arch/i386/mm/pgtable.c

   1 /*
   2  *  linux/arch/i386/mm/pgtable.c
   3  */
   4
   5 #include <linux/sched.h>
   6 #include <linux/kernel.h>
   7 #include <linux/errno.h>
   8 #include <linux/mm.h>
   9 #include <linux/swap.h>
  10 #include <linux/smp.h>
  11 #include <linux/highmem.h>
  12 #include <linux/slab.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/module.h>
  16 #include <linux/quicklist.h>
  17
  18 #include <asm/system.h>
  19 #include <asm/pgtable.h>
  20 #include <asm/pgalloc.h>
  21 #include <asm/fixmap.h>
  22 #include <asm/e820.h>
  23 #include <asm/tlb.h>
  24 #include <asm/tlbflush.h>
  25
  26 void show_mem(void)
  27 {
  28         int total = 0, reserved = 0;
  29         int shared = 0, cached = 0;
  30         int highmem = 0;
  31         struct page *page;
  32         pg_data_t *pgdat;
  33         unsigned long i;
  34         unsigned long flags;
  35
  36         printk(KERN_INFO "Mem-info:\n");
  37         show_free_areas();
  38         printk(KERN_INFO "Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
  39         for_each_online_pgdat(pgdat) {
  40                 pgdat_resize_lock(pgdat, &flags);
  41                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
  42                         page = pgdat_page_nr(pgdat, i);
  43                         total++;
  44                         if (PageHighMem(page))
  45                                 highmem++;
  46                         if (PageReserved(page))
  47                                 reserved++;
  48                         else if (PageSwapCache(page))
  49                                 cached++;
  50                         else if (page_count(page))
  51                                 shared += page_count(page) - 1;
  52                 }
  53                 pgdat_resize_unlock(pgdat, &flags);
  54         }
  55         printk(KERN_INFO "%d pages of RAM\n", total);
  56         printk(KERN_INFO "%d pages of HIGHMEM\n", highmem);
  57         printk(KERN_INFO "%d reserved pages\n", reserved);
  58         printk(KERN_INFO "%d pages shared\n", shared);
  59         printk(KERN_INFO "%d pages swap cached\n", cached);
  60
  61         printk(KERN_INFO "%lu pages dirty\n", global_page_state(NR_FILE_DIRTY));
  62         printk(KERN_INFO "%lu pages writeback\n",
  63                                         global_page_state(NR_WRITEBACK));
  64         printk(KERN_INFO "%lu pages mapped\n", global_page_state(NR_FILE_MAPPED));
  65         printk(KERN_INFO "%lu pages slab\n",
  66                 global_page_state(NR_SLAB_RECLAIMABLE) +
  67                 global_page_state(NR_SLAB_UNRECLAIMABLE));
  68         printk(KERN_INFO "%lu pages pagetables\n",
  69                                         global_page_state(NR_PAGETABLE));
  70 }
  71
  72 /*
  73  * Associate a virtual page frame with a given physical page frame
  74  * and protection flags for that frame.
  75  */
  76 static void set_pte_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
  77 {
  78         pgd_t *pgd;
  79         pud_t *pud;
  80         pmd_t *pmd;
  81         pte_t *pte;
  82
  83         pgd = swapper_pg_dir + pgd_index(vaddr);
  84         if (pgd_none(*pgd)) {
  85                 BUG();
  86                 return;
  87         }
  88         pud = pud_offset(pgd, vaddr);
  89         if (pud_none(*pud)) {
  90                 BUG();
  91                 return;
  92         }
  93         pmd = pmd_offset(pud, vaddr);
  94         if (pmd_none(*pmd)) {
  95                 BUG();
  96                 return;
  97         }
  98         pte = pte_offset_kernel(pmd, vaddr);
  99         if (pgprot_val(flags))
 100                 /* <pfn,flags> stored as-is, to permit clearing entries */
 101                 set_pte(pte, pfn_pte(pfn, flags));
 102         else
 103                 pte_clear(&init_mm, vaddr, pte);
 104
 105         /*
 106          * It's enough to flush this one mapping.
 107          * (PGE mappings get flushed as well)
 108          */
 109         __flush_tlb_one(vaddr);
 110 }
 111
 112 /*
 113  * Associate a large virtual page frame with a given physical page frame
 114  * and protection flags for that frame. pfn is for the base of the page,
 115  * vaddr is what the page gets mapped to - both must be properly aligned.
 116  * The pmd must already be instantiated. Assumes PAE mode.
