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 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), "i"(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 	ppc_after_tlbiel_barrier();
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 partition table cache if this is HV mode.
96 	 */
97 	if (early_cpu_has_feature(CPU_FTR_HVMODE))
98 		tlbiel_hash_set_isa300(0, is, 0, 2, 0);
99 
100 	/*
101 	 * Now invalidate the process table cache. UPRT=0 HPT modes (what
102 	 * current hardware implements) do not use the process table, but
103 	 * add the flushes anyway.
104 	 *
105 	 * From ISA v3.0B p. 1078:
106 	 *     The following forms are invalid.
107 	 *      * PRS=1, R=0, and RIC!=2 (The only process-scoped
108 	 *        HPT caching is of the Process Table.)
109 	 */
110 	tlbiel_hash_set_isa300(0, is, 0, 2, 1);
111 
112 	/*
113 	 * Then flush the sets of the TLB proper. Hash mode uses
114 	 * partition scoped TLB translations, which may be flushed
115 	 * in !HV mode.
116 	 */
117 	for (set = 0; set < num_sets; set++)
118 		tlbiel_hash_set_isa300(set, is, 0, 0, 0);
119 
120 	ppc_after_tlbiel_barrier();
121 
122 	asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT "; isync" : : :"memory");
123 }
124 
125 void hash__tlbiel_all(unsigned int action)
126 {
127 	unsigned int is;
128 
129 	switch (action) {
130 	case TLB_INVAL_SCOPE_GLOBAL:
131 		is = 3;
132 		break;
133 	case TLB_INVAL_SCOPE_LPID:
134 		is = 2;
135 		break;
136 	default:
137 		BUG();
138 	}
139 
140 	if (early_cpu_has_feature(CPU_FTR_ARCH_300))
141 		tlbiel_all_isa300(POWER9_TLB_SETS_HASH, is);
142 	else if (early_cpu_has_feature(CPU_FTR_ARCH_207S))
143 		tlbiel_all_isa206(POWER8_TLB_SETS, is);
144 	else if (early_cpu_has_feature(CPU_FTR_ARCH_206))
145 		tlbiel_all_isa206(POWER7_TLB_SETS, is);
146 	else
147 		WARN(1, "%s called on pre-POWER7 CPU\n", __func__);
148 }
149 
150 static inline unsigned long  ___tlbie(unsigned long vpn, int psize,
151 						int apsize, int ssize)
152 {
153 	unsigned long va;
154 	unsigned int penc;
155 	unsigned long sllp;
156 
157 	/*
158 	 * We need 14 to 65 bits of va for a tlibe of 4K page
159 	 * With vpn we ignore the lower VPN_SHIFT bits already.
160 	 * And top two bits are already ignored because we can
161 	 * only accomodate 76 bits in a 64 bit vpn with a VPN_SHIFT
162 	 * of 12.
163 	 */
164 	va = vpn << VPN_SHIFT;
165 	/*
166 	 * clear top 16 bits of 64bit va, non SLS segment
167 	 * Older versions of the architecture (2.02 and earler) require the
168 	 * masking of the top 16 bits.
169 	 */
170 	if (mmu_has_feature(MMU_FTR_TLBIE_CROP_VA))
171 		va &= ~(0xffffULL << 48);
172 
173 	switch (psize) {
174 	case MMU_PAGE_4K:
175 		/* clear out bits after (52) [0....52.....63] */
176 		va &= ~((1ul << (64 - 52)) - 1);
177 		va |= ssize << 8;
178 		sllp = get_sllp_encoding(apsize);
179 		va |= sllp << 5;
180 		asm volatile(ASM_FTR_IFCLR("tlbie %0,0", PPC_TLBIE(%1,%0), %2)
181 			     : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206)
182 			     : "memory");
183 		break;
184 	default:
185 		/* We need 14 to 14 + i bits of va */
186 		penc = mmu_psize_defs[psize].penc[apsize];
187 		va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1);
188 		va |= penc << 12;
189 		va |= ssize << 8;
190 		/*
191 		 * AVAL bits:
192 		 * We don't need all the bits, but rest of the bits
193 		 * must be ignored by the processor.
