arch/x86_64/mm/pageattr-xen.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5
   6 #include <linux/config.h>
   7 #include <linux/mm.h>
   8 #include <linux/sched.h>
   9 #include <linux/highmem.h>
  10 #include <linux/module.h>
  11 #include <linux/slab.h>
  12 #include <asm/uaccess.h>
  13 #include <asm/processor.h>
  14 #include <asm/tlbflush.h>
  15 #include <asm/io.h>
  16
  17 #ifdef CONFIG_XEN
  18 #include <asm/pgalloc.h>
  19 #include <asm/mmu_context.h>
  20
  21 LIST_HEAD(mm_unpinned);
  22 DEFINE_SPINLOCK(mm_unpinned_lock);
  23
  24 static inline void mm_walk_set_prot(void *pt, pgprot_t flags)
  25 {
  26         struct page *page = virt_to_page(pt);
  27         unsigned long pfn = page_to_pfn(page);
  28
  29         BUG_ON(HYPERVISOR_update_va_mapping(
  30                        (unsigned long)__va(pfn << PAGE_SHIFT),
  31                        pfn_pte(pfn, flags), 0));
  32 }
  33
  34 static void mm_walk(struct mm_struct *mm, pgprot_t flags)
  35 {
  36         pgd_t       *pgd;
  37         pud_t       *pud;
  38         pmd_t       *pmd;
  39         pte_t       *pte;
  40         int          g,u,m;
  41
  42         pgd = mm->pgd;
  43         /*
  44          * Cannot iterate up to USER_PTRS_PER_PGD as these pagetables may not
  45          * be the 'current' task's pagetables (e.g., current may be 32-bit,
  46          * but the pagetables may be for a 64-bit task).
  47          * Subtracting 1 from TASK_SIZE64 means the loop limit is correct
  48          * regardless of whether TASK_SIZE64 is a multiple of PGDIR_SIZE.
  49          */
  50         for (g = 0; g <= ((TASK_SIZE64-1) / PGDIR_SIZE); g++, pgd++) {
  51                 if (pgd_none(*pgd))
  52                         continue;
  53                 pud = pud_offset(pgd, 0);
  54                 if (PTRS_PER_PUD > 1) /* not folded */
  55                         mm_walk_set_prot(pud,flags);
  56                 for (u = 0; u < PTRS_PER_PUD; u++, pud++) {
  57                         if (pud_none(*pud))
  58                                 continue;
  59                         pmd = pmd_offset(pud, 0);
  60                         if (PTRS_PER_PMD > 1) /* not folded */
  61                                 mm_walk_set_prot(pmd,flags);
  62                         for (m = 0; m < PTRS_PER_PMD; m++, pmd++) {
  63                                 if (pmd_none(*pmd))
  64                                         continue;
  65                                 pte = pte_offset_kernel(pmd,0);
  66                                 mm_walk_set_prot(pte,flags);
  67                         }
  68                 }
  69         }
  70 }
  71
  72 void mm_pin(struct mm_struct *mm)
  73 {
  74         if (xen_feature(XENFEAT_writable_page_tables))
  75                 return;
  76
  77         spin_lock(&mm->page_table_lock);
  78
  79         mm_walk(mm, PAGE_KERNEL_RO);
  80         BUG_ON(HYPERVISOR_update_va_mapping(
  81                        (unsigned long)mm->pgd,
  82                        pfn_pte(virt_to_phys(mm->pgd)>>PAGE_SHIFT, PAGE_KERNEL_RO),
  83                        UVMF_TLB_FLUSH));
  84         BUG_ON(HYPERVISOR_update_va_mapping(
  85                        (unsigned long)__user_pgd(mm->pgd),
  86                        pfn_pte(virt_to_phys(__user_pgd(mm->pgd))>>PAGE_SHIFT, PAGE_KERNEL_RO),
  87                        UVMF_TLB_FLUSH));
  88         xen_pgd_pin(__pa(mm->pgd)); /* kernel */
  89         xen_pgd_pin(__pa(__user_pgd(mm->pgd))); /* user */
  90         mm->context.pinned = 1;
  91         spin_lock(&mm_unpinned_lock);
  92         list_del(&mm->context.unpinned);
  93         spin_unlock(&mm_unpinned_lock);
  94
  95         spin_unlock(&mm->page_table_lock);
  96 }
  97
  98 void mm_unpin(struct mm_struct *mm)
  99 {
