xref: /linux/arch/powerpc/mm/book3s64/hash_utils.c (revision 409c38d4f156740bf3165fd6ceae4fa6425eebf4)
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 		VM_WARN_ONCE(!(pteflags & _PAGE_RWX), "no-access mapping request");
314 	} else {
315 		if (pteflags & _PAGE_RWX)
316 			rflags |= 0x2;
317 		/*
318 		 * We should never hit this in normal fault handling because
319 		 * a permission check (check_pte_access()) will bubble this
320 		 * to higher level linux handler even for PAGE_NONE.
321 		 */
322 		VM_WARN_ONCE(!(pteflags & _PAGE_RWX), "no-access mapping request");
323 		if (!((pteflags & _PAGE_WRITE) && (pteflags & _PAGE_DIRTY)))
324 			rflags |= 0x1;
325 	}
326 	/*
327 	 * We can't allow hardware to update hpte bits. Hence always
328 	 * set 'R' bit and set 'C' if it is a write fault
329 	 */
330 	rflags |=  HPTE_R_R;
331 
332 	if (pteflags & _PAGE_DIRTY)
333 		rflags |= HPTE_R_C;
334 	/*
335 	 * Add in WIG bits
336 	 */
337 
338 	if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_TOLERANT)
339 		rflags |= HPTE_R_I;
340 	else if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_NON_IDEMPOTENT)
341 		rflags |= (HPTE_R_I | HPTE_R_G);
342 	else if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_SAO)
343 		rflags |= (HPTE_R_W | HPTE_R_I | HPTE_R_M);
344 	else
345 		/*
346 		 * Add memory coherence if cache inhibited is not set
347 		 */
348 		rflags |= HPTE_R_M;
349 
350 	rflags |= pte_to_hpte_pkey_bits(pteflags, flags);
351 	return rflags;
352 }
353 
354 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
355 		      unsigned long pstart, unsigned long prot,
356 		      int psize, int ssize)
357 {
358 	unsigned long vaddr, paddr;
359 	unsigned int step, shift;
360 	int ret = 0;
361 
362 	shift = mmu_psize_defs[psize].shift;
363 	step = 1 << shift;
364 
365 	prot = htab_convert_pte_flags(prot, HPTE_USE_KERNEL_KEY);
366 
367 	DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
368 	    vstart, vend, pstart, prot, psize, ssize);
369 
370 	/* Carefully map only the possible range */
371 	vaddr = ALIGN(vstart, step);
372 	paddr = ALIGN(pstart, step);
373 	vend  = ALIGN_DOWN(vend, step);
374 
375 	for (; vaddr < vend; vaddr += step, paddr += step) {
376 		unsigned long hash, hpteg;
377 		unsigned long vsid = get_kernel_vsid(vaddr, ssize);
378 		unsigned long vpn  = hpt_vpn(vaddr, vsid, ssize);
379 		unsigned long tprot = prot;
380 		bool secondary_hash = false;
381 
382 		/*
383 		 * If we hit a bad address return error.
384 		 */
385 		if (!vsid)
386 			return -1;
387 		/* Make kernel text executable */
388 		if (overlaps_kernel_text(vaddr, vaddr + step))
389 			tprot &= ~HPTE_R_N;
390 
391 		/*
392 		 * If relocatable, check if it overlaps interrupt vectors that
393 		 * are copied down to real 0. For relocatable kernel
394 		 * (e.g. kdump case) we copy interrupt vectors down to real
395 		 * address 0. Mark that region as executable. This is
396 		 * because on p8 system with relocation on exception feature
397 		 * enabled, exceptions are raised with MMU (IR=DR=1) ON. Hence
398 		 * in order to execute the interrupt handlers in virtual
399 		 * mode the vector region need to be marked as executable.
400 		 */
401 		if ((PHYSICAL_START > MEMORY_START) &&
402 			overlaps_interrupt_vector_text(vaddr, vaddr + step))
403 				tprot &= ~HPTE_R_N;
404 
405 		hash = hpt_hash(vpn, shift, ssize);
406 		hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
407 
408 		BUG_ON(!mmu_hash_ops.hpte_insert);
409 repeat:
410 		ret = mmu_hash_ops.hpte_insert(hpteg, vpn, paddr, tprot,
411 					       HPTE_V_BOLTED, psize, psize,
412 					       ssize);
413 		if (ret == -1) {
414 			/*
415 			 * Try to keep bolted entries in primary.
416 			 * Remove non bolted entries and try insert again
417 			 */
418 			ret = mmu_hash_ops.hpte_remove(hpteg);
419 			if (ret != -1)
420 				ret = mmu_hash_ops.hpte_insert(hpteg, vpn, paddr, tprot,
421 							       HPTE_V_BOLTED, psize, psize,
422 							       ssize);
423 			if (ret == -1 && !secondary_hash) {
424 				secondary_hash = true;
425 				hpteg = ((~hash & htab_hash_mask) * HPTES_PER_GROUP);
426 				goto repeat;
427 			}
428 		}
429 
430 		if (ret < 0)
431 			break;
432 
433 		cond_resched();
434 		if (debug_pagealloc_enabled_or_kfence() &&
435 			(paddr >> PAGE_SHIFT) < linear_map_hash_count)
436 			linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
437 	}
438 	return ret < 0 ? ret : 0;
439 }
440 
441 int htab_remove_mapping(unsigned long vstart, unsigned long vend,
442 		      int psize, int ssize)
443 {
444 	unsigned long vaddr, time_limit;
445 	unsigned int step, shift;
446 	int rc;
447 	int ret = 0;
448 
449 	shift = mmu_psize_defs[psize].shift;
450 	step = 1 << shift;
451 
452 	if (!mmu_hash_ops.hpte_removebolted)
453 		return -ENODEV;
454 
455 	/* Unmap the full range specificied */
456 	vaddr = ALIGN_DOWN(vstart, step);
457 	time_limit = jiffies + HZ;
458 
459 	for (;vaddr < vend; vaddr += step) {
460 		rc = mmu_hash_ops.hpte_removebolted(vaddr, psize, ssize);
461 
462 		/*
463 		 * For large number of mappings introduce a cond_resched()
464 		 * to prevent softlockup warnings.
465 		 */
466 		if (time_after(jiffies, time_limit)) {
467 			cond_resched();
468 			time_limit = jiffies + HZ;
469 		}
470 		if (rc == -ENOENT) {
471 			ret = -ENOENT;
472 			continue;
473 		}
474 		if (rc < 0)
475 			return rc;
476 	}
477 
478 	return ret;
479 }
480 
481 static bool disable_1tb_segments __ro_after_init;
482 
483 static int __init parse_disable_1tb_segments(char *p)
484 {
485 	disable_1tb_segments = true;
486 	return 0;
487 }
488 early_param("disable_1tb_segments", parse_disable_1tb_segments);
489 
490 bool stress_hpt_enabled __initdata;
491 
492 static int __init parse_stress_hpt(char *p)
493 {
494 	stress_hpt_enabled = true;
495 	return 0;
496 }
497 early_param("stress_hpt", parse_stress_hpt);
498 
499 __ro_after_init DEFINE_STATIC_KEY_FALSE(stress_hpt_key);
500 
501 /*
502  * per-CPU array allocated if we enable stress_hpt.
503  */
504 #define STRESS_MAX_GROUPS 16
505 struct stress_hpt_struct {
506 	unsigned long last_group[STRESS_MAX_GROUPS];
507 };
508 
509 static inline int stress_nr_groups(void)
510 {
511 	/*
512 	 * LPAR H_REMOVE flushes TLB, so need some number > 1 of entries
513 	 * to allow practical forward progress. Bare metal returns 1, which
514 	 * seems to help uncover more bugs.