194 		 * vpn cover upto 65 bits of va. (0...65) and we need
195 		 * 58..64 bits of va.
196 		 */
197 		va |= (vpn & 0xfe); /* AVAL */
198 		va |= 1; /* L */
199 		asm volatile(ASM_FTR_IFCLR("tlbie %0,1", PPC_TLBIE(%1,%0), %2)
200 			     : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206)
201 			     : "memory");
202 		break;
203 	}
204 	return va;
205 }
206 
207 static inline void fixup_tlbie_vpn(unsigned long vpn, int psize,
208 				   int apsize, int ssize)
209 {
210 	if (cpu_has_feature(CPU_FTR_P9_TLBIE_ERAT_BUG)) {
211 		/* Radix flush for a hash guest */
212 
213 		unsigned long rb,rs,prs,r,ric;
214 
215 		rb = PPC_BIT(52); /* IS = 2 */
216 		rs = 0;  /* lpid = 0 */
217 		prs = 0; /* partition scoped */
218 		r = 1;   /* radix format */
219 		ric = 0; /* RIC_FLSUH_TLB */
220 
221 		/*
222 		 * Need the extra ptesync to make sure we don't
223 		 * re-order the tlbie
224 		 */
225 		asm volatile("ptesync": : :"memory");
226 		asm volatile(PPC_TLBIE_5(%0, %4, %3, %2, %1)
227 			     : : "r"(rb), "i"(r), "i"(prs),
228 			       "i"(ric), "r"(rs) : "memory");
229 	}
230 
231 
232 	if (cpu_has_feature(CPU_FTR_P9_TLBIE_STQ_BUG)) {
233 		/* Need the extra ptesync to ensure we don't reorder tlbie*/
234 		asm volatile("ptesync": : :"memory");
235 		___tlbie(vpn, psize, apsize, ssize);
236 	}
237 }
238 
239 static inline void __tlbie(unsigned long vpn, int psize, int apsize, int ssize)
240 {
241 	unsigned long rb;
242 
243 	rb = ___tlbie(vpn, psize, apsize, ssize);
244 	trace_tlbie(0, 0, rb, 0, 0, 0, 0);
245 }
246 
247 static inline void __tlbiel(unsigned long vpn, int psize, int apsize, int ssize)
248 {
249 	unsigned long va;
250 	unsigned int penc;
251 	unsigned long sllp;
252 
253 	/* VPN_SHIFT can be atmost 12 */
254 	va = vpn << VPN_SHIFT;
255 	/*
256 	 * clear top 16 bits of 64 bit va, non SLS segment
257 	 * Older versions of the architecture (2.02 and earler) require the
258 	 * masking of the top 16 bits.
259 	 */
260 	if (mmu_has_feature(MMU_FTR_TLBIE_CROP_VA))
261 		va &= ~(0xffffULL << 48);
262 
263 	switch (psize) {
264 	case MMU_PAGE_4K:
265 		/* clear out bits after(52) [0....52.....63] */
266 		va &= ~((1ul << (64 - 52)) - 1);
267 		va |= ssize << 8;
268 		sllp = get_sllp_encoding(apsize);
269 		va |= sllp << 5;
270 		asm volatile(ASM_FTR_IFSET("tlbiel %0", "tlbiel %0,0", %1)
271 			     : : "r" (va), "i" (CPU_FTR_ARCH_206)
272 			     : "memory");
273 		break;
274 	default:
275 		/* We need 14 to 14 + i bits of va */
276 		penc = mmu_psize_defs[psize].penc[apsize];
277 		va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1);
278 		va |= penc << 12;
279 		va |= ssize << 8;
280 		/*
281 		 * AVAL bits:
282 		 * We don't need all the bits, but rest of the bits
283 		 * must be ignored by the processor.
284 		 * vpn cover upto 65 bits of va. (0...65) and we need
285 		 * 58..64 bits of va.
