arch/powerpc/mm/hash64_64k.c

   1 /*
   2  * Copyright IBM Corporation, 2015
   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 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 #include <linux/mm.h>
  16 #include <asm/machdep.h>
  17 #include <asm/mmu.h>
  18 /*
  19  * index from 0 - 15
  20  */
  21 bool __rpte_sub_valid(real_pte_t rpte, unsigned long index)
  22 {
  23         unsigned long g_idx;
  24         unsigned long ptev = pte_val(rpte.pte);
  25
  26         g_idx = (ptev & _PAGE_COMBO_VALID) >> _PAGE_F_GIX_SHIFT;
  27         index = index >> 2;
  28         if (g_idx & (0x1 << index))
  29                 return true;
  30         else
  31                 return false;
  32 }
  33 /*
  34  * index from 0 - 15
  35  */
  36 static unsigned long mark_subptegroup_valid(unsigned long ptev, unsigned long index)
  37 {
  38         unsigned long g_idx;
  39
  40         if (!(ptev & _PAGE_COMBO))
  41                 return ptev;
  42         index = index >> 2;
  43         g_idx = 0x1 << index;
  44
  45         return ptev | (g_idx << _PAGE_F_GIX_SHIFT);
  46 }
  47
  48 int __hash_page_4K(unsigned long ea, unsigned long access, unsigned long vsid,
  49                    pte_t *ptep, unsigned long trap, unsigned long flags,
  50                    int ssize, int subpg_prot)
  51 {
  52         real_pte_t rpte;
  53         unsigned long *hidxp;
  54         unsigned long hpte_group;
  55         unsigned int subpg_index;
  56         unsigned long rflags, pa, hidx;
  57         unsigned long old_pte, new_pte, subpg_pte;
  58         unsigned long vpn, hash, slot;
  59         unsigned long shift = mmu_psize_defs[MMU_PAGE_4K].shift;
  60
  61         /*
  62          * atomically mark the linux large page PTE busy and dirty
  63          */
  64         do {
  65                 pte_t pte = READ_ONCE(*ptep);
  66
  67                 old_pte = pte_val(pte);
  68                 /* If PTE busy, retry the access */
  69                 if (unlikely(old_pte & _PAGE_BUSY))
  70                         return 0;
  71                 /* If PTE permissions don't match, take page fault */
  72                 if (unlikely(access & ~old_pte))
  73                         return 1;
  74                 /*
  75                  * Try to lock the PTE, add ACCESSED and DIRTY if it was
  76                  * a write access. Since this is 4K insert of 64K page size
  77                  * also add _PAGE_COMBO
  78                  */
  79                 new_pte = old_pte | _PAGE_BUSY | _PAGE_ACCESSED | _PAGE_COMBO;
  80                 if (access & _PAGE_RW)
  81                         new_pte |= _PAGE_DIRTY;
  82         } while (old_pte != __cmpxchg_u64((unsigned long *)ptep,
  83                                           old_pte, new_pte));
  84         /*
  85          * Handle the subpage protection bits
  86          */
  87         subpg_pte = new_pte & ~subpg_prot;
  88         rflags = htab_convert_pte_flags(subpg_pte);
  89
  90         if (!cpu_has_feature(CPU_FTR_NOEXECUTE) &&
  91             !cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) {
  92
  93                 /*
  94                  * No CPU has hugepages but lacks no execute, so we
  95                  * don't need to worry about that case
  96                  */
  97                 rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
  98         }
  99
 100         subpg_index = (ea & (PAGE_SIZE - 1)) >> shift;
 101         vpn  = hpt_vpn(ea, vsid, ssize);
 102         rpte = __real_pte(__pte(old_pte), ptep);
 103         /*
 104          *None of the sub 4k page is hashed
 105          */
 106         if (!(old_pte & _PAGE_HASHPTE))
 107                 goto htab_insert_hpte;
 108         /*
 109          * Check if the pte was already inserted into the hash table
 110          * as a 64k HW page, and invalidate the 64k HPTE if so.
