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 permissions don't match, take page fault */
  43                 if (unlikely(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 | _PAGE_BUSY | _PAGE_ACCESSED;
  50                 if (access & _PAGE_RW)
  51                         new_pmd |= _PAGE_DIRTY;
  52         } while (old_pmd != __cmpxchg_u64((unsigned long *)pmdp,
  53                                           old_pmd, new_pmd));
  54         rflags = htab_convert_pte_flags(new_pmd);
  55
  56 #if 0
  57         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) {
  58
  59                 /*
  60                  * No CPU has hugepages but lacks no execute, so we
  61                  * don't need to worry about that case
  62                  */
  63                 rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
  64         }
  65 #endif
  66         /*
  67          * Find the slot index details for this ea, using base page size.
  68          */
  69         shift = mmu_psize_defs[psize].shift;
  70         index = (ea & ~HPAGE_PMD_MASK) >> shift;
  71         BUG_ON(index >= PTE_FRAG_SIZE);
  72
  73         vpn = hpt_vpn(ea, vsid, ssize);
  74         hpte_slot_array = get_hpte_slot_array(pmdp);
  75         if (psize == MMU_PAGE_4K) {
  76                 /*
  77                  * invalidate the old hpte entry if we have that mapped via 64K
  78                  * base page size. This is because demote_segment won't flush
  79                  * hash page table entries.
  80                  */
  81                 if ((old_pmd & _PAGE_HASHPTE) && !(old_pmd & _PAGE_COMBO)) {
  82                         flush_hash_hugepage(vsid, ea, pmdp, MMU_PAGE_64K,
  83                                             ssize, flags);
  84                         /*
  85                          * With THP, we also clear the slot information with
  86                          * respect to all the 64K hash pte mapping the 16MB
  87                          * page. They are all invalid now. This make sure we
  88                          * don't find the slot valid when we fault with 4k
  89                          * base page size.
  90                          *
  91                          */
  92                         memset(hpte_slot_array, 0, PTE_FRAG_SIZE);
  93                 }
  94         }
  95
  96         valid = hpte_valid(hpte_slot_array, index);
  97         if (valid) {
  98                 /* update the hpte bits */
  99                 hash = hpt_hash(vpn, shift, ssize);
 100                 hidx =  hpte_hash_index(hpte_slot_array, index);
 101                 if (hidx & _PTEIDX_SECONDARY)
 102                         hash = ~hash;
 103                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 104                 slot += hidx & _PTEIDX_GROUP_IX;
 105
 106                 ret = ppc_md.hpte_updatepp(slot, rflags, vpn,
 107                                            psize, lpsize, ssize, flags);
 108                 /*
 109                  * We failed to update, try to insert a new entry.
 110                  */
 111                 if (ret == -1) {
 112                         /*
 113                          * large pte is marked busy, so we can be sure
 114                          * nobody is looking at hpte_slot_array. hence we can
 115                          * safely update this here.
 116                          */
 117                         valid = 0;
 118                         hpte_slot_array[index] = 0;
 119                 }
 120         }
 121
 122         if (!valid) {
 123                 unsigned long hpte_group;
 124
 125                 hash = hpt_hash(vpn, shift, ssize);
 126                 /* insert new entry */
 127                 pa = pmd_pfn(__pmd(old_pmd)) << PAGE_SHIFT;
 128                 new_pmd |= _PAGE_HASHPTE;
 129
 130 repeat:
 131                 hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 132
 133                 /* Insert into the hash table, primary slot */
 134                 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, 0,
 135                                           psize, lpsize, ssize);
 136                 /*
 137                  * Primary is full, try the secondary
 138                  */
 139                 if (unlikely(slot == -1)) {
 140                         hpte_group = ((~hash & htab_hash_mask) *
 141                                       HPTES_PER_GROUP) & ~0x7UL;
 142                         slot = ppc_md.hpte_insert(hpte_group, vpn, pa,
 143                                                   rflags, HPTE_V_SECONDARY,
 144                                                   psize, lpsize, ssize);
 145                         if (slot == -1) {
 146                                 if (mftb() & 0x1)
 147                                         hpte_group = ((hash & htab_hash_mask) *
 148                                                       HPTES_PER_GROUP) & ~0x7UL;
 149
 150                                 ppc_md.hpte_remove(hpte_group);
 151                                 goto repeat;
 152                         }
 153                 }
 154                 /*
 155                  * Hypervisor failure. Restore old pmd and return -1
 156                  * similar to __hash_page_*
 157                  */
 158                 if (unlikely(slot == -2)) {
 159                         *pmdp = __pmd(old_pmd);
 160                         hash_failure_debug(ea, access, vsid, trap, ssize,
 161                                            psize, lpsize, old_pmd);
 162                         return -1;
 163                 }
 164                 /*
 165                  * large pte is marked busy, so we can be sure
 166                  * nobody is looking at hpte_slot_array. hence we can
 167                  * safely update this here.
 168                  */
 169                 mark_hpte_slot_valid(hpte_slot_array, index, slot);
 170         }
 171         /*
 172          * Mark the pte with _PAGE_COMBO, if we are trying to hash it with
 173          * base page size 4k.
 174          */
 175         if (psize == MMU_PAGE_4K)
 176                 new_pmd |= _PAGE_COMBO;
 177         /*
 178          * The hpte valid is stored in the pgtable whose address is in the
 179          * second half of the PMD. Order this against clearing of the busy bit in
 180          * huge pmd.
 181          */
 182         smp_wmb();
 183         *pmdp = __pmd(new_pmd & ~_PAGE_BUSY);
 184         return 0;
 185 }