286 		 */
287 		va |= (vpn & 0xfe);
288 		va |= 1; /* L */
289 		asm volatile(ASM_FTR_IFSET("tlbiel %0", "tlbiel %0,1", %1)
290 			     : : "r" (va), "i" (CPU_FTR_ARCH_206)
291 			     : "memory");
292 		break;
293 	}
294 	trace_tlbie(0, 1, va, 0, 0, 0, 0);
295 
296 }
297 
298 static inline void tlbie(unsigned long vpn, int psize, int apsize,
299 			 int ssize, int local)
300 {
301 	unsigned int use_local;
302 	int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
303 
304 	use_local = local && mmu_has_feature(MMU_FTR_TLBIEL) && !cxl_ctx_in_use();
305 
306 	if (use_local)
307 		use_local = mmu_psize_defs[psize].tlbiel;
308 	if (lock_tlbie && !use_local)
309 		raw_spin_lock(&native_tlbie_lock);
310 	asm volatile("ptesync": : :"memory");
311 	if (use_local) {
312 		__tlbiel(vpn, psize, apsize, ssize);
313 		ppc_after_tlbiel_barrier();
314 	} else {
315 		__tlbie(vpn, psize, apsize, ssize);
316 		fixup_tlbie_vpn(vpn, psize, apsize, ssize);
317 		asm volatile("eieio; tlbsync; ptesync": : :"memory");
318 	}
319 	if (lock_tlbie && !use_local)
320 		raw_spin_unlock(&native_tlbie_lock);
321 }
322 
323 static inline void native_lock_hpte(struct hash_pte *hptep)
324 {
325 	unsigned long *word = (unsigned long *)&hptep->v;
326 
327 	while (1) {
328 		if (!test_and_set_bit_lock(HPTE_LOCK_BIT, word))
329 			break;
330 		spin_begin();
331 		while(test_bit(HPTE_LOCK_BIT, word))
332 			spin_cpu_relax();
333 		spin_end();
334 	}
335 }
336 
337 static inline void native_unlock_hpte(struct hash_pte *hptep)
338 {
339 	unsigned long *word = (unsigned long *)&hptep->v;
340 
341 	clear_bit_unlock(HPTE_LOCK_BIT, word);
342 }
343 
344 static long native_hpte_insert(unsigned long hpte_group, unsigned long vpn,
345 			unsigned long pa, unsigned long rflags,
346 			unsigned long vflags, int psize, int apsize, int ssize)
347 {
348 	struct hash_pte *hptep = htab_address + hpte_group;
349 	unsigned long hpte_v, hpte_r;
350 	int i;
351 
352 	if (!(vflags & HPTE_V_BOLTED)) {
353 		DBG_LOW("    insert(group=%lx, vpn=%016lx, pa=%016lx,"
354 			" rflags=%lx, vflags=%lx, psize=%d)\n",
355 			hpte_group, vpn, pa, rflags, vflags, psize);
356 	}
357 
358 	for (i = 0; i < HPTES_PER_GROUP; i++) {
359 		if (! (be64_to_cpu(hptep->v) & HPTE_V_VALID)) {
360 			/* retry with lock held */
361 			native_lock_hpte(hptep);
362 			if (! (be64_to_cpu(hptep->v) & HPTE_V_VALID))
363 				break;
364 			native_unlock_hpte(hptep);
365 		}
366 
367 		hptep++;
368 	}
369 
370 	if (i == HPTES_PER_GROUP)
371 		return -1;
372 
373 	hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID;
374 	hpte_r = hpte_encode_r(pa, psize, apsize) | rflags;
375 
376 	if (!(vflags & HPTE_V_BOLTED)) {
377 		DBG_LOW(" i=%x hpte_v=%016lx, hpte_r=%016lx\n",
378 			i, hpte_v, hpte_r);
379 	}
380 
381 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
382 		hpte_r = hpte_old_to_new_r(hpte_v, hpte_r);
383 		hpte_v = hpte_old_to_new_v(hpte_v);
384 	}
385 
386 	hptep->r = cpu_to_be64(hpte_r);
387 	/* Guarantee the second dword is visible before the valid bit */
388 	eieio();
389 	/*
390 	 * Now set the first dword including the valid bit
391 	 * NOTE: this also unlocks the hpte
392 	 */
393 	hptep->v = cpu_to_be64(hpte_v);
394 
395 	__asm__ __volatile__ ("ptesync" : : : "memory");
396 
397 	return i | (!!(vflags & HPTE_V_SECONDARY) << 3);
398 }
399 
