62 {
63 	unsigned long rb;
64 	unsigned long rs;
65 	unsigned int r = 0; /* hash format */
66 
67 	rb = (set << PPC_BITLSHIFT(51)) | (is << PPC_BITLSHIFT(53));
68 	rs = ((unsigned long)pid << PPC_BITLSHIFT(31));
69 
70 	asm volatile(PPC_TLBIEL(%0, %1, %2, %3, %4)
71 		     : : "r"(rb), "r"(rs), "i"(ric), "i"(prs), "r"(r)
72 		     : "memory");
73 }
74 
75 
76 static void tlbiel_all_isa206(unsigned int num_sets, unsigned int is)
77 {
78 	unsigned int set;
79 
80 	asm volatile("ptesync": : :"memory");
81 
82 	for (set = 0; set < num_sets; set++)
83 		tlbiel_hash_set_isa206(set, is);
84 
85 	asm volatile("ptesync": : :"memory");
86 }
87 
88 static void tlbiel_all_isa300(unsigned int num_sets, unsigned int is)
89 {
90 	unsigned int set;
91 
92 	asm volatile("ptesync": : :"memory");
93 
94 	/*
95 	 * Flush the first set of the TLB, and any caching of partition table
96 	 * entries. Then flush the remaining sets of the TLB. Hash mode uses
97 	 * partition scoped TLB translations.
98 	 */
99 	tlbiel_hash_set_isa300(0, is, 0, 2, 0);
100 	for (set = 1; set < num_sets; set++)
101 		tlbiel_hash_set_isa300(set, is, 0, 0, 0);
102 
103 	/*
104 	 * Now invalidate the process table cache.
105 	 *
106 	 * From ISA v3.0B p. 1078:
107 	 *     The following forms are invalid.
108 	 *      * PRS=1, R=0, and RIC!=2 (The only process-scoped
109 	 *        HPT caching is of the Process Table.)
110 	 */
111 	tlbiel_hash_set_isa300(0, is, 0, 2, 1);
112 
113 	asm volatile("ptesync": : :"memory");
114 
115 	asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT "; isync" : : :"memory");
116 }
117 
118 void hash__tlbiel_all(unsigned int action)
119 {
120 	unsigned int is;
121 
122 	switch (action) {
123 	case TLB_INVAL_SCOPE_GLOBAL:
124 		is = 3;
125 		break;
126 	case TLB_INVAL_SCOPE_LPID:
127 		is = 2;
128 		break;
129 	default:
130 		BUG();
131 	}
132 
133 	if (early_cpu_has_feature(CPU_FTR_ARCH_300))
134 		tlbiel_all_isa300(POWER9_TLB_SETS_HASH, is);
135 	else if (early_cpu_has_feature(CPU_FTR_ARCH_207S))
136 		tlbiel_all_isa206(POWER8_TLB_SETS, is);
137 	else if (early_cpu_has_feature(CPU_FTR_ARCH_206))
138 		tlbiel_all_isa206(POWER7_TLB_SETS, is);
139 	else
140 		WARN(1, "%s called on pre-POWER7 CPU\n", __func__);
141 }
142 
143 static inline unsigned long  ___tlbie(unsigned long vpn, int psize,
144 						int apsize, int ssize)
145 {
146 	unsigned long va;
147 	unsigned int penc;
148 	unsigned long sllp;
149 
150 	/*
151 	 * We need 14 to 65 bits of va for a tlibe of 4K page
152 	 * With vpn we ignore the lower VPN_SHIFT bits already.
153 	 * And top two bits are already ignored because we can
154 	 * only accomodate 76 bits in a 64 bit vpn with a VPN_SHIFT
155 	 * of 12.
156 	 */
157 	va = vpn << VPN_SHIFT;
158 	/*
159 	 * clear top 16 bits of 64bit va, non SLS segment
160 	 * Older versions of the architecture (2.02 and earler) require the
161 	 * masking of the top 16 bits.
