arch/powerpc/mm/mmu_context_book3s64.c

   1 /*
   2  *  MMU context allocation for 64-bit kernels.
   3  *
   4  *  Copyright (C) 2004 Anton Blanchard, IBM Corp. <anton@samba.org>
   5  *
   6  *  This program is free software; you can redistribute it and/or
   7  *  modify it under the terms of the GNU General Public License
   8  *  as published by the Free Software Foundation; either version
   9  *  2 of the License, or (at your option) any later version.
  10  *
  11  */
  12
  13 #include <linux/sched.h>
  14 #include <linux/kernel.h>
  15 #include <linux/errno.h>
  16 #include <linux/string.h>
  17 #include <linux/types.h>
  18 #include <linux/mm.h>
  19 #include <linux/pkeys.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/idr.h>
  22 #include <linux/export.h>
  23 #include <linux/gfp.h>
  24 #include <linux/slab.h>
  25
  26 #include <asm/mmu_context.h>
  27 #include <asm/pgalloc.h>
  28
  29 static DEFINE_SPINLOCK(mmu_context_lock);
  30 static DEFINE_IDA(mmu_context_ida);
  31
  32 static int alloc_context_id(int min_id, int max_id)
  33 {
  34         int index, err;
  35
  36 again:
  37         if (!ida_pre_get(&mmu_context_ida, GFP_KERNEL))
  38                 return -ENOMEM;
  39
  40         spin_lock(&mmu_context_lock);
  41         err = ida_get_new_above(&mmu_context_ida, min_id, &index);
  42         spin_unlock(&mmu_context_lock);
  43
  44         if (err == -EAGAIN)
  45                 goto again;
  46         else if (err)
  47                 return err;
  48
  49         if (index > max_id) {
  50                 spin_lock(&mmu_context_lock);
  51                 ida_remove(&mmu_context_ida, index);
  52                 spin_unlock(&mmu_context_lock);
  53                 return -ENOMEM;
  54         }
  55
  56         return index;
  57 }
  58
  59 void hash__reserve_context_id(int id)
  60 {
  61         int rc, result = 0;
  62
  63         do {
  64                 if (!ida_pre_get(&mmu_context_ida, GFP_KERNEL))
  65                         break;
  66
  67                 spin_lock(&mmu_context_lock);
  68                 rc = ida_get_new_above(&mmu_context_ida, id, &result);
  69                 spin_unlock(&mmu_context_lock);
  70         } while (rc == -EAGAIN);
  71
  72         WARN(result != id, "mmu: Failed to reserve context id %d (rc %d)\n", id, result);
  73 }
  74
  75 int hash__alloc_context_id(void)
  76 {
  77         unsigned long max;
  78
  79         if (mmu_has_feature(MMU_FTR_68_BIT_VA))
  80                 max = MAX_USER_CONTEXT;
  81         else
  82                 max = MAX_USER_CONTEXT_65BIT_VA;
  83
  84         return alloc_context_id(MIN_USER_CONTEXT, max);
  85 }
  86 EXPORT_SYMBOL_GPL(hash__alloc_context_id);
  87
  88 static int hash__init_new_context(struct mm_struct *mm)
  89 {
  90         int index;
  91
  92         index = hash__alloc_context_id();
  93         if (index < 0)
  94                 return index;
  95
  96         /*
  97          * The old code would re-promote on fork, we don't do that when using
  98          * slices as it could cause problem promoting slices that have been
  99          * forced down to 4K.
 100          *
 101          * For book3s we have MMU_NO_CONTEXT set to be ~0. Hence check
 102          * explicitly against context.id == 0. This ensures that we properly
 103          * initialize context slice details for newly allocated mm's (which will
 104          * have id == 0) and don't alter context slice inherited via fork (which
 105          * will have id != 0).
 106          *
 107          * We should not be calling init_new_context() on init_mm. Hence a
 108          * check against 0 is OK.
 109          */
 110         if (mm->context.id == 0)
 111                 slice_init_new_context_exec(mm);
 112
 113         subpage_prot_init_new_context(mm);
 114
 115         pkey_mm_init(mm);
 116         return index;
 117 }
 118
 119 static int radix__init_new_context(struct mm_struct *mm)
 120 {
 121         unsigned long rts_field;
 122         int index, max_id;
 123
 124         max_id = (1 << mmu_pid_bits) - 1;
 125         index = alloc_context_id(mmu_base_pid, max_id);
 126         if (index < 0)
 127                 return index;
 128
 129         /*
 130          * set the process table entry,
 131          */
 132         rts_field = radix__get_tree_size();
 133         process_tb[index].prtb0 = cpu_to_be64(rts_field | __pa(mm->pgd) | RADIX_PGD_INDEX_SIZE);
 134
 135         /*
 136          * Order the above store with subsequent update of the PID
 137          * register (at which point HW can start loading/caching
 138          * the entry) and the corresponding load by the MMU from
 139          * the L2 cache.