400 static long native_hpte_remove(unsigned long hpte_group)
401 {
402 	struct hash_pte *hptep;
403 	int i;
404 	int slot_offset;
405 	unsigned long hpte_v;
406 
407 	DBG_LOW("    remove(group=%lx)\n", hpte_group);
408 
409 	/* pick a random entry to start at */
410 	slot_offset = mftb() & 0x7;
411 
412 	for (i = 0; i < HPTES_PER_GROUP; i++) {
413 		hptep = htab_address + hpte_group + slot_offset;
414 		hpte_v = be64_to_cpu(hptep->v);
415 
416 		if ((hpte_v & HPTE_V_VALID) && !(hpte_v & HPTE_V_BOLTED)) {
417 			/* retry with lock held */
418 			native_lock_hpte(hptep);
419 			hpte_v = be64_to_cpu(hptep->v);
420 			if ((hpte_v & HPTE_V_VALID)
421 			    && !(hpte_v & HPTE_V_BOLTED))
422 				break;
423 			native_unlock_hpte(hptep);
424 		}
425 
426 		slot_offset++;
427 		slot_offset &= 0x7;
428 	}
429 
430 	if (i == HPTES_PER_GROUP)
431 		return -1;
432 
433 	/* Invalidate the hpte. NOTE: this also unlocks it */
434 	hptep->v = 0;
435 
436 	return i;
437 }
438 
439 static long native_hpte_updatepp(unsigned long slot, unsigned long newpp,
440 				 unsigned long vpn, int bpsize,
441 				 int apsize, int ssize, unsigned long flags)
442 {
443 	struct hash_pte *hptep = htab_address + slot;
444 	unsigned long hpte_v, want_v;
445 	int ret = 0, local = 0;
446 
447 	want_v = hpte_encode_avpn(vpn, bpsize, ssize);
448 
449 	DBG_LOW("    update(vpn=%016lx, avpnv=%016lx, group=%lx, newpp=%lx)",
450 		vpn, want_v & HPTE_V_AVPN, slot, newpp);
451 
452 	hpte_v = hpte_get_old_v(hptep);
453 	/*
454 	 * We need to invalidate the TLB always because hpte_remove doesn't do
455 	 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less
456 	 * random entry from it. When we do that we don't invalidate the TLB
457 	 * (hpte_remove) because we assume the old translation is still
458 	 * technically "valid".
459 	 */
460 	if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID)) {
461 		DBG_LOW(" -> miss\n");
462 		ret = -1;
463 	} else {
464 		native_lock_hpte(hptep);
465 		/* recheck with locks held */
466 		hpte_v = hpte_get_old_v(hptep);
467 		if (unlikely(!HPTE_V_COMPARE(hpte_v, want_v) ||
468 			     !(hpte_v & HPTE_V_VALID))) {
469 			ret = -1;
470 		} else {
471 			DBG_LOW(" -> hit\n");
472 			/* Update the HPTE */
473 			hptep->r = cpu_to_be64((be64_to_cpu(hptep->r) &
474 						~(HPTE_R_PPP | HPTE_R_N)) |
475 					       (newpp & (HPTE_R_PPP | HPTE_R_N |
476 							 HPTE_R_C)));
477 		}
478 		native_unlock_hpte(hptep);
479 	}
480 
481 	if (flags & HPTE_LOCAL_UPDATE)
482 		local = 1;
483 	/*
484 	 * Ensure it is out of the tlb too if it is not a nohpte fault
485 	 */
486 	if (!(flags & HPTE_NOHPTE_UPDATE))
487 		tlbie(vpn, bpsize, apsize, ssize, local);
488 
489 	return ret;
490 }
491 
492 static long __native_hpte_find(unsigned long want_v, unsigned long slot)
493 {
494 	struct hash_pte *hptep;
495 	unsigned long hpte_v;
496 	unsigned long i;
497 
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 static long native_hpte_find(unsigned long vpn, int psize, int ssize)
512 {
513 	unsigned long hpte_group;
514 	unsigned long want_v;
515 	unsigned long hash;
516 	long slot;
517 
518 	hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize);
519 	want_v = hpte_encode_avpn(vpn, psize, ssize);
520 
521 	/*
522 	 * We try to keep bolted entries always in primary hash
523 	 * But in some case we can find them in secondary too.