162 	 */
163 	if (mmu_has_feature(MMU_FTR_TLBIE_CROP_VA))
164 		va &= ~(0xffffULL << 48);
165 
166 	switch (psize) {
167 	case MMU_PAGE_4K:
168 		/* clear out bits after (52) [0....52.....63] */
169 		va &= ~((1ul << (64 - 52)) - 1);
170 		va |= ssize << 8;
171 		sllp = get_sllp_encoding(apsize);
172 		va |= sllp << 5;
173 		asm volatile(ASM_FTR_IFCLR("tlbie %0,0", PPC_TLBIE(%1,%0), %2)
174 			     : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206)
175 			     : "memory");
176 		break;
177 	default:
178 		/* We need 14 to 14 + i bits of va */
179 		penc = mmu_psize_defs[psize].penc[apsize];
180 		va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1);
181 		va |= penc << 12;
182 		va |= ssize << 8;
183 		/*
184 		 * AVAL bits:
185 		 * We don't need all the bits, but rest of the bits
186 		 * must be ignored by the processor.
187 		 * vpn cover upto 65 bits of va. (0...65) and we need
188 		 * 58..64 bits of va.
189 		 */
190 		va |= (vpn & 0xfe); /* AVAL */
191 		va |= 1; /* L */
192 		asm volatile(ASM_FTR_IFCLR("tlbie %0,1", PPC_TLBIE(%1,%0), %2)
193 			     : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206)
194 			     : "memory");
195 		break;
196 	}
197 	return va;
198 }
199 
200 static inline void fixup_tlbie_vpn(unsigned long vpn, int psize,
201 				   int apsize, int ssize)
202 {
203 	if (cpu_has_feature(CPU_FTR_P9_TLBIE_ERAT_BUG)) {
204 		/* Radix flush for a hash guest */
205 
206 		unsigned long rb,rs,prs,r,ric;
207 
208 		rb = PPC_BIT(52); /* IS = 2 */
209 		rs = 0;  /* lpid = 0 */
210 		prs = 0; /* partition scoped */
211 		r = 1;   /* radix format */
212 		ric = 0; /* RIC_FLSUH_TLB */
213 
214 		/*
215 		 * Need the extra ptesync to make sure we don't
216 		 * re-order the tlbie
217 		 */
218 		asm volatile("ptesync": : :"memory");
219 		asm volatile(PPC_TLBIE_5(%0, %4, %3, %2, %1)
220 			     : : "r"(rb), "i"(r), "i"(prs),
221 			       "i"(ric), "r"(rs) : "memory");
222 	}
223 
224 
225 	if (cpu_has_feature(CPU_FTR_P9_TLBIE_STQ_BUG)) {
226 		/* Need the extra ptesync to ensure we don't reorder tlbie*/
227 		asm volatile("ptesync": : :"memory");
228 		___tlbie(vpn, psize, apsize, ssize);
229 	}
230 }
231 
232 static inline void __tlbie(unsigned long vpn, int psize, int apsize, int ssize)
233 {
234 	unsigned long rb;
235 
236 	rb = ___tlbie(vpn, psize, apsize, ssize);
237 	trace_tlbie(0, 0, rb, 0, 0, 0, 0);
238 }
239 
240 static inline void __tlbiel(unsigned long vpn, int psize, int apsize, int ssize)
241 {
242 	unsigned long va;
243 	unsigned int penc;
244 	unsigned long sllp;
245 
246 	/* VPN_SHIFT can be atmost 12 */
247 	va = vpn << VPN_SHIFT;
248 	/*
249 	 * clear top 16 bits of 64 bit va, non SLS segment
250 	 * Older versions of the architecture (2.02 and earler) require the
251 	 * masking of the top 16 bits.
252 	 */
253 	if (mmu_has_feature(MMU_FTR_TLBIE_CROP_VA))
254 		va &= ~(0xffffULL << 48);
255 
256 	switch (psize) {
257 	case MMU_PAGE_4K:
258 		/* clear out bits after(52) [0....52.....63] */
259 		va &= ~((1ul << (64 - 52)) - 1);
260 		va |= ssize << 8;
261 		sllp = get_sllp_encoding(apsize);
262 		va |= sllp << 5;
263 		asm volatile(ASM_FTR_IFSET("tlbiel %0", "tlbiel %0,0", %1)
264 			     : : "r" (va), "i" (CPU_FTR_ARCH_206)
265 			     : "memory");
266 		break;
267 	default:
268 		/* We need 14 to 14 + i bits of va */
269 		penc = mmu_psize_defs[psize].penc[apsize];
270 		va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1);
271 		va |= penc << 12;
272 		va |= ssize << 8;
273 		/*
274 		 * AVAL bits:
275 		 * We don't need all the bits, but rest of the bits
276 		 * must be ignored by the processor.