 111          */
 112         if (!(old_pte & _PAGE_COMBO)) {
 113                 flush_hash_page(vpn, rpte, MMU_PAGE_64K, ssize, flags);
 114                 old_pte &= ~_PAGE_HASHPTE | _PAGE_F_GIX | _PAGE_F_SECOND;
 115                 goto htab_insert_hpte;
 116         }
 117         /*
 118          * Check for sub page valid and update
 119          */
 120         if (__rpte_sub_valid(rpte, subpg_index)) {
 121                 int ret;
 122
 123                 hash = hpt_hash(vpn, shift, ssize);
 124                 hidx = __rpte_to_hidx(rpte, subpg_index);
 125                 if (hidx & _PTEIDX_SECONDARY)
 126                         hash = ~hash;
 127                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 128                 slot += hidx & _PTEIDX_GROUP_IX;
 129
 130                 ret = ppc_md.hpte_updatepp(slot, rflags, vpn,
 131                                            MMU_PAGE_4K, MMU_PAGE_4K,
 132                                            ssize, flags);
 133                 /*
 134                  *if we failed because typically the HPTE wasn't really here
 135                  * we try an insertion.
 136                  */
 137                 if (ret == -1)
 138                         goto htab_insert_hpte;
 139
 140                 *ptep = __pte(new_pte & ~_PAGE_BUSY);
 141                 return 0;
 142         }
 143
 144 htab_insert_hpte:
 145         /*
 146          * handle _PAGE_4K_PFN case
 147          */
 148         if (old_pte & _PAGE_4K_PFN) {
 149                 /*
 150                  * All the sub 4k page have the same
 151                  * physical address.
 152                  */
 153                 pa = pte_pfn(__pte(old_pte)) << HW_PAGE_SHIFT;
 154         } else {
 155                 pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
 156                 pa += (subpg_index << shift);
 157         }
 158         hash = hpt_hash(vpn, shift, ssize);
 159 repeat:
 160         hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 161
 162         /* Insert into the hash table, primary slot */
 163         slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, 0,
 164                                   MMU_PAGE_4K, MMU_PAGE_4K, ssize);
 165         /*
 166          * Primary is full, try the secondary
 167          */
 168         if (unlikely(slot == -1)) {
 169                 hpte_group = ((~hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 170                 slot = ppc_md.hpte_insert(hpte_group, vpn, pa,
 171                                           rflags, HPTE_V_SECONDARY,
 172                                           MMU_PAGE_4K, MMU_PAGE_4K, ssize);
 173                 if (slot == -1) {
 174                         if (mftb() & 0x1)
 175                                 hpte_group = ((hash & htab_hash_mask) *
 176                                               HPTES_PER_GROUP) & ~0x7UL;
 177                         ppc_md.hpte_remove(hpte_group);
 178                         /*
 179                          * FIXME!! Should be try the group from which we removed ?
 180                          */
 181                         goto repeat;
 182                 }
 183         }
 184         /*
 185          * Hypervisor failure. Restore old pmd and return -1
 186          * similar to __hash_page_*
 187          */
 188         if (unlikely(slot == -2)) {
 189                 *ptep = __pte(old_pte);
 190                 hash_failure_debug(ea, access, vsid, trap, ssize,
 191                                    MMU_PAGE_4K, MMU_PAGE_4K, old_pte);
 192                 return -1;
 193         }
 194         /*
 195          * Insert slot number & secondary bit in PTE second half,
 196          * clear _PAGE_BUSY and set appropriate HPTE slot bit
 197          * Since we have _PAGE_BUSY set on ptep, we can be sure
 198          * nobody is undating hidx.