515 	 */
516 	if (firmware_has_feature(FW_FEATURE_LPAR))
517 		return STRESS_MAX_GROUPS;
518 	else
519 		return 1;
520 }
521 
522 static struct stress_hpt_struct *stress_hpt_struct;
523 
524 static int __init htab_dt_scan_seg_sizes(unsigned long node,
525 					 const char *uname, int depth,
526 					 void *data)
527 {
528 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
529 	const __be32 *prop;
530 	int size = 0;
531 
532 	/* We are scanning "cpu" nodes only */
533 	if (type == NULL || strcmp(type, "cpu") != 0)
534 		return 0;
535 
536 	prop = of_get_flat_dt_prop(node, "ibm,processor-segment-sizes", &size);
537 	if (prop == NULL)
538 		return 0;
539 	for (; size >= 4; size -= 4, ++prop) {
540 		if (be32_to_cpu(prop[0]) == 40) {
541 			DBG("1T segment support detected\n");
542 
543 			if (disable_1tb_segments) {
544 				DBG("1T segments disabled by command line\n");
545 				break;
546 			}
547 
548 			cur_cpu_spec->mmu_features |= MMU_FTR_1T_SEGMENT;
549 			return 1;
550 		}
551 	}
552 	cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
553 	return 0;
554 }
555 
556 static int __init get_idx_from_shift(unsigned int shift)
557 {
558 	int idx = -1;
559 
560 	switch (shift) {
561 	case 0xc:
562 		idx = MMU_PAGE_4K;
563 		break;
564 	case 0x10:
565 		idx = MMU_PAGE_64K;
566 		break;
567 	case 0x14:
568 		idx = MMU_PAGE_1M;
569 		break;
570 	case 0x18:
571 		idx = MMU_PAGE_16M;
572 		break;
573 	case 0x22:
574 		idx = MMU_PAGE_16G;
575 		break;
576 	}
577 	return idx;
578 }
579 
580 static int __init htab_dt_scan_page_sizes(unsigned long node,
581 					  const char *uname, int depth,
582 					  void *data)
583 {
584 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
585 	const __be32 *prop;
586 	int size = 0;
587 
588 	/* We are scanning "cpu" nodes only */
589 	if (type == NULL || strcmp(type, "cpu") != 0)
590 		return 0;
591 
592 	prop = of_get_flat_dt_prop(node, "ibm,segment-page-sizes", &size);
593 	if (!prop)
594 		return 0;
595 
596 	pr_info("Page sizes from device-tree:\n");
597 	size /= 4;
598 	cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
599 	while(size > 0) {
600 		unsigned int base_shift = be32_to_cpu(prop[0]);
601 		unsigned int slbenc = be32_to_cpu(prop[1]);
602 		unsigned int lpnum = be32_to_cpu(prop[2]);
603 		struct mmu_psize_def *def;
604 		int idx, base_idx;
605 
606 		size -= 3; prop += 3;
607 		base_idx = get_idx_from_shift(base_shift);
608 		if (base_idx < 0) {
609 			/* skip the pte encoding also */
610 			prop += lpnum * 2; size -= lpnum * 2;
611 			continue;
612 		}
613 		def = &mmu_psize_defs[base_idx];
614 		if (base_idx == MMU_PAGE_16M)
615 			cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
616 
617 		def->shift = base_shift;
618 		if (base_shift <= 23)
619 			def->avpnm = 0;
620 		else
621 			def->avpnm = (1 << (base_shift - 23)) - 1;
622 		def->sllp = slbenc;
623 		/*
624 		 * We don't know for sure what's up with tlbiel, so
625 		 * for now we only set it for 4K and 64K pages
626 		 */
627 		if (base_idx == MMU_PAGE_4K || base_idx == MMU_PAGE_64K)
628 			def->tlbiel = 1;
629 		else
630 			def->tlbiel = 0;
631 
632 		while (size > 0 && lpnum) {
633 			unsigned int shift = be32_to_cpu(prop[0]);
634 			int penc  = be32_to_cpu(prop[1]);
635 
636 			prop += 2; size -= 2;
637 			lpnum--;
638 
639 			idx = get_idx_from_shift(shift);
640 			if (idx < 0)
641 				continue;
642 
643 			if (penc == -1)
644 				pr_err("Invalid penc for base_shift=%d "
645 				       "shift=%d\n", base_shift, shift);
646 
647 			def->penc[idx] = penc;
648 			pr_info("base_shift=%d: shift=%d, sllp=0x%04lx,"
649 				" avpnm=0x%08lx, tlbiel=%d, penc=%d\n",
650 				base_shift, shift, def->sllp,
651 				def->avpnm, def->tlbiel, def->penc[idx]);
652 		}
653 	}
654 
655 	return 1;
656 }
657 
658 #ifdef CONFIG_HUGETLB_PAGE
659 /*
660  * Scan for 16G memory blocks that have been set aside for huge pages
661  * and reserve those blocks for 16G huge pages.
662  */
663 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
664 					const char *uname, int depth,
665 					void *data) {
666 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
667 	const __be64 *addr_prop;
668 	const __be32 *page_count_prop;
669 	unsigned int expected_pages;
670 	long unsigned int phys_addr;
671 	long unsigned int block_size;
672 
673 	/* We are scanning "memory" nodes only */
674 	if (type == NULL || strcmp(type, "memory") != 0)
675 		return 0;
676 
677 	/*
678 	 * This property is the log base 2 of the number of virtual pages that
679 	 * will represent this memory block.
680 	 */
681 	page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
682 	if (page_count_prop == NULL)
683 		return 0;
684 	expected_pages = (1 << be32_to_cpu(page_count_prop[0]));
685 	addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
686 	if (addr_prop == NULL)
687 		return 0;
688 	phys_addr = be64_to_cpu(addr_prop[0]);
689 	block_size = be64_to_cpu(addr_prop[1]);
690 	if (block_size != (16 * GB))
691 		return 0;
692 	printk(KERN_INFO "Huge page(16GB) memory: "
693 			"addr = 0x%lX size = 0x%lX pages = %d\n",
694 			phys_addr, block_size, expected_pages);
695 	if (phys_addr + block_size * expected_pages <= memblock_end_of_DRAM()) {
696 		memblock_reserve(phys_addr, block_size * expected_pages);
697 		pseries_add_gpage(phys_addr, block_size, expected_pages);
698 	}
699 	return 0;
700 }
701 #endif /* CONFIG_HUGETLB_PAGE */
702 
703 static void __init mmu_psize_set_default_penc(void)
704 {
705 	int bpsize, apsize;
706 	for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
707 		for (apsize = 0; apsize < MMU_PAGE_COUNT; apsize++)
708 			mmu_psize_defs[bpsize].penc[apsize] = -1;
709 }
710 
711 #ifdef CONFIG_PPC_64K_PAGES
712 
713 static bool __init might_have_hea(void)
714 {
715 	/*
716 	 * The HEA ethernet adapter requires awareness of the
717 	 * GX bus. Without that awareness we can easily assume
718 	 * we will never see an HEA ethernet device.
719 	 */
720 #ifdef CONFIG_IBMEBUS
721 	return !cpu_has_feature(CPU_FTR_ARCH_207S) &&
722 		firmware_has_feature(FW_FEATURE_SPLPAR);
723 #else
724 	return false;
725 #endif
726 }
727 
728 #endif /* #ifdef CONFIG_PPC_64K_PAGES */
729 
730 static void __init htab_scan_page_sizes(void)
731 {
732 	int rc;
733 
734 	/* se the invalid penc to -1 */
735 	mmu_psize_set_default_penc();
736 
737 	/* Default to 4K pages only */
738 	memcpy(mmu_psize_defs, mmu_psize_defaults,
739 	       sizeof(mmu_psize_defaults));
740 
741 	/*
742 	 * Try to find the available page sizes in the device-tree
743 	 */
744 	rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
745 	if (rc == 0 && early_mmu_has_feature(MMU_FTR_16M_PAGE)) {
746 		/*
747 		 * Nothing in the device-tree, but the CPU supports 16M pages,
748 		 * so let's fallback on a known size list for 16M capable CPUs.
749 		 */
750 		memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
751 		       sizeof(mmu_psize_defaults_gp));
752 	}
753 
754 #ifdef CONFIG_HUGETLB_PAGE
755 	if (!hugetlb_disabled && !early_radix_enabled() ) {
756 		/* Reserve 16G huge page memory sections for huge pages */
757 		of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
758 	}
759 #endif /* CONFIG_HUGETLB_PAGE */
760 }
761 
762 /*
763  * Fill in the hpte_page_sizes[] array.
764  * We go through the mmu_psize_defs[] array looking for all the
765  * supported base/actual page size combinations.  Each combination
766  * has a unique pagesize encoding (penc) value in the low bits of
767  * the LP field of the HPTE.  For actual page sizes less than 1MB,
768  * some of the upper LP bits are used for RPN bits, meaning that
769  * we need to fill in several entries in hpte_page_sizes[].
770  *
771  * In diagrammatic form, with r = RPN bits and z = page size bits:
772  *        PTE LP     actual page size
773  *    rrrr rrrz		>=8KB
774  *    rrrr rrzz		>=16KB
775  *    rrrr rzzz		>=32KB
776  *    rrrr zzzz		>=64KB
777  *    ...
778  *
779  * The zzzz bits are implementation-specific but are chosen so that
780  * no encoding for a larger page size uses the same value in its
781  * low-order N bits as the encoding for the 2^(12+N) byte page size
782  * (if it exists).
783  */
784 static void __init init_hpte_page_sizes(void)
785 {
786 	long int ap, bp;
787 	long int shift, penc;
788 
789 	for (bp = 0; bp < MMU_PAGE_COUNT; ++bp) {
790 		if (!mmu_psize_defs[bp].shift)
791 			continue;	/* not a supported page size */
792 		for (ap = bp; ap < MMU_PAGE_COUNT; ++ap) {
793 			penc = mmu_psize_defs[bp].penc[ap];
794 			if (penc == -1 || !mmu_psize_defs[ap].shift)
795 				continue;
796 			shift = mmu_psize_defs[ap].shift - LP_SHIFT;
797 			if (shift <= 0)
798 				continue;	/* should never happen */
799 			/*
800 			 * For page sizes less than 1MB, this loop
801 			 * replicates the entry for all possible values
802 			 * of the rrrr bits.
803 			 */
804 			while (penc < (1 << LP_BITS)) {
805 				hpte_page_sizes[penc] = (ap << 4) | bp;
806 				penc += 1 << shift;
807 			}
808 		}
809 	}
810 }
811 
812 static void __init htab_init_page_sizes(void)
813 {
814 	bool aligned = true;
815 	init_hpte_page_sizes();
816 
817 	if (!debug_pagealloc_enabled_or_kfence()) {
818 		/*
819 		 * Pick a size for the linear mapping. Currently, we only
820 		 * support 16M, 1M and 4K which is the default
821 		 */
822 		if (IS_ENABLED(CONFIG_STRICT_KERNEL_RWX) &&
823 		    (unsigned long)_stext % 0x1000000) {
824 			if (mmu_psize_defs[MMU_PAGE_16M].shift)
825 				pr_warn("Kernel not 16M aligned, disabling 16M linear map alignment\n");
826 			aligned = false;
827 		}
828 
829 		if (mmu_psize_defs[MMU_PAGE_16M].shift && aligned)
830 			mmu_linear_psize = MMU_PAGE_16M;
831 		else if (mmu_psize_defs[MMU_PAGE_1M].shift)
832 			mmu_linear_psize = MMU_PAGE_1M;
833 	}
834 
835 #ifdef CONFIG_PPC_64K_PAGES
836 	/*
837 	 * Pick a size for the ordinary pages. Default is 4K, we support
838 	 * 64K for user mappings and vmalloc if supported by the processor.
