1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * PowerPC64 port by Mike Corrigan and Dave Engebretsen
4  *   {mikejc|engebret}@us.ibm.com
5  *
6  *    Copyright (c) 2000 Mike Corrigan <mikejc@us.ibm.com>
7  *
8  * SMP scalability work:
9  *    Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
10  *
11  *    Module name: htab.c
12  *
13  *    Description:
14  *      PowerPC Hashed Page Table functions
15  */
16 
17 #undef DEBUG
18 #undef DEBUG_LOW
19 
20 #define pr_fmt(fmt) "hash-mmu: " fmt
21 #include <linux/spinlock.h>
22 #include <linux/errno.h>
23 #include <linux/sched/mm.h>
24 #include <linux/proc_fs.h>
25 #include <linux/stat.h>
26 #include <linux/sysctl.h>
27 #include <linux/export.h>
28 #include <linux/ctype.h>
29 #include <linux/cache.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/memblock.h>
33 #include <linux/context_tracking.h>
34 #include <linux/libfdt.h>
35 #include <linux/pkeys.h>
36 #include <linux/hugetlb.h>
37 #include <linux/cpu.h>
38 #include <linux/pgtable.h>
39 #include <linux/debugfs.h>
40 #include <linux/random.h>
41 #include <linux/elf-randomize.h>
42 #include <linux/of_fdt.h>
43 
44 #include <asm/interrupt.h>
45 #include <asm/processor.h>
46 #include <asm/mmu.h>
47 #include <asm/mmu_context.h>
48 #include <asm/page.h>
49 #include <asm/types.h>
50 #include <linux/uaccess.h>
51 #include <asm/machdep.h>
52 #include <asm/io.h>
53 #include <asm/eeh.h>
54 #include <asm/tlb.h>
55 #include <asm/cacheflush.h>
56 #include <asm/cputable.h>
57 #include <asm/sections.h>
58 #include <asm/copro.h>
59 #include <asm/udbg.h>
60 #include <asm/code-patching.h>
61 #include <asm/fadump.h>
62 #include <asm/firmware.h>
63 #include <asm/tm.h>
64 #include <asm/trace.h>
65 #include <asm/ps3.h>
66 #include <asm/pte-walk.h>
67 #include <asm/asm-prototypes.h>
68 #include <asm/ultravisor.h>
69 
70 #include <mm/mmu_decl.h>
71 
72 #include "internal.h"
73 
74 
75 #ifdef DEBUG
76 #define DBG(fmt...) udbg_printf(fmt)
77 #else
78 #define DBG(fmt...)
79 #endif
80 
81 #ifdef DEBUG_LOW
82 #define DBG_LOW(fmt...) udbg_printf(fmt)
83 #else
84 #define DBG_LOW(fmt...)
85 #endif
86 
87 #define KB (1024)
88 #define MB (1024*KB)
89 #define GB (1024L*MB)
90 
91 /*
92  * Note:  pte   --> Linux PTE
93  *        HPTE  --> PowerPC Hashed Page Table Entry
94  *
95  * Execution context:
96  *   htab_initialize is called with the MMU off (of course), but
97  *   the kernel has been copied down to zero so it can directly
98  *   reference global data.  At this point it is very difficult
99  *   to print debug info.
100  *
101  */
102 
103 static unsigned long _SDR1;
104 
105 u8 hpte_page_sizes[1 << LP_BITS];
106 EXPORT_SYMBOL_GPL(hpte_page_sizes);
107 
108 struct hash_pte *htab_address;
109 unsigned long htab_size_bytes;
110 unsigned long htab_hash_mask;
111 EXPORT_SYMBOL_GPL(htab_hash_mask);
112 int mmu_linear_psize = MMU_PAGE_4K;
113 EXPORT_SYMBOL_GPL(mmu_linear_psize);
114 int mmu_virtual_psize = MMU_PAGE_4K;
115 int mmu_vmalloc_psize = MMU_PAGE_4K;
116 EXPORT_SYMBOL_GPL(mmu_vmalloc_psize);
117 int mmu_io_psize = MMU_PAGE_4K;
118 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
119 EXPORT_SYMBOL_GPL(mmu_kernel_ssize);
120 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
121 u16 mmu_slb_size = 64;
122 EXPORT_SYMBOL_GPL(mmu_slb_size);
123 #ifdef CONFIG_PPC_64K_PAGES
124 int mmu_ci_restrictions;
125 #endif
126 static u8 *linear_map_hash_slots;
127 static unsigned long linear_map_hash_count;
128 struct mmu_hash_ops mmu_hash_ops;
129 EXPORT_SYMBOL(mmu_hash_ops);
130 
131 /*
132  * These are definitions of page sizes arrays to be used when none
133  * is provided by the firmware.
134  */
135 
136 /*
137  * Fallback (4k pages only)
138  */
139 static struct mmu_psize_def mmu_psize_defaults[] = {
140 	[MMU_PAGE_4K] = {
141 		.shift	= 12,
142 		.sllp	= 0,
143 		.penc   = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
144 		.avpnm	= 0,
145 		.tlbiel = 0,
146 	},
147 };
148 
149 /*
150  * POWER4, GPUL, POWER5
151  *
152  * Support for 16Mb large pages
153  */
154 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
155 	[MMU_PAGE_4K] = {
156 		.shift	= 12,
157 		.sllp	= 0,
158 		.penc   = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
159 		.avpnm	= 0,
160 		.tlbiel = 1,
161 	},
162 	[MMU_PAGE_16M] = {
163 		.shift	= 24,
164 		.sllp	= SLB_VSID_L,
165 		.penc   = {[0 ... MMU_PAGE_16M - 1] = -1, [MMU_PAGE_16M] = 0,
166 			    [MMU_PAGE_16M + 1 ... MMU_PAGE_COUNT - 1] = -1 },
167 		.avpnm	= 0x1UL,
168 		.tlbiel = 0,
169 	},
170 };
171 
172 static inline void tlbiel_hash_set_isa206(unsigned int set, unsigned int is)
173 {
174 	unsigned long rb;
175 
176 	rb = (set << PPC_BITLSHIFT(51)) | (is << PPC_BITLSHIFT(53));
177 
178 	asm volatile("tlbiel %0" : : "r" (rb));
179 }
180 
181 /*
182  * tlbiel instruction for hash, set invalidation
183  * i.e., r=1 and is=01 or is=10 or is=11
184  */
185 static __always_inline void tlbiel_hash_set_isa300(unsigned int set, unsigned int is,
186 					unsigned int pid,
187 					unsigned int ric, unsigned int prs)
188 {
189 	unsigned long rb;
190 	unsigned long rs;
191 	unsigned int r = 0; /* hash format */
192 
193 	rb = (set << PPC_BITLSHIFT(51)) | (is << PPC_BITLSHIFT(53));
194 	rs = ((unsigned long)pid << PPC_BITLSHIFT(31));
195 
196 	asm volatile(PPC_TLBIEL(%0, %1, %2, %3, %4)
197 		     : : "r"(rb), "r"(rs), "i"(ric), "i"(prs), "i"(r)
198 		     : "memory");
199 }
200 
201 
202 static void tlbiel_all_isa206(unsigned int num_sets, unsigned int is)
203 {
204 	unsigned int set;
205 
206 	asm volatile("ptesync": : :"memory");
207 
208 	for (set = 0; set < num_sets; set++)
209 		tlbiel_hash_set_isa206(set, is);
210 
211 	ppc_after_tlbiel_barrier();
212 }
213 
214 static void tlbiel_all_isa300(unsigned int num_sets, unsigned int is)
215 {
216 	unsigned int set;
217 
218 	asm volatile("ptesync": : :"memory");
219 
220 	/*
221 	 * Flush the partition table cache if this is HV mode.
222 	 */
223 	if (early_cpu_has_feature(CPU_FTR_HVMODE))
224 		tlbiel_hash_set_isa300(0, is, 0, 2, 0);
225 
226 	/*
227 	 * Now invalidate the process table cache. UPRT=0 HPT modes (what
228 	 * current hardware implements) do not use the process table, but
229 	 * add the flushes anyway.
230 	 *
231 	 * From ISA v3.0B p. 1078:
232 	 *     The following forms are invalid.
233 	 *      * PRS=1, R=0, and RIC!=2 (The only process-scoped
234 	 *        HPT caching is of the Process Table.)
235 	 */
236 	tlbiel_hash_set_isa300(0, is, 0, 2, 1);
237 
238 	/*
239 	 * Then flush the sets of the TLB proper. Hash mode uses
240 	 * partition scoped TLB translations, which may be flushed
241 	 * in !HV mode.
242 	 */
243 	for (set = 0; set < num_sets; set++)
244 		tlbiel_hash_set_isa300(set, is, 0, 0, 0);
245 
246 	ppc_after_tlbiel_barrier();
247 
248 	asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT "; isync" : : :"memory");
249 }
250 
251 void hash__tlbiel_all(unsigned int action)
252 {
253 	unsigned int is;
254 
255 	switch (action) {
256 	case TLB_INVAL_SCOPE_GLOBAL:
257 		is = 3;
258 		break;
259 	case TLB_INVAL_SCOPE_LPID:
260 		is = 2;
261 		break;
262 	default:
263 		BUG();
264 	}
265 
266 	if (early_cpu_has_feature(CPU_FTR_ARCH_300))
267 		tlbiel_all_isa300(POWER9_TLB_SETS_HASH, is);
268 	else if (early_cpu_has_feature(CPU_FTR_ARCH_207S))
269 		tlbiel_all_isa206(POWER8_TLB_SETS, is);
270 	else if (early_cpu_has_feature(CPU_FTR_ARCH_206))
271 		tlbiel_all_isa206(POWER7_TLB_SETS, is);
272 	else
273 		WARN(1, "%s called on pre-POWER7 CPU\n", __func__);
274 }
275 
276 /*
277  * 'R' and 'C' update notes:
278  *  - Under pHyp or KVM, the updatepp path will not set C, thus it *will*
279  *     create writeable HPTEs without C set, because the hcall H_PROTECT
280  *     that we use in that case will not update C
281  *  - The above is however not a problem, because we also don't do that
282  *     fancy "no flush" variant of eviction and we use H_REMOVE which will
283  *     do the right thing and thus we don't have the race I described earlier
284  *
285  *    - Under bare metal,  we do have the race, so we need R and C set
286  *    - We make sure R is always set and never lost
287  *    - C is _PAGE_DIRTY, and *should* always be set for a writeable mapping
288  */
289 unsigned long htab_convert_pte_flags(unsigned long pteflags, unsigned long flags)
290 {
291 	unsigned long rflags = 0;
292 
293 	/* _PAGE_EXEC -> NOEXEC */
294 	if ((pteflags & _PAGE_EXEC) == 0)
295 		rflags |= HPTE_R_N;
296 	/*
297 	 * PPP bits:
298 	 * Linux uses slb key 0 for kernel and 1 for user.
299 	 * kernel RW areas are mapped with PPP=0b000
300 	 * User area is mapped with PPP=0b010 for read/write
301 	 * or PPP=0b011 for read-only (including writeable but clean pages).
