1 #ifndef _ASM_POWERPC_BOOK3S_64_HASH_64K_H
2 #define _ASM_POWERPC_BOOK3S_64_HASH_64K_H
3 
4 #define H_PTE_INDEX_SIZE  8
5 #define H_PMD_INDEX_SIZE  5
6 #define H_PUD_INDEX_SIZE  5
7 #define H_PGD_INDEX_SIZE  12
8 
9 /* With 4k base page size, hugepage PTEs go at the PMD level */
10 #define MIN_HUGEPTE_SHIFT	PAGE_SHIFT
11 
12 #define H_PAGE_COMBO	0x00001000 /* this is a combo 4k page */
13 #define H_PAGE_4K_PFN	0x00002000 /* PFN is for a single 4k page */
14 /*
15  * We need to differentiate between explicit huge page and THP huge
16  * page, since THP huge page also need to track real subpage details
17  */
18 #define H_PAGE_THP_HUGE  H_PAGE_4K_PFN
19 
20 /*
21  * Used to track subpage group valid if H_PAGE_COMBO is set
22  * This overloads H_PAGE_F_GIX and H_PAGE_F_SECOND
23  */
24 #define H_PAGE_COMBO_VALID	(H_PAGE_F_GIX | H_PAGE_F_SECOND)
25 
26 /* PTE flags to conserve for HPTE identification */
27 #define _PAGE_HPTEFLAGS (H_PAGE_BUSY | H_PAGE_F_SECOND | \
28 			 H_PAGE_F_GIX | H_PAGE_HASHPTE | H_PAGE_COMBO)
29 /*
30  * we support 16 fragments per PTE page of 64K size.
31  */
32 #define H_PTE_FRAG_NR	16
33 /*
34  * We use a 2K PTE page fragment and another 2K for storing
35  * real_pte_t hash index
36  */
37 #define H_PTE_FRAG_SIZE_SHIFT  12
38 #define PTE_FRAG_SIZE (1UL << PTE_FRAG_SIZE_SHIFT)
39 
40 #ifndef __ASSEMBLY__
41 #include <asm/errno.h>
42 
43 /*
44  * With 64K pages on hash table, we have a special PTE format that
45  * uses a second "half" of the page table to encode sub-page information
46  * in order to deal with 64K made of 4K HW pages. Thus we override the
47  * generic accessors and iterators here
48  */
49 #define __real_pte __real_pte
50 static inline real_pte_t __real_pte(pte_t pte, pte_t *ptep)
51 {
52 	real_pte_t rpte;
53 	unsigned long *hidxp;
54 
55 	rpte.pte = pte;
56 	rpte.hidx = 0;
57 	if (pte_val(pte) & H_PAGE_COMBO) {
58 		/*
59 		 * Make sure we order the hidx load against the H_PAGE_COMBO
60 		 * check. The store side ordering is done in __hash_page_4K
61 		 */
62 		smp_rmb();
63 		hidxp = (unsigned long *)(ptep + PTRS_PER_PTE);
64 		rpte.hidx = *hidxp;
65 	}
66 	return rpte;
67 }
68 
69 static inline unsigned long __rpte_to_hidx(real_pte_t rpte, unsigned long index)
70 {
71 	if ((pte_val(rpte.pte) & H_PAGE_COMBO))
72 		return (rpte.hidx >> (index<<2)) & 0xf;
73 	return (pte_val(rpte.pte) >> H_PAGE_F_GIX_SHIFT) & 0xf;
74 }
75 
76 #define __rpte_to_pte(r)	((r).pte)
77 extern bool __rpte_sub_valid(real_pte_t rpte, unsigned long index);
78 /*
79  * Trick: we set __end to va + 64k, which happens works for
80  * a 16M page as well as we want only one iteration
81  */
82 #define pte_iterate_hashed_subpages(rpte, psize, vpn, index, shift)	\
83 	do {								\
84 		unsigned long __end = vpn + (1UL << (PAGE_SHIFT - VPN_SHIFT));	\
85 		unsigned __split = (psize == MMU_PAGE_4K ||		\
86 				    psize == MMU_PAGE_64K_AP);		\
87 		shift = mmu_psize_defs[psize].shift;			\
88 		for (index = 0; vpn < __end; index++,			\
89 			     vpn += (1L << (shift - VPN_SHIFT))) {	\
90 			if (!__split || __rpte_sub_valid(rpte, index))	\
91 				do {
92 
93 #define pte_iterate_hashed_end() } while(0); } } while(0)
94 
95 #define pte_pagesize_index(mm, addr, pte)	\
96 	(((pte) & H_PAGE_COMBO)? MMU_PAGE_4K: MMU_PAGE_64K)
97 
98 extern int remap_pfn_range(struct vm_area_struct *, unsigned long addr,
99 			   unsigned long pfn, unsigned long size, pgprot_t);
100 static inline int hash__remap_4k_pfn(struct vm_area_struct *vma, unsigned long addr,
101 				 unsigned long pfn, pgprot_t prot)
102 {
103 	if (pfn > (PTE_RPN_MASK >> PAGE_SHIFT)) {
104 		WARN(1, "remap_4k_pfn called with wrong pfn value\n");
105 		return -EINVAL;
106 	}
107 	return remap_pfn_range(vma, addr, pfn, PAGE_SIZE,
108 			       __pgprot(pgprot_val(prot) | H_PAGE_4K_PFN));
109 }
110 
111 #define H_PTE_TABLE_SIZE	PTE_FRAG_SIZE
112 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
113 #define H_PMD_TABLE_SIZE	((sizeof(pmd_t) << PMD_INDEX_SIZE) + \
