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  10
6 #define H_PUD_INDEX_SIZE  7
7 #define H_PGD_INDEX_SIZE  8
8 
9 /*
10  * 64k aligned address free up few of the lower bits of RPN for us
11  * We steal that here. For more deatils look at pte_pfn/pfn_pte()
12  */
13 #define H_PAGE_COMBO	_RPAGE_RPN0 /* this is a combo 4k page */
14 #define H_PAGE_4K_PFN	_RPAGE_RPN1 /* PFN is for a single 4k page */
15 /*
16  * We need to differentiate between explicit huge page and THP huge
17  * page, since THP huge page also need to track real subpage details
18  */
19 #define H_PAGE_THP_HUGE  H_PAGE_4K_PFN
20 
21 /*
22  * Used to track subpage group valid if H_PAGE_COMBO is set
23  * This overloads H_PAGE_F_GIX and H_PAGE_F_SECOND
24  */
25 #define H_PAGE_COMBO_VALID	(H_PAGE_F_GIX | H_PAGE_F_SECOND)
26 
27 /* PTE flags to conserve for HPTE identification */
28 #define _PAGE_HPTEFLAGS (H_PAGE_BUSY | H_PAGE_F_SECOND | \
29 			 H_PAGE_F_GIX | H_PAGE_HASHPTE | H_PAGE_COMBO)
30 /*
31  * we support 16 fragments per PTE page of 64K size.
32  */
33 #define H_PTE_FRAG_NR	16
34 /*
35  * We use a 2K PTE page fragment and another 2K for storing
36  * real_pte_t hash index
37  */
38 #define H_PTE_FRAG_SIZE_SHIFT  12
39 #define PTE_FRAG_SIZE (1UL << PTE_FRAG_SIZE_SHIFT)
40 
41 #ifndef __ASSEMBLY__
42 #include <asm/errno.h>
43 
44 /*
45  * With 64K pages on hash table, we have a special PTE format that
46  * uses a second "half" of the page table to encode sub-page information
47  * in order to deal with 64K made of 4K HW pages. Thus we override the
48  * generic accessors and iterators here
49  */
50 #define __real_pte __real_pte
51 static inline real_pte_t __real_pte(pte_t pte, pte_t *ptep)
52 {
53 	real_pte_t rpte;
54 	unsigned long *hidxp;
55 
56 	rpte.pte = pte;
57 	rpte.hidx = 0;
58 	if (pte_val(pte) & H_PAGE_COMBO) {
59 		/*
60 		 * Make sure we order the hidx load against the H_PAGE_COMBO
61 		 * check. The store side ordering is done in __hash_page_4K
62 		 */
63 		smp_rmb();
64 		hidxp = (unsigned long *)(ptep + PTRS_PER_PTE);
65 		rpte.hidx = *hidxp;
66 	}
67 	return rpte;
68 }
69 
70 static inline unsigned long __rpte_to_hidx(real_pte_t rpte, unsigned long index)
71 {
72 	if ((pte_val(rpte.pte) & H_PAGE_COMBO))
73 		return (rpte.hidx >> (index<<2)) & 0xf;
74 	return (pte_val(rpte.pte) >> H_PAGE_F_GIX_SHIFT) & 0xf;
75 }
76 
77 #define __rpte_to_pte(r)	((r).pte)
78 extern bool __rpte_sub_valid(real_pte_t rpte, unsigned long index);
79 /*
80  * Trick: we set __end to va + 64k, which happens works for
81  * a 16M page as well as we want only one iteration
82  */
83 #define pte_iterate_hashed_subpages(rpte, psize, vpn, index, shift)	\
84 	do {								\
85 		unsigned long __end = vpn + (1UL << (PAGE_SHIFT - VPN_SHIFT));	\
86 		unsigned __split = (psize == MMU_PAGE_4K ||		\
87 				    psize == MMU_PAGE_64K_AP);		\
88 		shift = mmu_psize_defs[psize].shift;			\
89 		for (index = 0; vpn < __end; index++,			\
90 			     vpn += (1L << (shift - VPN_SHIFT))) {	\
91 			if (!__split || __rpte_sub_valid(rpte, index))	\
92 				do {
93 
94 #define pte_iterate_hashed_end() } while(0); } } while(0)
95 
96 #define pte_pagesize_index(mm, addr, pte)	\
97 	(((pte) & H_PAGE_COMBO)? MMU_PAGE_4K: MMU_PAGE_64K)
98 
99 extern int remap_pfn_range(struct vm_area_struct *, unsigned long addr,
100 			   unsigned long pfn, unsigned long size, pgprot_t);
101 static inline int hash__remap_4k_pfn(struct vm_area_struct *vma, unsigned long addr,
102 				 unsigned long pfn, pgprot_t prot)
103 {
104 	if (pfn > (PTE_RPN_MASK >> PAGE_SHIFT)) {
105 		WARN(1, "remap_4k_pfn called with wrong pfn value\n");
106 		return -EINVAL;
107 	}
108 	return remap_pfn_range(vma, addr, pfn, PAGE_SIZE,
109 			       __pgprot(pgprot_val(prot) | H_PAGE_4K_PFN));
110 }
111 
112 #define H_PTE_TABLE_SIZE	PTE_FRAG_SIZE
113 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
114 #define H_PMD_TABLE_SIZE	((sizeof(pmd_t) << PMD_INDEX_SIZE) + \
115 				 (sizeof(unsigned long) << PMD_INDEX_SIZE))
116 #else
117 #define H_PMD_TABLE_SIZE	(sizeof(pmd_t) << PMD_INDEX_SIZE)
118 #endif
119 #define H_PUD_TABLE_SIZE	(sizeof(pud_t) << PUD_INDEX_SIZE)
120 #define H_PGD_TABLE_SIZE	(sizeof(pgd_t) << PGD_INDEX_SIZE)
121 
122 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
123 static inline char *get_hpte_slot_array(pmd_t *pmdp)
124 {
125 	/*
126 	 * The hpte hindex is stored in the pgtable whose address is in the
127 	 * second half of the PMD
128 	 *
129 	 * Order this load with the test for pmd_trans_huge in the caller
130 	 */
131 	smp_rmb();
132 	return *(char **)(pmdp + PTRS_PER_PMD);
133 
134 
135 }
136 /*
137  * The linux hugepage PMD now include the pmd entries followed by the address
138  * to the stashed pgtable_t. The stashed pgtable_t contains the hpte bits.