302 	 */
303 	if (pteflags & _PAGE_PRIVILEGED) {
304 		/*
305 		 * Kernel read only mapped with ppp bits 0b110
306 		 */
307 		if (!(pteflags & _PAGE_WRITE)) {
308 			if (mmu_has_feature(MMU_FTR_KERNEL_RO))
309 				rflags |= (HPTE_R_PP0 | 0x2);
310 			else
311 				rflags |= 0x3;
312 		}
313 	} else {
314 		if (pteflags & _PAGE_RWX)
315 			rflags |= 0x2;
316 		if (!((pteflags & _PAGE_WRITE) && (pteflags & _PAGE_DIRTY)))
317 			rflags |= 0x1;
318 	}
319 	/*
320 	 * We can't allow hardware to update hpte bits. Hence always
321 	 * set 'R' bit and set 'C' if it is a write fault
322 	 */
323 	rflags |=  HPTE_R_R;
324 
325 	if (pteflags & _PAGE_DIRTY)
326 		rflags |= HPTE_R_C;
327 	/*
328 	 * Add in WIG bits
329 	 */
330 
331 	if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_TOLERANT)
332 		rflags |= HPTE_R_I;
333 	else if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_NON_IDEMPOTENT)
334 		rflags |= (HPTE_R_I | HPTE_R_G);
335 	else if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_SAO)
336 		rflags |= (HPTE_R_W | HPTE_R_I | HPTE_R_M);
337 	else
338 		/*
339 		 * Add memory coherence if cache inhibited is not set
340 		 */
341 		rflags |= HPTE_R_M;
342 
343 	rflags |= pte_to_hpte_pkey_bits(pteflags, flags);
344 	return rflags;
345 }
346 
347 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
348 		      unsigned long pstart, unsigned long prot,
349 		      int psize, int ssize)
350 {
351 	unsigned long vaddr, paddr;
352 	unsigned int step, shift;
353 	int ret = 0;
354 
355 	shift = mmu_psize_defs[psize].shift;
356 	step = 1 << shift;
357 
358 	prot = htab_convert_pte_flags(prot, HPTE_USE_KERNEL_KEY);
359 
360 	DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
361 	    vstart, vend, pstart, prot, psize, ssize);
362 
363 	/* Carefully map only the possible range */
364 	vaddr = ALIGN(vstart, step);
365 	paddr = ALIGN(pstart, step);
366 	vend  = ALIGN_DOWN(vend, step);
367 
368 	for (; vaddr < vend; vaddr += step, paddr += step) {
369 		unsigned long hash, hpteg;
370 		unsigned long vsid = get_kernel_vsid(vaddr, ssize);
371 		unsigned long vpn  = hpt_vpn(vaddr, vsid, ssize);
372 		unsigned long tprot = prot;
373 		bool secondary_hash = false;
374 
375 		/*
376 		 * If we hit a bad address return error.
377 		 */
378 		if (!vsid)
379 			return -1;
380 		/* Make kernel text executable */
381 		if (overlaps_kernel_text(vaddr, vaddr + step))
382 			tprot &= ~HPTE_R_N;
383 
384 		/*
385 		 * If relocatable, check if it overlaps interrupt vectors that
386 		 * are copied down to real 0. For relocatable kernel
387 		 * (e.g. kdump case) we copy interrupt vectors down to real
388 		 * address 0. Mark that region as executable. This is
389 		 * because on p8 system with relocation on exception feature
390 		 * enabled, exceptions are raised with MMU (IR=DR=1) ON. Hence
391 		 * in order to execute the interrupt handlers in virtual
392 		 * mode the vector region need to be marked as executable.
393 		 */
394 		if ((PHYSICAL_START > MEMORY_START) &&
395 			overlaps_interrupt_vector_text(vaddr, vaddr + step))
396 				tprot &= ~HPTE_R_N;
397 
398 		hash = hpt_hash(vpn, shift, ssize);
399 		hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
400 
401 		BUG_ON(!mmu_hash_ops.hpte_insert);
402 repeat:
403 		ret = mmu_hash_ops.hpte_insert(hpteg, vpn, paddr, tprot,
404 					       HPTE_V_BOLTED, psize, psize,
405 					       ssize);
406 		if (ret == -1) {
407 			/*
408 			 * Try to keep bolted entries in primary.
409 			 * Remove non bolted entries and try insert again
410 			 */
411 			ret = mmu_hash_ops.hpte_remove(hpteg);
412 			if (ret != -1)
413 				ret = mmu_hash_ops.hpte_insert(hpteg, vpn, paddr, tprot,
414 							       HPTE_V_BOLTED, psize, psize,
415 							       ssize);
416 			if (ret == -1 && !secondary_hash) {
417 				secondary_hash = true;
418 				hpteg = ((~hash & htab_hash_mask) * HPTES_PER_GROUP);
419 				goto repeat;
420 			}
421 		}
422 
423 		if (ret < 0)
424 			break;
425 
426 		cond_resched();
427 		if (debug_pagealloc_enabled_or_kfence() &&
428 			(paddr >> PAGE_SHIFT) < linear_map_hash_count)
429 			linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
430 	}
431 	return ret < 0 ? ret : 0;
432 }
433 
434 int htab_remove_mapping(unsigned long vstart, unsigned long vend,
435 		      int psize, int ssize)
436 {
437 	unsigned long vaddr, time_limit;
438 	unsigned int step, shift;
439 	int rc;
440 	int ret = 0;
441 
442 	shift = mmu_psize_defs[psize].shift;
443 	step = 1 << shift;
444 
445 	if (!mmu_hash_ops.hpte_removebolted)
446 		return -ENODEV;
447 
448 	/* Unmap the full range specificied */
449 	vaddr = ALIGN_DOWN(vstart, step);
450 	time_limit = jiffies + HZ;
451 
452 	for (;vaddr < vend; vaddr += step) {
453 		rc = mmu_hash_ops.hpte_removebolted(vaddr, psize, ssize);
454 
455 		/*
456 		 * For large number of mappings introduce a cond_resched()
457 		 * to prevent softlockup warnings.
458 		 */
459 		if (time_after(jiffies, time_limit)) {
460 			cond_resched();
461 			time_limit = jiffies + HZ;
462 		}
463 		if (rc == -ENOENT) {
464 			ret = -ENOENT;
465 			continue;
466 		}
467 		if (rc < 0)
468 			return rc;
469 	}
470 
471 	return ret;
472 }
473 
474 static bool disable_1tb_segments __ro_after_init;
475 
476 static int __init parse_disable_1tb_segments(char *p)
477 {
478 	disable_1tb_segments = true;
479 	return 0;
480 }
481 early_param("disable_1tb_segments", parse_disable_1tb_segments);
482 
483 bool stress_hpt_enabled __initdata;
484 
485 static int __init parse_stress_hpt(char *p)
486 {
487 	stress_hpt_enabled = true;
488 	return 0;
489 }
490 early_param("stress_hpt", parse_stress_hpt);
491 
492 __ro_after_init DEFINE_STATIC_KEY_FALSE(stress_hpt_key);
493 
494 /*
495  * per-CPU array allocated if we enable stress_hpt.
496  */
497 #define STRESS_MAX_GROUPS 16
498 struct stress_hpt_struct {
499 	unsigned long last_group[STRESS_MAX_GROUPS];
500 };
501 
502 static inline int stress_nr_groups(void)
503 {
504 	/*
505 	 * LPAR H_REMOVE flushes TLB, so need some number > 1 of entries
506 	 * to allow practical forward progress. Bare metal returns 1, which
507 	 * seems to help uncover more bugs.
508 	 */
509 	if (firmware_has_feature(FW_FEATURE_LPAR))
510 		return STRESS_MAX_GROUPS;
511 	else
512 		return 1;
513 }
514 
515 static struct stress_hpt_struct *stress_hpt_struct;
516 
517 static int __init htab_dt_scan_seg_sizes(unsigned long node,
518 					 const char *uname, int depth,
519 					 void *data)
520 {
521 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
522 	const __be32 *prop;
523 	int size = 0;
524 
525 	/* We are scanning "cpu" nodes only */
526 	if (type == NULL || strcmp(type, "cpu") != 0)
527 		return 0;
528 
529 	prop = of_get_flat_dt_prop(node, "ibm,processor-segment-sizes", &size);
530 	if (prop == NULL)
531 		return 0;
532 	for (; size >= 4; size -= 4, ++prop) {
533 		if (be32_to_cpu(prop[0]) == 40) {
534 			DBG("1T segment support detected\n");
535 
536 			if (disable_1tb_segments) {
537 				DBG("1T segments disabled by command line\n");
538 				break;
539 			}
540 
541 			cur_cpu_spec->mmu_features |= MMU_FTR_1T_SEGMENT;
542 			return 1;
543 		}
544 	}
545 	cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
546 	return 0;
547 }
548 
549 static int __init get_idx_from_shift(unsigned int shift)
550 {
551 	int idx = -1;
552 
553 	switch (shift) {
554 	case 0xc:
555 		idx = MMU_PAGE_4K;
556 		break;
557 	case 0x10:
558 		idx = MMU_PAGE_64K;
559 		break;
560 	case 0x14:
561 		idx = MMU_PAGE_1M;
562 		break;
563 	case 0x18:
564 		idx = MMU_PAGE_16M;
565 		break;
566 	case 0x22:
567 		idx = MMU_PAGE_16G;
568 		break;
569 	}
570 	return idx;
571 }
572 
573 static int __init htab_dt_scan_page_sizes(unsigned long node,
574 					  const char *uname, int depth,
575 					  void *data)
576 {
577 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
578 	const __be32 *prop;
579 	int size = 0;
580 
581 	/* We are scanning "cpu" nodes only */
582 	if (type == NULL || strcmp(type, "cpu") != 0)
583 		return 0;
584 
585 	prop = of_get_flat_dt_prop(node, "ibm,segment-page-sizes", &size);
586 	if (!prop)
587 		return 0;
588 
589 	pr_info("Page sizes from device-tree:\n");
590 	size /= 4;
591 	cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
592 	while(size > 0) {
593 		unsigned int base_shift = be32_to_cpu(prop[0]);
594 		unsigned int slbenc = be32_to_cpu(prop[1]);
595 		unsigned int lpnum = be32_to_cpu(prop[2]);
596 		struct mmu_psize_def *def;
597 		int idx, base_idx;
598 
599 		size -= 3; prop += 3;
600 		base_idx = get_idx_from_shift(base_shift);
601 		if (base_idx < 0) {
602 			/* skip the pte encoding also */
603 			prop += lpnum * 2; size -= lpnum * 2;
604 			continue;
605 		}
606 		def = &mmu_psize_defs[base_idx];
607 		if (base_idx == MMU_PAGE_16M)
608 			cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
609 
610 		def->shift = base_shift;
611 		if (base_shift <= 23)
612 			def->avpnm = 0;
613 		else
614 			def->avpnm = (1 << (base_shift - 23)) - 1;
615 		def->sllp = slbenc;
616 		/*
617 		 * We don't know for sure what's up with tlbiel, so
618 		 * for now we only set it for 4K and 64K pages
619 		 */
620 		if (base_idx == MMU_PAGE_4K || base_idx == MMU_PAGE_64K)
621 			def->tlbiel = 1;
622 		else
623 			def->tlbiel = 0;
624 
625 		while (size > 0 && lpnum) {
626 			unsigned int shift = be32_to_cpu(prop[0]);
627 			int penc  = be32_to_cpu(prop[1]);
628 
629 			prop += 2; size -= 2;
630 			lpnum--;
631 
632 			idx = get_idx_from_shift(shift);
633 			if (idx < 0)
634 				continue;
635 
636 			if (penc == -1)
637 				pr_err("Invalid penc for base_shift=%d "
638 				       "shift=%d\n", base_shift, shift);
639 
640 			def->penc[idx] = penc;
641 			pr_info("base_shift=%d: shift=%d, sllp=0x%04lx,"
642 				" avpnm=0x%08lx, tlbiel=%d, penc=%d\n",
643 				base_shift, shift, def->sllp,
644 				def->avpnm, def->tlbiel, def->penc[idx]);
645 		}
646 	}
647 
648 	return 1;
649 }
650 
651 #ifdef CONFIG_HUGETLB_PAGE
652 /*
653  * Scan for 16G memory blocks that have been set aside for huge pages
654  * and reserve those blocks for 16G huge pages.