 117  */
 118 void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
 119 {
 120         pgd_t *pgd;
 121         pud_t *pud;
 122         pmd_t *pmd;
 123
 124         if (vaddr & (PMD_SIZE-1)) {             /* vaddr is misaligned */
 125                 printk(KERN_WARNING "set_pmd_pfn: vaddr misaligned\n");
 126                 return; /* BUG(); */
 127         }
 128         if (pfn & (PTRS_PER_PTE-1)) {           /* pfn is misaligned */
 129                 printk(KERN_WARNING "set_pmd_pfn: pfn misaligned\n");
 130                 return; /* BUG(); */
 131         }
 132         pgd = swapper_pg_dir + pgd_index(vaddr);
 133         if (pgd_none(*pgd)) {
 134                 printk(KERN_WARNING "set_pmd_pfn: pgd_none\n");
 135                 return; /* BUG(); */
 136         }
 137         pud = pud_offset(pgd, vaddr);
 138         pmd = pmd_offset(pud, vaddr);
 139         set_pmd(pmd, pfn_pmd(pfn, flags));
 140         /*
 141          * It's enough to flush this one mapping.
 142          * (PGE mappings get flushed as well)
 143          */
 144         __flush_tlb_one(vaddr);
 145 }
 146
 147 static int fixmaps;
 148 unsigned long __FIXADDR_TOP = 0xfffff000;
 149 EXPORT_SYMBOL(__FIXADDR_TOP);
 150
 151 void __set_fixmap (enum fixed_addresses idx, unsigned long phys, pgprot_t flags)
 152 {
 153         unsigned long address = __fix_to_virt(idx);
 154
 155         if (idx >= __end_of_fixed_addresses) {
 156                 BUG();
 157                 return;
 158         }
 159         set_pte_pfn(address, phys >> PAGE_SHIFT, flags);
 160         fixmaps++;
 161 }
 162
 163 /**
 164  * reserve_top_address - reserves a hole in the top of kernel address space
 165  * @reserve - size of hole to reserve
 166  *
 167  * Can be used to relocate the fixmap area and poke a hole in the top
 168  * of kernel address space to make room for a hypervisor.
 169  */
 170 void reserve_top_address(unsigned long reserve)
 171 {
 172         BUG_ON(fixmaps > 0);
 173         printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
 174                (int)-reserve);
 175         __FIXADDR_TOP = -reserve - PAGE_SIZE;
 176         __VMALLOC_RESERVE += reserve;
 177 }
 178
 179 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
 180 {
 181         return (pte_t *)__get_free_page(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO);
 182 }
 183
 184 struct page *pte_alloc_one(struct mm_struct *mm, unsigned long address)
 185 {
 186         struct page *pte;
 187
 188 #ifdef CONFIG_HIGHPTE
 189         pte = alloc_pages(GFP_KERNEL|__GFP_HIGHMEM|__GFP_REPEAT|__GFP_ZERO, 0);
 190 #else
 191         pte = alloc_pages(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO, 0);
 192 #endif
 193         return pte;
 194 }
 195
 196 void pmd_ctor(void *pmd, struct kmem_cache *cache, unsigned long flags)
 197 {
 198         memset(pmd, 0, PTRS_PER_PMD*sizeof(pmd_t));
 199 }
 200
 201 /*
 202  * List of all pgd's needed for non-PAE so it can invalidate entries
 203  * in both cached and uncached pgd's; not needed for PAE since the
 204  * kernel pmd is shared. If PAE were not to share the pmd a similar
 205  * tactic would be needed. This is essentially codepath-based locking
 206  * against pageattr.c; it is the unique case in which a valid change
 207  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
 208  * vmalloc faults work because attached pagetables are never freed.
 209  * -- wli
 210  */
 211 DEFINE_SPINLOCK(pgd_lock);
 212 struct page *pgd_list;
 213
 214 static inline void pgd_list_add(pgd_t *pgd)
 215 {
 216         struct page *page = virt_to_page(pgd);
 217         page->index = (unsigned long)pgd_list;
 218         if (pgd_list)
 219                 set_page_private(pgd_list, (unsigned long)&page->index);
 220         pgd_list = page;
 221         set_page_private(page, (unsigned long)&pgd_list);
 222 }
 223
 224 static inline void pgd_list_del(pgd_t *pgd)
 225 {
 226         struct page *next, **pprev, *page = virt_to_page(pgd);
 227         next = (struct page *)page->index;
 228         pprev = (struct page **)page_private(page);
 229         *pprev = next;
 230         if (next)
 231                 set_page_private(next, (unsigned long)pprev);
 232 }
 233
 234
 235
 236 #if (PTRS_PER_PMD == 1)
 237 /* Non-PAE pgd constructor */
 238 static void pgd_ctor(void *pgd)
 239 {
 240         unsigned long flags;
 241
 242         /* !PAE, no pagetable sharing */
 243         memset(pgd, 0, USER_PTRS_PER_PGD*sizeof(pgd_t));
 244
 245         spin_lock_irqsave(&pgd_lock, flags);