277 		 * vpn cover upto 65 bits of va. (0...65) and we need
278 		 * 58..64 bits of va.
279 		 */
280 		va |= (vpn & 0xfe);
281 		va |= 1; /* L */
282 		asm volatile(ASM_FTR_IFSET("tlbiel %0", "tlbiel %0,1", %1)
283 			     : : "r" (va), "i" (CPU_FTR_ARCH_206)
284 			     : "memory");
285 		break;
286 	}
287 	trace_tlbie(0, 1, va, 0, 0, 0, 0);
288 
289 }
290 
291 static inline void tlbie(unsigned long vpn, int psize, int apsize,
292 			 int ssize, int local)
293 {
294 	unsigned int use_local;
295 	int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
296 
297 	use_local = local && mmu_has_feature(MMU_FTR_TLBIEL) && !cxl_ctx_in_use();
298 
299 	if (use_local)
300 		use_local = mmu_psize_defs[psize].tlbiel;
301 	if (lock_tlbie && !use_local)
302 		raw_spin_lock(&native_tlbie_lock);
303 	asm volatile("ptesync": : :"memory");
304 	if (use_local) {
305 		__tlbiel(vpn, psize, apsize, ssize);
306 		asm volatile("ptesync": : :"memory");
307 	} else {
308 		__tlbie(vpn, psize, apsize, ssize);
309 		fixup_tlbie_vpn(vpn, psize, apsize, ssize);
310 		asm volatile("eieio; tlbsync; ptesync": : :"memory");
311 	}
312 	if (lock_tlbie && !use_local)
313 		raw_spin_unlock(&native_tlbie_lock);
314 }
315 
316 static inline void native_lock_hpte(struct hash_pte *hptep)
317 {
318 	unsigned long *word = (unsigned long *)&hptep->v;
319 
320 	while (1) {
321 		if (!test_and_set_bit_lock(HPTE_LOCK_BIT, word))
322 			break;
323 		spin_begin();
324 		while(test_bit(HPTE_LOCK_BIT, word))
325 			spin_cpu_relax();
326 		spin_end();
327 	}
328 }
329 
330 static inline void native_unlock_hpte(struct hash_pte *hptep)
331 {
332 	unsigned long *word = (unsigned long *)&hptep->v;
333 
334 	clear_bit_unlock(HPTE_LOCK_BIT, word);
335 }
336 
337 static long native_hpte_insert(unsigned long hpte_group, unsigned long vpn,
338 			unsigned long pa, unsigned long rflags,
339 			unsigned long vflags, int psize, int apsize, int ssize)
340 {
341 	struct hash_pte *hptep = htab_address + hpte_group;
342 	unsigned long hpte_v, hpte_r;
343 	int i;
344 
345 	if (!(vflags & HPTE_V_BOLTED)) {
346 		DBG_LOW("    insert(group=%lx, vpn=%016lx, pa=%016lx,"
347 			" rflags=%lx, vflags=%lx, psize=%d)\n",
348 			hpte_group, vpn, pa, rflags, vflags, psize);
349 	}
350 
351 	for (i = 0; i < HPTES_PER_GROUP; i++) {
352 		if (! (be64_to_cpu(hptep->v) & HPTE_V_VALID)) {
353 			/* retry with lock held */
354 			native_lock_hpte(hptep);
355 			if (! (be64_to_cpu(hptep->v) & HPTE_V_VALID))
356 				break;
357 			native_unlock_hpte(hptep);
358 		}
359 
360 		hptep++;
361 	}
362 
363 	if (i == HPTES_PER_GROUP)
364 		return -1;
365 