839 	 * We only use 64k for ioremap if the processor
840 	 * (and firmware) support cache-inhibited large pages.
841 	 * If not, we use 4k and set mmu_ci_restrictions so that
842 	 * hash_page knows to switch processes that use cache-inhibited
843 	 * mappings to 4k pages.
844 	 */
845 	if (mmu_psize_defs[MMU_PAGE_64K].shift) {
846 		mmu_virtual_psize = MMU_PAGE_64K;
847 		mmu_vmalloc_psize = MMU_PAGE_64K;
848 		if (mmu_linear_psize == MMU_PAGE_4K)
849 			mmu_linear_psize = MMU_PAGE_64K;
850 		if (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
851 			/*
852 			 * When running on pSeries using 64k pages for ioremap
853 			 * would stop us accessing the HEA ethernet. So if we
854 			 * have the chance of ever seeing one, stay at 4k.
855 			 */
856 			if (!might_have_hea())
857 				mmu_io_psize = MMU_PAGE_64K;
858 		} else
859 			mmu_ci_restrictions = 1;
860 	}
861 #endif /* CONFIG_PPC_64K_PAGES */
862 
863 #ifdef CONFIG_SPARSEMEM_VMEMMAP
864 	/*
865 	 * We try to use 16M pages for vmemmap if that is supported
866 	 * and we have at least 1G of RAM at boot
867 	 */
868 	if (mmu_psize_defs[MMU_PAGE_16M].shift &&
869 	    memblock_phys_mem_size() >= 0x40000000)
870 		mmu_vmemmap_psize = MMU_PAGE_16M;
871 	else
872 		mmu_vmemmap_psize = mmu_virtual_psize;
873 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
874 
875 	printk(KERN_DEBUG "Page orders: linear mapping = %d, "
876 	       "virtual = %d, io = %d"
877 #ifdef CONFIG_SPARSEMEM_VMEMMAP
878 	       ", vmemmap = %d"
879 #endif
880 	       "\n",
881 	       mmu_psize_defs[mmu_linear_psize].shift,
882 	       mmu_psize_defs[mmu_virtual_psize].shift,
883 	       mmu_psize_defs[mmu_io_psize].shift
884 #ifdef CONFIG_SPARSEMEM_VMEMMAP
885 	       ,mmu_psize_defs[mmu_vmemmap_psize].shift
886 #endif
887 	       );
888 }
889 
890 static int __init htab_dt_scan_pftsize(unsigned long node,
891 				       const char *uname, int depth,
892 				       void *data)
893 {
894 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
895 	const __be32 *prop;
896 
897 	/* We are scanning "cpu" nodes only */
898 	if (type == NULL || strcmp(type, "cpu") != 0)
899 		return 0;
900 
901 	prop = of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
902 	if (prop != NULL) {
903 		/* pft_size[0] is the NUMA CEC cookie */
904 		ppc64_pft_size = be32_to_cpu(prop[1]);
905 		return 1;
906 	}
907 	return 0;
908 }
909 
910 unsigned htab_shift_for_mem_size(unsigned long mem_size)
911 {
912 	unsigned memshift = __ilog2(mem_size);
913 	unsigned pshift = mmu_psize_defs[mmu_virtual_psize].shift;
914 	unsigned pteg_shift;
915 
916 	/* round mem_size up to next power of 2 */
917 	if ((1UL << memshift) < mem_size)
918 		memshift += 1;
919 
920 	/* aim for 2 pages / pteg */
921 	pteg_shift = memshift - (pshift + 1);
922 
923 	/*
924 	 * 2^11 PTEGS of 128 bytes each, ie. 2^18 bytes is the minimum htab
925 	 * size permitted by the architecture.
926 	 */
927 	return max(pteg_shift + 7, 18U);
928 }
929 
930 static unsigned long __init htab_get_table_size(void)
931 {
932 	/*
933 	 * If hash size isn't already provided by the platform, we try to
934 	 * retrieve it from the device-tree. If it's not there neither, we
935 	 * calculate it now based on the total RAM size
936 	 */
937 	if (ppc64_pft_size == 0)
938 		of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
939 	if (ppc64_pft_size)
940 		return 1UL << ppc64_pft_size;
941 
942 	return 1UL << htab_shift_for_mem_size(memblock_phys_mem_size());
943 }
944 
945 #ifdef CONFIG_MEMORY_HOTPLUG
946 static int resize_hpt_for_hotplug(unsigned long new_mem_size)
947 {
948 	unsigned target_hpt_shift;
949 
950 	if (!mmu_hash_ops.resize_hpt)
951 		return 0;
952 
953 	target_hpt_shift = htab_shift_for_mem_size(new_mem_size);
954 
955 	/*
956 	 * To avoid lots of HPT resizes if memory size is fluctuating
957 	 * across a boundary, we deliberately have some hysterisis
958 	 * here: we immediately increase the HPT size if the target
959 	 * shift exceeds the current shift, but we won't attempt to
960 	 * reduce unless the target shift is at least 2 below the
961 	 * current shift
962 	 */
963 	if (target_hpt_shift > ppc64_pft_size ||
964 	    target_hpt_shift < ppc64_pft_size - 1)
965 		return mmu_hash_ops.resize_hpt(target_hpt_shift);
966 
967 	return 0;
968 }
969 
970 int hash__create_section_mapping(unsigned long start, unsigned long end,
971 				 int nid, pgprot_t prot)
972 {
973 	int rc;
974 
975 	if (end >= H_VMALLOC_START) {
976 		pr_warn("Outside the supported range\n");
977 		return -1;
978 	}
979 
980 	resize_hpt_for_hotplug(memblock_phys_mem_size());
981 
982 	rc = htab_bolt_mapping(start, end, __pa(start),
983 			       pgprot_val(prot), mmu_linear_psize,
984 			       mmu_kernel_ssize);
985 
986 	if (rc < 0) {
987 		int rc2 = htab_remove_mapping(start, end, mmu_linear_psize,
988 					      mmu_kernel_ssize);
989 		BUG_ON(rc2 && (rc2 != -ENOENT));
990 	}
991 	return rc;
992 }
993 
994 int hash__remove_section_mapping(unsigned long start, unsigned long end)
995 {
996 	int rc = htab_remove_mapping(start, end, mmu_linear_psize,
997 				     mmu_kernel_ssize);
998 
999 	if (resize_hpt_for_hotplug(memblock_phys_mem_size()) == -ENOSPC)
1000 		pr_warn("Hash collision while resizing HPT\n");
1001 
1002 	return rc;
1003 }
1004 #endif /* CONFIG_MEMORY_HOTPLUG */
1005 
1006 static void __init hash_init_partition_table(phys_addr_t hash_table,
1007 					     unsigned long htab_size)
1008 {
1009 	mmu_partition_table_init();
1010 
1011 	/*
1012 	 * PS field (VRMA page size) is not used for LPID 0, hence set to 0.
1013 	 * For now, UPRT is 0 and we have no segment table.
1014 	 */
1015 	htab_size =  __ilog2(htab_size) - 18;
1016 	mmu_partition_table_set_entry(0, hash_table | htab_size, 0, false);
1017 	pr_info("Partition table %p\n", partition_tb);
1018 }
1019 
1020 void hpt_clear_stress(void);
1021 static struct timer_list stress_hpt_timer;
1022 static void stress_hpt_timer_fn(struct timer_list *timer)
1023 {
1024 	int next_cpu;
1025 
1026 	hpt_clear_stress();
1027 	if (!firmware_has_feature(FW_FEATURE_LPAR))
1028 		tlbiel_all();
1029 
1030 	next_cpu = cpumask_next(raw_smp_processor_id(), cpu_online_mask);
1031 	if (next_cpu >= nr_cpu_ids)
1032 		next_cpu = cpumask_first(cpu_online_mask);
1033 	stress_hpt_timer.expires = jiffies + msecs_to_jiffies(10);
1034 	add_timer_on(&stress_hpt_timer, next_cpu);
1035 }
1036 
1037 static void __init htab_initialize(void)
1038 {
1039 	unsigned long table;
1040 	unsigned long pteg_count;
1041 	unsigned long prot;
1042 	phys_addr_t base = 0, size = 0, end;
1043 	u64 i;
1044 
1045 	DBG(" -> htab_initialize()\n");
1046 
1047 	if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
1048 		mmu_kernel_ssize = MMU_SEGSIZE_1T;
1049 		mmu_highuser_ssize = MMU_SEGSIZE_1T;
1050 		printk(KERN_INFO "Using 1TB segments\n");
1051 	}
1052 
1053 	if (stress_slb_enabled)
1054 		static_branch_enable(&stress_slb_key);
1055 
1056 	if (stress_hpt_enabled) {
1057 		unsigned long tmp;
1058 		static_branch_enable(&stress_hpt_key);
1059 		// Too early to use nr_cpu_ids, so use NR_CPUS
1060 		tmp = memblock_phys_alloc_range(sizeof(struct stress_hpt_struct) * NR_CPUS,
1061 						__alignof__(struct stress_hpt_struct),
1062 						0, MEMBLOCK_ALLOC_ANYWHERE);
1063 		memset((void *)tmp, 0xff, sizeof(struct stress_hpt_struct) * NR_CPUS);
1064 		stress_hpt_struct = __va(tmp);
1065 
1066 		timer_setup(&stress_hpt_timer, stress_hpt_timer_fn, 0);
1067 		stress_hpt_timer.expires = jiffies + msecs_to_jiffies(10);
1068 		add_timer(&stress_hpt_timer);
1069 	}
1070 
1071 	/*
1072 	 * Calculate the required size of the htab.  We want the number of
1073 	 * PTEGs to equal one half the number of real pages.
