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_minmax,
  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                 set_user_asce(mm);
  76         __tlb_flush_local();
  77 }
  78
  79 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
  80 {
  81         unsigned long *table, *pgd;
  82         int rc, notify;
  83
  84         /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
  85         VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE);
  86         rc = 0;
  87         notify = 0;
  88         while (mm->context.asce_limit < end) {
  89                 table = crst_table_alloc(mm);
  90                 if (!table) {
  91                         rc = -ENOMEM;
  92                         break;
  93                 }
  94                 spin_lock_bh(&mm->page_table_lock);
  95                 pgd = (unsigned long *) mm->pgd;
  96                 if (mm->context.asce_limit == _REGION2_SIZE) {
  97                         crst_table_init(table, _REGION2_ENTRY_EMPTY);
  98                         p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd);
  99                         mm->pgd = (pgd_t *) table;
 100                         mm->context.asce_limit = _REGION1_SIZE;
 101                         mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
 102                                 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
 103                         mm_inc_nr_puds(mm);
 104                 } else {
 105                         crst_table_init(table, _REGION1_ENTRY_EMPTY);
 106                         pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
 107                         mm->pgd = (pgd_t *) table;
 108                         mm->context.asce_limit = -PAGE_SIZE;
 109                         mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
 110                                 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
 111                 }
 112                 notify = 1;
 113                 spin_unlock_bh(&mm->page_table_lock);
 114         }
 115         if (notify)
 116                 on_each_cpu(__crst_table_upgrade, mm, 0);
 117         return rc;
 118 }
 119
 120 void crst_table_downgrade(struct mm_struct *mm)
 121 {
 122         pgd_t *pgd;
 123
 124         /* downgrade should only happen from 3 to 2 levels (compat only) */
 125         VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE);
 126
 127         if (current->active_mm == mm) {
 128                 clear_user_asce();
 129                 __tlb_flush_mm(mm);
 130         }
 131
 132         pgd = mm->pgd;
 133         mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
 134         mm->context.asce_limit = _REGION3_SIZE;
 135         mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
 136                            _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
 137         crst_table_free(mm, (unsigned long *) pgd);
 138
 139         if (current->active_mm == mm)
 140                 set_user_asce(mm);
 141 }
 142
 143 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
 144 {
 145         unsigned int old, new;
 146
 147         do {
 148                 old = atomic_read(v);
 149                 new = old ^ bits;
 150         } while (atomic_cmpxchg(v, old, new) != old);
 151         return new;
 152 }
 153
 154 #ifdef CONFIG_PGSTE
 155
 156 struct page *page_table_alloc_pgste(struct mm_struct *mm)
 157 {
 158         struct page *page;
 159         u64 *table;
 160
 161         page = alloc_page(GFP_KERNEL);
 162         if (page) {
 163                 table = (u64 *)page_to_phys(page);
 164                 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
 165                 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
 166         }
 167         return page;
 168 }
 169
 170 void page_table_free_pgste(struct page *page)
 171 {
 172         __free_page(page);
 173 }
 174
 175 #endif /* CONFIG_PGSTE */
 176
 177 /*
 178  * page table entry allocation/free routines.
 179  */
 180 unsigned long *page_table_alloc(struct mm_struct *mm)
 181 {
 182         unsigned long *table;
 183         struct page *page;
 184         unsigned int mask, bit;
 185
 186         /* Try to get a fragment of a 4K page as a 2K page table */
 187         if (!mm_alloc_pgste(mm)) {
 188                 table = NULL;
 189                 spin_lock_bh(&mm->context.lock);
 190                 if (!list_empty(&mm->context.pgtable_list)) {
 191                         page = list_first_entry(&mm->context.pgtable_list,
 192                                                 struct page, lru);
 193                         mask = atomic_read(&page->_mapcount);
 194                         mask = (mask | (mask >> 4)) & 3;
 195                         if (mask != 3) {
 196                                 table = (unsigned long *) page_to_phys(page);
 197                                 bit = mask & 1;         /* =1 -> second 2K */
 198                                 if (bit)
 199                                         table += PTRS_PER_PTE;
 200                                 atomic_xor_bits(&page->_mapcount, 1U << bit);
 201                                 list_del(&page->lru);
 202                         }
 203                 }
 204                 spin_unlock_bh(&mm->context.lock);
 205                 if (table)
 206                         return table;
 207         }
 208         /* Allocate a fresh page */
 209         page = alloc_page(GFP_KERNEL);
 210         if (!page)
 211                 return NULL;
 212         if (!pgtable_page_ctor(page)) {
 213                 __free_page(page);
 214                 return NULL;
 215         }
 216         arch_set_page_dat(page, 0);
 217         /* Initialize page table */
 218         table = (unsigned long *) page_to_phys(page);
