1/*
2 * PowerPC64 port by Mike Corrigan and Dave Engebretsen
3 *   {mikejc|engebret}@us.ibm.com
4 *
5 *    Copyright (c) 2000 Mike Corrigan <mikejc@us.ibm.com>
6 *
7 * SMP scalability work:
8 *    Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
9 *
10 *    Module name: htab.c
11 *
12 *    Description:
13 *      PowerPC Hashed Page Table functions
14 *
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
19 */
20
21#undef DEBUG
22#undef DEBUG_LOW
23
24#include <linux/spinlock.h>
25#include <linux/errno.h>
26#include <linux/sched.h>
27#include <linux/proc_fs.h>
28#include <linux/stat.h>
29#include <linux/sysctl.h>
30#include <linux/export.h>
31#include <linux/ctype.h>
32#include <linux/cache.h>
33#include <linux/init.h>
34#include <linux/signal.h>
35#include <linux/memblock.h>
36#include <linux/context_tracking.h>
37
38#include <asm/processor.h>
39#include <asm/pgtable.h>
40#include <asm/mmu.h>
41#include <asm/mmu_context.h>
42#include <asm/page.h>
43#include <asm/types.h>
44#include <asm/uaccess.h>
45#include <asm/machdep.h>
46#include <asm/prom.h>
47#include <asm/tlbflush.h>
48#include <asm/io.h>
49#include <asm/eeh.h>
50#include <asm/tlb.h>
51#include <asm/cacheflush.h>
52#include <asm/cputable.h>
53#include <asm/sections.h>
54#include <asm/copro.h>
55#include <asm/udbg.h>
56#include <asm/code-patching.h>
57#include <asm/fadump.h>
58#include <asm/firmware.h>
59#include <asm/tm.h>
60
61#ifdef DEBUG
62#define DBG(fmt...) udbg_printf(fmt)
63#else
64#define DBG(fmt...)
65#endif
66
67#ifdef DEBUG_LOW
68#define DBG_LOW(fmt...) udbg_printf(fmt)
69#else
70#define DBG_LOW(fmt...)
71#endif
72
73#define KB (1024)
74#define MB (1024*KB)
75#define GB (1024L*MB)
76
77/*
78 * Note:  pte   --> Linux PTE
79 *        HPTE  --> PowerPC Hashed Page Table Entry
80 *
81 * Execution context:
82 *   htab_initialize is called with the MMU off (of course), but
83 *   the kernel has been copied down to zero so it can directly
84 *   reference global data.  At this point it is very difficult
85 *   to print debug info.
86 *
87 */
88
89#ifdef CONFIG_U3_DART
90extern unsigned long dart_tablebase;
91#endif /* CONFIG_U3_DART */
92
93static unsigned long _SDR1;
94struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
95EXPORT_SYMBOL_GPL(mmu_psize_defs);
96
97struct hash_pte *htab_address;
98unsigned long htab_size_bytes;
99unsigned long htab_hash_mask;
100EXPORT_SYMBOL_GPL(htab_hash_mask);
101int mmu_linear_psize = MMU_PAGE_4K;
102EXPORT_SYMBOL_GPL(mmu_linear_psize);
103int mmu_virtual_psize = MMU_PAGE_4K;
104int mmu_vmalloc_psize = MMU_PAGE_4K;
105#ifdef CONFIG_SPARSEMEM_VMEMMAP
106int mmu_vmemmap_psize = MMU_PAGE_4K;
107#endif
108int mmu_io_psize = MMU_PAGE_4K;
109int mmu_kernel_ssize = MMU_SEGSIZE_256M;
110EXPORT_SYMBOL_GPL(mmu_kernel_ssize);
111int mmu_highuser_ssize = MMU_SEGSIZE_256M;
112u16 mmu_slb_size = 64;
113EXPORT_SYMBOL_GPL(mmu_slb_size);
114#ifdef CONFIG_PPC_64K_PAGES
115int mmu_ci_restrictions;
116#endif
117#ifdef CONFIG_DEBUG_PAGEALLOC
118static u8 *linear_map_hash_slots;
119static unsigned long linear_map_hash_count;
120static DEFINE_SPINLOCK(linear_map_hash_lock);
121#endif /* CONFIG_DEBUG_PAGEALLOC */
122
123/* There are definitions of page sizes arrays to be used when none
124 * is provided by the firmware.
125 */
126
127/* Pre-POWER4 CPUs (4k pages only)
128 */
129static struct mmu_psize_def mmu_psize_defaults_old[] = {
130	[MMU_PAGE_4K] = {
131		.shift	= 12,
132		.sllp	= 0,
133		.penc   = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
134		.avpnm	= 0,
135		.tlbiel = 0,
136	},
137};
138
139/* POWER4, GPUL, POWER5
140 *
141 * Support for 16Mb large pages
142 */
143static struct mmu_psize_def mmu_psize_defaults_gp[] = {
144	[MMU_PAGE_4K] = {
145		.shift	= 12,
146		.sllp	= 0,
147		.penc   = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
148		.avpnm	= 0,
149		.tlbiel = 1,
150	},
151	[MMU_PAGE_16M] = {
152		.shift	= 24,
153		.sllp	= SLB_VSID_L,
154		.penc   = {[0 ... MMU_PAGE_16M - 1] = -1, [MMU_PAGE_16M] = 0,
155			    [MMU_PAGE_16M + 1 ... MMU_PAGE_COUNT - 1] = -1 },
156		.avpnm	= 0x1UL,
157		.tlbiel = 0,
158	},
159};
160
161static unsigned long htab_convert_pte_flags(unsigned long pteflags)
162{
163	unsigned long rflags = pteflags & 0x1fa;
164
165	/* _PAGE_EXEC -> NOEXEC */
166	if ((pteflags & _PAGE_EXEC) == 0)
167		rflags |= HPTE_R_N;
168
169	/* PP bits. PAGE_USER is already PP bit 0x2, so we only
170	 * need to add in 0x1 if it's a read-only user page
171	 */
172	if ((pteflags & _PAGE_USER) && !((pteflags & _PAGE_RW) &&
173					 (pteflags & _PAGE_DIRTY)))
174		rflags |= 1;
175	/*
176	 * Always add "C" bit for perf. Memory coherence is always enabled
177	 */
178	return rflags | HPTE_R_C | HPTE_R_M;
179}
180
181int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
182		      unsigned long pstart, unsigned long prot,
183		      int psize, int ssize)
184{
185	unsigned long vaddr, paddr;
186	unsigned int step, shift;
187	int ret = 0;
188
189	shift = mmu_psize_defs[psize].shift;
190	step = 1 << shift;
191
192	prot = htab_convert_pte_flags(prot);
193
194	DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
195	    vstart, vend, pstart, prot, psize, ssize);
196
197	for (vaddr = vstart, paddr = pstart; vaddr < vend;
198	     vaddr += step, paddr += step) {
199		unsigned long hash, hpteg;
200		unsigned long vsid = get_kernel_vsid(vaddr, ssize);
201		unsigned long vpn  = hpt_vpn(vaddr, vsid, ssize);
202		unsigned long tprot = prot;
203
204		/*
205		 * If we hit a bad address return error.
