arch/s390/mm/pgalloc.c

   1 /*
   2  *  Page table allocation functions
   3  *
   4  *    Copyright IBM Corp. 2016
   5  *    Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sysctl.h>
  10 #include <asm/mmu_context.h>
  11 #include <asm/pgalloc.h>
  12 #include <asm/gmap.h>
  13 #include <asm/tlb.h>
  14 #include <asm/tlbflush.h>
  15
  16 #ifdef CONFIG_PGSTE
  17
  18 static int page_table_allocate_pgste_min = 0;
  19 static int page_table_allocate_pgste_max = 1;
  20 int page_table_allocate_pgste = 0;
  21 EXPORT_SYMBOL(page_table_allocate_pgste);
  22
  23 static struct ctl_table page_table_sysctl[] = {
  24         {
  25                 .procname       = "allocate_pgste",
  26                 .data           = &page_table_allocate_pgste,
  27                 .maxlen         = sizeof(int),
  28                 .mode           = S_IRUGO | S_IWUSR,
  29                 .proc_handler   = proc_dointvec,
  30                 .extra1         = &page_table_allocate_pgste_min,
  31                 .extra2         = &page_table_allocate_pgste_max,
  32         },
  33         { }
  34 };
  35
  36 static struct ctl_table page_table_sysctl_dir[] = {
  37         {
  38                 .procname       = "vm",
  39                 .maxlen         = 0,
  40                 .mode           = 0555,
  41                 .child          = page_table_sysctl,
  42         },
  43         { }
  44 };
  45
  46 static int __init page_table_register_sysctl(void)
  47 {
  48         return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
  49 }
  50 __initcall(page_table_register_sysctl);
  51
  52 #endif /* CONFIG_PGSTE */
  53
  54 unsigned long *crst_table_alloc(struct mm_struct *mm)
  55 {
  56         struct page *page = alloc_pages(GFP_KERNEL, 2);
  57
  58         if (!page)
  59                 return NULL;
  60         return (unsigned long *) page_to_phys(page);
  61 }
  62
  63 void crst_table_free(struct mm_struct *mm, unsigned long *table)
  64 {
  65         free_pages((unsigned long) table, 2);
  66 }
  67
  68 static void __crst_table_upgrade(void *arg)
  69 {
  70         struct mm_struct *mm = arg;
  71
  72         if (current->active_mm == mm) {
  73                 clear_user_asce();
  74                 set_user_asce(mm);
  75         }
  76         __tlb_flush_local();
  77 }
  78
  79 int crst_table_upgrade(struct mm_struct *mm, unsigned long limit)
  80 {
  81         unsigned long *table, *pgd;
  82         unsigned long entry;
  83         int flush;
  84
  85         BUG_ON(limit > TASK_MAX_SIZE);
  86         flush = 0;
  87 repeat:
  88         table = crst_table_alloc(mm);
  89         if (!table)
  90                 return -ENOMEM;
  91         spin_lock_bh(&mm->page_table_lock);
  92         if (mm->context.asce_limit < limit) {
  93                 pgd = (unsigned long *) mm->pgd;
  94                 if (mm->context.asce_limit <= (1UL << 31)) {
  95                         entry = _REGION3_ENTRY_EMPTY;
  96                         mm->context.asce_limit = 1UL << 42;
  97                         mm->context.asce_bits = _ASCE_TABLE_LENGTH |
  98                                                 _ASCE_USER_BITS |
  99                                                 _ASCE_TYPE_REGION3;
 100                 } else {
 101                         entry = _REGION2_ENTRY_EMPTY;
 102                         mm->context.asce_limit = 1UL << 53;
 103                         mm->context.asce_bits = _ASCE_TABLE_LENGTH |
 104                                                 _ASCE_USER_BITS |
 105                                                 _ASCE_TYPE_REGION2;
 106                 }
 107                 crst_table_init(table, entry);
 108                 pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd);
 109                 mm->pgd = (pgd_t *) table;
 110                 mm->task_size = mm->context.asce_limit;
 111                 table = NULL;
 112                 flush = 1;
 113         }
 114         spin_unlock_bh(&mm->page_table_lock);
 115         if (table)
 116                 crst_table_free(mm, table);
 117         if (mm->context.asce_limit < limit)
 118                 goto repeat;
 119         if (flush)
 120                 on_each_cpu(__crst_table_upgrade, mm, 0);
 121         return 0;
 122 }
 123
 124 void crst_table_downgrade(struct mm_struct *mm, unsigned long limit)
 125 {
 126         pgd_t *pgd;
 127
 128         if (current->active_mm == mm) {
 129                 clear_user_asce();
 130                 __tlb_flush_mm(mm);
 131         }
 132         while (mm->context.asce_limit > limit) {
 133                 pgd = mm->pgd;
