LXR linux/arch/powerpc/mm/book3s64/hash

   1/*
   2 * Copyright IBM Corporation, 2013
   3 * Author Aneesh Kumar K.V <aneesh.kumar@linux.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 & H_PAGE_BUSY))
  41                        return 0;
  42                /* If PMD permissions don't match, take page fault */
  43                if (unlikely(!check_pte_access(access, old_pmd)))
  44                        return 1;
  45                /*
  46                 * Try to lock the PTE, add ACCESSED and DIRTY if it was
  47                 * a write access
  48                 */
  49                new_pmd = old_pmd | H_PAGE_BUSY | _PAGE_ACCESSED;
  50                if (access & _PAGE_WRITE)
  51                        new_pmd |= _PAGE_DIRTY;
  52        } while (!pmd_xchg(pmdp, __pmd(old_pmd), __pmd(new_pmd)));
  53
  54        /*
  55         * Make sure this is thp or devmap entry
  56         */
  57        if (!(old_pmd & (H_PAGE_THP_HUGE | _PAGE_DEVMAP)))
  58                return 0;
  59
  60        rflags = htab_convert_pte_flags(new_pmd, flags);
  61
  62#if 0
  63        if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) {
  64
  65                /*
  66                 * No CPU has hugepages but lacks no execute, so we
  67                 * don't need to worry about that case
  68                 */
  69                rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
  70        }
  71#endif
  72        /*
  73         * Find the slot index details for this ea, using base page size.
  74         */
  75        shift = mmu_psize_defs[psize].shift;
  76        index = (ea & ~HPAGE_PMD_MASK) >> shift;
  77        BUG_ON(index >= PTE_FRAG_SIZE);
  78
  79        vpn = hpt_vpn(ea, vsid, ssize);
  80        hpte_slot_array = get_hpte_slot_array(pmdp);
  81        if (psize == MMU_PAGE_4K) {
  82                /*
  83                 * invalidate the old hpte entry if we have that mapped via 64K
  84                 * base page size. This is because demote_segment won't flush
  85                 * hash page table entries.
  86                 */
  87                if ((old_pmd & H_PAGE_HASHPTE) && !(old_pmd & H_PAGE_COMBO)) {
  88                        flush_hash_hugepage(vsid, ea, pmdp, MMU_PAGE_64K,
  89                                            ssize, flags);
  90                        /*
  91                         * With THP, we also clear the slot information with
  92                         * respect to all the 64K hash pte mapping the 16MB
  93                         * page. They are all invalid now. This make sure we
  94                         * don't find the slot valid when we fault with 4k
  95                         * base page size.
  96                         *
  97                         */
  98                        memset(hpte_slot_array, 0, PTE_FRAG_SIZE);
  99                }
 100        }
 101
 102        valid = hpte_valid(hpte_slot_array, index);
 103        if (valid) {
 104                /* update the hpte bits */
 105                hash = hpt_hash(vpn, shift, ssize);
 106                hidx =  hpte_hash_index(hpte_slot_array, index);
 107                if (hidx & _PTEIDX_SECONDARY)
 108                        hash = ~hash;
 109                slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 110                slot += hidx & _PTEIDX_GROUP_IX;
 111
 112                ret = mmu_hash_ops.hpte_updatepp(slot, rflags, vpn,
 113                                                 psize, lpsize, ssize, flags);
 114                /*
 115                 * We failed to update, try to insert a new entry.
 116                 */
 117                if (ret == -1) {
 118                        /*
 119                         * large pte is marked busy, so we can be sure
 120                         * nobody is looking at hpte_slot_array. hence we can
 121                         * safely update this here.
 122                         */
 123                        valid = 0;
 124                        hpte_slot_array[index] = 0;
 125                }
 126        }
 127
 128        if (!valid) {
 129                unsigned long hpte_group;
 130
 131                hash = hpt_hash(vpn, shift, ssize);
 132                /* insert new entry */
 133                pa = pmd_pfn(__pmd(old_pmd)) << PAGE_SHIFT;
 134                new_pmd |= H_PAGE_HASHPTE;
 135
 136repeat:
 137                hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 138
 139                /* Insert into the hash table, primary slot */
 140                slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags, 0,
 141                                                psize, lpsize, ssize);
 142                /*
 143                 * Primary is full, try the secondary
 144                 */
 145                if (unlikely(slot == -1)) {
 146                        hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
 147                        slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa,
 148                                                        rflags,
 149                                                        HPTE_V_SECONDARY,
 150                                                        psize, lpsize, ssize);
 151                        if (slot == -1) {
 152                                if (mftb() & 0x1)
 153                                        hpte_group = (hash & htab_hash_mask) *
 154                                                        HPTES_PER_GROUP;
 155
 156                                mmu_hash_ops.hpte_remove(hpte_group);
 157                                goto repeat;
 158                        }
 159                }
 160                /*
 161                 * Hypervisor failure. Restore old pmd and return -1
 162                 * similar to __hash_page_*
 163                 */
 164                if (unlikely(slot == -2)) {
 165                        *pmdp = __pmd(old_pmd);
 166                        hash_failure_debug(ea, access, vsid, trap, ssize,
 167                                           psize, lpsize, old_pmd);
 168                        return -1;
 169                }
 170                /*
 171                 * large pte is marked busy, so we can be sure
 172                 * nobody is looking at hpte_slot_array. hence we can
 173                 * safely update this here.
 174                 */
 175                mark_hpte_slot_valid(hpte_slot_array, index, slot);
 176        }
 177        /*
 178         * Mark the pte with H_PAGE_COMBO, if we are trying to hash it with
 179         * base page size 4k.
 180         */
 181        if (psize == MMU_PAGE_4K)
 182                new_pmd |= H_PAGE_COMBO;
 183        /*
 184         * The hpte valid is stored in the pgtable whose address is in the
 185         * second half of the PMD. Order this against clearing of the busy bit in
 186         * huge pmd.
 187         */
 188        smp_wmb();
 189        *pmdp = __pmd(new_pmd & ~H_PAGE_BUSY);
 190        return 0;
 191}
 192