655  */
656 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
657 					const char *uname, int depth,
658 					void *data) {
659 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
660 	const __be64 *addr_prop;
661 	const __be32 *page_count_prop;
662 	unsigned int expected_pages;
663 	long unsigned int phys_addr;
664 	long unsigned int block_size;
665 
666 	/* We are scanning "memory" nodes only */
667 	if (type == NULL || strcmp(type, "memory") != 0)
668 		return 0;
669 
670 	/*
671 	 * This property is the log base 2 of the number of virtual pages that
672 	 * will represent this memory block.
673 	 */
674 	page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
675 	if (page_count_prop == NULL)
676 		return 0;
677 	expected_pages = (1 << be32_to_cpu(page_count_prop[0]));
678 	addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
679 	if (addr_prop == NULL)
680 		return 0;
681 	phys_addr = be64_to_cpu(addr_prop[0]);
682 	block_size = be64_to_cpu(addr_prop[1]);
683 	if (block_size != (16 * GB))
684 		return 0;
685 	printk(KERN_INFO "Huge page(16GB) memory: "
686 			"addr = 0x%lX size = 0x%lX pages = %d\n",
687 			phys_addr, block_size, expected_pages);
688 	if (phys_addr + block_size * expected_pages <= memblock_end_of_DRAM()) {
689 		memblock_reserve(phys_addr, block_size * expected_pages);
690 		pseries_add_gpage(phys_addr, block_size, expected_pages);
691 	}
692 	return 0;
693 }
694 #endif /* CONFIG_HUGETLB_PAGE */
695 
696 static void __init mmu_psize_set_default_penc(void)
697 {
698 	int bpsize, apsize;
699 	for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
700 		for (apsize = 0; apsize < MMU_PAGE_COUNT; apsize++)
701 			mmu_psize_defs[bpsize].penc[apsize] = -1;
702 }
703 
704 #ifdef CONFIG_PPC_64K_PAGES
705 
706 static bool __init might_have_hea(void)
707 {
708 	/*
709 	 * The HEA ethernet adapter requires awareness of the
710 	 * GX bus. Without that awareness we can easily assume
711 	 * we will never see an HEA ethernet device.
712 	 */
713 #ifdef CONFIG_IBMEBUS
714 	return !cpu_has_feature(CPU_FTR_ARCH_207S) &&
715 		firmware_has_feature(FW_FEATURE_SPLPAR);
716 #else
717 	return false;
718 #endif
719 }
720 
721 #endif /* #ifdef CONFIG_PPC_64K_PAGES */
722 
723 static void __init htab_scan_page_sizes(void)
724 {
725 	int rc;
726 
727 	/* se the invalid penc to -1 */
728 	mmu_psize_set_default_penc();
729 
730 	/* Default to 4K pages only */
731 	memcpy(mmu_psize_defs, mmu_psize_defaults,
732 	       sizeof(mmu_psize_defaults));
733 
734 	/*
735 	 * Try to find the available page sizes in the device-tree
736 	 */
737 	rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
738 	if (rc == 0 && early_mmu_has_feature(MMU_FTR_16M_PAGE)) {
739 		/*
740 		 * Nothing in the device-tree, but the CPU supports 16M pages,
741 		 * so let's fallback on a known size list for 16M capable CPUs.
742 		 */
743 		memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
744 		       sizeof(mmu_psize_defaults_gp));
745 	}
746 
747 #ifdef CONFIG_HUGETLB_PAGE
748 	if (!hugetlb_disabled && !early_radix_enabled() ) {
749 		/* Reserve 16G huge page memory sections for huge pages */
750 		of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
751 	}
752 #endif /* CONFIG_HUGETLB_PAGE */
753 }
754 
755 /*
756  * Fill in the hpte_page_sizes[] array.
757  * We go through the mmu_psize_defs[] array looking for all the
758  * supported base/actual page size combinations.  Each combination
759  * has a unique pagesize encoding (penc) value in the low bits of
760  * the LP field of the HPTE.  For actual page sizes less than 1MB,
761  * some of the upper LP bits are used for RPN bits, meaning that
762  * we need to fill in several entries in hpte_page_sizes[].
763  *
764  * In diagrammatic form, with r = RPN bits and z = page size bits:
765  *        PTE LP     actual page size
766  *    rrrr rrrz		>=8KB
767  *    rrrr rrzz		>=16KB
768  *    rrrr rzzz		>=32KB
769  *    rrrr zzzz		>=64KB
770  *    ...
771  *
772  * The zzzz bits are implementation-specific but are chosen so that
773  * no encoding for a larger page size uses the same value in its
774  * low-order N bits as the encoding for the 2^(12+N) byte page size
775  * (if it exists).
776  */
777 static void __init init_hpte_page_sizes(void)
778 {
779 	long int ap, bp;
780 	long int shift, penc;
781 
782 	for (bp = 0; bp < MMU_PAGE_COUNT; ++bp) {
783 		if (!mmu_psize_defs[bp].shift)
784 			continue;	/* not a supported page size */
785 		for (ap = bp; ap < MMU_PAGE_COUNT; ++ap) {
786 			penc = mmu_psize_defs[bp].penc[ap];
787 			if (penc == -1 || !mmu_psize_defs[ap].shift)
788 				continue;
789 			shift = mmu_psize_defs[ap].shift - LP_SHIFT;
790 			if (shift <= 0)
791 				continue;	/* should never happen */
792 			/*
793 			 * For page sizes less than 1MB, this loop
794 			 * replicates the entry for all possible values
795 			 * of the rrrr bits.
796 			 */
797 			while (penc < (1 << LP_BITS)) {
798 				hpte_page_sizes[penc] = (ap << 4) | bp;
799 				penc += 1 << shift;
800 			}
801 		}
802 	}
803 }
804 
805 static void __init htab_init_page_sizes(void)
806 {
807 	bool aligned = true;
808 	init_hpte_page_sizes();
809 
810 	if (!debug_pagealloc_enabled_or_kfence()) {
811 		/*
812 		 * Pick a size for the linear mapping. Currently, we only
813 		 * support 16M, 1M and 4K which is the default
814 		 */
815 		if (IS_ENABLED(CONFIG_STRICT_KERNEL_RWX) &&
816 		    (unsigned long)_stext % 0x1000000) {
817 			if (mmu_psize_defs[MMU_PAGE_16M].shift)
818 				pr_warn("Kernel not 16M aligned, disabling 16M linear map alignment\n");
819 			aligned = false;
820 		}
821 
822 		if (mmu_psize_defs[MMU_PAGE_16M].shift && aligned)
823 			mmu_linear_psize = MMU_PAGE_16M;
824 		else if (mmu_psize_defs[MMU_PAGE_1M].shift)
825 			mmu_linear_psize = MMU_PAGE_1M;
826 	}
827 
828 #ifdef CONFIG_PPC_64K_PAGES
829 	/*
830 	 * Pick a size for the ordinary pages. Default is 4K, we support
831 	 * 64K for user mappings and vmalloc if supported by the processor.
832 	 * We only use 64k for ioremap if the processor
833 	 * (and firmware) support cache-inhibited large pages.
834 	 * If not, we use 4k and set mmu_ci_restrictions so that
835 	 * hash_page knows to switch processes that use cache-inhibited
836 	 * mappings to 4k pages.
837 	 */
838 	if (mmu_psize_defs[MMU_PAGE_64K].shift) {
839 		mmu_virtual_psize = MMU_PAGE_64K;
840 		mmu_vmalloc_psize = MMU_PAGE_64K;
841 		if (mmu_linear_psize == MMU_PAGE_4K)
842 			mmu_linear_psize = MMU_PAGE_64K;
843 		if (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
844 			/*
845 			 * When running on pSeries using 64k pages for ioremap
846 			 * would stop us accessing the HEA ethernet. So if we
847 			 * have the chance of ever seeing one, stay at 4k.
848 			 */
849 			if (!might_have_hea())
850 				mmu_io_psize = MMU_PAGE_64K;
851 		} else
852 			mmu_ci_restrictions = 1;
853 	}
854 #endif /* CONFIG_PPC_64K_PAGES */
855 
856 #ifdef CONFIG_SPARSEMEM_VMEMMAP
857 	/*
858 	 * We try to use 16M pages for vmemmap if that is supported
859 	 * and we have at least 1G of RAM at boot
860 	 */
861 	if (mmu_psize_defs[MMU_PAGE_16M].shift &&
862 	    memblock_phys_mem_size() >= 0x40000000)
863 		mmu_vmemmap_psize = MMU_PAGE_16M;
864 	else
865 		mmu_vmemmap_psize = mmu_virtual_psize;
866 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
867 
868 	printk(KERN_DEBUG "Page orders: linear mapping = %d, "
869 	       "virtual = %d, io = %d"
870 #ifdef CONFIG_SPARSEMEM_VMEMMAP
871 	       ", vmemmap = %d"
872 #endif
873 	       "\n",
874 	       mmu_psize_defs[mmu_linear_psize].shift,
875 	       mmu_psize_defs[mmu_virtual_psize].shift,
876 	       mmu_psize_defs[mmu_io_psize].shift
877 #ifdef CONFIG_SPARSEMEM_VMEMMAP
878 	       ,mmu_psize_defs[mmu_vmemmap_psize].shift
879 #endif
880 	       );
881 }
882 
883 static int __init htab_dt_scan_pftsize(unsigned long node,
884 				       const char *uname, int depth,
885 				       void *data)
886 {
887 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
888 	const __be32 *prop;
889 
890 	/* We are scanning "cpu" nodes only */
891 	if (type == NULL || strcmp(type, "cpu") != 0)
892 		return 0;
893 
894 	prop = of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
895 	if (prop != NULL) {
896 		/* pft_size[0] is the NUMA CEC cookie */
897 		ppc64_pft_size = be32_to_cpu(prop[1]);
898 		return 1;
899 	}
900 	return 0;
901 }
902 
903 unsigned htab_shift_for_mem_size(unsigned long mem_size)
904 {
905 	unsigned memshift = __ilog2(mem_size);
906 	unsigned pshift = mmu_psize_defs[mmu_virtual_psize].shift;
907 	unsigned pteg_shift;
908 
909 	/* round mem_size up to next power of 2 */
910 	if ((1UL << memshift) < mem_size)
911 		memshift += 1;
912 
913 	/* aim for 2 pages / pteg */
914 	pteg_shift = memshift - (pshift + 1);
915 
916 	/*
917 	 * 2^11 PTEGS of 128 bytes each, ie. 2^18 bytes is the minimum htab
918 	 * size permitted by the architecture.