366 	hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID;
367 	hpte_r = hpte_encode_r(pa, psize, apsize) | rflags;
368 
369 	if (!(vflags & HPTE_V_BOLTED)) {
370 		DBG_LOW(" i=%x hpte_v=%016lx, hpte_r=%016lx\n",
371 			i, hpte_v, hpte_r);
372 	}
373 
374 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
375 		hpte_r = hpte_old_to_new_r(hpte_v, hpte_r);
376 		hpte_v = hpte_old_to_new_v(hpte_v);
377 	}
378 
379 	hptep->r = cpu_to_be64(hpte_r);
380 	/* Guarantee the second dword is visible before the valid bit */
381 	eieio();
382 	/*
383 	 * Now set the first dword including the valid bit
384 	 * NOTE: this also unlocks the hpte
385 	 */
386 	hptep->v = cpu_to_be64(hpte_v);
387 
388 	__asm__ __volatile__ ("ptesync" : : : "memory");
389 
390 	return i | (!!(vflags & HPTE_V_SECONDARY) << 3);
391 }
392 
393 static long native_hpte_remove(unsigned long hpte_group)
394 {
395 	struct hash_pte *hptep;
396 	int i;
397 	int slot_offset;
398 	unsigned long hpte_v;
399 
400 	DBG_LOW("    remove(group=%lx)\n", hpte_group);
401 
402 	/* pick a random entry to start at */
403 	slot_offset = mftb() & 0x7;
404 
405 	for (i = 0; i < HPTES_PER_GROUP; i++) {
406 		hptep = htab_address + hpte_group + slot_offset;
407 		hpte_v = be64_to_cpu(hptep->v);
408 
409 		if ((hpte_v & HPTE_V_VALID) && !(hpte_v & HPTE_V_BOLTED)) {
410 			/* retry with lock held */
411 			native_lock_hpte(hptep);
412 			hpte_v = be64_to_cpu(hptep->v);
413 			if ((hpte_v & HPTE_V_VALID)
414 			    && !(hpte_v & HPTE_V_BOLTED))
415 				break;
416 			native_unlock_hpte(hptep);
417 		}
418 
419 		slot_offset++;
420 		slot_offset &= 0x7;
421 	}
422 
423 	if (i == HPTES_PER_GROUP)
424 		return -1;
425 
426 	/* Invalidate the hpte. NOTE: this also unlocks it */
427 	hptep->v = 0;
428 
429 	return i;
430 }
431 
432 static long native_hpte_updatepp(unsigned long slot, unsigned long newpp,
433 				 unsigned long vpn, int bpsize,
434 				 int apsize, int ssize, unsigned long flags)
435 {
436 	struct hash_pte *hptep = htab_address + slot;
437 	unsigned long hpte_v, want_v;
438 	int ret = 0, local = 0;
439 
440 	want_v = hpte_encode_avpn(vpn, bpsize, ssize);
441 
442 	DBG_LOW("    update(vpn=%016lx, avpnv=%016lx, group=%lx, newpp=%lx)",
443 		vpn, want_v & HPTE_V_AVPN, slot, newpp);
444 
445 	hpte_v = hpte_get_old_v(hptep);
446 	/*
447 	 * We need to invalidate the TLB always because hpte_remove doesn't do
448 	 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less
449 	 * random entry from it. When we do that we don't invalidate the TLB
450 	 * (hpte_remove) because we assume the old translation is still
451 	 * technically "valid".