1074 	 */
1075 	htab_size_bytes = htab_get_table_size();
1076 	pteg_count = htab_size_bytes >> 7;
1077 
1078 	htab_hash_mask = pteg_count - 1;
1079 
1080 	if (firmware_has_feature(FW_FEATURE_LPAR) ||
1081 	    firmware_has_feature(FW_FEATURE_PS3_LV1)) {
1082 		/* Using a hypervisor which owns the htab */
1083 		htab_address = NULL;
1084 		_SDR1 = 0;
1085 #ifdef CONFIG_FA_DUMP
1086 		/*
1087 		 * If firmware assisted dump is active firmware preserves
1088 		 * the contents of htab along with entire partition memory.
1089 		 * Clear the htab if firmware assisted dump is active so
1090 		 * that we dont end up using old mappings.
1091 		 */
1092 		if (is_fadump_active() && mmu_hash_ops.hpte_clear_all)
1093 			mmu_hash_ops.hpte_clear_all();
1094 #endif
1095 	} else {
1096 		unsigned long limit = MEMBLOCK_ALLOC_ANYWHERE;
1097 
1098 #ifdef CONFIG_PPC_CELL
1099 		/*
1100 		 * Cell may require the hash table down low when using the
1101 		 * Axon IOMMU in order to fit the dynamic region over it, see
1102 		 * comments in cell/iommu.c
1103 		 */
1104 		if (fdt_subnode_offset(initial_boot_params, 0, "axon") > 0) {
1105 			limit = 0x80000000;
1106 			pr_info("Hash table forced below 2G for Axon IOMMU\n");
1107 		}
1108 #endif /* CONFIG_PPC_CELL */
1109 
1110 		table = memblock_phys_alloc_range(htab_size_bytes,
1111 						  htab_size_bytes,
1112 						  0, limit);
1113 		if (!table)
1114 			panic("ERROR: Failed to allocate %pa bytes below %pa\n",
1115 			      &htab_size_bytes, &limit);
1116 
1117 		DBG("Hash table allocated at %lx, size: %lx\n", table,
1118 		    htab_size_bytes);
1119 
1120 		htab_address = __va(table);
1121 
1122 		/* htab absolute addr + encoded htabsize */
1123 		_SDR1 = table + __ilog2(htab_size_bytes) - 18;
1124 
1125 		/* Initialize the HPT with no entries */
1126 		memset((void *)table, 0, htab_size_bytes);
1127 
1128 		if (!cpu_has_feature(CPU_FTR_ARCH_300))
1129 			/* Set SDR1 */
1130 			mtspr(SPRN_SDR1, _SDR1);
1131 		else
1132 			hash_init_partition_table(table, htab_size_bytes);
1133 	}
1134 
1135 	prot = pgprot_val(PAGE_KERNEL);
1136 
1137 	if (debug_pagealloc_enabled_or_kfence()) {
1138 		linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
1139 		linear_map_hash_slots = memblock_alloc_try_nid(
1140 				linear_map_hash_count, 1, MEMBLOCK_LOW_LIMIT,
1141 				ppc64_rma_size,	NUMA_NO_NODE);
1142 		if (!linear_map_hash_slots)
1143 			panic("%s: Failed to allocate %lu bytes max_addr=%pa\n",
1144 			      __func__, linear_map_hash_count, &ppc64_rma_size);
1145 	}
1146 
1147 	/* create bolted the linear mapping in the hash table */
1148 	for_each_mem_range(i, &base, &end) {
1149 		size = end - base;
1150 		base = (unsigned long)__va(base);
1151 
1152 		DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
1153 		    base, size, prot);
1154 
1155 		if ((base + size) >= H_VMALLOC_START) {
1156 			pr_warn("Outside the supported range\n");
1157 			continue;
1158 		}
1159 
1160 		BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
1161 				prot, mmu_linear_psize, mmu_kernel_ssize));
1162 	}
1163 	memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
1164 
1165 	/*
1166 	 * If we have a memory_limit and we've allocated TCEs then we need to
1167 	 * explicitly map the TCE area at the top of RAM. We also cope with the
1168 	 * case that the TCEs start below memory_limit.
1169 	 * tce_alloc_start/end are 16MB aligned so the mapping should work
1170 	 * for either 4K or 16MB pages.
1171 	 */
1172 	if (tce_alloc_start) {
1173 		tce_alloc_start = (unsigned long)__va(tce_alloc_start);
1174 		tce_alloc_end = (unsigned long)__va(tce_alloc_end);
1175 
1176 		if (base + size >= tce_alloc_start)
1177 			tce_alloc_start = base + size + 1;
1178 
1179 		BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
1180 					 __pa(tce_alloc_start), prot,
1181 					 mmu_linear_psize, mmu_kernel_ssize));
1182 	}
1183 
1184 
1185 	DBG(" <- htab_initialize()\n");
1186 }
1187 #undef KB
1188 #undef MB
1189 
1190 void __init hash__early_init_devtree(void)
1191 {
1192 	/* Initialize segment sizes */
1193 	of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
1194 
1195 	/* Initialize page sizes */
1196 	htab_scan_page_sizes();
1197 }
1198 
1199 static struct hash_mm_context init_hash_mm_context;
1200 void __init hash__early_init_mmu(void)
1201 {
1202 #ifndef CONFIG_PPC_64K_PAGES
1203 	/*
1204 	 * We have code in __hash_page_4K() and elsewhere, which assumes it can
1205 	 * do the following:
1206 	 *   new_pte |= (slot << H_PAGE_F_GIX_SHIFT) & (H_PAGE_F_SECOND | H_PAGE_F_GIX);
1207 	 *
1208 	 * Where the slot number is between 0-15, and values of 8-15 indicate
1209 	 * the secondary bucket. For that code to work H_PAGE_F_SECOND and
1210 	 * H_PAGE_F_GIX must occupy four contiguous bits in the PTE, and
1211 	 * H_PAGE_F_SECOND must be placed above H_PAGE_F_GIX. Assert that here
1212 	 * with a BUILD_BUG_ON().
1213 	 */
1214 	BUILD_BUG_ON(H_PAGE_F_SECOND != (1ul  << (H_PAGE_F_GIX_SHIFT + 3)));
1215 #endif /* CONFIG_PPC_64K_PAGES */
1216 
1217 	htab_init_page_sizes();
1218 
1219 	/*
1220 	 * initialize page table size
1221 	 */
1222 	__pte_frag_nr = H_PTE_FRAG_NR;
1223 	__pte_frag_size_shift = H_PTE_FRAG_SIZE_SHIFT;
1224 	__pmd_frag_nr = H_PMD_FRAG_NR;
1225 	__pmd_frag_size_shift = H_PMD_FRAG_SIZE_SHIFT;
1226 
1227 	__pte_index_size = H_PTE_INDEX_SIZE;
1228 	__pmd_index_size = H_PMD_INDEX_SIZE;
1229 	__pud_index_size = H_PUD_INDEX_SIZE;
1230 	__pgd_index_size = H_PGD_INDEX_SIZE;
1231 	__pud_cache_index = H_PUD_CACHE_INDEX;
1232 	__pte_table_size = H_PTE_TABLE_SIZE;
1233 	__pmd_table_size = H_PMD_TABLE_SIZE;
1234 	__pud_table_size = H_PUD_TABLE_SIZE;
1235 	__pgd_table_size = H_PGD_TABLE_SIZE;
1236 	/*
1237 	 * 4k use hugepd format, so for hash set then to
1238 	 * zero
1239 	 */
1240 	__pmd_val_bits = HASH_PMD_VAL_BITS;
1241 	__pud_val_bits = HASH_PUD_VAL_BITS;
1242 	__pgd_val_bits = HASH_PGD_VAL_BITS;
1243 
1244 	__kernel_virt_start = H_KERN_VIRT_START;
1245 	__vmalloc_start = H_VMALLOC_START;
1246 	__vmalloc_end = H_VMALLOC_END;
1247 	__kernel_io_start = H_KERN_IO_START;
1248 	__kernel_io_end = H_KERN_IO_END;
1249 	vmemmap = (struct page *)H_VMEMMAP_START;
1250 	ioremap_bot = IOREMAP_BASE;
1251 
1252 #ifdef CONFIG_PCI
1253 	pci_io_base = ISA_IO_BASE;
1254 #endif
1255 
1256 	/* Select appropriate backend */
1257 	if (firmware_has_feature(FW_FEATURE_PS3_LV1))
1258 		ps3_early_mm_init();
1259 	else if (firmware_has_feature(FW_FEATURE_LPAR))
1260 		hpte_init_pseries();
1261 	else if (IS_ENABLED(CONFIG_PPC_HASH_MMU_NATIVE))
1262 		hpte_init_native();
1263 
1264 	if (!mmu_hash_ops.hpte_insert)
1265 		panic("hash__early_init_mmu: No MMU hash ops defined!\n");
1266 
1267 	/*
1268 	 * Initialize the MMU Hash table and create the linear mapping
1269 	 * of memory. Has to be done before SLB initialization as this is
1270 	 * currently where the page size encoding is obtained.