 219         if (mm_alloc_pgste(mm)) {
 220                 /* Return 4K page table with PGSTEs */
 221                 atomic_set(&page->_mapcount, 3);
 222                 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
 223                 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
 224         } else {
 225                 /* Return the first 2K fragment of the page */
 226                 atomic_set(&page->_mapcount, 1);
 227                 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
 228                 spin_lock_bh(&mm->context.lock);
 229                 list_add(&page->lru, &mm->context.pgtable_list);
 230                 spin_unlock_bh(&mm->context.lock);
 231         }
 232         return table;
 233 }
 234
 235 void page_table_free(struct mm_struct *mm, unsigned long *table)
 236 {
 237         struct page *page;
 238         unsigned int bit, mask;
 239
 240         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 241         if (!mm_alloc_pgste(mm)) {
 242                 /* Free 2K page table fragment of a 4K page */
 243                 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
 244                 spin_lock_bh(&mm->context.lock);
 245                 mask = atomic_xor_bits(&page->_mapcount, 1U << bit);
 246                 if (mask & 3)
 247                         list_add(&page->lru, &mm->context.pgtable_list);
 248                 else
 249                         list_del(&page->lru);
 250                 spin_unlock_bh(&mm->context.lock);
 251                 if (mask != 0)
 252                         return;
 253         }
 254
 255         pgtable_page_dtor(page);
 256         atomic_set(&page->_mapcount, -1);
 257         __free_page(page);
 258 }
 259
 260 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
 261                          unsigned long vmaddr)
 262 {
 263         struct mm_struct *mm;
 264         struct page *page;
 265         unsigned int bit, mask;
 266
 267         mm = tlb->mm;
 268         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 269         if (mm_alloc_pgste(mm)) {
 270                 gmap_unlink(mm, table, vmaddr);
 271                 table = (unsigned long *) (__pa(table) | 3);
 272                 tlb_remove_table(tlb, table);
 273                 return;
 274         }
 275         bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
 276         spin_lock_bh(&mm->context.lock);
 277         mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit);
 278         if (mask & 3)
 279                 list_add_tail(&page->lru, &mm->context.pgtable_list);
 280         else
 281                 list_del(&page->lru);
 282         spin_unlock_bh(&mm->context.lock);
 283         table = (unsigned long *) (__pa(table) | (1U << bit));
 284         tlb_remove_table(tlb, table);
 285 }
 286
 287 static void __tlb_remove_table(void *_table)
 288 {
 289         unsigned int mask = (unsigned long) _table & 3;
 290         void *table = (void *)((unsigned long) _table ^ mask);
 291         struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 292
 293         switch (mask) {
 294         case 0:         /* pmd, pud, or p4d */
 295                 free_pages((unsigned long) table, 2);
 296                 break;
 297         case 1:         /* lower 2K of a 4K page table */
 298         case 2:         /* higher 2K of a 4K page table */
 299                 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0)
 300                         break;
 301                 /* fallthrough */
 302         case 3:         /* 4K page table with pgstes */
 303                 pgtable_page_dtor(page);
 304                 atomic_set(&page->_mapcount, -1);
 305                 __free_page(page);
 306                 break;
 307         }
 308 }
 309
 310 static void tlb_remove_table_smp_sync(void *arg)
 311 {
 312         /* Simply deliver the interrupt */
 313 }
 314
 315 static void tlb_remove_table_one(void *table)
 316 {
 317         /*
 318          * This isn't an RCU grace period and hence the page-tables cannot be
 319          * assumed to be actually RCU-freed.
 320          *
 321          * It is however sufficient for software page-table walkers that rely
 322          * on IRQ disabling. See the comment near struct mmu_table_batch.
 323          */
 324         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 325         __tlb_remove_table(table);
 326 }
 327
 328 static void tlb_remove_table_rcu(struct rcu_head *head)
 329 {
 330         struct mmu_table_batch *batch;
 331         int i;
 332
 333         batch = container_of(head, struct mmu_table_batch, rcu);
 334
 335         for (i = 0; i < batch->nr; i++)
 336                 __tlb_remove_table(batch->tables[i]);
 337
 338         free_page((unsigned long)batch);
 339 }
 340
 341 void tlb_table_flush(struct mmu_gather *tlb)
 342 {
 343         struct mmu_table_batch **batch = &tlb->batch;
 344
 345         if (*batch) {
 346                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 347                 *batch = NULL;
 348         }
 349 }
 350
 351 void tlb_remove_table(struct mmu_gather *tlb, void *table)
 352 {
 353         struct mmu_table_batch **batch = &tlb->batch;
 354
 355         tlb->mm->context.flush_mm = 1;
 356         if (*batch == NULL) {
 357                 *batch = (struct mmu_table_batch *)
 358                         __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 359                 if (*batch == NULL) {
 360                         __tlb_flush_mm_lazy(tlb->mm);
 361                         tlb_remove_table_one(table);
 362                         return;
 363                 }
 364                 (*batch)->nr = 0;
 365         }
 366         (*batch)->tables[(*batch)->nr++] = table;
 367         if ((*batch)->nr == MAX_TABLE_BATCH)
 368                 tlb_flush_mmu(tlb);
 369 }