1/* 2 * Copyright IBM Corporation, 2013 3 * Author Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com> 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms of version 2.1 of the GNU Lesser General Public License 7 * as published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it would be useful, but 10 * WITHOUT ANY WARRANTY; without even the implied warranty of 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. 12 * 13 */ 14 15/* 16 * PPC64 THP Support for hash based MMUs 17 */ 18#include <linux/mm.h> 19#include <asm/machdep.h> 20 21int __hash_page_thp(unsigned long ea, unsigned long access, unsigned long vsid, 22 pmd_t *pmdp, unsigned long trap, unsigned long flags, 23 int ssize, unsigned int psize) 24{ 25 unsigned int index, valid; 26 unsigned char *hpte_slot_array; 27 unsigned long rflags, pa, hidx; 28 unsigned long old_pmd, new_pmd; 29 int ret, lpsize = MMU_PAGE_16M; 30 unsigned long vpn, hash, shift, slot; 31 32 /* 33 * atomically mark the linux large page PMD busy and dirty 34 */ 35 do { 36 pmd_t pmd = READ_ONCE(*pmdp); 37 38 old_pmd = pmd_val(pmd); 39 /* If PMD busy, retry the access */ 40 if (unlikely(old_pmd & _PAGE_BUSY)) 41 return 0; 42 /* If PMD is trans splitting retry the access */ 43 if (unlikely(old_pmd & _PAGE_SPLITTING)) 44 return 0; 45 /* If PMD permissions don't match, take page fault */ 46 if (unlikely(access & ~old_pmd)) 47 return 1; 48 /* 49 * Try to lock the PTE, add ACCESSED and DIRTY if it was 50 * a write access 51 */ 52 new_pmd = old_pmd | _PAGE_BUSY | _PAGE_ACCESSED; 53 if (access & _PAGE_RW) 54 new_pmd |= _PAGE_DIRTY; 55 } while (old_pmd != __cmpxchg_u64((unsigned long *)pmdp, 56 old_pmd, new_pmd)); 57 /* 58 * PP bits. _PAGE_USER is already PP bit 0x2, so we only 59 * need to add in 0x1 if it's a read-only user page 60 */ 61 rflags = new_pmd & _PAGE_USER; 62 if ((new_pmd & _PAGE_USER) && !((new_pmd & _PAGE_RW) && 63 (new_pmd & _PAGE_DIRTY))) 64 rflags |= 0x1; 65 /* 66 * _PAGE_EXEC -> HW_NO_EXEC since it's inverted 67 */ 68 rflags |= ((new_pmd & _PAGE_EXEC) ? 0 : HPTE_R_N); 69 70#if 0 71 if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) { 72 73 /* 74 * No CPU has hugepages but lacks no execute, so we 75 * don't need to worry about that case 76 */ 77 rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap); 78 } 79#endif 80 /* 81 * Find the slot index details for this ea, using base page size. 82 */ 83 shift = mmu_psize_defs[psize].shift; 84 index = (ea & ~HPAGE_PMD_MASK) >> shift; 85 BUG_ON(index >= 4096); 86 87 vpn = hpt_vpn(ea, vsid, ssize); 88 hpte_slot_array = get_hpte_slot_array(pmdp); 89 if (psize == MMU_PAGE_4K) { 90 /* 91 * invalidate the old hpte entry if we have that mapped via 64K 92 * base page size. This is because demote_segment won't flush 93 * hash page table entries. 94 */ 95 if ((old_pmd & _PAGE_HASHPTE) && !(old_pmd & _PAGE_COMBO)) 96 flush_hash_hugepage(vsid, ea, pmdp, MMU_PAGE_64K, 97 ssize, flags); 98 } 99 100 valid = hpte_valid(hpte_slot_array, index); 101 if (valid) { 102 /* update the hpte bits */ 103 hash = hpt_hash(vpn, shift, ssize); 104 hidx = hpte_hash_index(hpte_slot_array, index); 105 if (hidx & _PTEIDX_SECONDARY) 106 hash = ~hash; 107 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP; 108 slot += hidx & _PTEIDX_GROUP_IX; 109 110 ret = ppc_md.hpte_updatepp(slot, rflags, vpn, 111 psize, lpsize, ssize, flags); 112 /* 113 * We failed to update, try to insert a new entry. 114 */ 115 if (ret == -1) { 116 /* 117 * large pte is marked busy, so we can be sure 118 * nobody is looking at hpte_slot_array. hence we can 119 * safely update this here. 120 */ 121 valid = 0; 122 hpte_slot_array[index] = 0; 123 } 124 } 125 126 if (!valid) { 127 unsigned long hpte_group; 128 129 hash = hpt_hash(vpn, shift, ssize); 130 /* insert new entry */ 131 pa = pmd_pfn(__pmd(old_pmd)) << PAGE_SHIFT; 132 new_pmd |= _PAGE_HASHPTE; 133 134 /* Add in WIMG bits */ 135 rflags |= (new_pmd & (_PAGE_WRITETHRU | _PAGE_NO_CACHE | 136 _PAGE_GUARDED)); 137 /* 138 * enable the memory coherence always 139 */ 140 rflags |= HPTE_R_M; 141repeat: 142 hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL; 143 144 /* Insert into the hash table, primary slot */ 145 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, 0, 146 psize, lpsize, ssize); 147 /* 148 * Primary is full, try the secondary 149 */ 150 if (unlikely(slot == -1)) { 151 hpte_group = ((~hash & htab_hash_mask) * 152 HPTES_PER_GROUP) & ~0x7UL; 153 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, 154 rflags, HPTE_V_SECONDARY, 155 psize, lpsize, ssize); 156 if (slot == -1) { 157 if (mftb() & 0x1) 158 hpte_group = ((hash & htab_hash_mask) * 159 HPTES_PER_GROUP) & ~0x7UL; 160 161 ppc_md.hpte_remove(hpte_group); 162 goto repeat; 163 } 164 } 165 /* 166 * Hypervisor failure. Restore old pmd and return -1 167 * similar to __hash_page_* 168 */ 169 if (unlikely(slot == -2)) { 170 *pmdp = __pmd(old_pmd); 171 hash_failure_debug(ea, access, vsid, trap, ssize, 172 psize, lpsize, old_pmd); 173 return -1; 174 } 175 /* 176 * large pte is marked busy, so we can be sure 177 * nobody is looking at hpte_slot_array. hence we can 178 * safely update this here. 179 */ 180 mark_hpte_slot_valid(hpte_slot_array, index, slot); 181 } 182 /* 183 * Mark the pte with _PAGE_COMBO, if we are trying to hash it with 184 * base page size 4k. 185 */ 186 if (psize == MMU_PAGE_4K) 187 new_pmd |= _PAGE_COMBO; 188 /* 189 * The hpte valid is stored in the pgtable whose address is in the 190 * second half of the PMD. Order this against clearing of the busy bit in 191 * huge pmd. 192 */ 193 smp_wmb(); 194 *pmdp = __pmd(new_pmd & ~_PAGE_BUSY); 195 return 0; 196} 197