524 	 */
525 	hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
526 	slot = __native_hpte_find(want_v, hpte_group);
527 	if (slot < 0) {
528 		/* Try in secondary */
529 		hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
530 		slot = __native_hpte_find(want_v, hpte_group);
531 		if (slot < 0)
532 			return -1;
533 	}
534 
535 	return slot;
536 }
537 
538 /*
539  * Update the page protection bits. Intended to be used to create
540  * guard pages for kernel data structures on pages which are bolted
541  * in the HPT. Assumes pages being operated on will not be stolen.
542  *
543  * No need to lock here because we should be the only user.
544  */
545 static void native_hpte_updateboltedpp(unsigned long newpp, unsigned long ea,
546 				       int psize, int ssize)
547 {
548 	unsigned long vpn;
549 	unsigned long vsid;
550 	long slot;
551 	struct hash_pte *hptep;
552 
553 	vsid = get_kernel_vsid(ea, ssize);
554 	vpn = hpt_vpn(ea, vsid, ssize);
555 
556 	slot = native_hpte_find(vpn, psize, ssize);
557 	if (slot == -1)
558 		panic("could not find page to bolt\n");
559 	hptep = htab_address + slot;
560 
561 	/* Update the HPTE */
562 	hptep->r = cpu_to_be64((be64_to_cpu(hptep->r) &
563 				~(HPTE_R_PPP | HPTE_R_N)) |
564 			       (newpp & (HPTE_R_PPP | HPTE_R_N)));
565 	/*
566 	 * Ensure it is out of the tlb too. Bolted entries base and
567 	 * actual page size will be same.
568 	 */
569 	tlbie(vpn, psize, psize, ssize, 0);
570 }
571 
572 /*
573  * Remove a bolted kernel entry. Memory hotplug uses this.
574  *
575  * No need to lock here because we should be the only user.
576  */
577 static int native_hpte_removebolted(unsigned long ea, int psize, int ssize)
578 {
579 	unsigned long vpn;
580 	unsigned long vsid;
581 	long slot;
582 	struct hash_pte *hptep;
583 
584 	vsid = get_kernel_vsid(ea, ssize);
585 	vpn = hpt_vpn(ea, vsid, ssize);
586 
587 	slot = native_hpte_find(vpn, psize, ssize);
588 	if (slot == -1)
589 		return -ENOENT;
590 
591 	hptep = htab_address + slot;
592 
593 	VM_WARN_ON(!(be64_to_cpu(hptep->v) & HPTE_V_BOLTED));
594 
595 	/* Invalidate the hpte */
596 	hptep->v = 0;
597 
598 	/* Invalidate the TLB */
599 	tlbie(vpn, psize, psize, ssize, 0);
600 	return 0;
601 }
602 
603 
604 static void native_hpte_invalidate(unsigned long slot, unsigned long vpn,
605 				   int bpsize, int apsize, int ssize, int local)
606 {
607 	struct hash_pte *hptep = htab_address + slot;
608 	unsigned long hpte_v;
609 	unsigned long want_v;
610 	unsigned long flags;
611 
612 	local_irq_save(flags);
613 
614 	DBG_LOW("    invalidate(vpn=%016lx, hash: %lx)\n", vpn, slot);
615 
616 	want_v = hpte_encode_avpn(vpn, bpsize, ssize);
617 	hpte_v = hpte_get_old_v(hptep);
618 
619 	if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
620 		native_lock_hpte(hptep);
621 		/* recheck with locks held */
622 		hpte_v = hpte_get_old_v(hptep);
623 
624 		if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID))
625 			/* Invalidate the hpte. NOTE: this also unlocks it */
626 			hptep->v = 0;
627 		else
628 			native_unlock_hpte(hptep);
629 	}
630 	/*
631 	 * We need to invalidate the TLB always because hpte_remove doesn't do
632 	 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less
633 	 * random entry from it. When we do that we don't invalidate the TLB
634 	 * (hpte_remove) because we assume the old translation is still
635 	 * technically "valid".