 134                 switch (pgd_val(*pgd) & _REGION_ENTRY_TYPE_MASK) {
 135                 case _REGION_ENTRY_TYPE_R2:
 136                         mm->context.asce_limit = 1UL << 42;
 137                         mm->context.asce_bits = _ASCE_TABLE_LENGTH |
 138                                                 _ASCE_USER_BITS |
 139                                                 _ASCE_TYPE_REGION3;
 140                         break;
 141                 case _REGION_ENTRY_TYPE_R3:
 142                         mm->context.asce_limit = 1UL << 31;
 143                         mm->context.asce_bits = _ASCE_TABLE_LENGTH |
 144                                                 _ASCE_USER_BITS |
 145                                                 _ASCE_TYPE_SEGMENT;
 146                         break;
 147                 default:
 148                         BUG();
 149                 }
 150                 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
 151                 mm->task_size = mm->context.asce_limit;
 152                 crst_table_free(mm, (unsigned long *) pgd);
 153         }
 154         if (current->active_mm == mm)
 155                 set_user_asce(mm);
 156 }
 157
 158 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
 159 {
 160         unsigned int old, new;
 161
 162         do {
 163                 old = atomic_read(v);
 164                 new = old ^ bits;
 165         } while (atomic_cmpxchg(v, old, new) != old);
 166         return new;
 167 }
 168
 169 /*
 170  * page table entry allocation/free routines.
 171  */
 172 unsigned long *page_table_alloc(struct mm_struct *mm)
 173 {
 174         unsigned long *table;
 175         struct page *page;
 176         unsigned int mask, bit;
 177
 178         /* Try to get a fragment of a 4K page as a 2K page table */
 179         if (!mm_alloc_pgste(mm)) {
 180                 table = NULL;
 181                 spin_lock_bh(&mm->context.list_lock);
 182                 if (!list_empty(&mm->context.pgtable_list)) {
 183                         page = list_first_entry(&mm->context.pgtable_list,
 184                                                 struct page, lru);
 185                         mask = atomic_read(&page->_mapcount);
 186                         mask = (mask | (mask >> 4)) & 3;
 187                         if (mask != 3) {
 188                                 table = (unsigned long *) page_to_phys(page);
 189                                 bit = mask & 1;         /* =1 -> second 2K */
 190                                 if (bit)
 191                                         table += PTRS_PER_PTE;
 192                                 atomic_xor_bits(&page->_mapcount, 1U << bit);
 193                                 list_del(&page->lru);
 194                         }
 195                 }
 196                 spin_unlock_bh(&mm->context.list_lock);
 197                 if (table)
 198                         return table;
 199         }
 200         /* Allocate a fresh page */
 201         page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
 202         if (!page)
 203                 return NULL;
 204         if (!pgtable_page_ctor(page)) {
 205                 __free_page(page);
 206                 return NULL;
 207         }
 208         /* Initialize page table */
 209         table = (unsigned long *) page_to_phys(page);
 210         if (mm_alloc_pgste(mm)) {
 211                 /* Return 4K page table with PGSTEs */
 212                 atomic_set(&page->_mapcount, 3);
 213                 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2);
 214                 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2);
 215         } else {
 216                 /* Return the first 2K fragment of the page */
 217                 atomic_set(&page->_mapcount, 1);
 218                 clear_table(table, _PAGE_INVALID, PAGE_SIZE);
 219                 spin_lock_bh(&mm->context.list_lock);
 220                 list_add(&page->lru, &mm->context.pgtable_list);
 221                 spin_unlock_bh(&mm->context.list_lock);
 222         }
 223         return table;
 224 }
 225
 226 void page_table_free(struct mm_struct *mm, unsigned long *table)
 227 {
 228         struct page *page;
 229         unsigned int bit, mask;
 230
 231         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 232         if (!mm_alloc_pgste(mm)) {
 233                 /* Free 2K page table fragment of a 4K page */
 234                 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
 235                 spin_lock_bh(&mm->context.list_lock);
 236                 mask = atomic_xor_bits(&page->_mapcount, 1U << bit);