1271 	 */
1272 	htab_initialize();
1273 
1274 	init_mm.context.hash_context = &init_hash_mm_context;
1275 	mm_ctx_set_slb_addr_limit(&init_mm.context, SLB_ADDR_LIMIT_DEFAULT);
1276 
1277 	pr_info("Initializing hash mmu with SLB\n");
1278 	/* Initialize SLB management */
1279 	slb_initialize();
1280 
1281 	if (cpu_has_feature(CPU_FTR_ARCH_206)
1282 			&& cpu_has_feature(CPU_FTR_HVMODE))
1283 		tlbiel_all();
1284 }
1285 
1286 #ifdef CONFIG_SMP
1287 void hash__early_init_mmu_secondary(void)
1288 {
1289 	/* Initialize hash table for that CPU */
1290 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1291 
1292 		if (!cpu_has_feature(CPU_FTR_ARCH_300))
1293 			mtspr(SPRN_SDR1, _SDR1);
1294 		else
1295 			set_ptcr_when_no_uv(__pa(partition_tb) |
1296 					    (PATB_SIZE_SHIFT - 12));
1297 	}
1298 	/* Initialize SLB */
1299 	slb_initialize();
1300 
1301 	if (cpu_has_feature(CPU_FTR_ARCH_206)
1302 			&& cpu_has_feature(CPU_FTR_HVMODE))
1303 		tlbiel_all();
1304 
1305 #ifdef CONFIG_PPC_MEM_KEYS
1306 	if (mmu_has_feature(MMU_FTR_PKEY))
1307 		mtspr(SPRN_UAMOR, default_uamor);
1308 #endif
1309 }
1310 #endif /* CONFIG_SMP */
1311 
1312 /*
1313  * Called by asm hashtable.S for doing lazy icache flush
1314  */
1315 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
1316 {
1317 	struct folio *folio;
1318 
1319 	if (!pfn_valid(pte_pfn(pte)))
1320 		return pp;
1321 
1322 	folio = page_folio(pte_page(pte));
1323 
1324 	/* page is dirty */
1325 	if (!test_bit(PG_dcache_clean, &folio->flags) &&
1326 	    !folio_test_reserved(folio)) {
1327 		if (trap == INTERRUPT_INST_STORAGE) {
1328 			flush_dcache_icache_folio(folio);
1329 			set_bit(PG_dcache_clean, &folio->flags);
1330 		} else
1331 			pp |= HPTE_R_N;
1332 	}
1333 	return pp;
1334 }
1335 
1336 static unsigned int get_paca_psize(unsigned long addr)
1337 {
1338 	unsigned char *psizes;
1339 	unsigned long index, mask_index;
1340 
1341 	if (addr < SLICE_LOW_TOP) {
1342 		psizes = get_paca()->mm_ctx_low_slices_psize;
1343 		index = GET_LOW_SLICE_INDEX(addr);
1344 	} else {
1345 		psizes = get_paca()->mm_ctx_high_slices_psize;
1346 		index = GET_HIGH_SLICE_INDEX(addr);
1347 	}
1348 	mask_index = index & 0x1;
1349 	return (psizes[index >> 1] >> (mask_index * 4)) & 0xF;
1350 }
1351 
1352 
1353 /*
1354  * Demote a segment to using 4k pages.
1355  * For now this makes the whole process use 4k pages.
1356  */
1357 #ifdef CONFIG_PPC_64K_PAGES
1358 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
1359 {
1360 	if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
1361 		return;
1362 	slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
1363 	copro_flush_all_slbs(mm);
1364 	if ((get_paca_psize(addr) != MMU_PAGE_4K) && (current->mm == mm)) {
1365 
1366 		copy_mm_to_paca(mm);
1367 		slb_flush_and_restore_bolted();
1368 	}
1369 }
1370 #endif /* CONFIG_PPC_64K_PAGES */
1371 
1372 #ifdef CONFIG_PPC_SUBPAGE_PROT
1373 /*
1374  * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
1375  * Userspace sets the subpage permissions using the subpage_prot system call.
1376  *
1377  * Result is 0: full permissions, _PAGE_RW: read-only,
1378  * _PAGE_RWX: no access.
1379  */
1380 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
1381 {
1382 	struct subpage_prot_table *spt = mm_ctx_subpage_prot(&mm->context);
1383 	u32 spp = 0;
1384 	u32 **sbpm, *sbpp;
1385 
1386 	if (!spt)
1387 		return 0;
1388 
1389 	if (ea >= spt->maxaddr)
1390 		return 0;
1391 	if (ea < 0x100000000UL) {
1392 		/* addresses below 4GB use spt->low_prot */
1393 		sbpm = spt->low_prot;
1394 	} else {
1395 		sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
1396 		if (!sbpm)
1397 			return 0;
1398 	}
1399 	sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
1400 	if (!sbpp)
1401 		return 0;
1402 	spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
1403 
1404 	/* extract 2-bit bitfield for this 4k subpage */
1405 	spp >>= 30 - 2 * ((ea >> 12) & 0xf);
1406 
1407 	/*
1408 	 * 0 -> full permission
1409 	 * 1 -> Read only
1410 	 * 2 -> no access.
1411 	 * We return the flag that need to be cleared.
1412 	 */
1413 	spp = ((spp & 2) ? _PAGE_RWX : 0) | ((spp & 1) ? _PAGE_WRITE : 0);
1414 	return spp;
1415 }
1416 
1417 #else /* CONFIG_PPC_SUBPAGE_PROT */
1418 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
1419 {
1420 	return 0;
1421 }
1422 #endif
1423 
1424 void hash_failure_debug(unsigned long ea, unsigned long access,
1425 			unsigned long vsid, unsigned long trap,
1426 			int ssize, int psize, int lpsize, unsigned long pte)
1427 {
1428 	if (!printk_ratelimit())
1429 		return;
1430 	pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
1431 		ea, access, current->comm);
1432 	pr_info("    trap=0x%lx vsid=0x%lx ssize=%d base psize=%d psize %d pte=0x%lx\n",
1433 		trap, vsid, ssize, psize, lpsize, pte);
1434 }
1435 
1436 static void check_paca_psize(unsigned long ea, struct mm_struct *mm,
1437 			     int psize, bool user_region)
1438 {
1439 	if (user_region) {
1440 		if (psize != get_paca_psize(ea)) {
1441 			copy_mm_to_paca(mm);
1442 			slb_flush_and_restore_bolted();
1443 		}
1444 	} else if (get_paca()->vmalloc_sllp !=
1445 		   mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1446 		get_paca()->vmalloc_sllp =
1447 			mmu_psize_defs[mmu_vmalloc_psize].sllp;
1448 		slb_vmalloc_update();
1449 	}
1450 }
1451 
1452 /*
1453  * Result code is:
1454  *  0 - handled
1455  *  1 - normal page fault
1456  * -1 - critical hash insertion error
1457  * -2 - access not permitted by subpage protection mechanism
1458  */
1459 int hash_page_mm(struct mm_struct *mm, unsigned long ea,
1460 		 unsigned long access, unsigned long trap,
1461 		 unsigned long flags)
1462 {
1463 	bool is_thp;
1464 	pgd_t *pgdir;
1465 	unsigned long vsid;
1466 	pte_t *ptep;
1467 	unsigned hugeshift;
1468 	int rc, user_region = 0;
1469 	int psize, ssize;
1470 
1471 	DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
1472 		ea, access, trap);
1473 	trace_hash_fault(ea, access, trap);
1474 
1475 	/* Get region & vsid */
1476 	switch (get_region_id(ea)) {
1477 	case USER_REGION_ID:
1478 		user_region = 1;
1479 		if (! mm) {
1480 			DBG_LOW(" user region with no mm !\n");
1481 			rc = 1;
1482 			goto bail;
1483 		}
1484 		psize = get_slice_psize(mm, ea);
1485 		ssize = user_segment_size(ea);
1486 		vsid = get_user_vsid(&mm->context, ea, ssize);
1487 		break;
1488 	case VMALLOC_REGION_ID:
1489 		vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
1490 		psize = mmu_vmalloc_psize;
1491 		ssize = mmu_kernel_ssize;
1492 		flags |= HPTE_USE_KERNEL_KEY;
1493 		break;
1494 
1495 	case IO_REGION_ID:
1496 		vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
1497 		psize = mmu_io_psize;
1498 		ssize = mmu_kernel_ssize;
1499 		flags |= HPTE_USE_KERNEL_KEY;
1500 		break;
1501 	default:
1502 		/*
1503 		 * Not a valid range
1504 		 * Send the problem up to do_page_fault()
1505 		 */
1506 		rc = 1;
1507 		goto bail;
1508 	}
1509 	DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
1510 
1511 	/* Bad address. */
1512 	if (!vsid) {
1513 		DBG_LOW("Bad address!\n");
1514 		rc = 1;
1515 		goto bail;
1516 	}
1517 	/* Get pgdir */
1518 	pgdir = mm->pgd;
1519 	if (pgdir == NULL) {
1520 		rc = 1;
1521 		goto bail;
1522 	}
1523 
1524 	/* Check CPU locality */
1525 	if (user_region && mm_is_thread_local(mm))
1526 		flags |= HPTE_LOCAL_UPDATE;
1527 
1528 #ifndef CONFIG_PPC_64K_PAGES
1529 	/*
1530 	 * If we use 4K pages and our psize is not 4K, then we might
1531 	 * be hitting a special driver mapping, and need to align the
1532 	 * address before we fetch the PTE.
1533 	 *
1534 	 * It could also be a hugepage mapping, in which case this is
1535 	 * not necessary, but it's not harmful, either.
1536 	 */
1537 	if (psize != MMU_PAGE_4K)
1538 		ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
1539 #endif /* CONFIG_PPC_64K_PAGES */
1540 
1541 	/* Get PTE and page size from page tables */
1542 	ptep = find_linux_pte(pgdir, ea, &is_thp, &hugeshift);
1543 	if (ptep == NULL || !pte_present(*ptep)) {
1544 		DBG_LOW(" no PTE !\n");
1545 		rc = 1;
1546 		goto bail;
1547 	}
1548 
1549 	/*
1550 	 * Add _PAGE_PRESENT to the required access perm. If there are parallel
1551 	 * updates to the pte that can possibly clear _PAGE_PTE, catch that too.
1552 	 *
1553 	 * We can safely use the return pte address in rest of the function
1554 	 * because we do set H_PAGE_BUSY which prevents further updates to pte
1555 	 * from generic code.