452 	 */
453 	if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID)) {
454 		DBG_LOW(" -> miss\n");
455 		ret = -1;
456 	} else {
457 		native_lock_hpte(hptep);
458 		/* recheck with locks held */
459 		hpte_v = hpte_get_old_v(hptep);
460 		if (unlikely(!HPTE_V_COMPARE(hpte_v, want_v) ||
461 			     !(hpte_v & HPTE_V_VALID))) {
462 			ret = -1;
463 		} else {
464 			DBG_LOW(" -> hit\n");
465 			/* Update the HPTE */
466 			hptep->r = cpu_to_be64((be64_to_cpu(hptep->r) &
467 						~(HPTE_R_PPP | HPTE_R_N)) |
468 					       (newpp & (HPTE_R_PPP | HPTE_R_N |
469 							 HPTE_R_C)));
470 		}
471 		native_unlock_hpte(hptep);
472 	}
473 
474 	if (flags & HPTE_LOCAL_UPDATE)
475 		local = 1;
476 	/*
477 	 * Ensure it is out of the tlb too if it is not a nohpte fault
478 	 */
479 	if (!(flags & HPTE_NOHPTE_UPDATE))
480 		tlbie(vpn, bpsize, apsize, ssize, local);
481 
482 	return ret;
483 }
484 
485 static long native_hpte_find(unsigned long vpn, int psize, int ssize)
486 {
487 	struct hash_pte *hptep;
488 	unsigned long hash;
489 	unsigned long i;
490 	long slot;
491 	unsigned long want_v, hpte_v;
492 
493 	hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize);
494 	want_v = hpte_encode_avpn(vpn, psize, ssize);
495 
496 	/* Bolted mappings are only ever in the primary group */
497 	slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
498 	for (i = 0; i < HPTES_PER_GROUP; i++) {
499 
500 		hptep = htab_address + slot;
501 		hpte_v = hpte_get_old_v(hptep);
502 		if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID))
503 			/* HPTE matches */
504 			return slot;
505 		++slot;
506 	}
507 
508 	return -1;
509 }
510 
511 /*
512  * Update the page protection bits. Intended to be used to create
513  * guard pages for kernel data structures on pages which are bolted
514  * in the HPT. Assumes pages being operated on will not be stolen.
515  *
516  * No need to lock here because we should be the only user.
517  */
518 static void native_hpte_updateboltedpp(unsigned long newpp, unsigned long ea,
519 				       int psize, int ssize)
520 {
521 	unsigned long vpn;
522 	unsigned long vsid;
523 	long slot;
524 	struct hash_pte *hptep;
525 
526 	vsid = get_kernel_vsid(ea, ssize);
527 	vpn = hpt_vpn(ea, vsid, ssize);
528 
529 	slot = native_hpte_find(vpn, psize, ssize);
530 	if (slot == -1)
531 		panic("could not find page to bolt\n");
532 	hptep = htab_address + slot;
533 
534 	/* Update the HPTE */
535 	hptep->r = cpu_to_be64((be64_to_cpu(hptep->r) &
536 				~(HPTE_R_PPP | HPTE_R_N)) |
537 			       (newpp & (HPTE_R_PPP | HPTE_R_N)));
538 	/*
539 	 * Ensure it is out of the tlb too. Bolted entries base and
540 	 * actual page size will be same.
541 	 */
542 	tlbie(vpn, psize, psize, ssize, 0);
543 }
544 
545 /*
546  * Remove a bolted kernel entry. Memory hotplug uses this.
547  *
548  * No need to lock here because we should be the only user.
549  */
550 static int native_hpte_removebolted(unsigned long ea, int psize, int ssize)
551 {
552 	unsigned long vpn;
553 	unsigned long vsid;
554 	long slot;
555 	struct hash_pte *hptep;
556 
557 	vsid = get_kernel_vsid(ea, ssize);
558 	vpn = hpt_vpn(ea, vsid, ssize);
559 
560 	slot = native_hpte_find(vpn, psize, ssize);
561 	if (slot == -1)
562 		return -ENOENT;
563 
564 	hptep = htab_address + slot;
565 
566 	VM_WARN_ON(!(be64_to_cpu(hptep->v) & HPTE_V_BOLTED));
567 
568 	/* Invalidate the hpte */
569 	hptep->v = 0;
570 
571 	/* Invalidate the TLB */
572 	tlbie(vpn, psize, psize, ssize, 0);
573 	return 0;
574 }
575 
576 
577 static void native_hpte_invalidate(unsigned long slot, unsigned long vpn,
578 				   int bpsize, int apsize, int ssize, int local)
579 {
580 	struct hash_pte *hptep = htab_address + slot;
581 	unsigned long hpte_v;
582 	unsigned long want_v;
583 	unsigned long flags;
584 
585 	local_irq_save(flags);
586 
587 	DBG_LOW("    invalidate(vpn=%016lx, hash: %lx)\n", vpn, slot);
588 
589 	want_v = hpte_encode_avpn(vpn, bpsize, ssize);
590 	hpte_v = hpte_get_old_v(hptep);
591 
592 	if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
593 		native_lock_hpte(hptep);
594 		/* recheck with locks held */
595 		hpte_v = hpte_get_old_v(hptep);
596 
597 		if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID))
598 			/* Invalidate the hpte. NOTE: this also unlocks it */
599 			hptep->v = 0;
600 		else
601 			native_unlock_hpte(hptep);
602 	}
603 	/*
604 	 * We need to invalidate the TLB always because hpte_remove doesn't do
605 	 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less
606 	 * random entry from it. When we do that we don't invalidate the TLB
607 	 * (hpte_remove) because we assume the old translation is still
608 	 * technically "valid".