 100         if (xen_feature(XENFEAT_writable_page_tables))
 101                 return;
 102
 103         spin_lock(&mm->page_table_lock);
 104
 105         xen_pgd_unpin(__pa(mm->pgd));
 106         xen_pgd_unpin(__pa(__user_pgd(mm->pgd)));
 107         BUG_ON(HYPERVISOR_update_va_mapping(
 108                        (unsigned long)mm->pgd,
 109                        pfn_pte(virt_to_phys(mm->pgd)>>PAGE_SHIFT, PAGE_KERNEL), 0));
 110         BUG_ON(HYPERVISOR_update_va_mapping(
 111                        (unsigned long)__user_pgd(mm->pgd),
 112                        pfn_pte(virt_to_phys(__user_pgd(mm->pgd))>>PAGE_SHIFT, PAGE_KERNEL), 0));
 113         mm_walk(mm, PAGE_KERNEL);
 114         xen_tlb_flush();
 115         mm->context.pinned = 0;
 116         spin_lock(&mm_unpinned_lock);
 117         list_add(&mm->context.unpinned, &mm_unpinned);
 118         spin_unlock(&mm_unpinned_lock);
 119
 120         spin_unlock(&mm->page_table_lock);
 121 }
 122
 123 void mm_pin_all(void)
 124 {
 125         if (xen_feature(XENFEAT_writable_page_tables))
 126                 return;
 127
 128         while (!list_empty(&mm_unpinned))
 129                 mm_pin(list_entry(mm_unpinned.next, struct mm_struct,
 130                                   context.unpinned));
 131 }
 132
 133 void _arch_dup_mmap(struct mm_struct *mm)
 134 {
 135     if (!mm->context.pinned)
 136         mm_pin(mm);
 137 }
 138
 139 void _arch_exit_mmap(struct mm_struct *mm)
 140 {
 141     struct task_struct *tsk = current;
 142
 143     task_lock(tsk);
 144
 145     /*
 146      * We aggressively remove defunct pgd from cr3. We execute unmap_vmas()
 147      * *much* faster this way, as no tlb flushes means bigger wrpt batches.
 148      */
 149     if ( tsk->active_mm == mm )
 150     {
 151         tsk->active_mm = &init_mm;
 152         atomic_inc(&init_mm.mm_count);
 153
 154         switch_mm(mm, &init_mm, tsk);
 155
 156         atomic_dec(&mm->mm_count);
 157         BUG_ON(atomic_read(&mm->mm_count) == 0);
 158     }
 159
 160     task_unlock(tsk);
 161
 162     if ( mm->context.pinned && (atomic_read(&mm->mm_count) == 1) &&
 163          !mm->context.has_foreign_mappings )
 164         mm_unpin(mm);
 165 }
 166
 167 void pte_free(struct page *pte)
 168 {
 169         unsigned long va = (unsigned long)__va(page_to_pfn(pte)<<PAGE_SHIFT);
 170
 171         if (!pte_write(*virt_to_ptep(va)))
 172                 BUG_ON(HYPERVISOR_update_va_mapping(
 173                         va, pfn_pte(page_to_pfn(pte), PAGE_KERNEL), 0));
 174         __free_page(pte);
 175 }
 176 #endif  /* CONFIG_XEN */
 177
 178 static inline pte_t *lookup_address(unsigned long address)
 179 {
 180         pgd_t *pgd = pgd_offset_k(address);
 181         pud_t *pud;
 182         pmd_t *pmd;
 183         pte_t *pte;
 184         if (pgd_none(*pgd))
 185                 return NULL;
 186         pud = pud_offset(pgd, address);
 187         if (!pud_present(*pud))
 188                 return NULL;
 189         pmd = pmd_offset(pud, address);
 190         if (!pmd_present(*pmd))
 191                 return NULL;
 192         if (pmd_large(*pmd))
 193                 return (pte_t *)pmd;
 194         pte = pte_offset_kernel(pmd, address);
 195         if (pte && !pte_present(*pte))
 196                 pte = NULL;
 197         return pte;
 198 }
 199
 200 static struct page *split_large_page(unsigned long address, pgprot_t prot,
 201                                      pgprot_t ref_prot)
 202 {
 203         int i;
 204         unsigned long addr;
 205         struct page *base = alloc_pages(GFP_KERNEL, 0);
 206         pte_t *pbase;
 207         if (!base)
 208                 return NULL;
 209         /*
 210          * page_private is used to track the number of entries in
 211          * the page table page have non standard attributes.
