1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef _ASM_POWERPC_BOOK3S_64_HASH_64K_H 3 #define _ASM_POWERPC_BOOK3S_64_HASH_64K_H 4 5 #define H_PTE_INDEX_SIZE 8 6 #define H_PMD_INDEX_SIZE 10 7 #define H_PUD_INDEX_SIZE 10 8 #define H_PGD_INDEX_SIZE 8 9 10 /* 11 * Each context is 512TB size. SLB miss for first context/default context 12 * is handled in the hotpath. 13 */ 14 #define MAX_EA_BITS_PER_CONTEXT 49 15 16 /* 17 * 64k aligned address free up few of the lower bits of RPN for us 18 * We steal that here. For more deatils look at pte_pfn/pfn_pte() 19 */ 20 #define H_PAGE_COMBO _RPAGE_RPN0 /* this is a combo 4k page */ 21 #define H_PAGE_4K_PFN _RPAGE_RPN1 /* PFN is for a single 4k page */ 22 #define H_PAGE_BUSY _RPAGE_RPN44 /* software: PTE & hash are busy */ 23 #define H_PAGE_HASHPTE _RPAGE_RPN43 /* PTE has associated HPTE */ 24 25 /* memory key bits. */ 26 #define H_PTE_PKEY_BIT0 _RPAGE_RSV1 27 #define H_PTE_PKEY_BIT1 _RPAGE_RSV2 28 #define H_PTE_PKEY_BIT2 _RPAGE_RSV3 29 #define H_PTE_PKEY_BIT3 _RPAGE_RSV4 30 #define H_PTE_PKEY_BIT4 _RPAGE_RSV5 31 32 /* 33 * We need to differentiate between explicit huge page and THP huge 34 * page, since THP huge page also need to track real subpage details 35 */ 36 #define H_PAGE_THP_HUGE H_PAGE_4K_PFN 37 38 /* PTE flags to conserve for HPTE identification */ 39 #define _PAGE_HPTEFLAGS (H_PAGE_BUSY | H_PAGE_HASHPTE | H_PAGE_COMBO) 40 /* 41 * We use a 2K PTE page fragment and another 2K for storing 42 * real_pte_t hash index 43 * 8 bytes per each pte entry and another 8 bytes for storing 44 * slot details. 45 */ 46 #define H_PTE_FRAG_SIZE_SHIFT (H_PTE_INDEX_SIZE + 3 + 1) 47 #define H_PTE_FRAG_NR (PAGE_SIZE >> H_PTE_FRAG_SIZE_SHIFT) 48 49 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLB_PAGE) 50 #define H_PMD_FRAG_SIZE_SHIFT (H_PMD_INDEX_SIZE + 3 + 1) 51 #else 52 #define H_PMD_FRAG_SIZE_SHIFT (H_PMD_INDEX_SIZE + 3) 53 #endif 54 #define H_PMD_FRAG_NR (PAGE_SIZE >> H_PMD_FRAG_SIZE_SHIFT) 55 56 #ifndef __ASSEMBLY__ 57 #include <asm/errno.h> 58 59 /* 60 * With 64K pages on hash table, we have a special PTE format that 61 * uses a second "half" of the page table to encode sub-page information 62 * in order to deal with 64K made of 4K HW pages. Thus we override the 63 * generic accessors and iterators here 64 */ 65 #define __real_pte __real_pte 66 static inline real_pte_t __real_pte(pte_t pte, pte_t *ptep, int offset) 67 { 68 real_pte_t rpte; 69 unsigned long *hidxp; 70 71 rpte.pte = pte; 72 73 /* 74 * Ensure that we do not read the hidx before we read the PTE. Because 75 * the writer side is expected to finish writing the hidx first followed 76 * by the PTE, by using smp_wmb(). pte_set_hash_slot() ensures that. 77 */ 78 smp_rmb(); 79 80 hidxp = (unsigned long *)(ptep + offset); 81 rpte.hidx = *hidxp; 82 return rpte; 83 } 84 85 /* 86 * shift the hidx representation by one-modulo-0xf; i.e hidx 0 is respresented 87 * as 1, 1 as 2,... , and 0xf as 0. This convention lets us represent a 88 * invalid hidx 0xf with a 0x0 bit value. PTEs are anyway zero'd when 89 * allocated. We dont have to zero them gain; thus save on the initialization. 