 199          */
 200         hidxp = (unsigned long *)(ptep + PTRS_PER_PTE);
 201         rpte.hidx &= ~(0xfUL << (subpg_index << 2));
 202         *hidxp = rpte.hidx  | (slot << (subpg_index << 2));
 203         new_pte = mark_subptegroup_valid(new_pte, subpg_index);
 204         new_pte |=  _PAGE_HASHPTE;
 205         /*
 206          * check __real_pte for details on matching smp_rmb()
 207          */
 208         smp_wmb();
 209         *ptep = __pte(new_pte & ~_PAGE_BUSY);
 210         return 0;
 211 }
 212
 213 int __hash_page_64K(unsigned long ea, unsigned long access,
 214                     unsigned long vsid, pte_t *ptep, unsigned long trap,
 215                     unsigned long flags, int ssize)
 216 {
 217
 218         unsigned long hpte_group;
 219         unsigned long rflags, pa;
 220         unsigned long old_pte, new_pte;
 221         unsigned long vpn, hash, slot;
 222         unsigned long shift = mmu_psize_defs[MMU_PAGE_64K].shift;
 223
 224         /*
 225          * atomically mark the linux large page PTE busy and dirty
 226          */
 227         do {
 228                 pte_t pte = READ_ONCE(*ptep);
 229
 230                 old_pte = pte_val(pte);
 231                 /* If PTE busy, retry the access */
 232                 if (unlikely(old_pte & _PAGE_BUSY))
 233                         return 0;
 234                 /* If PTE permissions don't match, take page fault */
 235                 if (unlikely(access & ~old_pte))
 236                         return 1;
 237                 /*
 238                  * Check if PTE has the cache-inhibit bit set
 239                  * If so, bail out and refault as a 4k page
 240                  */
 241                 if (!mmu_has_feature(MMU_FTR_CI_LARGE_PAGE) &&
 242                     unlikely(old_pte & _PAGE_NO_CACHE))
 243                         return 0;
 244                 /*
 245                  * Try to lock the PTE, add ACCESSED and DIRTY if it was
 246                  * a write access. Since this is 4K insert of 64K page size
 247                  * also add _PAGE_COMBO
 248                  */
 249                 new_pte = old_pte | _PAGE_BUSY | _PAGE_ACCESSED;
 250                 if (access & _PAGE_RW)
 251                         new_pte |= _PAGE_DIRTY;
 252         } while (old_pte != __cmpxchg_u64((unsigned long *)ptep,
 253                                           old_pte, new_pte));
 254
 255         rflags = htab_convert_pte_flags(new_pte);
 256
 257         if (!cpu_has_feature(CPU_FTR_NOEXECUTE) &&
 258             !cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
 259                 rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
 260
 261         vpn  = hpt_vpn(ea, vsid, ssize);
 262         if (unlikely(old_pte & _PAGE_HASHPTE)) {
 263                 /*
 264                  * There MIGHT be an HPTE for this pte
 265                  */
 266                 hash = hpt_hash(vpn, shift, ssize);
 267                 if (old_pte & _PAGE_F_SECOND)
 268                         hash = ~hash;
 269                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 270                 slot += (old_pte & _PAGE_F_GIX) >> _PAGE_F_GIX_SHIFT;
 271
 272                 if (ppc_md.hpte_updatepp(slot, rflags, vpn, MMU_PAGE_64K,
 273                                          MMU_PAGE_64K, ssize, flags) == -1)
 274                         old_pte &= ~_PAGE_HPTEFLAGS;
 275         }
 276
 277         if (likely(!(old_pte & _PAGE_HASHPTE))) {
 278
 279                 pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
 280                 hash = hpt_hash(vpn, shift, ssize);
 281
 282 repeat:
 283                 hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 284
 285                 /* Insert into the hash table, primary slot */
 286                 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, 0,
 287                                   MMU_PAGE_64K, MMU_PAGE_64K, ssize);
 288                 /*
 289                  * Primary is full, try the secondary
 290                  */
 291                 if (unlikely(slot == -1)) {
 292                         hpte_group = ((~hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 293                         slot = ppc_md.hpte_insert(hpte_group, vpn, pa,
 294                                                   rflags, HPTE_V_SECONDARY,
 295                                                   MMU_PAGE_64K, MMU_PAGE_64K, ssize);
 296                         if (slot == -1) {
 297                                 if (mftb() & 0x1)
 298                                         hpte_group = ((hash & htab_hash_mask) *
 299                                                       HPTES_PER_GROUP) & ~0x7UL;
 300                                 ppc_md.hpte_remove(hpte_group);
 301                                 /*
 302                                  * FIXME!! Should be try the group from which we removed ?
 303                                  */
 304                                 goto repeat;
 305                         }
 306                 }
 307                 /*
 308                  * Hypervisor failure. Restore old pmd and return -1
 309                  * similar to __hash_page_*
 310                  */
 311                 if (unlikely(slot == -2)) {
 312                         *ptep = __pte(old_pte);
 313                         hash_failure_debug(ea, access, vsid, trap, ssize,
 314                                            MMU_PAGE_64K, MMU_PAGE_64K, old_pte);
 315                         return -1;
 316                 }
 317                 new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HASHPTE;
 318                 new_pte |= (slot << _PAGE_F_GIX_SHIFT) & (_PAGE_F_SECOND | _PAGE_F_GIX);
 319         }
 320         *ptep = __pte(new_pte & ~_PAGE_BUSY);
 321         return 0;
 322 }