arch/s390/mm/pgalloc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  Page table allocation functions
   4  *
   5  *    Copyright IBM Corp. 2016
   6  *    Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
   7  */
   8
   9 #include <linux/mm.h>
  10 #include <linux/sysctl.h>
  11 #include <asm/mmu_context.h>
  12 #include <asm/pgalloc.h>
  13 #include <asm/gmap.h>
  14 #include <asm/tlb.h>
  15 #include <asm/tlbflush.h>
  16
  17 #ifdef CONFIG_PGSTE
  18
  19 static int page_table_allocate_pgste_min = 0;
  20 static int page_table_allocate_pgste_max = 1;
  21 int page_table_allocate_pgste = 0;
  22 EXPORT_SYMBOL(page_table_allocate_pgste);
  23
  24 static struct ctl_table page_table_sysctl[] = {
  25         {
  26                 .procname       = "allocate_pgste",
  27                 .data           = &page_table_allocate_pgste,
  28                 .maxlen         = sizeof(int),
  29                 .mode           = S_IRUGO | S_IWUSR,
  30                 .proc_handler   = proc_dointvec,
  31                 .extra1         = &page_table_allocate_pgste_min,
  32                 .extra2         = &page_table_allocate_pgste_max,
  33         },
  34         { }
  35 };
  36
  37 static struct ctl_table page_table_sysctl_dir[] = {
  38         {
  39                 .procname       = "vm",
  40                 .maxlen         = 0,
  41                 .mode           = 0555,
  42                 .child          = page_table_sysctl,
  43         },
  44         { }
  45 };
  46
  47 static int __init page_table_register_sysctl(void)
  48 {
  49         return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
  50 }
  51 __initcall(page_table_register_sysctl);
  52
  53 #endif /* CONFIG_PGSTE */
  54
  55 unsigned long *crst_table_alloc(struct mm_struct *mm)
  56 {
  57         struct page *page = alloc_pages(GFP_KERNEL, 2);
  58
  59         if (!page)
  60                 return NULL;
  61         arch_set_page_dat(page, 2);
  62         return (unsigned long *) page_to_phys(page);
  63 }
  64
  65 void crst_table_free(struct mm_struct *mm, unsigned long *table)
  66 {
  67         free_pages((unsigned long) table, 2);
  68 }
  69
  70 static void __crst_table_upgrade(void *arg)
  71 {
  72         struct mm_struct *mm = arg;
  73
  74         if (current->active_mm == mm) {
  75                 clear_user_asce();
  76                 set_user_asce(mm);
  77         }
  78         __tlb_flush_local();
  79 }
  80
  81 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
  82 {
  83         unsigned long *table, *pgd;
  84         int rc, notify;
  85
  86         /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
  87         VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE);
  88         if (end >= TASK_SIZE_MAX)
  89                 return -ENOMEM;
  90         rc = 0;
  91         notify = 0;
  92         while (mm->context.asce_limit < end) {
  93                 table = crst_table_alloc(mm);
  94                 if (!table) {
  95                         rc = -ENOMEM;
  96                         break;
  97                 }
  98                 spin_lock_bh(&mm->page_table_lock);
  99                 pgd = (unsigned long *) mm->pgd;
 100                 if (mm->context.asce_limit == _REGION2_SIZE) {
 101                         crst_table_init(table, _REGION2_ENTRY_EMPTY);
 102                         p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd);
 103                         mm->pgd = (pgd_t *) table;
 104                         mm->context.asce_limit = _REGION1_SIZE;
 105                         mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
 106                                 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
 107                 } else {
 108                         crst_table_init(table, _REGION1_ENTRY_EMPTY);
 109                         pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
 110                         mm->pgd = (pgd_t *) table;
 111                         mm->context.asce_limit = -PAGE_SIZE;
 112                         mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
 113                                 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
 114                 }
 115                 notify = 1;
 116                 spin_unlock_bh(&mm->page_table_lock);
 117         }
 118         if (notify)
 119                 on_each_cpu(__crst_table_upgrade, mm, 0);
 120         return rc;
 121 }
 122
 123 void crst_table_downgrade(struct mm_struct *mm)
 124 {
 125         pgd_t *pgd;
 126
 127         /* downgrade should only happen from 3 to 2 levels (compat only) */