90 */ 91 #define HIDX_UNSHIFT_BY_ONE(x) ((x + 0xfUL) & 0xfUL) /* shift backward by one */ 92 #define HIDX_SHIFT_BY_ONE(x) ((x + 0x1UL) & 0xfUL) /* shift forward by one */ 93 #define HIDX_BITS(x, index) (x << (index << 2)) 94 #define BITS_TO_HIDX(x, index) ((x >> (index << 2)) & 0xfUL) 95 #define INVALID_RPTE_HIDX 0x0UL 96 97 static inline unsigned long __rpte_to_hidx(real_pte_t rpte, unsigned long index) 98 { 99 return HIDX_UNSHIFT_BY_ONE(BITS_TO_HIDX(rpte.hidx, index)); 100 } 101 102 /* 103 * Commit the hidx and return PTE bits that needs to be modified. The caller is 104 * expected to modify the PTE bits accordingly and commit the PTE to memory. 105 */ 106 static inline unsigned long pte_set_hidx(pte_t *ptep, real_pte_t rpte, 107 unsigned int subpg_index, 108 unsigned long hidx, int offset) 109 { 110 unsigned long *hidxp = (unsigned long *)(ptep + offset); 111 112 rpte.hidx &= ~HIDX_BITS(0xfUL, subpg_index); 113 *hidxp = rpte.hidx | HIDX_BITS(HIDX_SHIFT_BY_ONE(hidx), subpg_index); 114 115 /* 116 * Anyone reading PTE must ensure hidx bits are read after reading the 117 * PTE by using the read-side barrier smp_rmb(). __real_pte() can be 118 * used for that. 119 */ 120 smp_wmb(); 121 122 /* No PTE bits to be modified, return 0x0UL */ 123 return 0x0UL; 124 } 125 126 #define __rpte_to_pte(r) ((r).pte) 127 extern bool __rpte_sub_valid(real_pte_t rpte, unsigned long index); 128 /* 129 * Trick: we set __end to va + 64k, which happens works for 130 * a 16M page as well as we want only one iteration 131 */ 132 #define pte_iterate_hashed_subpages(rpte, psize, vpn, index, shift) \ 133 do { \ 134 unsigned long __end = vpn + (1UL << (PAGE_SHIFT - VPN_SHIFT)); \ 135 unsigned __split = (psize == MMU_PAGE_4K || \ 136 psize == MMU_PAGE_64K_AP); \ 137 shift = mmu_psize_defs[psize].shift; \ 138 for (index = 0; vpn < __end; index++, \ 139 vpn += (1L << (shift - VPN_SHIFT))) { \ 140 if (!__split || __rpte_sub_valid(rpte, index)) \ 141 do { 142 143 #define pte_iterate_hashed_end() } while(0); } } while(0) 144 145 #define pte_pagesize_index(mm, addr, pte) \ 146 (((pte) & H_PAGE_COMBO)? MMU_PAGE_4K: MMU_PAGE_64K) 147 148 extern int remap_pfn_range(struct vm_area_struct *, unsigned long addr, 149 unsigned long pfn, unsigned long size, pgprot_t); 150 static inline int hash__remap_4k_pfn(struct vm_area_struct *vma, unsigned long addr, 151 unsigned long pfn, pgprot_t prot) 152 { 153 if (pfn > (PTE_RPN_MASK >> PAGE_SHIFT)) { 154 WARN(1, "remap_4k_pfn called with wrong pfn value\n"); 155 return -EINVAL; 156 } 157 return remap_pfn_range(vma, addr, pfn, PAGE_SIZE, 158 __pgprot(pgprot_val(prot) | H_PAGE_4K_PFN)); 159 } 160 161 #define H_PTE_TABLE_SIZE PTE_FRAG_SIZE 162 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined (CONFIG_HUGETLB_PAGE) 163 #define H_PMD_TABLE_SIZE ((sizeof(pmd_t) << PMD_INDEX_SIZE) + \ 164 (sizeof(unsigned long) << PMD_INDEX_SIZE)) 165 #else 166 #define H_PMD_TABLE_SIZE (sizeof(pmd_t) << PMD_INDEX_SIZE) 167 #endif 168 #ifdef CONFIG_HUGETLB_PAGE 169 #define H_PUD_TABLE_SIZE ((sizeof(pud_t) << PUD_INDEX_SIZE) + \ 170 (sizeof(unsigned long) << PUD_INDEX_SIZE)) 171 #else 172 #define H_PUD_TABLE_SIZE (sizeof(pud_t) << PUD_INDEX_SIZE) 173 #endif 174 #define H_PGD_TABLE_SIZE (sizeof(pgd_t) << PGD_INDEX_SIZE) 175 176 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 177 static inline char *get_hpte_slot_array(pmd_t *pmdp) 178 { 179 /* 180 * The hpte hindex is stored in the pgtable