206		 */
207		if (!vsid)
208			return -1;
209		/* Make kernel text executable */
210		if (overlaps_kernel_text(vaddr, vaddr + step))
211			tprot &= ~HPTE_R_N;
212
213		/* Make kvm guest trampolines executable */
214		if (overlaps_kvm_tmp(vaddr, vaddr + step))
215			tprot &= ~HPTE_R_N;
216
217		/*
218		 * If relocatable, check if it overlaps interrupt vectors that
219		 * are copied down to real 0. For relocatable kernel
220		 * (e.g. kdump case) we copy interrupt vectors down to real
221		 * address 0. Mark that region as executable. This is
222		 * because on p8 system with relocation on exception feature
223		 * enabled, exceptions are raised with MMU (IR=DR=1) ON. Hence
224		 * in order to execute the interrupt handlers in virtual
225		 * mode the vector region need to be marked as executable.
226		 */
227		if ((PHYSICAL_START > MEMORY_START) &&
228			overlaps_interrupt_vector_text(vaddr, vaddr + step))
229				tprot &= ~HPTE_R_N;
230
231		hash = hpt_hash(vpn, shift, ssize);
232		hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
233
234		BUG_ON(!ppc_md.hpte_insert);
235		ret = ppc_md.hpte_insert(hpteg, vpn, paddr, tprot,
236					 HPTE_V_BOLTED, psize, psize, ssize);
237
238		if (ret < 0)
239			break;
240#ifdef CONFIG_DEBUG_PAGEALLOC
241		if ((paddr >> PAGE_SHIFT) < linear_map_hash_count)
242			linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
243#endif /* CONFIG_DEBUG_PAGEALLOC */
244	}
245	return ret < 0 ? ret : 0;
246}
247
248#ifdef CONFIG_MEMORY_HOTPLUG
249int htab_remove_mapping(unsigned long vstart, unsigned long vend,
250		      int psize, int ssize)
251{
252	unsigned long vaddr;
253	unsigned int step, shift;
254
255	shift = mmu_psize_defs[psize].shift;
256	step = 1 << shift;
257
258	if (!ppc_md.hpte_removebolted) {
259		printk(KERN_WARNING "Platform doesn't implement "
260				"hpte_removebolted\n");
261		return -EINVAL;
262	}
263
264	for (vaddr = vstart; vaddr < vend; vaddr += step)
265		ppc_md.hpte_removebolted(vaddr, psize, ssize);
266
267	return 0;
268}
269#endif /* CONFIG_MEMORY_HOTPLUG */
270
271static int __init htab_dt_scan_seg_sizes(unsigned long node,
272					 const char *uname, int depth,
273					 void *data)
274{
275	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
276	const __be32 *prop;
277	int size = 0;
278
279	/* We are scanning "cpu" nodes only */
280	if (type == NULL || strcmp(type, "cpu") != 0)
281		return 0;
282
283	prop = of_get_flat_dt_prop(node, "ibm,processor-segment-sizes", &size);
284	if (prop == NULL)
285		return 0;
286	for (; size >= 4; size -= 4, ++prop) {
287		if (be32_to_cpu(prop[0]) == 40) {
288			DBG("1T segment support detected\n");
289			cur_cpu_spec->mmu_features |= MMU_FTR_1T_SEGMENT;
290			return 1;
291		}
292	}
293	cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
294	return 0;
295}
296
297static void __init htab_init_seg_sizes(void)
298{
299	of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
300}
301
302static int __init get_idx_from_shift(unsigned int shift)
303{
304	int idx = -1;
305
306	switch (shift) {
307	case 0xc:
308		idx = MMU_PAGE_4K;
309		break;
310	case 0x10:
311		idx = MMU_PAGE_64K;
312		break;
313	case 0x14:
314		idx = MMU_PAGE_1M;
315		break;
316	case 0x18:
317		idx = MMU_PAGE_16M;
318		break;
319	case 0x22:
320		idx = MMU_PAGE_16G;
321		break;
322	}
323	return idx;
324}
325
326static int __init htab_dt_scan_page_sizes(unsigned long node,
327					  const char *uname, int depth,
328					  void *data)
329{
330	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
331	const __be32 *prop;
332	int size = 0;
333
334	/* We are scanning "cpu" nodes only */
335	if (type == NULL || strcmp(type, "cpu") != 0)
336		return 0;
337
338	prop = of_get_flat_dt_prop(node, "ibm,segment-page-sizes", &size);
339	if (!prop)
340		return 0;
341
342	pr_info("Page sizes from device-tree:\n");
343	size /= 4;
344	cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
345	while(size > 0) {
346		unsigned int base_shift = be32_to_cpu(prop[0]);
347		unsigned int slbenc = be32_to_cpu(prop[1]);
348		unsigned int lpnum = be32_to_cpu(prop[2]);
349		struct mmu_psize_def *def;
350		int idx, base_idx;
351
352		size -= 3; prop += 3;
353		base_idx = get_idx_from_shift(base_shift);
354		if (base_idx < 0) {
355			/* skip the pte encoding also */
356			prop += lpnum * 2; size -= lpnum * 2;
357			continue;
358		}
359		def = &mmu_psize_defs[base_idx];
360		if (base_idx == MMU_PAGE_16M)
361			cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
362
363		def->shift = base_shift;
364		if (base_shift <= 23)
365			def->avpnm = 0;
366		else
367			def->avpnm = (1 << (base_shift - 23)) - 1;
368		def->sllp = slbenc;
369		/*
370		 * We don't know for sure what's up with tlbiel, so
371		 * for now we only set it for 4K and 64K pages
372		 */
373		if (base_idx == MMU_PAGE_4K || base_idx == MMU_PAGE_64K)
374			def->tlbiel = 1;
375		else
376			def->tlbiel = 0;
377
378		while (size > 0 && lpnum) {
379			unsigned int shift = be32_to_cpu(prop[0]);
380			int penc  = be32_to_cpu(prop[1]);
381
382			prop += 2; size -= 2;
383			lpnum--;
384
385			idx = get_idx_from_shift(shift);
386			if (idx < 0)
387				continue;
388
389			if (penc == -1)
390				pr_err("Invalid penc for base_shift=%d "
391				       "shift=%d\n", base_shift, shift);
392
393			def->penc[idx] = penc;
394			pr_info("base_shift=%d: shift=%d, sllp=0x%04lx,"
395				" avpnm=0x%08lx, tlbiel=%d, penc=%d\n",
396				base_shift, shift, def->sllp,
397				def->avpnm, def->tlbiel, def->penc[idx]);
398		}
399	}
400
401	return 1;
402}
403
404#ifdef CONFIG_HUGETLB_PAGE
405/* Scan for 16G memory blocks that have been set aside for huge pages
406 * and reserve those blocks for 16G huge pages.