636 	 */
637 	tlbie(vpn, bpsize, apsize, ssize, local);
638 
639 	local_irq_restore(flags);
640 }
641 
642 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
643 static void native_hugepage_invalidate(unsigned long vsid,
644 				       unsigned long addr,
645 				       unsigned char *hpte_slot_array,
646 				       int psize, int ssize, int local)
647 {
648 	int i;
649 	struct hash_pte *hptep;
650 	int actual_psize = MMU_PAGE_16M;
651 	unsigned int max_hpte_count, valid;
652 	unsigned long flags, s_addr = addr;
653 	unsigned long hpte_v, want_v, shift;
654 	unsigned long hidx, vpn = 0, hash, slot;
655 
656 	shift = mmu_psize_defs[psize].shift;
657 	max_hpte_count = 1U << (PMD_SHIFT - shift);
658 
659 	local_irq_save(flags);
660 	for (i = 0; i < max_hpte_count; i++) {
661 		valid = hpte_valid(hpte_slot_array, i);
662 		if (!valid)
663 			continue;
664 		hidx =  hpte_hash_index(hpte_slot_array, i);
665 
666 		/* get the vpn */
667 		addr = s_addr + (i * (1ul << shift));
668 		vpn = hpt_vpn(addr, vsid, ssize);
669 		hash = hpt_hash(vpn, shift, ssize);
670 		if (hidx & _PTEIDX_SECONDARY)
671 			hash = ~hash;
672 
673 		slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
674 		slot += hidx & _PTEIDX_GROUP_IX;
675 
676 		hptep = htab_address + slot;
677 		want_v = hpte_encode_avpn(vpn, psize, ssize);
678 		hpte_v = hpte_get_old_v(hptep);
679 
680 		/* Even if we miss, we need to invalidate the TLB */
681 		if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
682 			/* recheck with locks held */
683 			native_lock_hpte(hptep);
684 			hpte_v = hpte_get_old_v(hptep);
685 
686 			if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) {
687 				/*
688 				 * Invalidate the hpte. NOTE: this also unlocks it
689 				 */
690 
691 				hptep->v = 0;
692 			} else
693 				native_unlock_hpte(hptep);
694 		}
695 		/*
696 		 * We need to do tlb invalidate for all the address, tlbie
697 		 * instruction compares entry_VA in tlb with the VA specified
698 		 * here
699 		 */
700 		tlbie(vpn, psize, actual_psize, ssize, local);
701 	}
702 	local_irq_restore(flags);
703 }
704 #else
705 static void native_hugepage_invalidate(unsigned long vsid,
706 				       unsigned long addr,
707 				       unsigned char *hpte_slot_array,
708 				       int psize, int ssize, int local)
709 {
710 	WARN(1, "%s called without THP support\n", __func__);
711 }
712 #endif
713 
714 static void hpte_decode(struct hash_pte *hpte, unsigned long slot,
715 			int *psize, int *apsize, int *ssize, unsigned long *vpn)
716 {
717 	unsigned long avpn, pteg, vpi;
718 	unsigned long hpte_v = be64_to_cpu(hpte->v);
719 	unsigned long hpte_r = be64_to_cpu(hpte->r);
720 	unsigned long vsid, seg_off;
721 	int size, a_size, shift;
722 	/* Look at the 8 bit LP value */
723 	unsigned int lp = (hpte_r >> LP_SHIFT) & ((1 << LP_BITS) - 1);
724 
725 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
726 		hpte_v = hpte_new_to_old_v(hpte_v, hpte_r);
727 		hpte_r = hpte_new_to_old_r(hpte_r);