919 	 */
920 	return max(pteg_shift + 7, 18U);
921 }
922 
923 static unsigned long __init htab_get_table_size(void)
924 {
925 	/*
926 	 * If hash size isn't already provided by the platform, we try to
927 	 * retrieve it from the device-tree. If it's not there neither, we
928 	 * calculate it now based on the total RAM size
929 	 */
930 	if (ppc64_pft_size == 0)
931 		of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
932 	if (ppc64_pft_size)
933 		return 1UL << ppc64_pft_size;
934 
935 	return 1UL << htab_shift_for_mem_size(memblock_phys_mem_size());
936 }
937 
938 #ifdef CONFIG_MEMORY_HOTPLUG
939 static int resize_hpt_for_hotplug(unsigned long new_mem_size)
940 {
941 	unsigned target_hpt_shift;
942 
943 	if (!mmu_hash_ops.resize_hpt)
944 		return 0;
945 
946 	target_hpt_shift = htab_shift_for_mem_size(new_mem_size);
947 
948 	/*
949 	 * To avoid lots of HPT resizes if memory size is fluctuating
950 	 * across a boundary, we deliberately have some hysterisis
951 	 * here: we immediately increase the HPT size if the target
952 	 * shift exceeds the current shift, but we won't attempt to
953 	 * reduce unless the target shift is at least 2 below the
954 	 * current shift
955 	 */
956 	if (target_hpt_shift > ppc64_pft_size ||
957 	    target_hpt_shift < ppc64_pft_size - 1)
958 		return mmu_hash_ops.resize_hpt(target_hpt_shift);
959 
960 	return 0;
961 }
962 
963 int hash__create_section_mapping(unsigned long start, unsigned long end,
964 				 int nid, pgprot_t prot)
965 {
966 	int rc;
967 
968 	if (end >= H_VMALLOC_START) {
969 		pr_warn("Outside the supported range\n");
970 		return -1;
971 	}
972 
973 	resize_hpt_for_hotplug(memblock_phys_mem_size());
974 
975 	rc = htab_bolt_mapping(start, end, __pa(start),
976 			       pgprot_val(prot), mmu_linear_psize,
977 			       mmu_kernel_ssize);
978 
979 	if (rc < 0) {
980 		int rc2 = htab_remove_mapping(start, end, mmu_linear_psize,
981 					      mmu_kernel_ssize);
982 		BUG_ON(rc2 && (rc2 != -ENOENT));
983 	}
984 	return rc;
985 }
986 
987 int hash__remove_section_mapping(unsigned long start, unsigned long end)
988 {
989 	int rc = htab_remove_mapping(start, end, mmu_linear_psize,
990 				     mmu_kernel_ssize);
991 
992 	if (resize_hpt_for_hotplug(memblock_phys_mem_size()) == -ENOSPC)
993 		pr_warn("Hash collision while resizing HPT\n");
994 
995 	return rc;
996 }
997 #endif /* CONFIG_MEMORY_HOTPLUG */
998 
999 static void __init hash_init_partition_table(phys_addr_t hash_table,
1000 					     unsigned long htab_size)
1001 {
1002 	mmu_partition_table_init();
1003 
1004 	/*
1005 	 * PS field (VRMA page size) is not used for LPID 0, hence set to 0.
1006 	 * For now, UPRT is 0 and we have no segment table.
1007 	 */
1008 	htab_size =  __ilog2(htab_size) - 18;
1009 	mmu_partition_table_set_entry(0, hash_table | htab_size, 0, false);
1010 	pr_info("Partition table %p\n", partition_tb);
1011 }
1012 
1013 void hpt_clear_stress(void);
1014 static struct timer_list stress_hpt_timer;
1015 static void stress_hpt_timer_fn(struct timer_list *timer)
1016 {
1017 	int next_cpu;
1018 
1019 	hpt_clear_stress();
1020 	if (!firmware_has_feature(FW_FEATURE_LPAR))
1021 		tlbiel_all();
1022 
1023 	next_cpu = cpumask_next(raw_smp_processor_id(), cpu_online_mask);
1024 	if (next_cpu >= nr_cpu_ids)
1025 		next_cpu = cpumask_first(cpu_online_mask);
1026 	stress_hpt_timer.expires = jiffies + msecs_to_jiffies(10);
1027 	add_timer_on(&stress_hpt_timer, next_cpu);
1028 }
1029 
1030 static void __init htab_initialize(void)
1031 {
1032 	unsigned long table;
1033 	unsigned long pteg_count;
1034 	unsigned long prot;
1035 	phys_addr_t base = 0, size = 0, end;
1036 	u64 i;
1037 
1038 	DBG(" -> htab_initialize()\n");
1039 
1040 	if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
1041 		mmu_kernel_ssize = MMU_SEGSIZE_1T;
1042 		mmu_highuser_ssize = MMU_SEGSIZE_1T;
1043 		printk(KERN_INFO "Using 1TB segments\n");
1044 	}
1045 
1046 	if (stress_slb_enabled)
1047 		static_branch_enable(&stress_slb_key);
1048 
1049 	if (stress_hpt_enabled) {
1050 		unsigned long tmp;
1051 		static_branch_enable(&stress_hpt_key);
1052 		// Too early to use nr_cpu_ids, so use NR_CPUS
1053 		tmp = memblock_phys_alloc_range(sizeof(struct stress_hpt_struct) * NR_CPUS,
1054 						0, 0, MEMBLOCK_ALLOC_ANYWHERE);
1055 		memset((void *)tmp, 0xff, sizeof(struct stress_hpt_struct) * NR_CPUS);
1056 		stress_hpt_struct = __va(tmp);
1057 
1058 		timer_setup(&stress_hpt_timer, stress_hpt_timer_fn, 0);
1059 		stress_hpt_timer.expires = jiffies + msecs_to_jiffies(10);
1060 		add_timer(&stress_hpt_timer);
1061 	}
1062 
1063 	/*
1064 	 * Calculate the required size of the htab.  We want the number of
1065 	 * PTEGs to equal one half the number of real pages.
1066 	 */
1067 	htab_size_bytes = htab_get_table_size();
1068 	pteg_count = htab_size_bytes >> 7;
1069 
1070 	htab_hash_mask = pteg_count - 1;
1071 
1072 	if (firmware_has_feature(FW_FEATURE_LPAR) ||
1073 	    firmware_has_feature(FW_FEATURE_PS3_LV1)) {
1074 		/* Using a hypervisor which owns the htab */
1075 		htab_address = NULL;
1076 		_SDR1 = 0;
1077 #ifdef CONFIG_FA_DUMP
1078 		/*
1079 		 * If firmware assisted dump is active firmware preserves
1080 		 * the contents of htab along with entire partition memory.
1081 		 * Clear the htab if firmware assisted dump is active so
1082 		 * that we dont end up using old mappings.
1083 		 */
1084 		if (is_fadump_active() && mmu_hash_ops.hpte_clear_all)
1085 			mmu_hash_ops.hpte_clear_all();
1086 #endif
1087 	} else {
1088 		unsigned long limit = MEMBLOCK_ALLOC_ANYWHERE;
1089 
1090 #ifdef CONFIG_PPC_CELL
1091 		/*
1092 		 * Cell may require the hash table down low when using the
1093 		 * Axon IOMMU in order to fit the dynamic region over it, see
1094 		 * comments in cell/iommu.c
1095 		 */
1096 		if (fdt_subnode_offset(initial_boot_params, 0, "axon") > 0) {
1097 			limit = 0x80000000;
1098 			pr_info("Hash table forced below 2G for Axon IOMMU\n");
1099 		}
1100 #endif /* CONFIG_PPC_CELL */
1101 
1102 		table = memblock_phys_alloc_range(htab_size_bytes,
1103 						  htab_size_bytes,
1104 						  0, limit);
1105 		if (!table)
1106 			panic("ERROR: Failed to allocate %pa bytes below %pa\n",
1107 			      &htab_size_bytes, &limit);
1108 
1109 		DBG("Hash table allocated at %lx, size: %lx\n", table,
1110 		    htab_size_bytes);
1111 
1112 		htab_address = __va(table);
1113 
1114 		/* htab absolute addr + encoded htabsize */
1115 		_SDR1 = table + __ilog2(htab_size_bytes) - 18;
1116 
1117 		/* Initialize the HPT with no entries */
1118 		memset((void *)table, 0, htab_size_bytes);
1119 
1120 		if (!cpu_has_feature(CPU_FTR_ARCH_300))
1121 			/* Set SDR1 */
1122 			mtspr(SPRN_SDR1, _SDR1);
1123 		else
1124 			hash_init_partition_table(table, htab_size_bytes);
1125 	}
1126 
1127 	prot = pgprot_val(PAGE_KERNEL);
1128 
1129 	if (debug_pagealloc_enabled_or_kfence()) {
1130 		linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
1131 		linear_map_hash_slots = memblock_alloc_try_nid(
1132 				linear_map_hash_count, 1, MEMBLOCK_LOW_LIMIT,
1133 				ppc64_rma_size,	NUMA_NO_NODE);
1134 		if (!linear_map_hash_slots)
1135 			panic("%s: Failed to allocate %lu bytes max_addr=%pa\n",
1136 			      __func__, linear_map_hash_count, &ppc64_rma_size);
1137 	}
1138 
1139 	/* create bolted the linear mapping in the hash table */
1140 	for_each_mem_range(i, &base, &end) {
1141 		size = end - base;
1142 		base = (unsigned long)__va(base);
1143 
1144 		DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
1145 		    base, size, prot);
1146 
1147 		if ((base + size) >= H_VMALLOC_START) {
1148 			pr_warn("Outside the supported range\n");
1149 			continue;
1150 		}
1151 
1152 		BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
1153 				prot, mmu_linear_psize, mmu_kernel_ssize));
1154 	}
1155 	memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
1156 
1157 	/*
1158 	 * If we have a memory_limit and we've allocated TCEs then we need to
1159 	 * explicitly map the TCE area at the top of RAM. We also cope with the
1160 	 * case that the TCEs start below memory_limit.
1161 	 * tce_alloc_start/end are 16MB aligned so the mapping should work
1162 	 * for either 4K or 16MB pages.