 237                 if (mask & 3)
 238                         list_add(&page->lru, &mm->context.pgtable_list);
 239                 else
 240                         list_del(&page->lru);
 241                 spin_unlock_bh(&mm->context.list_lock);
 242                 if (mask != 0)
 243                         return;
 244         }
 245
 246         pgtable_page_dtor(page);
 247         atomic_set(&page->_mapcount, -1);
 248         __free_page(page);
 249 }
 250
 251 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
 252                          unsigned long vmaddr)
 253 {
 254         struct mm_struct *mm;
 255         struct page *page;
 256         unsigned int bit, mask;
 257
 258         mm = tlb->mm;
 259         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 260         if (mm_alloc_pgste(mm)) {
 261                 gmap_unlink(mm, table, vmaddr);
 262                 table = (unsigned long *) (__pa(table) | 3);
 263                 tlb_remove_table(tlb, table);
 264                 return;
 265         }
 266         bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
 267         spin_lock_bh(&mm->context.list_lock);
 268         mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit);
 269         if (mask & 3)
 270                 list_add_tail(&page->lru, &mm->context.pgtable_list);
 271         else
 272                 list_del(&page->lru);
 273         spin_unlock_bh(&mm->context.list_lock);
 274         table = (unsigned long *) (__pa(table) | (1U << bit));
 275         tlb_remove_table(tlb, table);
 276 }
 277
 278 static void __tlb_remove_table(void *_table)
 279 {
 280         unsigned int mask = (unsigned long) _table & 3;
 281         void *table = (void *)((unsigned long) _table ^ mask);
 282         struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 283
 284         switch (mask) {
 285         case 0:         /* pmd or pud */
 286                 free_pages((unsigned long) table, 2);
 287                 break;
 288         case 1:         /* lower 2K of a 4K page table */
 289         case 2:         /* higher 2K of a 4K page table */
 290                 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0)
 291                         break;
 292                 /* fallthrough */
 293         case 3:         /* 4K page table with pgstes */
 294                 pgtable_page_dtor(page);
 295                 atomic_set(&page->_mapcount, -1);
 296                 __free_page(page);
 297                 break;
 298         }
 299 }
 300
 301 static void tlb_remove_table_smp_sync(void *arg)
 302 {
 303         /* Simply deliver the interrupt */
 304 }
 305
 306 static void tlb_remove_table_one(void *table)
 307 {
 308         /*
 309          * This isn't an RCU grace period and hence the page-tables cannot be
 310          * assumed to be actually RCU-freed.
 311          *
 312          * It is however sufficient for software page-table walkers that rely
 313          * on IRQ disabling. See the comment near struct mmu_table_batch.
 314          */
 315         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 316         __tlb_remove_table(table);
 317 }
 318
 319 static void tlb_remove_table_rcu(struct rcu_head *head)
 320 {
 321         struct mmu_table_batch *batch;
 322         int i;
 323
 324         batch = container_of(head, struct mmu_table_batch, rcu);
 325
 326         for (i = 0; i < batch->nr; i++)
 327                 __tlb_remove_table(batch->tables[i]);
 328
 329         free_page((unsigned long)batch);
 330 }
 331
 332 void tlb_table_flush(struct mmu_gather *tlb)
 333 {
 334         struct mmu_table_batch **batch = &tlb->batch;
 335
 336         if (*batch) {
 337                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 338                 *batch = NULL;
 339         }
 340 }
 341
 342 void tlb_remove_table(struct mmu_gather *tlb, void *table)
 343 {
 344         struct mmu_table_batch **batch = &tlb->batch;
 345
 346         tlb->mm->context.flush_mm = 1;
 347         if (*batch == NULL) {
 348                 *batch = (struct mmu_table_batch *)
 349                         __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 350                 if (*batch == NULL) {
 351                         __tlb_flush_mm_lazy(tlb->mm);
 352                         tlb_remove_table_one(table);
 353                         return;
 354                 }
 355                 (*batch)->nr = 0;
 356         }
 357         (*batch)->tables[(*batch)->nr++] = table;
 358         if ((*batch)->nr == MAX_TABLE_BATCH)
 359                 tlb_flush_mmu(tlb);
 360 }