407 */
408static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
409					const char *uname, int depth,
410					void *data) {
411	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
412	const __be64 *addr_prop;
413	const __be32 *page_count_prop;
414	unsigned int expected_pages;
415	long unsigned int phys_addr;
416	long unsigned int block_size;
417
418	/* We are scanning "memory" nodes only */
419	if (type == NULL || strcmp(type, "memory") != 0)
420		return 0;
421
422	/* This property is the log base 2 of the number of virtual pages that
423	 * will represent this memory block. */
424	page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
425	if (page_count_prop == NULL)
426		return 0;
427	expected_pages = (1 << be32_to_cpu(page_count_prop[0]));
428	addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
429	if (addr_prop == NULL)
430		return 0;
431	phys_addr = be64_to_cpu(addr_prop[0]);
432	block_size = be64_to_cpu(addr_prop[1]);
433	if (block_size != (16 * GB))
434		return 0;
435	printk(KERN_INFO "Huge page(16GB) memory: "
436			"addr = 0x%lX size = 0x%lX pages = %d\n",
437			phys_addr, block_size, expected_pages);
438	if (phys_addr + (16 * GB) <= memblock_end_of_DRAM()) {
439		memblock_reserve(phys_addr, block_size * expected_pages);
440		add_gpage(phys_addr, block_size, expected_pages);
441	}
442	return 0;
443}
444#endif /* CONFIG_HUGETLB_PAGE */
445
446static void mmu_psize_set_default_penc(void)
447{
448	int bpsize, apsize;
449	for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
450		for (apsize = 0; apsize < MMU_PAGE_COUNT; apsize++)
451			mmu_psize_defs[bpsize].penc[apsize] = -1;
452}
453
454#ifdef CONFIG_PPC_64K_PAGES
455
456static bool might_have_hea(void)
457{
458	/*
459	 * The HEA ethernet adapter requires awareness of the
460	 * GX bus. Without that awareness we can easily assume
461	 * we will never see an HEA ethernet device.
462	 */
463#ifdef CONFIG_IBMEBUS
464	return !cpu_has_feature(CPU_FTR_ARCH_207S);
465#else
466	return false;
467#endif
468}
469
470#endif /* #ifdef CONFIG_PPC_64K_PAGES */
471
472static void __init htab_init_page_sizes(void)
473{
474	int rc;
475
476	/* se the invalid penc to -1 */
477	mmu_psize_set_default_penc();
478
479	/* Default to 4K pages only */
480	memcpy(mmu_psize_defs, mmu_psize_defaults_old,
481	       sizeof(mmu_psize_defaults_old));
482
483	/*
484	 * Try to find the available page sizes in the device-tree
485	 */
486	rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
487	if (rc != 0)  /* Found */
488		goto found;
489
490	/*
491	 * Not in the device-tree, let's fallback on known size
492	 * list for 16M capable GP & GR
493	 */
494	if (mmu_has_feature(MMU_FTR_16M_PAGE))
495		memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
496		       sizeof(mmu_psize_defaults_gp));
497 found:
498#ifndef CONFIG_DEBUG_PAGEALLOC
499	/*
500	 * Pick a size for the linear mapping. Currently, we only support
501	 * 16M, 1M and 4K which is the default
502	 */
503	if (mmu_psize_defs[MMU_PAGE_16M].shift)
504		mmu_linear_psize = MMU_PAGE_16M;
505	else if (mmu_psize_defs[MMU_PAGE_1M].shift)
506		mmu_linear_psize = MMU_PAGE_1M;
507#endif /* CONFIG_DEBUG_PAGEALLOC */
508
509#ifdef CONFIG_PPC_64K_PAGES
510	/*
511	 * Pick a size for the ordinary pages. Default is 4K, we support
512	 * 64K for user mappings and vmalloc if supported by the processor.
513	 * We only use 64k for ioremap if the processor
514	 * (and firmware) support cache-inhibited large pages.
515	 * If not, we use 4k and set mmu_ci_restrictions so that
516	 * hash_page knows to switch processes that use cache-inhibited
517	 * mappings to 4k pages.
518	 */
519	if (mmu_psize_defs[MMU_PAGE_64K].shift) {
520		mmu_virtual_psize = MMU_PAGE_64K;
521		mmu_vmalloc_psize = MMU_PAGE_64K;
522		if (mmu_linear_psize == MMU_PAGE_4K)
523			mmu_linear_psize = MMU_PAGE_64K;
524		if (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
525			/*
526			 * When running on pSeries using 64k pages for ioremap
527			 * would stop us accessing the HEA ethernet. So if we
528			 * have the chance of ever seeing one, stay at 4k.
529			 */
530			if (!might_have_hea() || !machine_is(pseries))
531				mmu_io_psize = MMU_PAGE_64K;
532		} else
533			mmu_ci_restrictions = 1;
534	}
535#endif /* CONFIG_PPC_64K_PAGES */
536
537#ifdef CONFIG_SPARSEMEM_VMEMMAP
538	/* We try to use 16M pages for vmemmap if that is supported
539	 * and we have at least 1G of RAM at boot
540	 */
541	if (mmu_psize_defs[MMU_PAGE_16M].shift &&
542	    memblock_phys_mem_size() >= 0x40000000)
543		mmu_vmemmap_psize = MMU_PAGE_16M;
544	else if (mmu_psize_defs[MMU_PAGE_64K].shift)
545		mmu_vmemmap_psize = MMU_PAGE_64K;
546	else
547		mmu_vmemmap_psize = MMU_PAGE_4K;
548#endif /* CONFIG_SPARSEMEM_VMEMMAP */
549
550	printk(KERN_DEBUG "Page orders: linear mapping = %d, "
551	       "virtual = %d, io = %d"
552#ifdef CONFIG_SPARSEMEM_VMEMMAP
553	       ", vmemmap = %d"
554#endif
555	       "\n",
556	       mmu_psize_defs[mmu_linear_psize].shift,
557	       mmu_psize_defs[mmu_virtual_psize].shift,
558	       mmu_psize_defs[mmu_io_psize].shift
559#ifdef CONFIG_SPARSEMEM_VMEMMAP
560	       ,mmu_psize_defs[mmu_vmemmap_psize].shift
561#endif
562	       );
563
564#ifdef CONFIG_HUGETLB_PAGE
565	/* Reserve 16G huge page memory sections for huge pages */
566	of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
567#endif /* CONFIG_HUGETLB_PAGE */
568}
569
570static int __init htab_dt_scan_pftsize(unsigned long node,
571				       const char *uname, int depth,
572				       void *data)
573{
574	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
575	const __be32 *prop;
576
577	/* We are scanning "cpu" nodes only */
578	if (type == NULL || strcmp(type, "cpu") != 0)
579		return 0;
580
581	prop = of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
582	if (prop != NULL) {
583		/* pft_size[0] is the NUMA CEC cookie */
584		ppc64_pft_size = be32_to_cpu(prop[1]);
585		return 1;
586	}
587	return 0;
588}
589