 140          */
 141         asm volatile("ptesync;isync" : : : "memory");
 142
 143         mm->context.npu_context = NULL;
 144
 145         return index;
 146 }
 147
 148 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 149 {
 150         int index;
 151
 152         if (radix_enabled())
 153                 index = radix__init_new_context(mm);
 154         else
 155                 index = hash__init_new_context(mm);
 156
 157         if (index < 0)
 158                 return index;
 159
 160         mm->context.id = index;
 161
 162         mm->context.pte_frag = NULL;
 163         mm->context.pmd_frag = NULL;
 164 #ifdef CONFIG_SPAPR_TCE_IOMMU
 165         mm_iommu_init(mm);
 166 #endif
 167         atomic_set(&mm->context.active_cpus, 0);
 168         atomic_set(&mm->context.copros, 0);
 169
 170         return 0;
 171 }
 172
 173 void __destroy_context(int context_id)
 174 {
 175         spin_lock(&mmu_context_lock);
 176         ida_remove(&mmu_context_ida, context_id);
 177         spin_unlock(&mmu_context_lock);
 178 }
 179 EXPORT_SYMBOL_GPL(__destroy_context);
 180
 181 static void destroy_contexts(mm_context_t *ctx)
 182 {
 183         int index, context_id;
 184
 185         spin_lock(&mmu_context_lock);
 186         for (index = 0; index < ARRAY_SIZE(ctx->extended_id); index++) {
 187                 context_id = ctx->extended_id[index];
 188                 if (context_id)
 189                         ida_remove(&mmu_context_ida, context_id);
 190         }
 191         spin_unlock(&mmu_context_lock);
 192 }
 193
 194 static void pte_frag_destroy(void *pte_frag)
 195 {
 196         int count;
 197         struct page *page;
 198
 199         page = virt_to_page(pte_frag);
 200         /* drop all the pending references */
 201         count = ((unsigned long)pte_frag & ~PAGE_MASK) >> PTE_FRAG_SIZE_SHIFT;
 202         /* We allow PTE_FRAG_NR fragments from a PTE page */
 203         if (atomic_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
 204                 pgtable_page_dtor(page);
 205                 __free_page(page);
 206         }
 207 }
 208
 209 static void pmd_frag_destroy(void *pmd_frag)
 210 {
 211         int count;
 212         struct page *page;
 213
 214         page = virt_to_page(pmd_frag);
 215         /* drop all the pending references */
 216         count = ((unsigned long)pmd_frag & ~PAGE_MASK) >> PMD_FRAG_SIZE_SHIFT;
 217         /* We allow PTE_FRAG_NR fragments from a PTE page */
 218         if (atomic_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
 219                 pgtable_pmd_page_dtor(page);
 220                 __free_page(page);
 221         }
 222 }
 223
 224 static void destroy_pagetable_page(struct mm_struct *mm)
 225 {
 226         void *frag;
 227
 228         frag = mm->context.pte_frag;
 229         if (frag)
 230                 pte_frag_destroy(frag);
 231
 232         frag = mm->context.pmd_frag;
 233         if (frag)
 234                 pmd_frag_destroy(frag);
 235         return;
 236 }
 237
 238 void destroy_context(struct mm_struct *mm)
 239 {
 240 #ifdef CONFIG_SPAPR_TCE_IOMMU
 241         WARN_ON_ONCE(!list_empty(&mm->context.iommu_group_mem_list));
 242 #endif
 243         if (radix_enabled())
 244                 WARN_ON(process_tb[mm->context.id].prtb0 != 0);
 245         else
 246                 subpage_prot_free(mm);
 247         destroy_pagetable_page(mm);
 248         destroy_contexts(&mm->context);
 249         mm->context.id = MMU_NO_CONTEXT;
 250 }
 251
 252 void arch_exit_mmap(struct mm_struct *mm)
 253 {
 254         if (radix_enabled()) {
 255                 /*
 256                  * Radix doesn't have a valid bit in the process table
 257                  * entries. However we know that at least P9 implementation
 258                  * will avoid caching an entry with an invalid RTS field,
 259                  * and 0 is invalid. So this will do.
 260                  *
 261                  * This runs before the "fullmm" tlb flush in exit_mmap,
 262                  * which does a RIC=2 tlbie to clear the process table
 263                  * entry. See the "fullmm" comments in tlb-radix.c.
 264                  *
 265                  * No barrier required here after the store because
 266                  * this process will do the invalidate, which starts with
 267                  * ptesync.
 268                  */
 269                 process_tb[mm->context.id].prtb0 = 0;
 270         }
 271 }
 272
 273 #ifdef CONFIG_PPC_RADIX_MMU
 274 void radix__switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
 275 {
 276
 277         if (cpu_has_feature(CPU_FTR_POWER9_DD1)) {
 278                 isync();
 279                 mtspr(SPRN_PID, next->context.id);
 280                 isync();
 281                 asm volatile(PPC_INVALIDATE_ERAT : : :"memory");
 282         } else {
 283                 mtspr(SPRN_PID, next->context.id);
 284                 isync();
 285         }
 286 }
 287 #endif