1556 	 */
1557 	access |= _PAGE_PRESENT | _PAGE_PTE;
1558 
1559 	/*
1560 	 * Pre-check access permissions (will be re-checked atomically
1561 	 * in __hash_page_XX but this pre-check is a fast path
1562 	 */
1563 	if (!check_pte_access(access, pte_val(*ptep))) {
1564 		DBG_LOW(" no access !\n");
1565 		rc = 1;
1566 		goto bail;
1567 	}
1568 
1569 	if (hugeshift) {
1570 		if (is_thp)
1571 			rc = __hash_page_thp(ea, access, vsid, (pmd_t *)ptep,
1572 					     trap, flags, ssize, psize);
1573 #ifdef CONFIG_HUGETLB_PAGE
1574 		else
1575 			rc = __hash_page_huge(ea, access, vsid, ptep, trap,
1576 					      flags, ssize, hugeshift, psize);
1577 #else
1578 		else {
1579 			/*
1580 			 * if we have hugeshift, and is not transhuge with
1581 			 * hugetlb disabled, something is really wrong.
1582 			 */
1583 			rc = 1;
1584 			WARN_ON(1);
1585 		}
1586 #endif
1587 		if (current->mm == mm)
1588 			check_paca_psize(ea, mm, psize, user_region);
1589 
1590 		goto bail;
1591 	}
1592 
1593 #ifndef CONFIG_PPC_64K_PAGES
1594 	DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1595 #else
1596 	DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1597 		pte_val(*(ptep + PTRS_PER_PTE)));
1598 #endif
1599 	/* Do actual hashing */
1600 #ifdef CONFIG_PPC_64K_PAGES
1601 	/* If H_PAGE_4K_PFN is set, make sure this is a 4k segment */
1602 	if ((pte_val(*ptep) & H_PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1603 		demote_segment_4k(mm, ea);
1604 		psize = MMU_PAGE_4K;
1605 	}
1606 
1607 	/*
1608 	 * If this PTE is non-cacheable and we have restrictions on
1609 	 * using non cacheable large pages, then we switch to 4k
1610 	 */
1611 	if (mmu_ci_restrictions && psize == MMU_PAGE_64K && pte_ci(*ptep)) {
1612 		if (user_region) {
1613 			demote_segment_4k(mm, ea);
1614 			psize = MMU_PAGE_4K;
1615 		} else if (ea < VMALLOC_END) {
1616 			/*
1617 			 * some driver did a non-cacheable mapping
1618 			 * in vmalloc space, so switch vmalloc
1619 			 * to 4k pages
1620 			 */
1621 			printk(KERN_ALERT "Reducing vmalloc segment "
1622 			       "to 4kB pages because of "
1623 			       "non-cacheable mapping\n");
1624 			psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1625 			copro_flush_all_slbs(mm);
1626 		}
1627 	}
1628 
1629 #endif /* CONFIG_PPC_64K_PAGES */
1630 
1631 	if (current->mm == mm)
1632 		check_paca_psize(ea, mm, psize, user_region);
1633 
1634 #ifdef CONFIG_PPC_64K_PAGES
1635 	if (psize == MMU_PAGE_64K)
1636 		rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1637 				     flags, ssize);
1638 	else
1639 #endif /* CONFIG_PPC_64K_PAGES */
1640 	{
1641 		int spp = subpage_protection(mm, ea);
1642 		if (access & spp)
1643 			rc = -2;
1644 		else
1645 			rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1646 					    flags, ssize, spp);
1647 	}
1648 
1649 	/*
1650 	 * Dump some info in case of hash insertion failure, they should
1651 	 * never happen so it is really useful to know if/when they do
1652 	 */
1653 	if (rc == -1)
1654 		hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1655 				   psize, pte_val(*ptep));
1656 #ifndef CONFIG_PPC_64K_PAGES
1657 	DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1658 #else
1659 	DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1660 		pte_val(*(ptep + PTRS_PER_PTE)));
1661 #endif
1662 	DBG_LOW(" -> rc=%d\n", rc);
1663 
1664 bail:
1665 	return rc;
1666 }
1667 EXPORT_SYMBOL_GPL(hash_page_mm);
1668 
1669 int hash_page(unsigned long ea, unsigned long access, unsigned long trap,
1670 	      unsigned long dsisr)
1671 {
1672 	unsigned long flags = 0;
1673 	struct mm_struct *mm = current->mm;
1674 
1675 	if ((get_region_id(ea) == VMALLOC_REGION_ID) ||
1676 	    (get_region_id(ea) == IO_REGION_ID))
1677 		mm = &init_mm;
1678 
1679 	if (dsisr & DSISR_NOHPTE)
1680 		flags |= HPTE_NOHPTE_UPDATE;
1681 
1682 	return hash_page_mm(mm, ea, access, trap, flags);
1683 }
1684 EXPORT_SYMBOL_GPL(hash_page);
1685 
1686 DEFINE_INTERRUPT_HANDLER(do_hash_fault)
1687 {
1688 	unsigned long ea = regs->dar;
1689 	unsigned long dsisr = regs->dsisr;
1690 	unsigned long access = _PAGE_PRESENT | _PAGE_READ;
1691 	unsigned long flags = 0;
1692 	struct mm_struct *mm;
1693 	unsigned int region_id;
1694 	long err;
1695 
1696 	if (unlikely(dsisr & (DSISR_BAD_FAULT_64S | DSISR_KEYFAULT))) {
1697 		hash__do_page_fault(regs);
1698 		return;
1699 	}
1700 
1701 	region_id = get_region_id(ea);
1702 	if ((region_id == VMALLOC_REGION_ID) || (region_id == IO_REGION_ID))
1703 		mm = &init_mm;
1704 	else
1705 		mm = current->mm;
1706 
1707 	if (dsisr & DSISR_NOHPTE)
1708 		flags |= HPTE_NOHPTE_UPDATE;
1709 
1710 	if (dsisr & DSISR_ISSTORE)
1711 		access |= _PAGE_WRITE;
1712 	/*
1713 	 * We set _PAGE_PRIVILEGED only when
1714 	 * kernel mode access kernel space.
1715 	 *
1716 	 * _PAGE_PRIVILEGED is NOT set
1717 	 * 1) when kernel mode access user space
1718 	 * 2) user space access kernel space.
1719 	 */
1720 	access |= _PAGE_PRIVILEGED;
1721 	if (user_mode(regs) || (region_id == USER_REGION_ID))
1722 		access &= ~_PAGE_PRIVILEGED;
1723 
1724 	if (TRAP(regs) == INTERRUPT_INST_STORAGE)
1725 		access |= _PAGE_EXEC;
1726 
1727 	err = hash_page_mm(mm, ea, access, TRAP(regs), flags);
1728 	if (unlikely(err < 0)) {
1729 		// failed to insert a hash PTE due to an hypervisor error
1730 		if (user_mode(regs)) {
1731 			if (IS_ENABLED(CONFIG_PPC_SUBPAGE_PROT) && err == -2)
1732 				_exception(SIGSEGV, regs, SEGV_ACCERR, ea);
1733 			else
1734 				_exception(SIGBUS, regs, BUS_ADRERR, ea);
1735 		} else {
1736 			bad_page_fault(regs, SIGBUS);
1737 		}
1738 		err = 0;
1739 
1740 	} else if (err) {
1741 		hash__do_page_fault(regs);
1742 	}
1743 }
1744 
1745 static bool should_hash_preload(struct mm_struct *mm, unsigned long ea)
1746 {
1747 	int psize = get_slice_psize(mm, ea);
1748 
1749 	/* We only prefault standard pages for now */
1750 	if (unlikely(psize != mm_ctx_user_psize(&mm->context)))
1751 		return false;
1752 
1753 	/*
1754 	 * Don't prefault if subpage protection is enabled for the EA.
1755 	 */
1756 	if (unlikely((psize == MMU_PAGE_4K) && subpage_protection(mm, ea)))
1757 		return false;
1758 
1759 	return true;
1760 }
1761 
1762 static void hash_preload(struct mm_struct *mm, pte_t *ptep, unsigned long ea,
1763 			 bool is_exec, unsigned long trap)
1764 {
1765 	unsigned long vsid;
1766 	pgd_t *pgdir;
1767 	int rc, ssize, update_flags = 0;
1768 	unsigned long access = _PAGE_PRESENT | _PAGE_READ | (is_exec ? _PAGE_EXEC : 0);
1769 	unsigned long flags;
1770 
1771 	BUG_ON(get_region_id(ea) != USER_REGION_ID);
1772 
1773 	if (!should_hash_preload(mm, ea))
1774 		return;
1775 
1776 	DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1777 		" trap=%lx\n", mm, mm->pgd, ea, access, trap);
1778 
1779 	/* Get Linux PTE if available */
1780 	pgdir = mm->pgd;
1781 	if (pgdir == NULL)
1782 		return;
1783 
1784 	/* Get VSID */
1785 	ssize = user_segment_size(ea);
1786 	vsid = get_user_vsid(&mm->context, ea, ssize);
1787 	if (!vsid)
1788 		return;
1789 
1790 #ifdef CONFIG_PPC_64K_PAGES
1791 	/* If either H_PAGE_4K_PFN or cache inhibited is set (and we are on
1792 	 * a 64K kernel), then we don't preload, hash_page() will take
1793 	 * care of it once we actually try to access the page.
1794 	 * That way we don't have to duplicate all of the logic for segment
1795 	 * page size demotion here
1796 	 * Called with  PTL held, hence can be sure the value won't change in
1797 	 * between.
1798 	 */
1799 	if ((pte_val(*ptep) & H_PAGE_4K_PFN) || pte_ci(*ptep))
1800 		return;
1801 #endif /* CONFIG_PPC_64K_PAGES */
1802 
1803 	/*
1804 	 * __hash_page_* must run with interrupts off, including PMI interrupts
1805 	 * off, as it sets the H_PAGE_BUSY bit.