 128         VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE);
 129
 130         if (current->active_mm == mm) {
 131                 clear_user_asce();
 132                 __tlb_flush_mm(mm);
 133         }
 134
 135         pgd = mm->pgd;
 136         mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
 137         mm->context.asce_limit = _REGION3_SIZE;
 138         mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
 139                            _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
 140         crst_table_free(mm, (unsigned long *) pgd);
 141
 142         if (current->active_mm == mm)
 143                 set_user_asce(mm);
 144 }
 145
 146 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
 147 {
 148         unsigned int old, new;
 149
 150         do {
 151                 old = atomic_read(v);
 152                 new = old ^ bits;
 153         } while (atomic_cmpxchg(v, old, new) != old);
 154         return new;
 155 }
 156
 157 #ifdef CONFIG_PGSTE
 158
 159 struct page *page_table_alloc_pgste(struct mm_struct *mm)
 160 {
 161         struct page *page;
 162         unsigned long *table;
 163
 164         page = alloc_page(GFP_KERNEL);
 165         if (page) {
 166                 table = (unsigned long *) page_to_phys(page);
 167                 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2);
 168                 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2);
 169         }
 170         return page;
 171 }
 172
 173 void page_table_free_pgste(struct page *page)
 174 {
 175         __free_page(page);
 176 }
 177
 178 #endif /* CONFIG_PGSTE */
 179
 180 /*
 181  * page table entry allocation/free routines.
 182  */
 183 unsigned long *page_table_alloc(struct mm_struct *mm)
 184 {
 185         unsigned long *table;
 186         struct page *page;
 187         unsigned int mask, bit;
 188
 189         /* Try to get a fragment of a 4K page as a 2K page table */
 190         if (!mm_alloc_pgste(mm)) {
 191                 table = NULL;
 192                 spin_lock_bh(&mm->context.lock);
 193                 if (!list_empty(&mm->context.pgtable_list)) {
 194                         page = list_first_entry(&mm->context.pgtable_list,
 195                                                 struct page, lru);
 196                         mask = atomic_read(&page->_mapcount);
 197                         mask = (mask | (mask >> 4)) & 3;
 198                         if (mask != 3) {
 199                                 table = (unsigned long *) page_to_phys(page);
 200                                 bit = mask & 1;         /* =1 -> second 2K */
 201                                 if (bit)
 202                                         table += PTRS_PER_PTE;
 203                                 atomic_xor_bits(&page->_mapcount, 1U << bit);
 204                                 list_del(&page->lru);
 205                         }
 206                 }
 207                 spin_unlock_bh(&mm->context.lock);
 208                 if (table)
 209                         return table;
 210         }
 211         /* Allocate a fresh page */
 212         page = alloc_page(GFP_KERNEL);
 213         if (!page)
 214                 return NULL;
 215         if (!pgtable_page_ctor(page)) {
 216                 __free_page(page);
 217                 return NULL;
 218         }
 219         arch_set_page_dat(page, 0);
 220         /* Initialize page table */
 221         table = (unsigned long *) page_to_phys(page);
 222         if (mm_alloc_pgste(mm)) {
 223                 /* Return 4K page table with PGSTEs */
 224                 atomic_set(&page->_mapcount, 3);
 225                 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2);
 226                 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2);
 227         } else {
 228                 /* Return the first 2K fragment of the page */
 229                 atomic_set(&page->_mapcount, 1);
 230                 clear_table(table, _PAGE_INVALID, PAGE_SIZE);
 231                 spin_lock_bh(&mm->context.lock);
 232                 list_add(&page->lru, &mm->context.pgtable_list);
 233                 spin_unlock_bh(&mm->context.lock);
 234         }
 235         return table;
 236 }
 237
 238 void page_table_free(struct mm_struct *mm, unsigned long *table)
 239 {
 240         struct page *page;
 241         unsigned int bit, mask;
 242
 243         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 244         if (!mm_alloc_pgste(mm)) {
 245                 /* Free 2K page table fragment of a 4K page */
 246                 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
 247                 spin_lock_bh(&mm->context.lock);