609 	 */
610 	tlbie(vpn, bpsize, apsize, ssize, local);
611 
612 	local_irq_restore(flags);
613 }
614 
615 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
616 static void native_hugepage_invalidate(unsigned long vsid,
617 				       unsigned long addr,
618 				       unsigned char *hpte_slot_array,
619 				       int psize, int ssize, int local)
620 {
621 	int i;
622 	struct hash_pte *hptep;
623 	int actual_psize = MMU_PAGE_16M;
624 	unsigned int max_hpte_count, valid;
625 	unsigned long flags, s_addr = addr;
626 	unsigned long hpte_v, want_v, shift;
627 	unsigned long hidx, vpn = 0, hash, slot;
628 
629 	shift = mmu_psize_defs[psize].shift;
630 	max_hpte_count = 1U << (PMD_SHIFT - shift);
631 
632 	local_irq_save(flags);
633 	for (i = 0; i < max_hpte_count; i++) {
634 		valid = hpte_valid(hpte_slot_array, i);
635 		if (!valid)
636 			continue;
637 		hidx =  hpte_hash_index(hpte_slot_array, i);
638 
639 		/* get the vpn */
640 		addr = s_addr + (i * (1ul << shift));
641 		vpn = hpt_vpn(addr, vsid, ssize);
642 		hash = hpt_hash(vpn, shift, ssize);
643 		if (hidx & _PTEIDX_SECONDARY)
644 			hash = ~hash;
645 
646 		slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
647 		slot += hidx & _PTEIDX_GROUP_IX;
648 
649 		hptep = htab_address + slot;
650 		want_v = hpte_encode_avpn(vpn, psize, ssize);
651 		hpte_v = hpte_get_old_v(hptep);
652 
653 		/* Even if we miss, we need to invalidate the TLB */
654 		if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
655 			/* recheck with locks held */
656 			native_lock_hpte(hptep);
657 			hpte_v = hpte_get_old_v(hptep);
658 
659 			if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
660 				/*
661 				 * Invalidate the hpte. NOTE: this also unlocks it
662 				 */
663 
664 				hptep->v = 0;
665 			} else
666 				native_unlock_hpte(hptep);
667 		}
668 		/*
669 		 * We need to do tlb invalidate for all the address, tlbie
670 		 * instruction compares entry_VA in tlb with the VA specified
671 		 * here
672 		 */
673 		tlbie(vpn, psize, actual_psize, ssize, local);
674 	}
675 	local_irq_restore(flags);
676 }
677 #else
678 static void native_hugepage_invalidate(unsigned long vsid,
679 				       unsigned long addr,
680 				       unsigned char *hpte_slot_array,
681 				       int psize, int ssize, int local)
682 {
683 	WARN(1, "%s called without THP support\n", __func__);
684 }
685 #endif
686 
687 static void hpte_decode(struct hash_pte *hpte, unsigned long slot,
688 			int *psize, int *apsize, int *ssize, unsigned long *vpn)
689 {
690 	unsigned long avpn, pteg, vpi;
691 	unsigned long hpte_v = be64_to_cpu(hpte->v);
692 	unsigned long hpte_r = be64_to_cpu(hpte->r);
693 	unsigned long vsid, seg_off;
694 	int size, a_size, shift;
695 	/* Look at the 8 bit LP value */
696 	unsigned int lp = (hpte_r >> LP_SHIFT) & ((1 << LP_BITS) - 1);
697 
698 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
699 		hpte_v = hpte_new_to_old_v(hpte_v, hpte_r);
700 		hpte_r = hpte_new_to_old_r(hpte_r);