1163 	 */
1164 	if (tce_alloc_start) {
1165 		tce_alloc_start = (unsigned long)__va(tce_alloc_start);
1166 		tce_alloc_end = (unsigned long)__va(tce_alloc_end);
1167 
1168 		if (base + size >= tce_alloc_start)
1169 			tce_alloc_start = base + size + 1;
1170 
1171 		BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
1172 					 __pa(tce_alloc_start), prot,
1173 					 mmu_linear_psize, mmu_kernel_ssize));
1174 	}
1175 
1176 
1177 	DBG(" <- htab_initialize()\n");
1178 }
1179 #undef KB
1180 #undef MB
1181 
1182 void __init hash__early_init_devtree(void)
1183 {
1184 	/* Initialize segment sizes */
1185 	of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
1186 
1187 	/* Initialize page sizes */
1188 	htab_scan_page_sizes();
1189 }
1190 
1191 static struct hash_mm_context init_hash_mm_context;
1192 void __init hash__early_init_mmu(void)
1193 {
1194 #ifndef CONFIG_PPC_64K_PAGES
1195 	/*
1196 	 * We have code in __hash_page_4K() and elsewhere, which assumes it can
1197 	 * do the following:
1198 	 *   new_pte |= (slot << H_PAGE_F_GIX_SHIFT) & (H_PAGE_F_SECOND | H_PAGE_F_GIX);
1199 	 *
1200 	 * Where the slot number is between 0-15, and values of 8-15 indicate
1201 	 * the secondary bucket. For that code to work H_PAGE_F_SECOND and
1202 	 * H_PAGE_F_GIX must occupy four contiguous bits in the PTE, and
1203 	 * H_PAGE_F_SECOND must be placed above H_PAGE_F_GIX. Assert that here
1204 	 * with a BUILD_BUG_ON().
1205 	 */
1206 	BUILD_BUG_ON(H_PAGE_F_SECOND != (1ul  << (H_PAGE_F_GIX_SHIFT + 3)));
1207 #endif /* CONFIG_PPC_64K_PAGES */
1208 
1209 	htab_init_page_sizes();
1210 
1211 	/*
1212 	 * initialize page table size
1213 	 */
1214 	__pte_frag_nr = H_PTE_FRAG_NR;
1215 	__pte_frag_size_shift = H_PTE_FRAG_SIZE_SHIFT;
1216 	__pmd_frag_nr = H_PMD_FRAG_NR;
1217 	__pmd_frag_size_shift = H_PMD_FRAG_SIZE_SHIFT;
1218 
1219 	__pte_index_size = H_PTE_INDEX_SIZE;
1220 	__pmd_index_size = H_PMD_INDEX_SIZE;
1221 	__pud_index_size = H_PUD_INDEX_SIZE;
1222 	__pgd_index_size = H_PGD_INDEX_SIZE;
1223 	__pud_cache_index = H_PUD_CACHE_INDEX;
1224 	__pte_table_size = H_PTE_TABLE_SIZE;
1225 	__pmd_table_size = H_PMD_TABLE_SIZE;
1226 	__pud_table_size = H_PUD_TABLE_SIZE;
1227 	__pgd_table_size = H_PGD_TABLE_SIZE;
1228 	/*
1229 	 * 4k use hugepd format, so for hash set then to
1230 	 * zero
1231 	 */
1232 	__pmd_val_bits = HASH_PMD_VAL_BITS;
1233 	__pud_val_bits = HASH_PUD_VAL_BITS;
1234 	__pgd_val_bits = HASH_PGD_VAL_BITS;
1235 
1236 	__kernel_virt_start = H_KERN_VIRT_START;
1237 	__vmalloc_start = H_VMALLOC_START;
1238 	__vmalloc_end = H_VMALLOC_END;
1239 	__kernel_io_start = H_KERN_IO_START;
1240 	__kernel_io_end = H_KERN_IO_END;
1241 	vmemmap = (struct page *)H_VMEMMAP_START;
1242 	ioremap_bot = IOREMAP_BASE;
1243 
1244 #ifdef CONFIG_PCI
1245 	pci_io_base = ISA_IO_BASE;
1246 #endif
1247 
1248 	/* Select appropriate backend */
1249 	if (firmware_has_feature(FW_FEATURE_PS3_LV1))
1250 		ps3_early_mm_init();
1251 	else if (firmware_has_feature(FW_FEATURE_LPAR))
1252 		hpte_init_pseries();
1253 	else if (IS_ENABLED(CONFIG_PPC_HASH_MMU_NATIVE))
1254 		hpte_init_native();
1255 
1256 	if (!mmu_hash_ops.hpte_insert)
1257 		panic("hash__early_init_mmu: No MMU hash ops defined!\n");
1258 
1259 	/*
1260 	 * Initialize the MMU Hash table and create the linear mapping
1261 	 * of memory. Has to be done before SLB initialization as this is
1262 	 * currently where the page size encoding is obtained.
1263 	 */
1264 	htab_initialize();
1265 
1266 	init_mm.context.hash_context = &init_hash_mm_context;
1267 	mm_ctx_set_slb_addr_limit(&init_mm.context, SLB_ADDR_LIMIT_DEFAULT);
1268 
1269 	pr_info("Initializing hash mmu with SLB\n");
1270 	/* Initialize SLB management */
1271 	slb_initialize();
1272 
1273 	if (cpu_has_feature(CPU_FTR_ARCH_206)
1274 			&& cpu_has_feature(CPU_FTR_HVMODE))
1275 		tlbiel_all();
1276 }
1277 
1278 #ifdef CONFIG_SMP
1279 void hash__early_init_mmu_secondary(void)
1280 {
1281 	/* Initialize hash table for that CPU */
1282 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1283 
1284 		if (!cpu_has_feature(CPU_FTR_ARCH_300))
1285 			mtspr(SPRN_SDR1, _SDR1);
1286 		else
1287 			set_ptcr_when_no_uv(__pa(partition_tb) |
1288 					    (PATB_SIZE_SHIFT - 12));
1289 	}
1290 	/* Initialize SLB */
1291 	slb_initialize();
1292 
1293 	if (cpu_has_feature(CPU_FTR_ARCH_206)
1294 			&& cpu_has_feature(CPU_FTR_HVMODE))
1295 		tlbiel_all();
1296 
1297 #ifdef CONFIG_PPC_MEM_KEYS
1298 	if (mmu_has_feature(MMU_FTR_PKEY))
1299 		mtspr(SPRN_UAMOR, default_uamor);
1300 #endif
1301 }
1302 #endif /* CONFIG_SMP */
1303 
1304 /*
1305  * Called by asm hashtable.S for doing lazy icache flush
1306  */
1307 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
1308 {
1309 	struct page *page;
1310 
1311 	if (!pfn_valid(pte_pfn(pte)))
1312 		return pp;
1313 
1314 	page = pte_page(pte);
1315 
1316 	/* page is dirty */
1317 	if (!test_bit(PG_dcache_clean, &page->flags) && !PageReserved(page)) {
1318 		if (trap == INTERRUPT_INST_STORAGE) {
1319 			flush_dcache_icache_page(page);
1320 			set_bit(PG_dcache_clean, &page->flags);
1321 		} else
1322 			pp |= HPTE_R_N;
1323 	}
1324 	return pp;
1325 }
1326 
1327 static unsigned int get_paca_psize(unsigned long addr)
1328 {
1329 	unsigned char *psizes;
1330 	unsigned long index, mask_index;
1331 
1332 	if (addr < SLICE_LOW_TOP) {
1333 		psizes = get_paca()->mm_ctx_low_slices_psize;
1334 		index = GET_LOW_SLICE_INDEX(addr);
1335 	} else {
1336 		psizes = get_paca()->mm_ctx_high_slices_psize;
1337 		index = GET_HIGH_SLICE_INDEX(addr);
1338 	}
1339 	mask_index = index & 0x1;
1340 	return (psizes[index >> 1] >> (mask_index * 4)) & 0xF;
1341 }
1342 
1343 
1344 /*
1345  * Demote a segment to using 4k pages.
1346  * For now this makes the whole process use 4k pages.
1347  */
1348 #ifdef CONFIG_PPC_64K_PAGES
1349 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
1350 {
1351 	if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
1352 		return;
1353 	slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
1354 	copro_flush_all_slbs(mm);
1355 	if ((get_paca_psize(addr) != MMU_PAGE_4K) && (current->mm == mm)) {
1356 
1357 		copy_mm_to_paca(mm);
1358 		slb_flush_and_restore_bolted();
1359 	}
1360 }
1361 #endif /* CONFIG_PPC_64K_PAGES */
1362 
1363 #ifdef CONFIG_PPC_SUBPAGE_PROT
1364 /*
1365  * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
1366  * Userspace sets the subpage permissions using the subpage_prot system call.
1367  *
1368  * Result is 0: full permissions, _PAGE_RW: read-only,
1369  * _PAGE_RWX: no access.
1370  */
1371 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
1372 {
1373 	struct subpage_prot_table *spt = mm_ctx_subpage_prot(&mm->context);
1374 	u32 spp = 0;
1375 	u32 **sbpm, *sbpp;
1376 
1377 	if (!spt)
1378 		return 0;
1379 
1380 	if (ea >= spt->maxaddr)
1381 		return 0;
1382 	if (ea < 0x100000000UL) {
1383 		/* addresses below 4GB use spt->low_prot */
1384 		sbpm = spt->low_prot;
1385 	} else {
1386 		sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
1387 		if (!sbpm)
1388 			return 0;
1389 	}
1390 	sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
1391 	if (!sbpp)
1392 		return 0;
1393 	spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
1394 
1395 	/* extract 2-bit bitfield for this 4k subpage */
1396 	spp >>= 30 - 2 * ((ea >> 12) & 0xf);
1397 
1398 	/*
1399 	 * 0 -> full permission
1400 	 * 1 -> Read only
1401 	 * 2 -> no access.
1402 	 * We return the flag that need to be cleared.