590static unsigned long __init htab_get_table_size(void)
591{
592	unsigned long mem_size, rnd_mem_size, pteg_count, psize;
593
594	/* If hash size isn't already provided by the platform, we try to
595	 * retrieve it from the device-tree. If it's not there neither, we
596	 * calculate it now based on the total RAM size
597	 */
598	if (ppc64_pft_size == 0)
599		of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
600	if (ppc64_pft_size)
601		return 1UL << ppc64_pft_size;
602
603	/* round mem_size up to next power of 2 */
604	mem_size = memblock_phys_mem_size();
605	rnd_mem_size = 1UL << __ilog2(mem_size);
606	if (rnd_mem_size < mem_size)
607		rnd_mem_size <<= 1;
608
609	/* # pages / 2 */
610	psize = mmu_psize_defs[mmu_virtual_psize].shift;
611	pteg_count = max(rnd_mem_size >> (psize + 1), 1UL << 11);
612
613	return pteg_count << 7;
614}
615
616#ifdef CONFIG_MEMORY_HOTPLUG
617int create_section_mapping(unsigned long start, unsigned long end)
618{
619	return htab_bolt_mapping(start, end, __pa(start),
620				 pgprot_val(PAGE_KERNEL), mmu_linear_psize,
621				 mmu_kernel_ssize);
622}
623
624int remove_section_mapping(unsigned long start, unsigned long end)
625{
626	return htab_remove_mapping(start, end, mmu_linear_psize,
627			mmu_kernel_ssize);
628}
629#endif /* CONFIG_MEMORY_HOTPLUG */
630
631extern u32 htab_call_hpte_insert1[];
632extern u32 htab_call_hpte_insert2[];
633extern u32 htab_call_hpte_remove[];
634extern u32 htab_call_hpte_updatepp[];
635extern u32 ht64_call_hpte_insert1[];
636extern u32 ht64_call_hpte_insert2[];
637extern u32 ht64_call_hpte_remove[];
638extern u32 ht64_call_hpte_updatepp[];
639
640static void __init htab_finish_init(void)
641{
642#ifdef CONFIG_PPC_HAS_HASH_64K
643	patch_branch(ht64_call_hpte_insert1,
644		ppc_function_entry(ppc_md.hpte_insert),
645		BRANCH_SET_LINK);
646	patch_branch(ht64_call_hpte_insert2,
647		ppc_function_entry(ppc_md.hpte_insert),
648		BRANCH_SET_LINK);
649	patch_branch(ht64_call_hpte_remove,
650		ppc_function_entry(ppc_md.hpte_remove),
651		BRANCH_SET_LINK);
652	patch_branch(ht64_call_hpte_updatepp,
653		ppc_function_entry(ppc_md.hpte_updatepp),
654		BRANCH_SET_LINK);
655#endif /* CONFIG_PPC_HAS_HASH_64K */
656
657	patch_branch(htab_call_hpte_insert1,
658		ppc_function_entry(ppc_md.hpte_insert),
659		BRANCH_SET_LINK);
660	patch_branch(htab_call_hpte_insert2,
661		ppc_function_entry(ppc_md.hpte_insert),
662		BRANCH_SET_LINK);
663	patch_branch(htab_call_hpte_remove,
664		ppc_function_entry(ppc_md.hpte_remove),
665		BRANCH_SET_LINK);
666	patch_branch(htab_call_hpte_updatepp,
667		ppc_function_entry(ppc_md.hpte_updatepp),
668		BRANCH_SET_LINK);
669}
670
671static void __init htab_initialize(void)
672{
673	unsigned long table;
674	unsigned long pteg_count;
675	unsigned long prot;
676	unsigned long base = 0, size = 0, limit;
677	struct memblock_region *reg;
678
679	DBG(" -> htab_initialize()\n");
680
681	/* Initialize segment sizes */
682	htab_init_seg_sizes();
683
684	/* Initialize page sizes */
685	htab_init_page_sizes();
686
687	if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
688		mmu_kernel_ssize = MMU_SEGSIZE_1T;
689		mmu_highuser_ssize = MMU_SEGSIZE_1T;
690		printk(KERN_INFO "Using 1TB segments\n");
691	}
692
693	/*
694	 * Calculate the required size of the htab.  We want the number of
695	 * PTEGs to equal one half the number of real pages.
696	 */
697	htab_size_bytes = htab_get_table_size();
698	pteg_count = htab_size_bytes >> 7;
699
700	htab_hash_mask = pteg_count - 1;
701
702	if (firmware_has_feature(FW_FEATURE_LPAR)) {
703		/* Using a hypervisor which owns the htab */
704		htab_address = NULL;
705		_SDR1 = 0;
706#ifdef CONFIG_FA_DUMP
707		/*
708		 * If firmware assisted dump is active firmware preserves
709		 * the contents of htab along with entire partition memory.
710		 * Clear the htab if firmware assisted dump is active so
711		 * that we dont end up using old mappings.
712		 */
713		if (is_fadump_active() && ppc_md.hpte_clear_all)
714			ppc_md.hpte_clear_all();
715#endif
716	} else {
717		/* Find storage for the HPT.  Must be contiguous in
718		 * the absolute address space. On cell we want it to be
719		 * in the first 2 Gig so we can use it for IOMMU hacks.
720		 */
721		if (machine_is(cell))
722			limit = 0x80000000;
723		else
724			limit = MEMBLOCK_ALLOC_ANYWHERE;
725
726		table = memblock_alloc_base(htab_size_bytes, htab_size_bytes, limit);
727
728		DBG("Hash table allocated at %lx, size: %lx\n", table,
729		    htab_size_bytes);
730
731		htab_address = __va(table);
732
733		/* htab absolute addr + encoded htabsize */
734		_SDR1 = table + __ilog2(pteg_count) - 11;
735
736		/* Initialize the HPT with no entries */
737		memset((void *)table, 0, htab_size_bytes);
738
739		/* Set SDR1 */
740		mtspr(SPRN_SDR1, _SDR1);
741	}
742
743	prot = pgprot_val(PAGE_KERNEL);
744
745#ifdef CONFIG_DEBUG_PAGEALLOC
746	linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
747	linear_map_hash_slots = __va(memblock_alloc_base(linear_map_hash_count,
748						    1, ppc64_rma_size));
749	memset(linear_map_hash_slots, 0, linear_map_hash_count);
750#endif /* CONFIG_DEBUG_PAGEALLOC */
751
752	/* On U3 based machines, we need to reserve the DART area and
753	 * _NOT_ map it to avoid cache paradoxes as it's remapped non
754	 * cacheable later on
755	 */
756
757	/* create bolted the linear mapping in the hash table */
758	for_each_memblock(memory, reg) {
759		base = (unsigned long)__va(reg->base);
760		size = reg->size;
761
762		DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
763		    base, size, prot);
764
765#ifdef CONFIG_U3_DART
766		/* Do not map the DART space. Fortunately, it will be aligned
767		 * in such a way that it will not cross two memblock regions and
768		 * will fit within a single 16Mb page.
769		 * The DART space is assumed to be a full 16Mb region even if
770		 * we only use 2Mb of that space. We will use more of it later
771		 * for AGP GART. We have to use a full 16Mb large page.