139  * [ 000 | 1 bit secondary | 3 bit hidx | 1 bit valid]. We use one byte per
140  * each HPTE entry. With 16MB hugepage and 64K HPTE we need 256 entries and
141  * with 4K HPTE we need 4096 entries. Both will fit in a 4K pgtable_t.
142  *
143  * The top three bits are intentionally left as zero. This memory location
144  * are also used as normal page PTE pointers. So if we have any pointers
145  * left around while we collapse a hugepage, we need to make sure
146  * _PAGE_PRESENT bit of that is zero when we look at them
147  */
148 static inline unsigned int hpte_valid(unsigned char *hpte_slot_array, int index)
149 {
150 	return hpte_slot_array[index] & 0x1;
151 }
152 
153 static inline unsigned int hpte_hash_index(unsigned char *hpte_slot_array,
154 					   int index)
155 {
156 	return hpte_slot_array[index] >> 1;
157 }
158 
159 static inline void mark_hpte_slot_valid(unsigned char *hpte_slot_array,
160 					unsigned int index, unsigned int hidx)
161 {
162 	hpte_slot_array[index] = (hidx << 1) | 0x1;
163 }
164 
165 /*
166  *
167  * For core kernel code by design pmd_trans_huge is never run on any hugetlbfs
168  * page. The hugetlbfs page table walking and mangling paths are totally
169  * separated form the core VM paths and they're differentiated by
170  *  VM_HUGETLB being set on vm_flags well before any pmd_trans_huge could run.
171  *
172  * pmd_trans_huge() is defined as false at build time if
173  * CONFIG_TRANSPARENT_HUGEPAGE=n to optimize away code blocks at build
174  * time in such case.
175  *
176  * For ppc64 we need to differntiate from explicit hugepages from THP, because
177  * for THP we also track the subpage details at the pmd level. We don't do
178  * that for explicit huge pages.
179  *
180  */
181 static inline int hash__pmd_trans_huge(pmd_t pmd)
182 {
183 	return !!((pmd_val(pmd) & (_PAGE_PTE | H_PAGE_THP_HUGE)) ==
184 		  (_PAGE_PTE | H_PAGE_THP_HUGE));
185 }
186 
187 static inline int hash__pmd_same(pmd_t pmd_a, pmd_t pmd_b)
188 {
189 	return (((pmd_raw(pmd_a) ^ pmd_raw(pmd_b)) & ~cpu_to_be64(_PAGE_HPTEFLAGS)) == 0);
190 }
191 
192 static inline pmd_t hash__pmd_mkhuge(pmd_t pmd)
193 {
194 	return __pmd(pmd_val(pmd) | (_PAGE_PTE | H_PAGE_THP_HUGE));
195 }
196 
197 extern unsigned long hash__pmd_hugepage_update(struct mm_struct *mm,
198 					   unsigned long addr, pmd_t *pmdp,
199 					   unsigned long clr, unsigned long set);
200 extern pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma,
201 				   unsigned long address, pmd_t *pmdp);
202 extern void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
203 					 pgtable_t pgtable);
204 extern pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp);
205 extern void hash__pmdp_huge_split_prepare(struct vm_area_struct *vma,
206 				      unsigned long address, pmd_t *pmdp);
207 extern pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm,
208 				       unsigned long addr, pmd_t *pmdp);
209 extern int hash__has_transparent_hugepage(void);
210 #endif /*  CONFIG_TRANSPARENT_HUGEPAGE */
211 #endif	/* __ASSEMBLY__ */
212 
213 #endif /* _ASM_POWERPC_BOOK3S_64_HASH_64K_H */
214