1403 	 */
1404 	spp = ((spp & 2) ? _PAGE_RWX : 0) | ((spp & 1) ? _PAGE_WRITE : 0);
1405 	return spp;
1406 }
1407 
1408 #else /* CONFIG_PPC_SUBPAGE_PROT */
1409 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
1410 {
1411 	return 0;
1412 }
1413 #endif
1414 
1415 void hash_failure_debug(unsigned long ea, unsigned long access,
1416 			unsigned long vsid, unsigned long trap,
1417 			int ssize, int psize, int lpsize, unsigned long pte)
1418 {
1419 	if (!printk_ratelimit())
1420 		return;
1421 	pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
1422 		ea, access, current->comm);
1423 	pr_info("    trap=0x%lx vsid=0x%lx ssize=%d base psize=%d psize %d pte=0x%lx\n",
1424 		trap, vsid, ssize, psize, lpsize, pte);
1425 }
1426 
1427 static void check_paca_psize(unsigned long ea, struct mm_struct *mm,
1428 			     int psize, bool user_region)
1429 {
1430 	if (user_region) {
1431 		if (psize != get_paca_psize(ea)) {
1432 			copy_mm_to_paca(mm);
1433 			slb_flush_and_restore_bolted();
1434 		}
1435 	} else if (get_paca()->vmalloc_sllp !=
1436 		   mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1437 		get_paca()->vmalloc_sllp =
1438 			mmu_psize_defs[mmu_vmalloc_psize].sllp;
1439 		slb_vmalloc_update();
1440 	}
1441 }
1442 
1443 /*
1444  * Result code is:
1445  *  0 - handled
1446  *  1 - normal page fault
1447  * -1 - critical hash insertion error
1448  * -2 - access not permitted by subpage protection mechanism
1449  */
1450 int hash_page_mm(struct mm_struct *mm, unsigned long ea,
1451 		 unsigned long access, unsigned long trap,
1452 		 unsigned long flags)
1453 {
1454 	bool is_thp;
1455 	pgd_t *pgdir;
1456 	unsigned long vsid;
1457 	pte_t *ptep;
1458 	unsigned hugeshift;
1459 	int rc, user_region = 0;
1460 	int psize, ssize;
1461 
1462 	DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
1463 		ea, access, trap);
1464 	trace_hash_fault(ea, access, trap);
1465 
1466 	/* Get region & vsid */
1467 	switch (get_region_id(ea)) {
1468 	case USER_REGION_ID:
1469 		user_region = 1;
1470 		if (! mm) {
1471 			DBG_LOW(" user region with no mm !\n");
1472 			rc = 1;
1473 			goto bail;
1474 		}
1475 		psize = get_slice_psize(mm, ea);
1476 		ssize = user_segment_size(ea);
1477 		vsid = get_user_vsid(&mm->context, ea, ssize);
1478 		break;
1479 	case VMALLOC_REGION_ID:
1480 		vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
1481 		psize = mmu_vmalloc_psize;
1482 		ssize = mmu_kernel_ssize;
1483 		flags |= HPTE_USE_KERNEL_KEY;
1484 		break;
1485 
1486 	case IO_REGION_ID:
1487 		vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
1488 		psize = mmu_io_psize;
1489 		ssize = mmu_kernel_ssize;
1490 		flags |= HPTE_USE_KERNEL_KEY;
1491 		break;
1492 	default:
1493 		/*
1494 		 * Not a valid range
1495 		 * Send the problem up to do_page_fault()
1496 		 */
1497 		rc = 1;
1498 		goto bail;
1499 	}
1500 	DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
1501 
1502 	/* Bad address. */
1503 	if (!vsid) {
1504 		DBG_LOW("Bad address!\n");
1505 		rc = 1;
1506 		goto bail;
1507 	}
1508 	/* Get pgdir */
1509 	pgdir = mm->pgd;
1510 	if (pgdir == NULL) {
1511 		rc = 1;
1512 		goto bail;
1513 	}
1514 
1515 	/* Check CPU locality */
1516 	if (user_region && mm_is_thread_local(mm))
1517 		flags |= HPTE_LOCAL_UPDATE;
1518 
1519 #ifndef CONFIG_PPC_64K_PAGES
1520 	/*
1521 	 * If we use 4K pages and our psize is not 4K, then we might
1522 	 * be hitting a special driver mapping, and need to align the
1523 	 * address before we fetch the PTE.
1524 	 *
1525 	 * It could also be a hugepage mapping, in which case this is
1526 	 * not necessary, but it's not harmful, either.
1527 	 */
1528 	if (psize != MMU_PAGE_4K)
1529 		ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
1530 #endif /* CONFIG_PPC_64K_PAGES */
1531 
1532 	/* Get PTE and page size from page tables */
1533 	ptep = find_linux_pte(pgdir, ea, &is_thp, &hugeshift);
1534 	if (ptep == NULL || !pte_present(*ptep)) {
1535 		DBG_LOW(" no PTE !\n");
1536 		rc = 1;
1537 		goto bail;
1538 	}
1539 
1540 	/*
1541 	 * Add _PAGE_PRESENT to the required access perm. If there are parallel
1542 	 * updates to the pte that can possibly clear _PAGE_PTE, catch that too.
1543 	 *
1544 	 * We can safely use the return pte address in rest of the function
1545 	 * because we do set H_PAGE_BUSY which prevents further updates to pte
1546 	 * from generic code.
1547 	 */
1548 	access |= _PAGE_PRESENT | _PAGE_PTE;
1549 
1550 	/*
1551 	 * Pre-check access permissions (will be re-checked atomically
1552 	 * in __hash_page_XX but this pre-check is a fast path
1553 	 */
1554 	if (!check_pte_access(access, pte_val(*ptep))) {
1555 		DBG_LOW(" no access !\n");
1556 		rc = 1;
1557 		goto bail;
1558 	}
1559 
1560 	if (hugeshift) {
1561 		if (is_thp)
1562 			rc = __hash_page_thp(ea, access, vsid, (pmd_t *)ptep,
1563 					     trap, flags, ssize, psize);
1564 #ifdef CONFIG_HUGETLB_PAGE
1565 		else
1566 			rc = __hash_page_huge(ea, access, vsid, ptep, trap,
1567 					      flags, ssize, hugeshift, psize);
1568 #else
1569 		else {
1570 			/*
1571 			 * if we have hugeshift, and is not transhuge with
1572 			 * hugetlb disabled, something is really wrong.
1573 			 */
1574 			rc = 1;
1575 			WARN_ON(1);
1576 		}
1577 #endif
1578 		if (current->mm == mm)
1579 			check_paca_psize(ea, mm, psize, user_region);
1580 
1581 		goto bail;
1582 	}
1583 
1584 #ifndef CONFIG_PPC_64K_PAGES
1585 	DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1586 #else
1587 	DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1588 		pte_val(*(ptep + PTRS_PER_PTE)));
1589 #endif
1590 	/* Do actual hashing */
1591 #ifdef CONFIG_PPC_64K_PAGES
1592 	/* If H_PAGE_4K_PFN is set, make sure this is a 4k segment */
1593 	if ((pte_val(*ptep) & H_PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1594 		demote_segment_4k(mm, ea);
1595 		psize = MMU_PAGE_4K;
1596 	}
1597 
1598 	/*
1599 	 * If this PTE is non-cacheable and we have restrictions on
1600 	 * using non cacheable large pages, then we switch to 4k
1601 	 */
1602 	if (mmu_ci_restrictions && psize == MMU_PAGE_64K && pte_ci(*ptep)) {
1603 		if (user_region) {
1604 			demote_segment_4k(mm, ea);
1605 			psize = MMU_PAGE_4K;
1606 		} else if (ea < VMALLOC_END) {
1607 			/*
1608 			 * some driver did a non-cacheable mapping
1609 			 * in vmalloc space, so switch vmalloc
1610 			 * to 4k pages
1611 			 */
1612 			printk(KERN_ALERT "Reducing vmalloc segment "
1613 			       "to 4kB pages because of "
1614 			       "non-cacheable mapping\n");
1615 			psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1616 			copro_flush_all_slbs(mm);
1617 		}
1618 	}
1619 
1620 #endif /* CONFIG_PPC_64K_PAGES */
1621 
1622 	if (current->mm == mm)
1623 		check_paca_psize(ea, mm, psize, user_region);
1624 
1625 #ifdef CONFIG_PPC_64K_PAGES
1626 	if (psize == MMU_PAGE_64K)
1627 		rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1628 				     flags, ssize);
1629 	else
1630 #endif /* CONFIG_PPC_64K_PAGES */
1631 	{
1632 		int spp = subpage_protection(mm, ea);
1633 		if (access & spp)
1634 			rc = -2;
1635 		else
1636 			rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1637 					    flags, ssize, spp);
1638 	}
1639 
1640 	/*
1641 	 * Dump some info in case of hash insertion failure, they should
1642 	 * never happen so it is really useful to know if/when they do
1643 	 */
1644 	if (rc == -1)
1645 		hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1646 				   psize, pte_val(*ptep));
1647 #ifndef CONFIG_PPC_64K_PAGES
1648 	DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1649 #else
1650 	DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1651 		pte_val(*(ptep + PTRS_PER_PTE)));
1652 #endif
1653 	DBG_LOW(" -> rc=%d\n", rc);
1654 
1655 bail:
1656 	return rc;
1657 }
1658 EXPORT_SYMBOL_GPL(hash_page_mm);
1659 
1660 int hash_page(unsigned long ea, unsigned long access, unsigned long trap,
1661 	      unsigned long dsisr)
1662 {
1663 	unsigned long flags = 0;
1664 	struct mm_struct *mm = current->mm;
1665 
1666 	if ((get_region_id(ea) == VMALLOC_REGION_ID) ||
1667 	    (get_region_id(ea) == IO_REGION_ID))
1668 		mm = &init_mm;
1669 
1670 	if (dsisr & DSISR_NOHPTE)
1671 		flags |= HPTE_NOHPTE_UPDATE;
1672 
1673 	return hash_page_mm(mm, ea, access, trap, flags);
1674 }
1675 EXPORT_SYMBOL_GPL(hash_page);
1676 
1677 DEFINE_INTERRUPT_HANDLER(do_hash_fault)
1678 {
1679 	unsigned long ea = regs->dar;
1680 	unsigned long dsisr = regs->dsisr;
1681 	unsigned long access = _PAGE_PRESENT | _PAGE_READ;
1682 	unsigned long flags = 0;
1683 	struct mm_struct *mm;
1684 	unsigned int region_id;
1685 	long err;
1686 
1687 	if (unlikely(dsisr & (DSISR_BAD_FAULT_64S | DSISR_KEYFAULT))) {
1688 		hash__do_page_fault(regs);
1689 		return;
1690 	}
1691 
1692 	region_id = get_region_id(ea);
1693 	if ((region_id == VMALLOC_REGION_ID) || (region_id == IO_REGION_ID))
1694 		mm = &init_mm;
1695 	else
1696 		mm = current->mm;
1697 
1698 	if (dsisr & DSISR_NOHPTE)
1699 		flags |= HPTE_NOHPTE_UPDATE;
1700 
1701 	if (dsisr & DSISR_ISSTORE)
1702 		access |= _PAGE_WRITE;
1703 	/*
1704 	 * We set _PAGE_PRIVILEGED only when
1705 	 * kernel mode access kernel space.
1706 	 *
1707 	 * _PAGE_PRIVILEGED is NOT set
1708 	 * 1) when kernel mode access user space
1709 	 * 2) user space access kernel space.
1710 	 */
1711 	access |= _PAGE_PRIVILEGED;
1712 	if (user_mode(regs) || (region_id == USER_REGION_ID))
1713 		access &= ~_PAGE_PRIVILEGED;
1714 
1715 	if (TRAP(regs) == INTERRUPT_INST_STORAGE)
1716 		access |= _PAGE_EXEC;
1717 
1718 	err = hash_page_mm(mm, ea, access, TRAP(regs), flags);
1719 	if (unlikely(err < 0)) {
1720 		// failed to insert a hash PTE due to an hypervisor error
1721 		if (user_mode(regs)) {
1722 			if (IS_ENABLED(CONFIG_PPC_SUBPAGE_PROT) && err == -2)
1723 				_exception(SIGSEGV, regs, SEGV_ACCERR, ea);
1724 			else
1725 				_exception(SIGBUS, regs, BUS_ADRERR, ea);
1726 		} else {
1727 			bad_page_fault(regs, SIGBUS);
1728 		}
1729 		err = 0;
1730 
1731 	} else if (err) {
1732 		hash__do_page_fault(regs);
1733 	}
1734 }
1735 
1736 static bool should_hash_preload(struct mm_struct *mm, unsigned long ea)
1737 {
1738 	int psize = get_slice_psize(mm, ea);
1739 
1740 	/* We only prefault standard pages for now */
1741 	if (unlikely(psize != mm_ctx_user_psize(&mm->context)))
1742 		return false;
1743 
1744 	/*
1745 	 * Don't prefault if subpage protection is enabled for the EA.