772		 */
773		DBG("DART base: %lx\n", dart_tablebase);
774
775		if (dart_tablebase != 0 && dart_tablebase >= base
776		    && dart_tablebase < (base + size)) {
777			unsigned long dart_table_end = dart_tablebase + 16 * MB;
778			if (base != dart_tablebase)
779				BUG_ON(htab_bolt_mapping(base, dart_tablebase,
780							__pa(base), prot,
781							mmu_linear_psize,
782							mmu_kernel_ssize));
783			if ((base + size) > dart_table_end)
784				BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
785							base + size,
786							__pa(dart_table_end),
787							 prot,
788							 mmu_linear_psize,
789							 mmu_kernel_ssize));
790			continue;
791		}
792#endif /* CONFIG_U3_DART */
793		BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
794				prot, mmu_linear_psize, mmu_kernel_ssize));
795	}
796	memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
797
798	/*
799	 * If we have a memory_limit and we've allocated TCEs then we need to
800	 * explicitly map the TCE area at the top of RAM. We also cope with the
801	 * case that the TCEs start below memory_limit.
802	 * tce_alloc_start/end are 16MB aligned so the mapping should work
803	 * for either 4K or 16MB pages.
804	 */
805	if (tce_alloc_start) {
806		tce_alloc_start = (unsigned long)__va(tce_alloc_start);
807		tce_alloc_end = (unsigned long)__va(tce_alloc_end);
808
809		if (base + size >= tce_alloc_start)
810			tce_alloc_start = base + size + 1;
811
812		BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
813					 __pa(tce_alloc_start), prot,
814					 mmu_linear_psize, mmu_kernel_ssize));
815	}
816
817	htab_finish_init();
818
819	DBG(" <- htab_initialize()\n");
820}
821#undef KB
822#undef MB
823
824void __init early_init_mmu(void)
825{
826	/* Initialize the MMU Hash table and create the linear mapping
827	 * of memory. Has to be done before SLB initialization as this is
828	 * currently where the page size encoding is obtained.
829	 */
830	htab_initialize();
831
832	/* Initialize SLB management */
833	slb_initialize();
834}
835
836#ifdef CONFIG_SMP
837void early_init_mmu_secondary(void)
838{
839	/* Initialize hash table for that CPU */
840	if (!firmware_has_feature(FW_FEATURE_LPAR))
841		mtspr(SPRN_SDR1, _SDR1);
842
843	/* Initialize SLB */
844	slb_initialize();
845}
846#endif /* CONFIG_SMP */
847
848/*
849 * Called by asm hashtable.S for doing lazy icache flush
850 */
851unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
852{
853	struct page *page;
854
855	if (!pfn_valid(pte_pfn(pte)))
856		return pp;
857
858	page = pte_page(pte);
859
860	/* page is dirty */
861	if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
862		if (trap == 0x400) {
863			flush_dcache_icache_page(page);
864			set_bit(PG_arch_1, &page->flags);
865		} else
866			pp |= HPTE_R_N;
867	}
868	return pp;
869}
870
871#ifdef CONFIG_PPC_MM_SLICES
872static unsigned int get_paca_psize(unsigned long addr)
873{
874	u64 lpsizes;
875	unsigned char *hpsizes;
876	unsigned long index, mask_index;
877
878	if (addr < SLICE_LOW_TOP) {
879		lpsizes = get_paca()->context.low_slices_psize;
880		index = GET_LOW_SLICE_INDEX(addr);
881		return (lpsizes >> (index * 4)) & 0xF;
882	}
883	hpsizes = get_paca()->context.high_slices_psize;
884	index = GET_HIGH_SLICE_INDEX(addr);
885	mask_index = index & 0x1;
886	return (hpsizes[index >> 1] >> (mask_index * 4)) & 0xF;
887}
888
889#else
890unsigned int get_paca_psize(unsigned long addr)
891{
892	return get_paca()->context.user_psize;
893}
894#endif
895
896/*
897 * Demote a segment to using 4k pages.
898 * For now this makes the whole process use 4k pages.
899 */
900#ifdef CONFIG_PPC_64K_PAGES
901void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
902{
903	if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
904		return;
905	slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
906	copro_flush_all_slbs(mm);
907	if ((get_paca_psize(addr) != MMU_PAGE_4K) && (current->mm == mm)) {
908		get_paca()->context = mm->context;
909		slb_flush_and_rebolt();
910	}
911}
912#endif /* CONFIG_PPC_64K_PAGES */
913
914#ifdef CONFIG_PPC_SUBPAGE_PROT
915/*
916 * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
917 * Userspace sets the subpage permissions using the subpage_prot system call.
918 *
919 * Result is 0: full permissions, _PAGE_RW: read-only,
920 * _PAGE_USER or _PAGE_USER|_PAGE_RW: no access.
921 */
922static int subpage_protection(struct mm_struct *mm, unsigned long ea)
923{
924	struct subpage_prot_table *spt = &mm->context.spt;
925	u32 spp = 0;
926	u32 **sbpm, *sbpp;
927
928	if (ea >= spt->maxaddr)
929		return 0;
930	if (ea < 0x100000000UL) {
931		/* addresses below 4GB use spt->low_prot */
932		sbpm = spt->low_prot;
933	} else {
934		sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
935		if (!sbpm)
936			return 0;
937	}
938	sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
939	if (!sbpp)
940		return 0;
941	spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
942
943	/* extract 2-bit bitfield for this 4k subpage */
944	spp >>= 30 - 2 * ((ea >> 12) & 0xf);
945
946	/* turn 0,1,2,3 into combination of _PAGE_USER and _PAGE_RW */
947	spp = ((spp & 2) ? _PAGE_USER : 0) | ((spp & 1) ? _PAGE_RW : 0);
948	return spp;
949}
950
951#else /* CONFIG_PPC_SUBPAGE_PROT */
952static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
953{
954	return 0;
955}
956#endif
957
958void hash_failure_debug(unsigned long ea, unsigned long access,
959			unsigned long vsid, unsigned long trap,
960			int ssize, int psize, int lpsize, unsigned long pte)
961{
962	if (!printk_ratelimit())
963		return;
964	pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
965		ea, access, current->comm);
966	pr_info("    trap=0x%lx vsid=0x%lx ssize=%d base psize=%d psize %d pte=0x%lx\n",
967		trap, vsid, ssize, psize, lpsize, pte);
968}
969
970static void check_paca_psize(unsigned long ea, struct mm_struct *mm,
971			     int psize, bool user_region)
972{
973	if (user_region) {
974		if (psize != get_paca_psize(ea)) {
975			get_paca()->context = mm->context;
976			slb_flush_and_rebolt();
977		}
978	} else if (get_paca()->vmalloc_sllp !=
979		   mmu_psize_defs[mmu_vmalloc_psize].sllp) {
980		get_paca()->vmalloc_sllp =
981			mmu_psize_defs[mmu_vmalloc_psize].sllp;
982		slb_vmalloc_update();
983	}
984}
985
986/* Result code is:
987 *  0 - handled
988 *  1 - normal page fault
989 * -1 - critical hash insertion error
990 * -2 - access not permitted by subpage protection mechanism
991 */
992int hash_page_mm(struct mm_struct *mm, unsigned long ea,
993		 unsigned long access, unsigned long trap,
994		 unsigned long flags)
995{
996	enum ctx_state prev_state = exception_enter();
997	pgd_t *pgdir;
998	unsigned long vsid;
999	pte_t *ptep;
1000	unsigned hugeshift;
1001	const struct cpumask *tmp;
1002	int rc, user_region = 0;
1003	int psize, ssize;
1004