1806 	 *
1807 	 * It's otherwise possible for perf interrupts to hit at any time and
1808 	 * may take a hash fault reading the user stack, which could take a
1809 	 * hash miss and deadlock on the same H_PAGE_BUSY bit.
1810 	 *
1811 	 * Interrupts must also be off for the duration of the
1812 	 * mm_is_thread_local test and update, to prevent preempt running the
1813 	 * mm on another CPU (XXX: this may be racy vs kthread_use_mm).
1814 	 */
1815 	powerpc_local_irq_pmu_save(flags);
1816 
1817 	/* Is that local to this CPU ? */
1818 	if (mm_is_thread_local(mm))
1819 		update_flags |= HPTE_LOCAL_UPDATE;
1820 
1821 	/* Hash it in */
1822 #ifdef CONFIG_PPC_64K_PAGES
1823 	if (mm_ctx_user_psize(&mm->context) == MMU_PAGE_64K)
1824 		rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1825 				     update_flags, ssize);
1826 	else
1827 #endif /* CONFIG_PPC_64K_PAGES */
1828 		rc = __hash_page_4K(ea, access, vsid, ptep, trap, update_flags,
1829 				    ssize, subpage_protection(mm, ea));
1830 
1831 	/* Dump some info in case of hash insertion failure, they should
1832 	 * never happen so it is really useful to know if/when they do
1833 	 */
1834 	if (rc == -1)
1835 		hash_failure_debug(ea, access, vsid, trap, ssize,
1836 				   mm_ctx_user_psize(&mm->context),
1837 				   mm_ctx_user_psize(&mm->context),
1838 				   pte_val(*ptep));
1839 
1840 	powerpc_local_irq_pmu_restore(flags);
1841 }
1842 
1843 /*
1844  * This is called at the end of handling a user page fault, when the
1845  * fault has been handled by updating a PTE in the linux page tables.
1846  * We use it to preload an HPTE into the hash table corresponding to
1847  * the updated linux PTE.
1848  *
1849  * This must always be called with the pte lock held.
1850  */
1851 void __update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
1852 		      pte_t *ptep)
1853 {
1854 	/*
1855 	 * We don't need to worry about _PAGE_PRESENT here because we are
1856 	 * called with either mm->page_table_lock held or ptl lock held
1857 	 */
1858 	unsigned long trap;
1859 	bool is_exec;
1860 
1861 	/* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
1862 	if (!pte_young(*ptep) || address >= TASK_SIZE)
1863 		return;
1864 
1865 	/*
1866 	 * We try to figure out if we are coming from an instruction
1867 	 * access fault and pass that down to __hash_page so we avoid
1868 	 * double-faulting on execution of fresh text. We have to test
1869 	 * for regs NULL since init will get here first thing at boot.
1870 	 *
1871 	 * We also avoid filling the hash if not coming from a fault.
1872 	 */
1873 
1874 	trap = current->thread.regs ? TRAP(current->thread.regs) : 0UL;
1875 	switch (trap) {
1876 	case 0x300:
1877 		is_exec = false;
1878 		break;
1879 	case 0x400:
1880 		is_exec = true;
1881 		break;
1882 	default:
1883 		return;
1884 	}
1885 
1886 	hash_preload(vma->vm_mm, ptep, address, is_exec, trap);
1887 }
1888 
1889 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1890 static inline void tm_flush_hash_page(int local)
1891 {
1892 	/*
1893 	 * Transactions are not aborted by tlbiel, only tlbie. Without, syncing a
1894 	 * page back to a block device w/PIO could pick up transactional data
1895 	 * (bad!) so we force an abort here. Before the sync the page will be
1896 	 * made read-only, which will flush_hash_page. BIG ISSUE here: if the
1897 	 * kernel uses a page from userspace without unmapping it first, it may
1898 	 * see the speculated version.
1899 	 */
1900 	if (local && cpu_has_feature(CPU_FTR_TM) && current->thread.regs &&
1901 	    MSR_TM_ACTIVE(current->thread.regs->msr)) {
1902 		tm_enable();
1903 		tm_abort(TM_CAUSE_TLBI);
1904 	}
1905 }
1906 #else
1907 static inline void tm_flush_hash_page(int local)
1908 {
1909 }
1910 #endif
1911 
1912 /*
1913  * Return the global hash slot, corresponding to the given PTE, which contains
1914  * the HPTE.
1915  */
1916 unsigned long pte_get_hash_gslot(unsigned long vpn, unsigned long shift,
1917 		int ssize, real_pte_t rpte, unsigned int subpg_index)
1918 {
1919 	unsigned long hash, gslot, hidx;
1920 
1921 	hash = hpt_hash(vpn, shift, ssize);
1922 	hidx = __rpte_to_hidx(rpte, subpg_index);
1923 	if (hidx & _PTEIDX_SECONDARY)
1924 		hash = ~hash;
1925 	gslot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1926 	gslot += hidx & _PTEIDX_GROUP_IX;
1927 	return gslot;
1928 }
1929 
1930 void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1931 		     unsigned long flags)
1932 {
1933 	unsigned long index, shift, gslot;
1934 	int local = flags & HPTE_LOCAL_UPDATE;
1935 
1936 	DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1937 	pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1938 		gslot = pte_get_hash_gslot(vpn, shift, ssize, pte, index);
1939 		DBG_LOW(" sub %ld: gslot=%lx\n", index, gslot);
1940 		/*
1941 		 * We use same base page size and actual psize, because we don't
1942 		 * use these functions for hugepage
1943 		 */
1944 		mmu_hash_ops.hpte_invalidate(gslot, vpn, psize, psize,
1945 					     ssize, local);
1946 	} pte_iterate_hashed_end();
1947 
1948 	tm_flush_hash_page(local);
1949 }
1950 
1951 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1952 void flush_hash_hugepage(unsigned long vsid, unsigned long addr,
1953 			 pmd_t *pmdp, unsigned int psize, int ssize,
1954 			 unsigned long flags)
1955 {
1956 	int i, max_hpte_count, valid;
1957 	unsigned long s_addr;
1958 	unsigned char *hpte_slot_array;
1959 	unsigned long hidx, shift, vpn, hash, slot;
1960 	int local = flags & HPTE_LOCAL_UPDATE;
1961 
1962 	s_addr = addr & HPAGE_PMD_MASK;
1963 	hpte_slot_array = get_hpte_slot_array(pmdp);
1964 	/*
1965 	 * IF we try to do a HUGE PTE update after a withdraw is done.
1966 	 * we will find the below NULL. This happens when we do
1967 	 * split_huge_pmd
1968 	 */
1969 	if (!hpte_slot_array)
1970 		return;
1971 
1972 	if (mmu_hash_ops.hugepage_invalidate) {
1973 		mmu_hash_ops.hugepage_invalidate(vsid, s_addr, hpte_slot_array,
1974 						 psize, ssize, local);
1975 		goto tm_abort;
1976 	}
1977 	/*
1978 	 * No bluk hpte removal support, invalidate each entry
1979 	 */
1980 	shift = mmu_psize_defs[psize].shift;
1981 	max_hpte_count = HPAGE_PMD_SIZE >> shift;
1982 	for (i = 0; i < max_hpte_count; i++) {
1983 		/*
1984 		 * 8 bits per each hpte entries
1985 		 * 000| [ secondary group (one bit) | hidx (3 bits) | valid bit]
1986 		 */
1987 		valid = hpte_valid(hpte_slot_array, i);
1988 		if (!valid)
1989 			continue;
1990 		hidx =  hpte_hash_index(hpte_slot_array, i);
1991 
1992 		/* get the vpn */
1993 		addr = s_addr + (i * (1ul << shift));
1994 		vpn = hpt_vpn(addr, vsid, ssize);
1995 		hash = hpt_hash(vpn, shift, ssize);
1996 		if (hidx & _PTEIDX_SECONDARY)
1997 			hash = ~hash;
1998 
1999 		slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
2000 		slot += hidx & _PTEIDX_GROUP_IX;
2001 		mmu_hash_ops.hpte_invalidate(slot, vpn, psize,
2002 					     MMU_PAGE_16M, ssize, local);
2003 	}
2004 tm_abort:
2005 	tm_flush_hash_page(local);
2006 }
2007 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2008 
2009 void flush_hash_range(unsigned long number, int local)
2010 {
2011 	if (mmu_hash_ops.flush_hash_range)
2012 		mmu_hash_ops.flush_hash_range(number, local);
2013 	else {
2014 		int i;
2015 		struct ppc64_tlb_batch *batch =
2016 			this_cpu_ptr(&ppc64_tlb_batch);
2017 
2018 		for (i = 0; i < number; i++)
2019 			flush_hash_page(batch->vpn[i], batch->pte[i],
2020 					batch->psize, batch->ssize, local);
2021 	}
2022 }
2023 
2024 long hpte_insert_repeating(unsigned long hash, unsigned long vpn,
2025 			   unsigned long pa, unsigned long rflags,
2026 			   unsigned long vflags, int psize, int ssize)
2027 {
2028 	unsigned long hpte_group;
2029 	long slot;
2030 
2031 repeat:
2032 	hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
2033 
2034 	/* Insert into the hash table, primary slot */
2035 	slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags, vflags,
2036 					psize, psize, ssize);
2037 
2038 	/* Primary is full, try the secondary */
2039 	if (unlikely(slot == -1)) {
2040 		hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
2041 		slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags,
2042 						vflags | HPTE_V_SECONDARY,
2043 						psize, psize, ssize);
2044 		if (slot == -1) {
2045 			if (mftb() & 0x1)
2046 				hpte_group = (hash & htab_hash_mask) *
2047 						HPTES_PER_GROUP;
2048 
2049 			mmu_hash_ops.hpte_remove(hpte_group);
2050 			goto repeat;
2051 		}
2052 	}
2053 
2054 	return slot;
2055 }
2056 
2057 void hpt_clear_stress(void)
2058 {
2059 	int cpu = raw_smp_processor_id();
2060 	int g;
2061 
2062 	for (g = 0; g < stress_nr_groups(); g++) {
2063 		unsigned long last_group;
2064 		last_group = stress_hpt_struct[cpu].last_group[g];
2065 
2066 		if (last_group != -1UL) {
2067 			int i;
2068 			for (i = 0; i < HPTES_PER_GROUP; i++) {
2069 				if (mmu_hash_ops.hpte_remove(last_group) == -1)
2070 					break;
2071 			}
2072 			stress_hpt_struct[cpu].last_group[g] = -1;
2073 		}
2074 	}
2075 }
2076 
2077 void hpt_do_stress(unsigned long ea, unsigned long hpte_group)
2078 {
2079 	unsigned long last_group;
2080 	int cpu = raw_smp_processor_id();
2081 
2082 	last_group = stress_hpt_struct[cpu].last_group[stress_nr_groups() - 1];
2083 	if (hpte_group == last_group)
2084 		return;
2085 
2086 	if (last_group != -1UL) {
2087 		int i;
2088 		/*
2089 		 * Concurrent CPUs might be inserting into this group, so
2090 		 * give up after a number of iterations, to prevent a live
2091 		 * lock.