 248                 mask = atomic_xor_bits(&page->_mapcount, 1U << bit);
 249                 if (mask & 3)
 250                         list_add(&page->lru, &mm->context.pgtable_list);
 251                 else
 252                         list_del(&page->lru);
 253                 spin_unlock_bh(&mm->context.lock);
 254                 if (mask != 0)
 255                         return;
 256         }
 257
 258         pgtable_page_dtor(page);
 259         atomic_set(&page->_mapcount, -1);
 260         __free_page(page);
 261 }
 262
 263 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
 264                          unsigned long vmaddr)
 265 {
 266         struct mm_struct *mm;
 267         struct page *page;
 268         unsigned int bit, mask;
 269
 270         mm = tlb->mm;
 271         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 272         if (mm_alloc_pgste(mm)) {
 273                 gmap_unlink(mm, table, vmaddr);
 274                 table = (unsigned long *) (__pa(table) | 3);
 275                 tlb_remove_table(tlb, table);
 276                 return;
 277         }
 278         bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
 279         spin_lock_bh(&mm->context.lock);
 280         mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit);
 281         if (mask & 3)
 282                 list_add_tail(&page->lru, &mm->context.pgtable_list);
 283         else
 284                 list_del(&page->lru);
 285         spin_unlock_bh(&mm->context.lock);
 286         table = (unsigned long *) (__pa(table) | (1U << bit));
 287         tlb_remove_table(tlb, table);
 288 }
 289
 290 static void __tlb_remove_table(void *_table)
 291 {
 292         unsigned int mask = (unsigned long) _table & 3;
 293         void *table = (void *)((unsigned long) _table ^ mask);
 294         struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 295
 296         switch (mask) {
 297         case 0:         /* pmd, pud, or p4d */
 298                 free_pages((unsigned long) table, 2);
 299                 break;
 300         case 1:         /* lower 2K of a 4K page table */
 301         case 2:         /* higher 2K of a 4K page table */
 302                 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0)
 303                         break;
 304                 /* fallthrough */
 305         case 3:         /* 4K page table with pgstes */
 306                 pgtable_page_dtor(page);
 307                 atomic_set(&page->_mapcount, -1);
 308                 __free_page(page);
 309                 break;
 310         }
 311 }
 312
 313 static void tlb_remove_table_smp_sync(void *arg)
 314 {
 315         /* Simply deliver the interrupt */
 316 }
 317
 318 static void tlb_remove_table_one(void *table)
 319 {
 320         /*
 321          * This isn't an RCU grace period and hence the page-tables cannot be
 322          * assumed to be actually RCU-freed.
 323          *
 324          * It is however sufficient for software page-table walkers that rely
 325          * on IRQ disabling. See the comment near struct mmu_table_batch.
 326          */
 327         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 328         __tlb_remove_table(table);
 329 }
 330
 331 static void tlb_remove_table_rcu(struct rcu_head *head)
 332 {
 333         struct mmu_table_batch *batch;
 334         int i;
 335
 336         batch = container_of(head, struct mmu_table_batch, rcu);
 337
 338         for (i = 0; i < batch->nr; i++)
 339                 __tlb_remove_table(batch->tables[i]);
 340
 341         free_page((unsigned long)batch);
 342 }
 343
 344 void tlb_table_flush(struct mmu_gather *tlb)
 345 {
 346         struct mmu_table_batch **batch = &tlb->batch;
 347
 348         if (*batch) {
 349                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 350                 *batch = NULL;
 351         }
 352 }
 353
 354 void tlb_remove_table(struct mmu_gather *tlb, void *table)
 355 {
 356         struct mmu_table_batch **batch = &tlb->batch;
 357
 358         tlb->mm->context.flush_mm = 1;
 359         if (*batch == NULL) {
 360                 *batch = (struct mmu_table_batch *)
 361                         __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 362                 if (*batch == NULL) {
 363                         __tlb_flush_mm_lazy(tlb->mm);
 364                         tlb_remove_table_one(table);
 365                         return;
 366                 }
 367                 (*batch)->nr = 0;
 368         }
 369         (*batch)->tables[(*batch)->nr++] = table;
 370         if ((*batch)->nr == MAX_TABLE_BATCH)
 371                 tlb_flush_mmu(tlb);
 372 }