1005	DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
1006		ea, access, trap);
1007
1008	/* Get region & vsid */
1009 	switch (REGION_ID(ea)) {
1010	case USER_REGION_ID:
1011		user_region = 1;
1012		if (! mm) {
1013			DBG_LOW(" user region with no mm !\n");
1014			rc = 1;
1015			goto bail;
1016		}
1017		psize = get_slice_psize(mm, ea);
1018		ssize = user_segment_size(ea);
1019		vsid = get_vsid(mm->context.id, ea, ssize);
1020		break;
1021	case VMALLOC_REGION_ID:
1022		vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
1023		if (ea < VMALLOC_END)
1024			psize = mmu_vmalloc_psize;
1025		else
1026			psize = mmu_io_psize;
1027		ssize = mmu_kernel_ssize;
1028		break;
1029	default:
1030		/* Not a valid range
1031		 * Send the problem up to do_page_fault
1032		 */
1033		rc = 1;
1034		goto bail;
1035	}
1036	DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
1037
1038	/* Bad address. */
1039	if (!vsid) {
1040		DBG_LOW("Bad address!\n");
1041		rc = 1;
1042		goto bail;
1043	}
1044	/* Get pgdir */
1045	pgdir = mm->pgd;
1046	if (pgdir == NULL) {
1047		rc = 1;
1048		goto bail;
1049	}
1050
1051	/* Check CPU locality */
1052	tmp = cpumask_of(smp_processor_id());
1053	if (user_region && cpumask_equal(mm_cpumask(mm), tmp))
1054		flags |= HPTE_LOCAL_UPDATE;
1055
1056#ifndef CONFIG_PPC_64K_PAGES
1057	/* If we use 4K pages and our psize is not 4K, then we might
1058	 * be hitting a special driver mapping, and need to align the
1059	 * address before we fetch the PTE.
1060	 *
1061	 * It could also be a hugepage mapping, in which case this is
1062	 * not necessary, but it's not harmful, either.
1063	 */
1064	if (psize != MMU_PAGE_4K)
1065		ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
1066#endif /* CONFIG_PPC_64K_PAGES */
1067
1068	/* Get PTE and page size from page tables */
1069	ptep = __find_linux_pte_or_hugepte(pgdir, ea, &hugeshift);
1070	if (ptep == NULL || !pte_present(*ptep)) {
1071		DBG_LOW(" no PTE !\n");
1072		rc = 1;
1073		goto bail;
1074	}
1075
1076	/* Add _PAGE_PRESENT to the required access perm */
1077	access |= _PAGE_PRESENT;
1078
1079	/* Pre-check access permissions (will be re-checked atomically
1080	 * in __hash_page_XX but this pre-check is a fast path
1081	 */
1082	if (access & ~pte_val(*ptep)) {
1083		DBG_LOW(" no access !\n");
1084		rc = 1;
1085		goto bail;
1086	}
1087
1088	if (hugeshift) {
1089		if (pmd_trans_huge(*(pmd_t *)ptep))
1090			rc = __hash_page_thp(ea, access, vsid, (pmd_t *)ptep,
1091					     trap, flags, ssize, psize);
1092#ifdef CONFIG_HUGETLB_PAGE
1093		else
1094			rc = __hash_page_huge(ea, access, vsid, ptep, trap,
1095					      flags, ssize, hugeshift, psize);
1096#else
1097		else {
1098			/*
1099			 * if we have hugeshift, and is not transhuge with
1100			 * hugetlb disabled, something is really wrong.
1101			 */
1102			rc = 1;
1103			WARN_ON(1);
1104		}
1105#endif
1106		if (current->mm == mm)
1107			check_paca_psize(ea, mm, psize, user_region);
1108
1109		goto bail;
1110	}
1111
1112#ifndef CONFIG_PPC_64K_PAGES
1113	DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1114#else
1115	DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1116		pte_val(*(ptep + PTRS_PER_PTE)));
1117#endif
1118	/* Do actual hashing */
1119#ifdef CONFIG_PPC_64K_PAGES
1120	/* If _PAGE_4K_PFN is set, make sure this is a 4k segment */
1121	if ((pte_val(*ptep) & _PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1122		demote_segment_4k(mm, ea);
1123		psize = MMU_PAGE_4K;
1124	}
1125
1126	/* If this PTE is non-cacheable and we have restrictions on
1127	 * using non cacheable large pages, then we switch to 4k
1128	 */
1129	if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
1130	    (pte_val(*ptep) & _PAGE_NO_CACHE)) {
1131		if (user_region) {
1132			demote_segment_4k(mm, ea);
1133			psize = MMU_PAGE_4K;
1134		} else if (ea < VMALLOC_END) {
1135			/*
1136			 * some driver did a non-cacheable mapping
1137			 * in vmalloc space, so switch vmalloc
1138			 * to 4k pages
1139			 */
1140			printk(KERN_ALERT "Reducing vmalloc segment "
1141			       "to 4kB pages because of "
1142			       "non-cacheable mapping\n");
1143			psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1144			copro_flush_all_slbs(mm);
1145		}
1146	}
1147
1148	if (current->mm == mm)
1149		check_paca_psize(ea, mm, psize, user_region);
1150#endif /* CONFIG_PPC_64K_PAGES */
1151
1152#ifdef CONFIG_PPC_HAS_HASH_64K
1153	if (psize == MMU_PAGE_64K)
1154		rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1155				     flags, ssize);
1156	else
1157#endif /* CONFIG_PPC_HAS_HASH_64K */
1158	{
1159		int spp = subpage_protection(mm, ea);
1160		if (access & spp)
1161			rc = -2;
1162		else
1163			rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1164					    flags, ssize, spp);
1165	}
1166
1167	/* Dump some info in case of hash insertion failure, they should
1168	 * never happen so it is really useful to know if/when they do
1169	 */
1170	if (rc == -1)
1171		hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1172				   psize, pte_val(*ptep));
1173#ifndef CONFIG_PPC_64K_PAGES
1174	DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1175#else
1176	DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1177		pte_val(*(ptep + PTRS_PER_PTE)));
1178#endif
1179	DBG_LOW(" -> rc=%d\n", rc);
1180
1181bail:
1182	exception_exit(prev_state);
1183	return rc;
1184}
1185EXPORT_SYMBOL_GPL(hash_page_mm);
1186
1187int hash_page(unsigned long ea, unsigned long access, unsigned long trap,
1188	      unsigned long dsisr)
1189{
1190	unsigned long flags = 0;
1191	struct mm_struct *mm = current->mm;
1192
1193	if (REGION_ID(ea) == VMALLOC_REGION_ID)
1194		mm = &init_mm;
1195
1196	if (dsisr & DSISR_NOHPTE)
1197		flags |= HPTE_NOHPTE_UPDATE;
1198
1199	return hash_page_mm(mm, ea, access, trap, flags);
1200}
1201EXPORT_SYMBOL_GPL(hash_page);
1202
1203void hash_preload(struct mm_struct *mm, unsigned long ea,
1204		  unsigned long access, unsigned long trap)
1205{
1206	int hugepage_shift;
1207	unsigned long vsid;
1208	pgd_t *pgdir;
1209	pte_t *ptep;
1210	unsigned long flags;
1211	int rc, ssize, update_flags = 0;
1212
1213	BUG_ON(REGION_ID(ea) != USER_REGION_ID);
1214
1215#ifdef CONFIG_PPC_MM_SLICES
1216	/* We only prefault standard pages for now */
1217	if (unlikely(get_slice_psize(mm, ea) != mm->context.user_psize))
1218		return;
1219#endif
1220
1221	DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1222		" trap=%lx\n", mm, mm->pgd, ea, access, trap);
1223
1224	/* Get Linux PTE if available */
1225	pgdir = mm->pgd;
1226	if (pgdir == NULL)
1227		return;
1228
1229	/* Get VSID */
1230	ssize = user_segment_size(ea);
1231	vsid = get_vsid(mm->context.id, ea, ssize);
1232	if (!vsid)
1233		return;
1234	/*
1235	 * Hash doesn't like irqs. Walking linux page table with irq disabled
1236	 * saves us from holding multiple locks.