701 	}
702 	if (!(hpte_v & HPTE_V_LARGE)) {
703 		size   = MMU_PAGE_4K;
704 		a_size = MMU_PAGE_4K;
705 	} else {
706 		size = hpte_page_sizes[lp] & 0xf;
707 		a_size = hpte_page_sizes[lp] >> 4;
708 	}
709 	/* This works for all page sizes, and for 256M and 1T segments */
710 	*ssize = hpte_v >> HPTE_V_SSIZE_SHIFT;
711 	shift = mmu_psize_defs[size].shift;
712 
713 	avpn = (HPTE_V_AVPN_VAL(hpte_v) & ~mmu_psize_defs[size].avpnm);
714 	pteg = slot / HPTES_PER_GROUP;
715 	if (hpte_v & HPTE_V_SECONDARY)
716 		pteg = ~pteg;
717 
718 	switch (*ssize) {
719 	case MMU_SEGSIZE_256M:
720 		/* We only have 28 - 23 bits of seg_off in avpn */
721 		seg_off = (avpn & 0x1f) << 23;
722 		vsid    =  avpn >> 5;
723 		/* We can find more bits from the pteg value */
724 		if (shift < 23) {
725 			vpi = (vsid ^ pteg) & htab_hash_mask;
726 			seg_off |= vpi << shift;
727 		}
728 		*vpn = vsid << (SID_SHIFT - VPN_SHIFT) | seg_off >> VPN_SHIFT;
729 		break;
730 	case MMU_SEGSIZE_1T:
731 		/* We only have 40 - 23 bits of seg_off in avpn */
732 		seg_off = (avpn & 0x1ffff) << 23;
733 		vsid    = avpn >> 17;
734 		if (shift < 23) {
735 			vpi = (vsid ^ (vsid << 25) ^ pteg) & htab_hash_mask;
736 			seg_off |= vpi << shift;
737 		}
738 		*vpn = vsid << (SID_SHIFT_1T - VPN_SHIFT) | seg_off >> VPN_SHIFT;
739 		break;
740 	default:
741 		*vpn = size = 0;
742 	}
743 	*psize  = size;
744 	*apsize = a_size;
745 }
746 
747 /*
748  * clear all mappings on kexec.  All cpus are in real mode (or they will
749  * be when they isi), and we are the only one left.  We rely on our kernel
750  * mapping being 0xC0's and the hardware ignoring those two real bits.
751  *
752  * This must be called with interrupts disabled.
753  *
754  * Taking the native_tlbie_lock is unsafe here due to the possibility of
755  * lockdep being on. On pre POWER5 hardware, not taking the lock could
756  * cause deadlock. POWER5 and newer not taking the lock is fine. This only
757  * gets called during boot before secondary CPUs have come up and during
758  * crashdump and all bets are off anyway.
759  *
760  * TODO: add batching support when enabled.  remember, no dynamic memory here,
761  * although there is the control page available...
762  */
763 static void native_hpte_clear(void)
764 {
765 	unsigned long vpn = 0;
766 	unsigned long slot, slots;
767 	struct hash_pte *hptep = htab_address;
768 	unsigned long hpte_v;
769 	unsigned long pteg_count;
770 	int psize, apsize, ssize;
771 
772 	pteg_count = htab_hash_mask + 1;
773 
774 	slots = pteg_count * HPTES_PER_GROUP;
775 
776 	for (slot = 0; slot < slots; slot++, hptep++) {
777 		/*
778 		 * we could lock the pte here, but we are the only cpu
779 		 * running,  right?  and for crash dump, we probably
780 		 * don't want to wait for a maybe bad cpu.
781 		 */
782 		hpte_v = be64_to_cpu(hptep->v);
783 
784 		/*
785 		 * Call __tlbie() here rather than tlbie() since we can't take the
786 		 * native_tlbie_lock.
