arch/powerpc/mm/hugepage-hash64.c

   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
  21 int __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         rflags = htab_convert_pte_flags(new_pmd);
  58
  59 #if 0
  60         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) {
  61
  62                 /*
  63                  * No CPU has hugepages but lacks no execute, so we
  64                  * don't need to worry about that case
  65                  */
  66                 rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
  67         }
  68 #endif
  69         /*
  70          * Find the slot index details for this ea, using base page size.
  71          */
  72         shift = mmu_psize_defs[psize].shift;
  73         index = (ea & ~HPAGE_PMD_MASK) >> shift;
  74         BUG_ON(index >= PTE_FRAG_SIZE);
  75
  76         vpn = hpt_vpn(ea, vsid, ssize);
  77         hpte_slot_array = get_hpte_slot_array(pmdp);
  78         if (psize == MMU_PAGE_4K) {
  79                 /*
  80                  * invalidate the old hpte entry if we have that mapped via 64K
  81                  * base page size. This is because demote_segment won't flush
  82                  * hash page table entries.
  83                  */
  84                 if ((old_pmd & _PAGE_HASHPTE) && !(old_pmd & _PAGE_COMBO))
  85                         flush_hash_hugepage(vsid, ea, pmdp, MMU_PAGE_64K,
  86                                             ssize, flags);
  87         }
  88
  89         valid = hpte_valid(hpte_slot_array, index);
  90         if (valid) {
  91                 /* update the hpte bits */
  92                 hash = hpt_hash(vpn, shift, ssize);
  93                 hidx =  hpte_hash_index(hpte_slot_array, index);
  94                 if (hidx & _PTEIDX_SECONDARY)
  95                         hash = ~hash;
  96                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
  97                 slot += hidx & _PTEIDX_GROUP_IX;
  98
  99                 ret = ppc_md.hpte_updatepp(slot, rflags, vpn,
 100                                            psize, lpsize, ssize, flags);
 101                 /*
 102                  * We failed to update, try to insert a new entry.
 103                  */
 104                 if (ret == -1) {
 105                         /*
 106                          * large pte is marked busy, so we can be sure
 107                          * nobody is looking at hpte_slot_array. hence we can
 108                          * safely update this here.
 109                          */
 110                         valid = 0;
 111                         hpte_slot_array[index] = 0;
 112                 }
 113         }
 114
 115         if (!valid) {
 116                 unsigned long hpte_group;
 117
 118                 hash = hpt_hash(vpn, shift, ssize);
 119                 /* insert new entry */
 120                 pa = pmd_pfn(__pmd(old_pmd)) << PAGE_SHIFT;
 121                 new_pmd |= _PAGE_HASHPTE;
 122
 123 repeat:
 124                 hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 125
 126                 /* Insert into the hash table, primary slot */
 127                 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, 0,
 128                                           psize, lpsize, ssize);
 129                 /*
 130                  * Primary is full, try the secondary
 131                  */
 132                 if (unlikely(slot == -1)) {
 133                         hpte_group = ((~hash & htab_hash_mask) *
 134                                       HPTES_PER_GROUP) & ~0x7UL;
 135                         slot = ppc_md.hpte_insert(hpte_group, vpn, pa,
 136                                                   rflags, HPTE_V_SECONDARY,
 137                                                   psize, lpsize, ssize);
 138                         if (slot == -1) {
 139                                 if (mftb() & 0x1)
 140                                         hpte_group = ((hash & htab_hash_mask) *
 141                                                       HPTES_PER_GROUP) & ~0x7UL;
 142
 143                                 ppc_md.hpte_remove(hpte_group);
 144                                 goto repeat;
 145                         }
 146                 }
 147                 /*
 148                  * Hypervisor failure. Restore old pmd and return -1
 149                  * similar to __hash_page_*
 150                  */
 151                 if (unlikely(slot == -2)) {
 152                         *pmdp = __pmd(old_pmd);
 153                         hash_failure_debug(ea, access, vsid, trap, ssize,
 154                                            psize, lpsize, old_pmd);
 155                         return -1;
 156                 }
 157                 /*
 158                  * large pte is marked busy, so we can be sure
 159                  * nobody is looking at hpte_slot_array. hence we can
 160                  * safely update this here.
 161                  */
 162                 mark_hpte_slot_valid(hpte_slot_array, index, slot);
 163         }
 164         /*
 165          * Mark the pte with _PAGE_COMBO, if we are trying to hash it with
 166          * base page size 4k.
 167          */
 168         if (psize == MMU_PAGE_4K)
 169                 new_pmd |= _PAGE_COMBO;
 170         /*
 171          * The hpte valid is stored in the pgtable whose address is in the
 172          * second half of the PMD. Order this against clearing of the busy bit in
 173          * huge pmd.
 174          */
 175         smp_wmb();
 176         *pmdp = __pmd(new_pmd & ~_PAGE_BUSY);
 177         return 0;
 178 }