1237	 */
1238	local_irq_save(flags);
1239
1240	/*
1241	 * THP pages use update_mmu_cache_pmd. We don't do
1242	 * hash preload there. Hence can ignore THP here
1243	 */
1244	ptep = find_linux_pte_or_hugepte(pgdir, ea, &hugepage_shift);
1245	if (!ptep)
1246		goto out_exit;
1247
1248	WARN_ON(hugepage_shift);
1249#ifdef CONFIG_PPC_64K_PAGES
1250	/* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
1251	 * a 64K kernel), then we don't preload, hash_page() will take
1252	 * care of it once we actually try to access the page.
1253	 * That way we don't have to duplicate all of the logic for segment
1254	 * page size demotion here
1255	 */
1256	if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
1257		goto out_exit;
1258#endif /* CONFIG_PPC_64K_PAGES */
1259
1260	/* Is that local to this CPU ? */
1261	if (cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
1262		update_flags |= HPTE_LOCAL_UPDATE;
1263
1264	/* Hash it in */
1265#ifdef CONFIG_PPC_HAS_HASH_64K
1266	if (mm->context.user_psize == MMU_PAGE_64K)
1267		rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1268				     update_flags, ssize);
1269	else
1270#endif /* CONFIG_PPC_HAS_HASH_64K */
1271		rc = __hash_page_4K(ea, access, vsid, ptep, trap, update_flags,
1272				    ssize, subpage_protection(mm, ea));
1273
1274	/* Dump some info in case of hash insertion failure, they should
1275	 * never happen so it is really useful to know if/when they do
1276	 */
1277	if (rc == -1)
1278		hash_failure_debug(ea, access, vsid, trap, ssize,
1279				   mm->context.user_psize,
1280				   mm->context.user_psize,
1281				   pte_val(*ptep));
1282out_exit:
1283	local_irq_restore(flags);
1284}
1285
1286/* WARNING: This is called from hash_low_64.S, if you change this prototype,
1287 *          do not forget to update the assembly call site !
1288 */
1289void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1290		     unsigned long flags)
1291{
1292	unsigned long hash, index, shift, hidx, slot;
1293	int local = flags & HPTE_LOCAL_UPDATE;
1294
1295	DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1296	pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1297		hash = hpt_hash(vpn, shift, ssize);
1298		hidx = __rpte_to_hidx(pte, index);
1299		if (hidx & _PTEIDX_SECONDARY)
1300			hash = ~hash;
1301		slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1302		slot += hidx & _PTEIDX_GROUP_IX;
1303		DBG_LOW(" sub %ld: hash=%lx, hidx=%lx\n", index, slot, hidx);
1304		/*
1305		 * We use same base page size and actual psize, because we don't
1306		 * use these functions for hugepage
1307		 */
1308		ppc_md.hpte_invalidate(slot, vpn, psize, psize, ssize, local);
1309	} pte_iterate_hashed_end();
1310
1311#ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1312	/* Transactions are not aborted by tlbiel, only tlbie.
1313	 * Without, syncing a page back to a block device w/ PIO could pick up
1314	 * transactional data (bad!) so we force an abort here.  Before the
1315	 * sync the page will be made read-only, which will flush_hash_page.
1316	 * BIG ISSUE here: if the kernel uses a page from userspace without
1317	 * unmapping it first, it may see the speculated version.
1318	 */
1319	if (local && cpu_has_feature(CPU_FTR_TM) &&
1320	    current->thread.regs &&
1321	    MSR_TM_ACTIVE(current->thread.regs->msr)) {
1322		tm_enable();
1323		tm_abort(TM_CAUSE_TLBI);
1324	}
1325#endif
1326}
1327
1328#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1329void flush_hash_hugepage(unsigned long vsid, unsigned long addr,
1330			 pmd_t *pmdp, unsigned int psize, int ssize,
1331			 unsigned long flags)
1332{
1333	int i, max_hpte_count, valid;
1334	unsigned long s_addr;
1335	unsigned char *hpte_slot_array;
1336	unsigned long hidx, shift, vpn, hash, slot;
1337	int local = flags & HPTE_LOCAL_UPDATE;
1338
1339	s_addr = addr & HPAGE_PMD_MASK;
1340	hpte_slot_array = get_hpte_slot_array(pmdp);
1341	/*
1342	 * IF we try to do a HUGE PTE update after a withdraw is done.
1343	 * we will find the below NULL. This happens when we do
1344	 * split_huge_page_pmd
1345	 */
1346	if (!hpte_slot_array)
1347		return;
1348
1349	if (ppc_md.hugepage_invalidate) {
1350		ppc_md.hugepage_invalidate(vsid, s_addr, hpte_slot_array,
1351					   psize, ssize, local);
1352		goto tm_abort;
1353	}
1354	/*
1355	 * No bluk hpte removal support, invalidate each entry
1356	 */
1357	shift = mmu_psize_defs[psize].shift;
1358	max_hpte_count = HPAGE_PMD_SIZE >> shift;
1359	for (i = 0; i < max_hpte_count; i++) {
1360		/*
1361		 * 8 bits per each hpte entries
1362		 * 000| [ secondary group (one bit) | hidx (3 bits) | valid bit]
1363		 */
1364		valid = hpte_valid(hpte_slot_array, i);
1365		if (!valid)
1366			continue;
1367		hidx =  hpte_hash_index(hpte_slot_array, i);
1368
1369		/* get the vpn */
1370		addr = s_addr + (i * (1ul << shift));
1371		vpn = hpt_vpn(addr, vsid, ssize);
1372		hash = hpt_hash(vpn, shift, ssize);
1373		if (hidx & _PTEIDX_SECONDARY)
1374			hash = ~hash;
1375
1376		slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1377		slot += hidx & _PTEIDX_GROUP_IX;
1378		ppc_md.hpte_invalidate(slot, vpn, psize,
1379				       MMU_PAGE_16M, ssize, local);
1380	}
1381tm_abort:
1382#ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1383	/* Transactions are not aborted by tlbiel, only tlbie.
