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