 212          */
 213         SetPagePrivate(base);
 214         page_private(base) = 0;
 215
 216         address = __pa(address);
 217         addr = address & LARGE_PAGE_MASK;
 218         pbase = (pte_t *)page_address(base);
 219         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
 220                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
 221                                    addr == address ? prot : ref_prot);
 222         }
 223         return base;
 224 }
 225
 226
 227 static void flush_kernel_map(void *address)
 228 {
 229         if (0 && address && cpu_has_clflush) {
 230                 /* is this worth it? */
 231                 int i;
 232                 for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
 233                         asm volatile("clflush (%0)" :: "r" (address + i));
 234         } else
 235                 asm volatile("wbinvd":::"memory");
 236         if (address)
 237                 __flush_tlb_one(address);
 238         else
 239                 __flush_tlb_all();
 240 }
 241
 242
 243 static inline void flush_map(unsigned long address)
 244 {
 245         on_each_cpu(flush_kernel_map, (void *)address, 1, 1);
 246 }
 247
 248 static struct page *deferred_pages; /* protected by init_mm.mmap_sem */
 249
 250 static inline void save_page(struct page *fpage)
 251 {
 252         fpage->lru.next = (struct list_head *)deferred_pages;
 253         deferred_pages = fpage;
 254 }
 255
 256 /*
 257  * No more special protections in this 2/4MB area - revert to a
 258  * large page again.
 259  */
 260 static void revert_page(unsigned long address, pgprot_t ref_prot)
 261 {
 262         pgd_t *pgd;
 263         pud_t *pud;
 264         pmd_t *pmd;
 265         pte_t large_pte;
 266
 267         pgd = pgd_offset_k(address);
 268         BUG_ON(pgd_none(*pgd));
 269         pud = pud_offset(pgd,address);
 270         BUG_ON(pud_none(*pud));
 271         pmd = pmd_offset(pud, address);
 272         BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 273         pgprot_val(ref_prot) |= _PAGE_PSE;
 274         large_pte = mk_pte_phys(__pa(address) & LARGE_PAGE_MASK, ref_prot);
 275         set_pte((pte_t *)pmd, large_pte);
 276 }
 277
 278 static int
 279 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot,
 280                                    pgprot_t ref_prot)
 281 {
 282         pte_t *kpte;
 283         struct page *kpte_page;
 284         unsigned kpte_flags;
 285         pgprot_t ref_prot2;
 286         kpte = lookup_address(address);
 287         if (!kpte) return 0;
 288         kpte_page = virt_to_page(((unsigned long)kpte) & PAGE_MASK);
 289         kpte_flags = pte_val(*kpte);
 290         if (pgprot_val(prot) != pgprot_val(ref_prot)) {
 291                 if ((kpte_flags & _PAGE_PSE) == 0) {
 292                         set_pte(kpte, pfn_pte(pfn, prot));
 293                 } else {
 294                         /*
 295                          * split_large_page will take the reference for this
 296                          * change_page_attr on the split page.