whose address is in the 181 * second half of the PMD 182 * 183 * Order this load with the test for pmd_trans_huge in the caller 184 */ 185 smp_rmb(); 186 return *(char **)(pmdp + PTRS_PER_PMD); 187 188 189 } 190 /* 191 * The linux hugepage PMD now include the pmd entries followed by the address 192 * to the stashed pgtable_t. The stashed pgtable_t contains the hpte bits. 193 * [ 000 | 1 bit secondary | 3 bit hidx | 1 bit valid]. We use one byte per 194 * each HPTE entry. With 16MB hugepage and 64K HPTE we need 256 entries and 195 * with 4K HPTE we need 4096 entries. Both will fit in a 4K pgtable_t. 196 * 197 * The top three bits are intentionally left as zero. This memory location 198 * are also used as normal page PTE pointers. So if we have any pointers 199 * left around while we collapse a hugepage, we need to make sure 200 * _PAGE_PRESENT bit of that is zero when we look at them 201 */ 202 static inline unsigned int hpte_valid(unsigned char *hpte_slot_array, int index) 203 { 204 return hpte_slot_array[index] & 0x1; 205 } 206 207 static inline unsigned int hpte_hash_index(unsigned char *hpte_slot_array, 208 int index) 209 { 210 return hpte_slot_array[index] >> 1; 211 } 212 213 static inline void mark_hpte_slot_valid(unsigned char *hpte_slot_array, 214 unsigned int index, unsigned int hidx) 215 { 216 hpte_slot_array[index] = (hidx << 1) | 0x1; 217 } 218 219 /* 220 * 221 * For core kernel code by design pmd_trans_huge is never run on any hugetlbfs 222 * page. The hugetlbfs page table walking and mangling paths are totally 223 * separated form the core VM paths and they're differentiated by 224 * VM_HUGETLB being set on vm_flags well before any pmd_trans_huge could run. 225 * 226 * pmd_trans_huge() is defined as false at build time if 227 * CONFIG_TRANSPARENT_HUGEPAGE=n to optimize away code blocks at build 228 * time in such case. 229 * 230 * For ppc64 we need to differntiate from explicit hugepages from THP, because 231 * for THP we also track the subpage details at the pmd level. We don't do 232 * that for explicit huge pages. 233 * 234 */ 235 static inline int hash__pmd_trans_huge(pmd_t pmd) 236 { 237 return !!((pmd_val(pmd) & (_PAGE_PTE | H_PAGE_THP_HUGE)) == 238 (_PAGE_PTE | H_PAGE_THP_HUGE)); 239 } 240 241 static inline int hash__pmd_same(pmd_t pmd_a, pmd_t pmd_b) 242 { 243 return (((pmd_raw(pmd_a) ^ pmd_raw(pmd_b)) & ~cpu_to_be64(_PAGE_HPTEFLAGS)) == 0); 244 } 245 246 static inline pmd_t hash__pmd_mkhuge(pmd_t pmd) 247 { 248 return __pmd(pmd_val(pmd) | (_PAGE_PTE | H_PAGE_THP_HUGE)); 249 } 250 251 extern unsigned long hash__pmd_hugepage_update(struct mm_struct *mm, 252 unsigned long addr, pmd_t *pmdp, 253 unsigned long clr, unsigned long set); 254 extern pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma, 255 unsigned long address, pmd_t *pmdp); 256 extern void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp, 257 pgtable_t pgtable); 258 extern pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp); 259 extern pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm, 260 unsigned long addr, pmd_t *pmdp); 261 extern int hash__has_transparent_hugepage(void); 262 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 263 #endif /* __ASSEMBLY__ */ 264 265 #endif /* _ASM_POWERPC_BOOK3S_64_HASH_64K_H */ 266