 246
 247         /* must happen under lock */
 248         clone_pgd_range((pgd_t *)pgd + USER_PTRS_PER_PGD,
 249                         swapper_pg_dir + USER_PTRS_PER_PGD,
 250                         KERNEL_PGD_PTRS);
 251         paravirt_alloc_pd_clone(__pa(pgd) >> PAGE_SHIFT,
 252                                 __pa(swapper_pg_dir) >> PAGE_SHIFT,
 253                                 USER_PTRS_PER_PGD,
 254                                 KERNEL_PGD_PTRS);
 255         pgd_list_add(pgd);
 256         spin_unlock_irqrestore(&pgd_lock, flags);
 257 }
 258 #else  /* PTRS_PER_PMD > 1 */
 259 /* PAE pgd constructor */
 260 static void pgd_ctor(void *pgd)
 261 {
 262         /* PAE, kernel PMD may be shared */
 263
 264         if (SHARED_KERNEL_PMD) {
 265                 clone_pgd_range((pgd_t *)pgd + USER_PTRS_PER_PGD,
 266                                 swapper_pg_dir + USER_PTRS_PER_PGD,
 267                                 KERNEL_PGD_PTRS);
 268         } else {
 269                 unsigned long flags;
 270
 271                 memset(pgd, 0, USER_PTRS_PER_PGD*sizeof(pgd_t));
 272                 spin_lock_irqsave(&pgd_lock, flags);
 273                 pgd_list_add(pgd);
 274                 spin_unlock_irqrestore(&pgd_lock, flags);
 275         }
 276 }
 277 #endif  /* PTRS_PER_PMD */
 278
 279 static void pgd_dtor(void *pgd)
 280 {
 281         unsigned long flags; /* can be called from interrupt context */
 282
 283         if (SHARED_KERNEL_PMD)
 284                 return;
 285
 286         paravirt_release_pd(__pa(pgd) >> PAGE_SHIFT);
 287         spin_lock_irqsave(&pgd_lock, flags);
 288         pgd_list_del(pgd);
 289         spin_unlock_irqrestore(&pgd_lock, flags);
 290 }
 291
 292 #define UNSHARED_PTRS_PER_PGD                           \
 293         (SHARED_KERNEL_PMD ? USER_PTRS_PER_PGD : PTRS_PER_PGD)
 294
 295 /* If we allocate a pmd for part of the kernel address space, then
 296    make sure its initialized with the appropriate kernel mappings.
 297    Otherwise use a cached zeroed pmd.  */
 298 static pmd_t *pmd_cache_alloc(int idx)
 299 {
 300         pmd_t *pmd;
 301
 302         if (idx >= USER_PTRS_PER_PGD) {
 303                 pmd = (pmd_t *)__get_free_page(GFP_KERNEL);
 304
 305                 if (pmd)
 306                         memcpy(pmd,
 307                                (void *)pgd_page_vaddr(swapper_pg_dir[idx]),
 308                                sizeof(pmd_t) * PTRS_PER_PMD);
 309         } else
 310                 pmd = kmem_cache_alloc(pmd_cache, GFP_KERNEL);
 311
 312         return pmd;
 313 }
 314
 315 static void pmd_cache_free(pmd_t *pmd, int idx)
 316 {
 317         if (idx >= USER_PTRS_PER_PGD)
 318                 free_page((unsigned long)pmd);
 319         else
 320                 kmem_cache_free(pmd_cache, pmd);
 321 }
 322
 323 pgd_t *pgd_alloc(struct mm_struct *mm)
 324 {
 325         int i;
 326         pgd_t *pgd = quicklist_alloc(0, GFP_KERNEL, pgd_ctor);
 327
 328         if (PTRS_PER_PMD == 1 || !pgd)
 329                 return pgd;
 330
 331         for (i = 0; i < UNSHARED_PTRS_PER_PGD; ++i) {
 332                 pmd_t *pmd = pmd_cache_alloc(i);
 333
 334                 if (!pmd)
 335                         goto out_oom;
 336
 337                 paravirt_alloc_pd(__pa(pmd) >> PAGE_SHIFT);
 338                 set_pgd(&pgd[i], __pgd(1 + __pa(pmd)));
 339         }
 340         return pgd;
 341
 342 out_oom:
 343         for (i--; i >= 0; i--) {
 344                 pgd_t pgdent = pgd[i];
 345                 void* pmd = (void *)__va(pgd_val(pgdent)-1);
 346                 paravirt_release_pd(__pa(pmd) >> PAGE_SHIFT);
 347                 pmd_cache_free(pmd, i);
 348         }
 349         quicklist_free(0, pgd_dtor, pgd);
 350         return NULL;
 351 }
 352
 353 void pgd_free(pgd_t *pgd)
 354 {
 355         int i;
 356
 357         /* in the PAE case user pgd entries are overwritten before usage */
 358         if (PTRS_PER_PMD > 1)
 359                 for (i = 0; i < UNSHARED_PTRS_PER_PGD; ++i) {
 360                         pgd_t pgdent = pgd[i];
 361                         void* pmd = (void *)__va(pgd_val(pgdent)-1);
 362                         paravirt_release_pd(__pa(pmd) >> PAGE_SHIFT);
 363                         pmd_cache_free(pmd, i);
 364                 }
 365         /* in the non-PAE case, free_pgtables() clears user pgd entries */
 366         quicklist_free(0, pgd_dtor, pgd);
 367 }
 368
 369 void check_pgt_cache(void)
 370 {
 371         quicklist_trim(0, pgd_dtor, 25, 16);
 372 }
 373