1746 	 */
1747 	if (unlikely((psize == MMU_PAGE_4K) && subpage_protection(mm, ea)))
1748 		return false;
1749 
1750 	return true;
1751 }
1752 
1753 static void hash_preload(struct mm_struct *mm, pte_t *ptep, unsigned long ea,
1754 			 bool is_exec, unsigned long trap)
1755 {
1756 	unsigned long vsid;
1757 	pgd_t *pgdir;
1758 	int rc, ssize, update_flags = 0;
1759 	unsigned long access = _PAGE_PRESENT | _PAGE_READ | (is_exec ? _PAGE_EXEC : 0);
1760 	unsigned long flags;
1761 
1762 	BUG_ON(get_region_id(ea) != USER_REGION_ID);
1763 
1764 	if (!should_hash_preload(mm, ea))
1765 		return;
1766 
1767 	DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1768 		" trap=%lx\n", mm, mm->pgd, ea, access, trap);
1769 
1770 	/* Get Linux PTE if available */
1771 	pgdir = mm->pgd;
1772 	if (pgdir == NULL)
1773 		return;
1774 
1775 	/* Get VSID */
1776 	ssize = user_segment_size(ea);
1777 	vsid = get_user_vsid(&mm->context, ea, ssize);
1778 	if (!vsid)
1779 		return;
1780 
1781 #ifdef CONFIG_PPC_64K_PAGES
1782 	/* If either H_PAGE_4K_PFN or cache inhibited is set (and we are on
1783 	 * a 64K kernel), then we don't preload, hash_page() will take
1784 	 * care of it once we actually try to access the page.
1785 	 * That way we don't have to duplicate all of the logic for segment
1786 	 * page size demotion here
1787 	 * Called with  PTL held, hence can be sure the value won't change in
1788 	 * between.
1789 	 */
1790 	if ((pte_val(*ptep) & H_PAGE_4K_PFN) || pte_ci(*ptep))
1791 		return;
1792 #endif /* CONFIG_PPC_64K_PAGES */
1793 
1794 	/*
1795 	 * __hash_page_* must run with interrupts off, including PMI interrupts
1796 	 * off, as it sets the H_PAGE_BUSY bit.
1797 	 *
1798 	 * It's otherwise possible for perf interrupts to hit at any time and
1799 	 * may take a hash fault reading the user stack, which could take a
1800 	 * hash miss and deadlock on the same H_PAGE_BUSY bit.
1801 	 *
1802 	 * Interrupts must also be off for the duration of the
1803 	 * mm_is_thread_local test and update, to prevent preempt running the
1804 	 * mm on another CPU (XXX: this may be racy vs kthread_use_mm).
1805 	 */
1806 	powerpc_local_irq_pmu_save(flags);
1807 
1808 	/* Is that local to this CPU ? */
1809 	if (mm_is_thread_local(mm))
1810 		update_flags |= HPTE_LOCAL_UPDATE;
1811 
1812 	/* Hash it in */
1813 #ifdef CONFIG_PPC_64K_PAGES
1814 	if (mm_ctx_user_psize(&mm->context) == MMU_PAGE_64K)
1815 		rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1816 				     update_flags, ssize);
1817 	else
1818 #endif /* CONFIG_PPC_64K_PAGES */
1819 		rc = __hash_page_4K(ea, access, vsid, ptep, trap, update_flags,
1820 				    ssize, subpage_protection(mm, ea));
1821 
1822 	/* Dump some info in case of hash insertion failure, they should
1823 	 * never happen so it is really useful to know if/when they do
1824 	 */
1825 	if (rc == -1)
1826 		hash_failure_debug(ea, access, vsid, trap, ssize,
1827 				   mm_ctx_user_psize(&mm->context),
1828 				   mm_ctx_user_psize(&mm->context),
1829 				   pte_val(*ptep));
1830 
1831 	powerpc_local_irq_pmu_restore(flags);
1832 }
1833 
1834 /*
1835  * This is called at the end of handling a user page fault, when the
1836  * fault has been handled by updating a PTE in the linux page tables.
1837  * We use it to preload an HPTE into the hash table corresponding to
1838  * the updated linux PTE.
1839  *
1840  * This must always be called with the pte lock held.
1841  */
1842 void __update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
1843 		      pte_t *ptep)
1844 {
1845 	/*
1846 	 * We don't need to worry about _PAGE_PRESENT here because we are
1847 	 * called with either mm->page_table_lock held or ptl lock held
1848 	 */
1849 	unsigned long trap;
1850 	bool is_exec;
1851 
1852 	/* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
1853 	if (!pte_young(*ptep) || address >= TASK_SIZE)
1854 		return;
1855 
1856 	/*
1857 	 * We try to figure out if we are coming from an instruction
1858 	 * access fault and pass that down to __hash_page so we avoid
1859 	 * double-faulting on execution of fresh text. We have to test
1860 	 * for regs NULL since init will get here first thing at boot.
1861 	 *
1862 	 * We also avoid filling the hash if not coming from a fault.
1863 	 */
1864 
1865 	trap = current->thread.regs ? TRAP(current->thread.regs) : 0UL;
1866 	switch (trap) {
1867 	case 0x300:
1868 		is_exec = false;
1869 		break;
1870 	case 0x400:
1871 		is_exec = true;
1872 		break;
1873 	default:
1874 		return;
1875 	}
1876 
1877 	hash_preload(vma->vm_mm, ptep, address, is_exec, trap);
1878 }
1879 
1880 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1881 static inline void tm_flush_hash_page(int local)
1882 {
1883 	/*
1884 	 * Transactions are not aborted by tlbiel, only tlbie. Without, syncing a
1885 	 * page back to a block device w/PIO could pick up transactional data
1886 	 * (bad!) so we force an abort here. Before the sync the page will be
1887 	 * made read-only, which will flush_hash_page. BIG ISSUE here: if the
1888 	 * kernel uses a page from userspace without unmapping it first, it may
1889 	 * see the speculated version.
1890 	 */
1891 	if (local && cpu_has_feature(CPU_FTR_TM) && current->thread.regs &&
1892 	    MSR_TM_ACTIVE(current->thread.regs->msr)) {
1893 		tm_enable();
1894 		tm_abort(TM_CAUSE_TLBI);
1895 	}
1896 }
1897 #else
1898 static inline void tm_flush_hash_page(int local)
1899 {
1900 }
1901 #endif
1902 
1903 /*
1904  * Return the global hash slot, corresponding to the given PTE, which contains
1905  * the HPTE.
1906  */
1907 unsigned long pte_get_hash_gslot(unsigned long vpn, unsigned long shift,
1908 		int ssize, real_pte_t rpte, unsigned int subpg_index)
1909 {
1910 	unsigned long hash, gslot, hidx;
1911 
1912 	hash = hpt_hash(vpn, shift, ssize);
1913 	hidx = __rpte_to_hidx(rpte, subpg_index);
1914 	if (hidx & _PTEIDX_SECONDARY)
1915 		hash = ~hash;
1916 	gslot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1917 	gslot += hidx & _PTEIDX_GROUP_IX;
1918 	return gslot;
1919 }
1920 
1921 void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1922 		     unsigned long flags)
1923 {
1924 	unsigned long index, shift, gslot;
1925 	int local = flags & HPTE_LOCAL_UPDATE;
1926 
1927 	DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1928 	pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1929 		gslot = pte_get_hash_gslot(vpn, shift, ssize, pte, index);
1930 		DBG_LOW(" sub %ld: gslot=%lx\n", index, gslot);
1931 		/*
1932 		 * We use same base page size and actual psize, because we don't
1933 		 * use these functions for hugepage
1934 		 */
1935 		mmu_hash_ops.hpte_invalidate(gslot, vpn, psize, psize,
1936 					     ssize, local);
1937 	} pte_iterate_hashed_end();
1938 
1939 	tm_flush_hash_page(local);
1940 }
1941 
1942 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1943 void flush_hash_hugepage(unsigned long vsid, unsigned long addr,
1944 			 pmd_t *pmdp, unsigned int psize, int ssize,
1945 			 unsigned long flags)
1946 {
1947 	int i, max_hpte_count, valid;
1948 	unsigned long s_addr;
1949 	unsigned char *hpte_slot_array;
1950 	unsigned long hidx, shift, vpn, hash, slot;
1951 	int local = flags & HPTE_LOCAL_UPDATE;
1952 
1953 	s_addr = addr & HPAGE_PMD_MASK;
1954 	hpte_slot_array = get_hpte_slot_array(pmdp);
1955 	/*
1956 	 * IF we try to do a HUGE PTE update after a withdraw is done.