728 	}
729 	if (!(hpte_v & HPTE_V_LARGE)) {
730 		size   = MMU_PAGE_4K;
731 		a_size = MMU_PAGE_4K;
732 	} else {
733 		size = hpte_page_sizes[lp] & 0xf;
734 		a_size = hpte_page_sizes[lp] >> 4;
735 	}
736 	/* This works for all page sizes, and for 256M and 1T segments */
737 	*ssize = hpte_v >> HPTE_V_SSIZE_SHIFT;
738 	shift = mmu_psize_defs[size].shift;
739 
740 	avpn = (HPTE_V_AVPN_VAL(hpte_v) & ~mmu_psize_defs[size].avpnm);
741 	pteg = slot / HPTES_PER_GROUP;
742 	if (hpte_v & HPTE_V_SECONDARY)
743 		pteg = ~pteg;
744 
745 	switch (*ssize) {
746 	case MMU_SEGSIZE_256M:
747 		/* We only have 28 - 23 bits of seg_off in avpn */
748 		seg_off = (avpn & 0x1f) << 23;
749 		vsid    =  avpn >> 5;
750 		/* We can find more bits from the pteg value */
751 		if (shift < 23) {
752 			vpi = (vsid ^ pteg) & htab_hash_mask;
753 			seg_off |= vpi << shift;
754 		}
755 		*vpn = vsid << (SID_SHIFT - VPN_SHIFT) | seg_off >> VPN_SHIFT;
756 		break;
757 	case MMU_SEGSIZE_1T:
758 		/* We only have 40 - 23 bits of seg_off in avpn */
759 		seg_off = (avpn & 0x1ffff) << 23;
760 		vsid    = avpn >> 17;
761 		if (shift < 23) {
762 			vpi = (vsid ^ (vsid << 25) ^ pteg) & htab_hash_mask;
763 			seg_off |= vpi << shift;
764 		}
765 		*vpn = vsid << (SID_SHIFT_1T - VPN_SHIFT) | seg_off >> VPN_SHIFT;
766 		break;
767 	default:
768 		*vpn = size = 0;
769 	}
770 	*psize  = size;
771 	*apsize = a_size;
772 }
773 
774 /*
775  * clear all mappings on kexec.  All cpus are in real mode (or they will
776  * be when they isi), and we are the only one left.  We rely on our kernel
777  * mapping being 0xC0's and the hardware ignoring those two real bits.
778  *
779  * This must be called with interrupts disabled.
780  *
781  * Taking the native_tlbie_lock is unsafe here due to the possibility of
782  * lockdep being on. On pre POWER5 hardware, not taking the lock could
783  * cause deadlock. POWER5 and newer not taking the lock is fine. This only
784  * gets called during boot before secondary CPUs have come up and during
785  * crashdump and all bets are off anyway.
786  *
787  * TODO: add batching support when enabled.  remember, no dynamic memory here,
788  * although there is the control page available...
789  */
790 static notrace void native_hpte_clear(void)
791 {
792 	unsigned long vpn = 0;
793 	unsigned long slot, slots;
794 	struct hash_pte *hptep = htab_address;
795 	unsigned long hpte_v;
796 	unsigned long pteg_count;
797 	int psize, apsize, ssize;
798 
799 	pteg_count = htab_hash_mask + 1;
800 
801 	slots = pteg_count * HPTES_PER_GROUP;
802 
803 	for (slot = 0; slot < slots; slot++, hptep++) {
804 		/*
805 		 * we could lock the pte here, but we are the only cpu
806 		 * running,  right?  and for crash dump, we probably
807 		 * don't want to wait for a maybe bad cpu.
808 		 */
809 		hpte_v = be64_to_cpu(hptep->v);
810 
811 		/*
812 		 * Call __tlbie() here rather than tlbie() since we can't take the
813 		 * native_tlbie_lock.