1384	 * Without, syncing a page back to a block device w/ PIO could pick up
1385	 * transactional data (bad!) so we force an abort here.  Before the
1386	 * sync the page will be made read-only, which will flush_hash_page.
1387	 * BIG ISSUE here: if the kernel uses a page from userspace without
1388	 * unmapping it first, it may see the speculated version.
1389	 */
1390	if (local && cpu_has_feature(CPU_FTR_TM) &&
1391	    current->thread.regs &&
1392	    MSR_TM_ACTIVE(current->thread.regs->msr)) {
1393		tm_enable();
1394		tm_abort(TM_CAUSE_TLBI);
1395	}
1396#endif
1397	return;
1398}
1399#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1400
1401void flush_hash_range(unsigned long number, int local)
1402{
1403	if (ppc_md.flush_hash_range)
1404		ppc_md.flush_hash_range(number, local);
1405	else {
1406		int i;
1407		struct ppc64_tlb_batch *batch =
1408			this_cpu_ptr(&ppc64_tlb_batch);
1409
1410		for (i = 0; i < number; i++)
1411			flush_hash_page(batch->vpn[i], batch->pte[i],
1412					batch->psize, batch->ssize, local);
1413	}
1414}
1415
1416/*
1417 * low_hash_fault is called when we the low level hash code failed
1418 * to instert a PTE due to an hypervisor error
1419 */
1420void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1421{
1422	enum ctx_state prev_state = exception_enter();
1423
1424	if (user_mode(regs)) {
1425#ifdef CONFIG_PPC_SUBPAGE_PROT
1426		if (rc == -2)
1427			_exception(SIGSEGV, regs, SEGV_ACCERR, address);
1428		else
1429#endif
1430			_exception(SIGBUS, regs, BUS_ADRERR, address);
1431	} else
1432		bad_page_fault(regs, address, SIGBUS);
1433
1434	exception_exit(prev_state);
1435}
1436
1437long hpte_insert_repeating(unsigned long hash, unsigned long vpn,
1438			   unsigned long pa, unsigned long rflags,
1439			   unsigned long vflags, int psize, int ssize)
1440{
1441	unsigned long hpte_group;
1442	long slot;
1443
1444repeat:
1445	hpte_group = ((hash & htab_hash_mask) *
1446		       HPTES_PER_GROUP) & ~0x7UL;
1447
1448	/* Insert into the hash table, primary slot */
1449	slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, vflags,
1450				  psize, psize, ssize);
1451
1452	/* Primary is full, try the secondary */
1453	if (unlikely(slot == -1)) {
1454		hpte_group = ((~hash & htab_hash_mask) *
1455			      HPTES_PER_GROUP) & ~0x7UL;
1456		slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags,
1457					  vflags | HPTE_V_SECONDARY,
1458					  psize, psize, ssize);
1459		if (slot == -1) {
1460			if (mftb() & 0x1)
1461				hpte_group = ((hash & htab_hash_mask) *
1462					      HPTES_PER_GROUP)&~0x7UL;
1463
1464			ppc_md.hpte_remove(hpte_group);
1465			goto repeat;
1466		}
1467	}
1468
1469	return slot;
1470}
1471
1472#ifdef CONFIG_DEBUG_PAGEALLOC
1473static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1474{
1475	unsigned long hash;
1476	unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1477	unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1478	unsigned long mode = htab_convert_pte_flags(PAGE_KERNEL);
1479	long ret;
1480
1481	hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1482
1483	/* Don't create HPTE entries for bad address */
1484	if (!vsid)
1485		return;
1486
1487	ret = hpte_insert_repeating(hash, vpn, __pa(vaddr), mode,
1488				    HPTE_V_BOLTED,
1489				    mmu_linear_psize, mmu_kernel_ssize);
1490
1491	BUG_ON (ret < 0);
1492	spin_lock(&linear_map_hash_lock);
1493	BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1494	linear_map_hash_slots[lmi] = ret | 0x80;
1495	spin_unlock(&linear_map_hash_lock);
1496}
1497
1498static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1499{
1500	unsigned long hash, hidx, slot;
1501	unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1502	unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1503
1504	hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1505	spin_lock(&linear_map_hash_lock);
1506	BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1507	hidx = linear_map_hash_slots[lmi] & 0x7f;
1508	linear_map_hash_slots[lmi] = 0;
1509	spin_unlock(&linear_map_hash_lock);
1510	if (hidx & _PTEIDX_SECONDARY)
1511		hash = ~hash;
1512	slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1513	slot += hidx & _PTEIDX_GROUP_IX;
1514	ppc_md.hpte_invalidate(slot, vpn, mmu_linear_psize, mmu_linear_psize,
1515			       mmu_kernel_ssize, 0);
1516}
1517
1518void __kernel_map_pages(struct page *page, int numpages, int enable)
1519{
1520	unsigned long flags, vaddr, lmi;
1521	int i;
1522
1523	local_irq_save(flags);
1524	for (i = 0; i < numpages; i++, page++) {
1525		vaddr = (unsigned long)page_address(page);
1526		lmi = __pa(vaddr) >> PAGE_SHIFT;
1527		if (lmi >= linear_map_hash_count)
1528			continue;
1529		if (enable)
1530			kernel_map_linear_page(vaddr, lmi);
1531		else
1532			kernel_unmap_linear_page(vaddr, lmi);
1533	}
1534	local_irq_restore(flags);
1535}
1536#endif /* CONFIG_DEBUG_PAGEALLOC */
1537
1538void setup_initial_memory_limit(phys_addr_t first_memblock_base,
1539				phys_addr_t first_memblock_size)
1540{
1541	/* We don't currently support the first MEMBLOCK not mapping 0
1542	 * physical on those processors
1543	 */
1544	BUG_ON(first_memblock_base != 0);
1545
1546	/* On LPAR systems, the first entry is our RMA region,
1547	 * non-LPAR 64-bit hash MMU systems don't have a limitation
1548	 * on real mode access, but using the first entry works well
1549	 * enough. We also clamp it to 1G to avoid some funky things
1550	 * such as RTAS bugs etc...
1551	 */
1552	ppc64_rma_size = min_t(u64, first_memblock_size, 0x40000000);
1553
1554	/* Finally limit subsequent allocations */
1555	memblock_set_current_limit(ppc64_rma_size);
1556}
1557