114 				 (sizeof(unsigned long) << PMD_INDEX_SIZE))
115 #else
116 #define H_PMD_TABLE_SIZE	(sizeof(pmd_t) << PMD_INDEX_SIZE)
117 #endif
118 #define H_PUD_TABLE_SIZE	(sizeof(pud_t) << PUD_INDEX_SIZE)
119 #define H_PGD_TABLE_SIZE	(sizeof(pgd_t) << PGD_INDEX_SIZE)
120 
121 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
122 static inline char *get_hpte_slot_array(pmd_t *pmdp)
123 {
124 	/*
125 	 * The hpte hindex is stored in the pgtable whose address is in the
126 	 * second half of the PMD
127 	 *
128 	 * Order this load with the test for pmd_trans_huge in the caller
129 	 */
130 	smp_rmb();
131 	return *(char **)(pmdp + PTRS_PER_PMD);
132 
133 
134 }
135 /*
136  * The linux hugepage PMD now include the pmd entries followed by the address
137  * to the stashed pgtable_t. The stashed pgtable_t contains the hpte bits.
138  * [ 000 | 1 bit secondary | 3 bit hidx | 1 bit valid]. We use one byte per
139  * each HPTE entry. With 16MB hugepage and 64K HPTE we need 256 entries and
140  * with 4K HPTE we need 4096 entries. Both will fit in a 4K pgtable_t.
141  *
142  * The top three bits are intentionally left as zero. This memory location
143  * are also used as normal page PTE pointers. So if we have any pointers
144  * left around while we collapse a hugepage, we need to make sure
145  * _PAGE_PRESENT bit of that is zero when we look at them
146  */
147 static inline unsigned int hpte_valid(unsigned char *hpte_slot_array, int index)
148 {
149 	return hpte_slot_array[index] & 0x1;
150 }
151 
152 static inline unsigned int hpte_hash_index(unsigned char *hpte_slot_array,
153 					   int index)
154 {
155 	return hpte_slot_array[index] >> 1;
156 }
157 
158 static inline void mark_hpte_slot_valid(unsigned char *hpte_slot_array,
159 					unsigned int index, unsigned int hidx)
160 {
161 	hpte_slot_array[index] = (hidx << 1) | 0x1;
162 }
163 
164 /*
165  *
166  * For core kernel code by design pmd_trans_huge is never run on any hugetlbfs
167  * page. The hugetlbfs page table walking and mangling paths are totally
168  * separated form the core VM paths and they're differentiated by
169  *  VM_HUGETLB being set on vm_flags well before any pmd_trans_huge could run.
170  *
171  * pmd_trans_huge() is defined as false at build time if
172  * CONFIG_TRANSPARENT_HUGEPAGE=n to optimize away code blocks at build
173  * time in such case.
174  *
175  * For ppc64 we need to differntiate from explicit hugepages from THP, because
176  * for THP we also track the subpage details at the pmd level. We don't do
177  * that for explicit huge pages.
178  *
179  */
180 static inline int hash__pmd_trans_huge(pmd_t pmd)
181 {
182 	return !!((pmd_val(pmd) & (_PAGE_PTE | H_PAGE_THP_HUGE)) ==
183 		  (_PAGE_PTE | H_PAGE_THP_HUGE));
184 }
185 
186 static inline int hash__pmd_same(pmd_t pmd_a, pmd_t pmd_b)
187 {
188 	return (((pmd_raw(pmd_a) ^ pmd_raw(pmd_b)) & ~cpu_to_be64(_PAGE_HPTEFLAGS)) == 0);
189 }
190 
191 static inline pmd_t hash__pmd_mkhuge(pmd_t pmd)
192 {
193 	return __pmd(pmd_val(pmd) | (_PAGE_PTE | H_PAGE_THP_HUGE));
194 }
195 
196 extern unsigned long hash__pmd_hugepage_update(struct mm_struct *mm,
197 					   unsigned long addr, pmd_t *pmdp,
198 					   unsigned long clr, unsigned long set);
199 extern pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma,
200 				   unsigned long address, pmd_t *pmdp);
201 extern void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
202 					 pgtable_t pgtable);
203 extern pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp);
204 extern void hash__pmdp_huge_split_prepare(struct vm_area_struct *vma,
205 				      unsigned long address, pmd_t *pmdp);
206 extern pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm,
207 				       unsigned long addr, pmd_t *pmdp);
208 extern int hash__has_transparent_hugepage(void);
209 #endif /*  CONFIG_TRANSPARENT_HUGEPAGE */
210 #endif	/* __ASSEMBLY__ */
211 
212 #endif /* _ASM_POWERPC_BOOK3S_64_HASH_64K_H */
213