 297                          */
 298
 299                         struct page *split;
 300                         ref_prot2 = __pgprot(pgprot_val(pte_pgprot(*lookup_address(address))) & ~(1<<_PAGE_BIT_PSE));
 301
 302                         split = split_large_page(address, prot, ref_prot2);
 303                         if (!split)
 304                                 return -ENOMEM;
 305                         set_pte(kpte,mk_pte(split, ref_prot2));
 306                         kpte_page = split;
 307                 }
 308                 page_private(kpte_page)++;
 309         } else if ((kpte_flags & _PAGE_PSE) == 0) {
 310                 set_pte(kpte, pfn_pte(pfn, ref_prot));
 311                 BUG_ON(page_private(kpte_page) == 0);
 312                 page_private(kpte_page)--;
 313         } else
 314                 BUG();
 315
 316         /* on x86-64 the direct mapping set at boot is not using 4k pages */
 317         /*
 318          * ..., but the XEN guest kernels (currently) do:
 319          * If the pte was reserved, it means it was created at boot
 320          * time (not via split_large_page) and in turn we must not
 321          * replace it with a large page.
 322          */
 323 #ifndef CONFIG_XEN
 324         BUG_ON(PageReserved(kpte_page));
 325 #endif
 326         if (page_private(kpte_page) == 0) {
 327                 save_page(kpte_page);
 328                 revert_page(address, ref_prot);
 329         }
 330         return 0;
 331 }
 332
 333 /*
 334  * Change the page attributes of an page in the linear mapping.
 335  *
 336  * This should be used when a page is mapped with a different caching policy
 337  * than write-back somewhere - some CPUs do not like it when mappings with
 338  * different caching policies exist. This changes the page attributes of the
 339  * in kernel linear mapping too.
 340  *
 341  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 342  * This function only deals with the kernel linear map.
 343  *
 344  * Caller must call global_flush_tlb() after this.
 345  */
 346 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 347 {
 348         int err = 0;
 349         int i;
 350
 351         down_write(&init_mm.mmap_sem);
 352         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 353                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 354
 355                 err = __change_page_attr(address, pfn, prot, PAGE_KERNEL);
 356                 if (err)
 357                         break;
 358                 /* Handle kernel mapping too which aliases part of the
 359                  * lowmem */
 360                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 361                         unsigned long addr2;
 362                         pgprot_t prot2 = prot;
 363                         addr2 = __START_KERNEL_map + __pa(address);
 364                         pgprot_val(prot2) &= ~_PAGE_NX;
 365                         err = __change_page_attr(addr2, pfn, prot2, PAGE_KERNEL_EXEC);
 366                 }
 367         }
 368         up_write(&init_mm.mmap_sem);
 369         return err;
 370 }
 371
 372 /* Don't call this for MMIO areas that may not have a mem_map entry */
 373 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 374 {
 375         unsigned long addr = (unsigned long)page_address(page);
 376         return change_page_attr_addr(addr, numpages, prot);
 377 }
 378
 379 void global_flush_tlb(void)
 380 {
 381         struct page *dpage;
 382
 383         down_read(&init_mm.mmap_sem);
 384         dpage = xchg(&deferred_pages, NULL);
 385         up_read(&init_mm.mmap_sem);
 386
 387         flush_map((dpage && !dpage->lru.next) ? (unsigned long)page_address(dpage) : 0);
 388         while (dpage) {
 389                 struct page *tmp = dpage;
 390                 dpage = (struct page *)dpage->lru.next;
 391                 ClearPagePrivate(tmp);
 392                 __free_page(tmp);
 393         }
 394 }
 395
 396 EXPORT_SYMBOL(change_page_attr);
 397 EXPORT_SYMBOL(global_flush_tlb);