814 		 */
815 		if (hpte_v & HPTE_V_VALID) {
816 			hpte_decode(hptep, slot, &psize, &apsize, &ssize, &vpn);
817 			hptep->v = 0;
818 			___tlbie(vpn, psize, apsize, ssize);
819 		}
820 	}
821 
822 	asm volatile("eieio; tlbsync; ptesync":::"memory");
823 }
824 
825 /*
826  * Batched hash table flush, we batch the tlbie's to avoid taking/releasing
827  * the lock all the time
828  */
829 static void native_flush_hash_range(unsigned long number, int local)
830 {
831 	unsigned long vpn = 0;
832 	unsigned long hash, index, hidx, shift, slot;
833 	struct hash_pte *hptep;
834 	unsigned long hpte_v;
835 	unsigned long want_v;
836 	unsigned long flags;
837 	real_pte_t pte;
838 	struct ppc64_tlb_batch *batch = this_cpu_ptr(&ppc64_tlb_batch);
839 	unsigned long psize = batch->psize;
840 	int ssize = batch->ssize;
841 	int i;
842 	unsigned int use_local;
843 
844 	use_local = local && mmu_has_feature(MMU_FTR_TLBIEL) &&
845 		mmu_psize_defs[psize].tlbiel && !cxl_ctx_in_use();
846 
847 	local_irq_save(flags);
848 
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, vpn, index, shift) {
854 			hash = hpt_hash(vpn, shift, ssize);
855 			hidx = __rpte_to_hidx(pte, index);
856 			if (hidx & _PTEIDX_SECONDARY)
857 				hash = ~hash;
858 			slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
859 			slot += hidx & _PTEIDX_GROUP_IX;
860 			hptep = htab_address + slot;
861 			want_v = hpte_encode_avpn(vpn, psize, ssize);
862 			hpte_v = hpte_get_old_v(hptep);
863 
864 			if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID))
865 				continue;
866 			/* lock and try again */
867 			native_lock_hpte(hptep);
868 			hpte_v = hpte_get_old_v(hptep);
869 
870 			if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID))
871 				native_unlock_hpte(hptep);
872 			else
873 				hptep->v = 0;
874 
875 		} pte_iterate_hashed_end();
876 	}
877 
878 	if (use_local) {
879 		asm volatile("ptesync":::"memory");
880 		for (i = 0; i < number; i++) {
881 			vpn = batch->vpn[i];
882 			pte = batch->pte[i];
883 
884 			pte_iterate_hashed_subpages(pte, psize,
885 						    vpn, index, shift) {
886 				__tlbiel(vpn, psize, psize, ssize);
887 			} pte_iterate_hashed_end();
888 		}
889 		ppc_after_tlbiel_barrier();
890 	} else {
891 		int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
892 
893 		if (lock_tlbie)
894 			raw_spin_lock(&native_tlbie_lock);
895 
896 		asm volatile("ptesync":::"memory");
897 		for (i = 0; i < number; i++) {
898 			vpn = batch->vpn[i];
899 			pte = batch->pte[i];
900 
901 			pte_iterate_hashed_subpages(pte, psize,
902 						    vpn, index, shift) {
903 				__tlbie(vpn, psize, psize, ssize);
904 			} pte_iterate_hashed_end();
905 		}
906 		/*
907 		 * Just do one more with the last used values.
908 		 */
909 		fixup_tlbie_vpn(vpn, psize, psize, ssize);
910 		asm volatile("eieio; tlbsync; ptesync":::"memory");
911 
912 		if (lock_tlbie)
913 			raw_spin_unlock(&native_tlbie_lock);
914 	}
915 
916 	local_irq_restore(flags);
917 }
918 
919 void __init hpte_init_native(void)
920 {
921 	mmu_hash_ops.hpte_invalidate	= native_hpte_invalidate;
922 	mmu_hash_ops.hpte_updatepp	= native_hpte_updatepp;
923 	mmu_hash_ops.hpte_updateboltedpp = native_hpte_updateboltedpp;
924 	mmu_hash_ops.hpte_removebolted = native_hpte_removebolted;
925 	mmu_hash_ops.hpte_insert	= native_hpte_insert;
926 	mmu_hash_ops.hpte_remove	= native_hpte_remove;
927 	mmu_hash_ops.hpte_clear_all	= native_hpte_clear;
928 	mmu_hash_ops.flush_hash_range = native_flush_hash_range;
929 	mmu_hash_ops.hugepage_invalidate   = native_hugepage_invalidate;
930 }
931