1957 	 * we will find the below NULL. This happens when we do
1958 	 * split_huge_pmd
1959 	 */
1960 	if (!hpte_slot_array)
1961 		return;
1962 
1963 	if (mmu_hash_ops.hugepage_invalidate) {
1964 		mmu_hash_ops.hugepage_invalidate(vsid, s_addr, hpte_slot_array,
1965 						 psize, ssize, local);
1966 		goto tm_abort;
1967 	}
1968 	/*
1969 	 * No bluk hpte removal support, invalidate each entry
1970 	 */
1971 	shift = mmu_psize_defs[psize].shift;
1972 	max_hpte_count = HPAGE_PMD_SIZE >> shift;
1973 	for (i = 0; i < max_hpte_count; i++) {
1974 		/*
1975 		 * 8 bits per each hpte entries
1976 		 * 000| [ secondary group (one bit) | hidx (3 bits) | valid bit]
1977 		 */
1978 		valid = hpte_valid(hpte_slot_array, i);
1979 		if (!valid)
1980 			continue;
1981 		hidx =  hpte_hash_index(hpte_slot_array, i);
1982 
1983 		/* get the vpn */
1984 		addr = s_addr + (i * (1ul << shift));
1985 		vpn = hpt_vpn(addr, vsid, ssize);
1986 		hash = hpt_hash(vpn, shift, ssize);
1987 		if (hidx & _PTEIDX_SECONDARY)
1988 			hash = ~hash;
1989 
1990 		slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1991 		slot += hidx & _PTEIDX_GROUP_IX;
1992 		mmu_hash_ops.hpte_invalidate(slot, vpn, psize,
1993 					     MMU_PAGE_16M, ssize, local);
1994 	}
1995 tm_abort:
1996 	tm_flush_hash_page(local);
1997 }
1998 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1999 
2000 void flush_hash_range(unsigned long number, int local)
2001 {
2002 	if (mmu_hash_ops.flush_hash_range)
2003 		mmu_hash_ops.flush_hash_range(number, local);
2004 	else {
2005 		int i;
2006 		struct ppc64_tlb_batch *batch =
2007 			this_cpu_ptr(&ppc64_tlb_batch);
2008 
2009 		for (i = 0; i < number; i++)
2010 			flush_hash_page(batch->vpn[i], batch->pte[i],
2011 					batch->psize, batch->ssize, local);
2012 	}
2013 }
2014 
2015 long hpte_insert_repeating(unsigned long hash, unsigned long vpn,
2016 			   unsigned long pa, unsigned long rflags,
2017 			   unsigned long vflags, int psize, int ssize)
2018 {
2019 	unsigned long hpte_group;
2020 	long slot;
2021 
2022 repeat:
2023 	hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
2024 
2025 	/* Insert into the hash table, primary slot */
2026 	slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags, vflags,
2027 					psize, psize, ssize);
2028 
2029 	/* Primary is full, try the secondary */
2030 	if (unlikely(slot == -1)) {
2031 		hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
2032 		slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags,
2033 						vflags | HPTE_V_SECONDARY,
2034 						psize, psize, ssize);
2035 		if (slot == -1) {
2036 			if (mftb() & 0x1)
2037 				hpte_group = (hash & htab_hash_mask) *
2038 						HPTES_PER_GROUP;
2039 
2040 			mmu_hash_ops.hpte_remove(hpte_group);
2041 			goto repeat;
2042 		}
2043 	}
2044 
2045 	return slot;
2046 }
2047 
2048 void hpt_clear_stress(void)
2049 {
2050 	int cpu = raw_smp_processor_id();
2051 	int g;
2052 
2053 	for (g = 0; g < stress_nr_groups(); g++) {
2054 		unsigned long last_group;
2055 		last_group = stress_hpt_struct[cpu].last_group[g];
2056 
2057 		if (last_group != -1UL) {
2058 			int i;
2059 			for (i = 0; i < HPTES_PER_GROUP; i++) {
2060 				if (mmu_hash_ops.hpte_remove(last_group) == -1)
2061 					break;
2062 			}
2063 			stress_hpt_struct[cpu].last_group[g] = -1;
2064 		}
2065 	}
2066 }
2067 
2068 void hpt_do_stress(unsigned long ea, unsigned long hpte_group)
2069 {
2070 	unsigned long last_group;
2071 	int cpu = raw_smp_processor_id();
2072 
2073 	last_group = stress_hpt_struct[cpu].last_group[stress_nr_groups() - 1];
2074 	if (hpte_group == last_group)
2075 		return;
2076 
2077 	if (last_group != -1UL) {
2078 		int i;
2079 		/*
2080 		 * Concurrent CPUs might be inserting into this group, so
2081 		 * give up after a number of iterations, to prevent a live
2082 		 * lock.
2083 		 */
2084 		for (i = 0; i < HPTES_PER_GROUP; i++) {
2085 			if (mmu_hash_ops.hpte_remove(last_group) == -1)
2086 				break;
2087 		}
2088 		stress_hpt_struct[cpu].last_group[stress_nr_groups() - 1] = -1;
2089 	}
2090 
2091 	if (ea >= PAGE_OFFSET) {
2092 		/*
2093 		 * We would really like to prefetch to get the TLB loaded, then
2094 		 * remove the PTE before returning from fault interrupt, to
2095 		 * increase the hash fault rate.
2096 		 *
2097 		 * Unfortunately QEMU TCG does not model the TLB in a way that
2098 		 * makes this possible, and systemsim (mambo) emulator does not
2099 		 * bring in TLBs with prefetches (although loads/stores do
2100 		 * work for non-CI PTEs).
2101 		 *
2102 		 * So remember this PTE and clear it on the next hash fault.
2103 		 */
2104 		memmove(&stress_hpt_struct[cpu].last_group[1],
2105 			&stress_hpt_struct[cpu].last_group[0],
2106 			(stress_nr_groups() - 1) * sizeof(unsigned long));
2107 		stress_hpt_struct[cpu].last_group[0] = hpte_group;
2108 	}
2109 }
2110 
2111 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KFENCE)
2112 static DEFINE_RAW_SPINLOCK(linear_map_hash_lock);
2113 
2114 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
2115 {
2116 	unsigned long hash;
2117 	unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
2118 	unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
2119 	unsigned long mode = htab_convert_pte_flags(pgprot_val(PAGE_KERNEL), HPTE_USE_KERNEL_KEY);
2120 	long ret;
2121 
2122 	hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
2123 
2124 	/* Don't create HPTE entries for bad address */
2125 	if (!vsid)
2126 		return;
2127 
2128 	if (linear_map_hash_slots[lmi] & 0x80)
2129 		return;
2130 
2131 	ret = hpte_insert_repeating(hash, vpn, __pa(vaddr), mode,
2132 				    HPTE_V_BOLTED,
2133 				    mmu_linear_psize, mmu_kernel_ssize);
2134 
2135 	BUG_ON (ret < 0);
2136 	raw_spin_lock(&linear_map_hash_lock);
2137 	BUG_ON(linear_map_hash_slots[lmi] & 0x80);
2138 	linear_map_hash_slots[lmi] = ret | 0x80;
2139 	raw_spin_unlock(&linear_map_hash_lock);
2140 }
2141 
2142 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
2143 {
2144 	unsigned long hash, hidx, slot;
2145 	unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
2146 	unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
2147 
2148 	hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
2149 	raw_spin_lock(&linear_map_hash_lock);
2150 	if (!(linear_map_hash_slots[lmi] & 0x80)) {
2151 		raw_spin_unlock(&linear_map_hash_lock);
2152 		return;
2153 	}
2154 	hidx = linear_map_hash_slots[lmi] & 0x7f;
2155 	linear_map_hash_slots[lmi] = 0;
2156 	raw_spin_unlock(&linear_map_hash_lock);
2157 	if (hidx & _PTEIDX_SECONDARY)
2158 		hash = ~hash;
2159 	slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
2160 	slot += hidx & _PTEIDX_GROUP_IX;
2161 	mmu_hash_ops.hpte_invalidate(slot, vpn, mmu_linear_psize,
2162 				     mmu_linear_psize,
2163 				     mmu_kernel_ssize, 0);
2164 }
2165 
2166 void hash__kernel_map_pages(struct page *page, int numpages, int enable)
2167 {
2168 	unsigned long flags, vaddr, lmi;
2169 	int i;
2170 
2171 	local_irq_save(flags);
2172 	for (i = 0; i < numpages; i++, page++) {
2173 		vaddr = (unsigned long)page_address(page);
2174 		lmi = __pa(vaddr) >> PAGE_SHIFT;
2175 		if (lmi >= linear_map_hash_count)
2176 			continue;
2177 		if (enable)
2178 			kernel_map_linear_page(vaddr, lmi);
2179 		else
2180 			kernel_unmap_linear_page(vaddr, lmi);
2181 	}
2182 	local_irq_restore(flags);
2183 }
2184 #endif /* CONFIG_DEBUG_PAGEALLOC || CONFIG_KFENCE */
2185 
2186 void hash__setup_initial_memory_limit(phys_addr_t first_memblock_base,
2187 				phys_addr_t first_memblock_size)
2188 {
2189 	/*
2190 	 * We don't currently support the first MEMBLOCK not mapping 0
2191 	 * physical on those processors
2192 	 */
2193 	BUG_ON(first_memblock_base != 0);
2194 
2195 	/*
2196 	 * On virtualized systems the first entry is our RMA region aka VRMA,
2197 	 * non-virtualized 64-bit hash MMU systems don't have a limitation
2198 	 * on real mode access.
2199 	 *
2200 	 * For guests on platforms before POWER9, we clamp the it limit to 1G
2201 	 * to avoid some funky things such as RTAS bugs etc...
2202 	 *
2203 	 * On POWER9 we limit to 1TB in case the host erroneously told us that
2204 	 * the RMA was >1TB. Effective address bits 0:23 are treated as zero
2205 	 * (meaning the access is aliased to zero i.e. addr = addr % 1TB)
2206 	 * for virtual real mode addressing and so it doesn't make sense to
2207 	 * have an area larger than 1TB as it can't be addressed.
2208 	 */
2209 	if (!early_cpu_has_feature(CPU_FTR_HVMODE)) {
2210 		ppc64_rma_size = first_memblock_size;
2211 		if (!early_cpu_has_feature(CPU_FTR_ARCH_300))
2212 			ppc64_rma_size = min_t(u64, ppc64_rma_size, 0x40000000);
2213 		else
2214 			ppc64_rma_size = min_t(u64, ppc64_rma_size,
2215 					       1UL << SID_SHIFT_1T);
2216 
2217 		/* Finally limit subsequent allocations */
2218 		memblock_set_current_limit(ppc64_rma_size);
2219 	} else {
2220 		ppc64_rma_size = ULONG_MAX;
2221 	}
2222 }
2223 
2224 #ifdef CONFIG_DEBUG_FS
2225 
2226 static int hpt_order_get(void *data, u64 *val)
2227 {
2228 	*val = ppc64_pft_size;
2229 	return 0;
2230 }
2231 
2232 static int hpt_order_set(void *data, u64 val)
2233 {
2234 	int ret;
2235 
2236 	if (!mmu_hash_ops.resize_hpt)
2237 		return -ENODEV;
2238 
2239 	cpus_read_lock();
2240 	ret = mmu_hash_ops.resize_hpt(val);
2241 	cpus_read_unlock();
2242 
2243 	return ret;
2244 }
2245 
2246 DEFINE_DEBUGFS_ATTRIBUTE(fops_hpt_order, hpt_order_get, hpt_order_set, "%llu\n");
2247 
2248 static int __init hash64_debugfs(void)
2249 {
2250 	debugfs_create_file("hpt_order", 0600, arch_debugfs_dir, NULL,
2251 			    &fops_hpt_order);
2252 	return 0;
2253 }
2254 machine_device_initcall(pseries, hash64_debugfs);
2255 #endif /* CONFIG_DEBUG_FS */
2256 
2257 void __init print_system_hash_info(void)
2258 {
2259 	pr_info("ppc64_pft_size    = 0x%llx\n", ppc64_pft_size);
2260 
2261 	if (htab_hash_mask)
2262 		pr_info("htab_hash_mask    = 0x%lx\n", htab_hash_mask);
2263 }
2264 
2265 unsigned long arch_randomize_brk(struct mm_struct *mm)
2266 {
2267 	/*
2268 	 * If we are using 1TB segments and we are allowed to randomise
2269 	 * the heap, we can put it above 1TB so it is backed by a 1TB
2270 	 * segment. Otherwise the heap will be in the bottom 1TB
2271 	 * which always uses 256MB segments and this may result in a
2272 	 * performance penalty.
2273 	 */
2274 	if (is_32bit_task())
2275 		return randomize_page(mm->brk, SZ_32M);
2276 	else if (!radix_enabled() && mmu_highuser_ssize == MMU_SEGSIZE_1T)
2277 		return randomize_page(max_t(unsigned long, mm->brk, SZ_1T), SZ_1G);
2278 	else
2279 		return randomize_page(mm->brk, SZ_1G);
2280 }
2281