2092 		 */
2093 		for (i = 0; i < HPTES_PER_GROUP; i++) {
2094 			if (mmu_hash_ops.hpte_remove(last_group) == -1)
2095 				break;
2096 		}
2097 		stress_hpt_struct[cpu].last_group[stress_nr_groups() - 1] = -1;
2098 	}
2099 
2100 	if (ea >= PAGE_OFFSET) {
2101 		/*
2102 		 * We would really like to prefetch to get the TLB loaded, then
2103 		 * remove the PTE before returning from fault interrupt, to
2104 		 * increase the hash fault rate.
2105 		 *
2106 		 * Unfortunately QEMU TCG does not model the TLB in a way that
2107 		 * makes this possible, and systemsim (mambo) emulator does not
2108 		 * bring in TLBs with prefetches (although loads/stores do
2109 		 * work for non-CI PTEs).
2110 		 *
2111 		 * So remember this PTE and clear it on the next hash fault.
2112 		 */
2113 		memmove(&stress_hpt_struct[cpu].last_group[1],
2114 			&stress_hpt_struct[cpu].last_group[0],
2115 			(stress_nr_groups() - 1) * sizeof(unsigned long));
2116 		stress_hpt_struct[cpu].last_group[0] = hpte_group;
2117 	}
2118 }
2119 
2120 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KFENCE)
2121 static DEFINE_RAW_SPINLOCK(linear_map_hash_lock);
2122 
2123 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
2124 {
2125 	unsigned long hash;
2126 	unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
2127 	unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
2128 	unsigned long mode = htab_convert_pte_flags(pgprot_val(PAGE_KERNEL), HPTE_USE_KERNEL_KEY);
2129 	long ret;
2130 
2131 	hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
2132 
2133 	/* Don't create HPTE entries for bad address */
2134 	if (!vsid)
2135 		return;
2136 
2137 	if (linear_map_hash_slots[lmi] & 0x80)
2138 		return;
2139 
2140 	ret = hpte_insert_repeating(hash, vpn, __pa(vaddr), mode,
2141 				    HPTE_V_BOLTED,
2142 				    mmu_linear_psize, mmu_kernel_ssize);
2143 
2144 	BUG_ON (ret < 0);
2145 	raw_spin_lock(&linear_map_hash_lock);
2146 	BUG_ON(linear_map_hash_slots[lmi] & 0x80);
2147 	linear_map_hash_slots[lmi] = ret | 0x80;
2148 	raw_spin_unlock(&linear_map_hash_lock);
2149 }
2150 
2151 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
2152 {
2153 	unsigned long hash, hidx, slot;
2154 	unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
2155 	unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
2156 
2157 	hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
2158 	raw_spin_lock(&linear_map_hash_lock);
2159 	if (!(linear_map_hash_slots[lmi] & 0x80)) {
2160 		raw_spin_unlock(&linear_map_hash_lock);
2161 		return;
2162 	}
2163 	hidx = linear_map_hash_slots[lmi] & 0x7f;
2164 	linear_map_hash_slots[lmi] = 0;
2165 	raw_spin_unlock(&linear_map_hash_lock);
2166 	if (hidx & _PTEIDX_SECONDARY)
2167 		hash = ~hash;
2168 	slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
2169 	slot += hidx & _PTEIDX_GROUP_IX;
2170 	mmu_hash_ops.hpte_invalidate(slot, vpn, mmu_linear_psize,
2171 				     mmu_linear_psize,
2172 				     mmu_kernel_ssize, 0);
2173 }
2174 
2175 void hash__kernel_map_pages(struct page *page, int numpages, int enable)
2176 {
2177 	unsigned long flags, vaddr, lmi;
2178 	int i;
2179 
2180 	local_irq_save(flags);
2181 	for (i = 0; i < numpages; i++, page++) {
2182 		vaddr = (unsigned long)page_address(page);
2183 		lmi = __pa(vaddr) >> PAGE_SHIFT;
2184 		if (lmi >= linear_map_hash_count)
2185 			continue;
2186 		if (enable)
2187 			kernel_map_linear_page(vaddr, lmi);
2188 		else
2189 			kernel_unmap_linear_page(vaddr, lmi);
2190 	}
2191 	local_irq_restore(flags);
2192 }
2193 #endif /* CONFIG_DEBUG_PAGEALLOC || CONFIG_KFENCE */
2194 
2195 void hash__setup_initial_memory_limit(phys_addr_t first_memblock_base,
2196 				phys_addr_t first_memblock_size)
2197 {
2198 	/*
2199 	 * We don't currently support the first MEMBLOCK not mapping 0
2200 	 * physical on those processors
2201 	 */
2202 	BUG_ON(first_memblock_base != 0);
2203 
2204 	/*
2205 	 * On virtualized systems the first entry is our RMA region aka VRMA,
2206 	 * non-virtualized 64-bit hash MMU systems don't have a limitation
2207 	 * on real mode access.
2208 	 *
2209 	 * For guests on platforms before POWER9, we clamp the it limit to 1G
2210 	 * to avoid some funky things such as RTAS bugs etc...
2211 	 *
2212 	 * On POWER9 we limit to 1TB in case the host erroneously told us that
2213 	 * the RMA was >1TB. Effective address bits 0:23 are treated as zero
2214 	 * (meaning the access is aliased to zero i.e. addr = addr % 1TB)
2215 	 * for virtual real mode addressing and so it doesn't make sense to
2216 	 * have an area larger than 1TB as it can't be addressed.
2217 	 */
2218 	if (!early_cpu_has_feature(CPU_FTR_HVMODE)) {
2219 		ppc64_rma_size = first_memblock_size;
2220 		if (!early_cpu_has_feature(CPU_FTR_ARCH_300))
2221 			ppc64_rma_size = min_t(u64, ppc64_rma_size, 0x40000000);
2222 		else
2223 			ppc64_rma_size = min_t(u64, ppc64_rma_size,
2224 					       1UL << SID_SHIFT_1T);
2225 
2226 		/* Finally limit subsequent allocations */
2227 		memblock_set_current_limit(ppc64_rma_size);
2228 	} else {
2229 		ppc64_rma_size = ULONG_MAX;
2230 	}
2231 }
2232 
2233 #ifdef CONFIG_DEBUG_FS
2234 
2235 static int hpt_order_get(void *data, u64 *val)
2236 {
2237 	*val = ppc64_pft_size;
2238 	return 0;
2239 }
2240 
2241 static int hpt_order_set(void *data, u64 val)
2242 {
2243 	int ret;
2244 
2245 	if (!mmu_hash_ops.resize_hpt)
2246 		return -ENODEV;
2247 
2248 	cpus_read_lock();
2249 	ret = mmu_hash_ops.resize_hpt(val);
2250 	cpus_read_unlock();
2251 
2252 	return ret;
2253 }
2254 
2255 DEFINE_DEBUGFS_ATTRIBUTE(fops_hpt_order, hpt_order_get, hpt_order_set, "%llu\n");
2256 
2257 static int __init hash64_debugfs(void)
2258 {
2259 	debugfs_create_file("hpt_order", 0600, arch_debugfs_dir, NULL,
2260 			    &fops_hpt_order);
2261 	return 0;
2262 }
2263 machine_device_initcall(pseries, hash64_debugfs);
2264 #endif /* CONFIG_DEBUG_FS */
2265 
2266 void __init print_system_hash_info(void)
2267 {
2268 	pr_info("ppc64_pft_size    = 0x%llx\n", ppc64_pft_size);
2269 
2270 	if (htab_hash_mask)
2271 		pr_info("htab_hash_mask    = 0x%lx\n", htab_hash_mask);
2272 }
2273 
2274 unsigned long arch_randomize_brk(struct mm_struct *mm)
2275 {
2276 	/*
2277 	 * If we are using 1TB segments and we are allowed to randomise
2278 	 * the heap, we can put it above 1TB so it is backed by a 1TB
2279 	 * segment. Otherwise the heap will be in the bottom 1TB
2280 	 * which always uses 256MB segments and this may result in a
2281 	 * performance penalty.
2282 	 */
2283 	if (is_32bit_task())
2284 		return randomize_page(mm->brk, SZ_32M);
2285 	else if (!radix_enabled() && mmu_highuser_ssize == MMU_SEGSIZE_1T)
2286 		return randomize_page(max_t(unsigned long, mm->brk, SZ_1T), SZ_1G);
2287 	else
2288 		return randomize_page(mm->brk, SZ_1G);
2289 }
2290