787 		 */
788 		if (hpte_v & HPTE_V_VALID) {
789 			hpte_decode(hptep, slot, &psize, &apsize, &ssize, &vpn);
790 			hptep->v = 0;
791 			___tlbie(vpn, psize, apsize, ssize);
792 		}
793 	}
794 
795 	asm volatile("eieio; tlbsync; ptesync":::"memory");
796 }
797 
798 /*
799  * Batched hash table flush, we batch the tlbie's to avoid taking/releasing
800  * the lock all the time
801  */
802 static void native_flush_hash_range(unsigned long number, int local)
803 {
804 	unsigned long vpn = 0;
805 	unsigned long hash, index, hidx, shift, slot;
806 	struct hash_pte *hptep;
807 	unsigned long hpte_v;
808 	unsigned long want_v;
809 	unsigned long flags;
810 	real_pte_t pte;
811 	struct ppc64_tlb_batch *batch = this_cpu_ptr(&ppc64_tlb_batch);
812 	unsigned long psize = batch->psize;
813 	int ssize = batch->ssize;
814 	int i;
815 	unsigned int use_local;
816 
817 	use_local = local && mmu_has_feature(MMU_FTR_TLBIEL) &&
818 		mmu_psize_defs[psize].tlbiel && !cxl_ctx_in_use();
819 
820 	local_irq_save(flags);
821 
822 	for (i = 0; i < number; i++) {
823 		vpn = batch->vpn[i];
824 		pte = batch->pte[i];
825 
826 		pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
827 			hash = hpt_hash(vpn, shift, ssize);
828 			hidx = __rpte_to_hidx(pte, index);
829 			if (hidx & _PTEIDX_SECONDARY)
830 				hash = ~hash;
831 			slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
832 			slot += hidx & _PTEIDX_GROUP_IX;
833 			hptep = htab_address + slot;
834 			want_v = hpte_encode_avpn(vpn, psize, ssize);
835 			hpte_v = hpte_get_old_v(hptep);
836 
837 			if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID))
838 				continue;
839 			/* lock and try again */
840 			native_lock_hpte(hptep);
841 			hpte_v = hpte_get_old_v(hptep);
842 
843 			if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID))
844 				native_unlock_hpte(hptep);
845 			else
846 				hptep->v = 0;
847 
848 		} pte_iterate_hashed_end();
849 	}
850 
851 	if (use_local) {
852 		asm volatile("ptesync":::"memory");
853 		for (i = 0; i < number; i++) {
854 			vpn = batch->vpn[i];
855 			pte = batch->pte[i];
856 
857 			pte_iterate_hashed_subpages(pte, psize,
858 						    vpn, index, shift) {
859 				__tlbiel(vpn, psize, psize, ssize);
860 			} pte_iterate_hashed_end();
861 		}
862 		asm volatile("ptesync":::"memory");
863 	} else {
864 		int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
865 
866 		if (lock_tlbie)
867 			raw_spin_lock(&native_tlbie_lock);
868 
869 		asm volatile("ptesync":::"memory");
870 		for (i = 0; i < number; i++) {
871 			vpn = batch->vpn[i];
872 			pte = batch->pte[i];
873 
874 			pte_iterate_hashed_subpages(pte, psize,
875 						    vpn, index, shift) {
876 				__tlbie(vpn, psize, psize, ssize);
877 			} pte_iterate_hashed_end();
878 		}
879 		/*
880 		 * Just do one more with the last used values.
881 		 */
882 		fixup_tlbie_vpn(vpn, psize, psize, ssize);
883 		asm volatile("eieio; tlbsync; ptesync":::"memory");
884 
885 		if (lock_tlbie)
886 			raw_spin_unlock(&native_tlbie_lock);
887 	}
888 
889 	local_irq_restore(flags);
890 }
891 
892 void __init hpte_init_native(void)
893 {
894 	mmu_hash_ops.hpte_invalidate	= native_hpte_invalidate;
895 	mmu_hash_ops.hpte_updatepp	= native_hpte_updatepp;
896 	mmu_hash_ops.hpte_updateboltedpp = native_hpte_updateboltedpp;
897 	mmu_hash_ops.hpte_removebolted = native_hpte_removebolted;
898 	mmu_hash_ops.hpte_insert	= native_hpte_insert;
899 	mmu_hash_ops.hpte_remove	= native_hpte_remove;
900 	mmu_hash_ops.hpte_clear_all	= native_hpte_clear;
901 	mmu_hash_ops.flush_hash_range = native_flush_hash_range;
902 	mmu_hash_ops.hugepage_invalidate   = native_hugepage_invalidate;
903 }
904