SAFE public projects git trees. - safe/jmp/linux-2.6/blob - arch/x86/mm/pgtable_32.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/nmi.h>
  10 #include <linux/swap.h>
  11 #include <linux/smp.h>
  12 #include <linux/highmem.h>
  13 #include <linux/slab.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/module.h>
  17 #include <linux/quicklist.h>
  18
  19 #include <asm/system.h>
  20 #include <asm/pgtable.h>
  21 #include <asm/pgalloc.h>
  22 #include <asm/fixmap.h>
  23 #include <asm/e820.h>
  24 #include <asm/tlb.h>
  25 #include <asm/tlbflush.h>
  26
  27 void show_mem(void)
  28 {
  29         int total = 0, reserved = 0;
  30         int shared = 0, cached = 0;
  31         int highmem = 0;
  32         struct page *page;
  33         pg_data_t *pgdat;
  34         unsigned long i;
  35         unsigned long flags;
  36
  37         printk(KERN_INFO "Mem-info:\n");
  38         show_free_areas();
  39         for_each_online_pgdat(pgdat) {
  40                 pgdat_resize_lock(pgdat, &flags);
  41                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
  42                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
  43                                 touch_nmi_watchdog();
  44                         page = pgdat_page_nr(pgdat, i);
  45                         total++;
  46                         if (PageHighMem(page))
  47                                 highmem++;
  48                         if (PageReserved(page))
  49                                 reserved++;
  50                         else if (PageSwapCache(page))
  51                                 cached++;
  52                         else if (page_count(page))
  53                                 shared += page_count(page) - 1;
  54                 }
  55                 pgdat_resize_unlock(pgdat, &flags);
  56         }
  57         printk(KERN_INFO "%d pages of RAM\n", total);
  58         printk(KERN_INFO "%d pages of HIGHMEM\n", highmem);
  59         printk(KERN_INFO "%d reserved pages\n", reserved);
  60         printk(KERN_INFO "%d pages shared\n", shared);
  61         printk(KERN_INFO "%d pages swap cached\n", cached);
  62
  63         printk(KERN_INFO "%lu pages dirty\n", global_page_state(NR_FILE_DIRTY));
  64         printk(KERN_INFO "%lu pages writeback\n",
  65                                         global_page_state(NR_WRITEBACK));
  66         printk(KERN_INFO "%lu pages mapped\n", global_page_state(NR_FILE_MAPPED));
  67         printk(KERN_INFO "%lu pages slab\n",
  68                 global_page_state(NR_SLAB_RECLAIMABLE) +
  69                 global_page_state(NR_SLAB_UNRECLAIMABLE));
  70         printk(KERN_INFO "%lu pages pagetables\n",
  71                                         global_page_state(NR_PAGETABLE));
  72 }
  73
  74 /*
  75  * Associate a virtual page frame with a given physical page frame
  76  * and protection flags for that frame.
  77  */
  78 static void set_pte_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
  79 {
  80         pgd_t *pgd;
  81         pud_t *pud;
  82         pmd_t *pmd;
  83         pte_t *pte;
  84
  85         pgd = swapper_pg_dir + pgd_index(vaddr);
  86         if (pgd_none(*pgd)) {
  87                 BUG();
  88                 return;
  89         }
  90         pud = pud_offset(pgd, vaddr);
  91         if (pud_none(*pud)) {
  92                 BUG();
  93                 return;
  94         }
  95         pmd = pmd_offset(pud, vaddr);
  96         if (pmd_none(*pmd)) {
  97                 BUG();
  98                 return;
  99         }
 100         pte = pte_offset_kernel(pmd, vaddr);
 101         if (pgprot_val(flags))
 102                 set_pte_present(&init_mm, vaddr, pte, pfn_pte(pfn, flags));
 103         else
 104                 pte_clear(&init_mm, vaddr, pte);
 105
 106         /*
 107          * It's enough to flush this one mapping.
 108          * (PGE mappings get flushed as well)
 109          */
 110         __flush_tlb_one(vaddr);
 111 }
 112
 113 /*
 114  * Associate a large virtual page frame with a given physical page frame
 115  * and protection flags for that frame. pfn is for the base of the page,
 116  * vaddr is what the page gets mapped to - both must be properly aligned.
 117  * The pmd must already be instantiated. Assumes PAE mode.
 118  */
 119 void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
 120 {
 121         pgd_t *pgd;
 122         pud_t *pud;
 123         pmd_t *pmd;
 124
 125         if (vaddr & (PMD_SIZE-1)) {             /* vaddr is misaligned */
 126                 printk(KERN_WARNING "set_pmd_pfn: vaddr misaligned\n");
 127                 return; /* BUG(); */
 128         }
 129         if (pfn & (PTRS_PER_PTE-1)) {           /* pfn is misaligned */
 130                 printk(KERN_WARNING "set_pmd_pfn: pfn misaligned\n");
 131                 return; /* BUG(); */
 132         }
 133         pgd = swapper_pg_dir + pgd_index(vaddr);
 134         if (pgd_none(*pgd)) {
 135                 printk(KERN_WARNING "set_pmd_pfn: pgd_none\n");
 136                 return; /* BUG(); */
 137         }
 138         pud = pud_offset(pgd, vaddr);
 139         pmd = pmd_offset(pud, vaddr);
 140         set_pmd(pmd, pfn_pmd(pfn, flags));
 141         /*
 142          * It's enough to flush this one mapping.
 143          * (PGE mappings get flushed as well)
 144          */
 145         __flush_tlb_one(vaddr);
 146 }
 147
 148 static int fixmaps;
 149 unsigned long __FIXADDR_TOP = 0xfffff000;
 150 EXPORT_SYMBOL(__FIXADDR_TOP);
 151
 152 void __set_fixmap (enum fixed_addresses idx, unsigned long phys, pgprot_t flags)
 153 {
 154         unsigned long address = __fix_to_virt(idx);
 155
 156         if (idx >= __end_of_fixed_addresses) {
 157                 BUG();
 158                 return;
 159         }
 160         set_pte_pfn(address, phys >> PAGE_SHIFT, flags);
 161         fixmaps++;
 162 }
 163
 164 /**
 165  * reserve_top_address - reserves a hole in the top of kernel address space
 166  * @reserve - size of hole to reserve
 167  *
 168  * Can be used to relocate the fixmap area and poke a hole in the top
 169  * of kernel address space to make room for a hypervisor.
 170  */
 171 void reserve_top_address(unsigned long reserve)
 172 {
 173         BUG_ON(fixmaps > 0);
 174         printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
 175                (int)-reserve);
 176         __FIXADDR_TOP = -reserve - PAGE_SIZE;
 177         __VMALLOC_RESERVE += reserve;
 178 }
 179
 180 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
 181 {
 182         return (pte_t *)__get_free_page(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO);
 183 }
 184
 185 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
 186 {
 187         struct page *pte;
 188
 189 #ifdef CONFIG_HIGHPTE
 190         pte = alloc_pages(GFP_KERNEL|__GFP_HIGHMEM|__GFP_REPEAT|__GFP_ZERO, 0);
 191 #else
 192         pte = alloc_pages(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO, 0);
 193 #endif
 194         if (pte)
 195                 pgtable_page_ctor(pte);
 196         return pte;
 197 }
 198
 199 /*
 200  * List of all pgd's needed for non-PAE so it can invalidate entries
 201  * in both cached and uncached pgd's; not needed for PAE since the
 202  * kernel pmd is shared. If PAE were not to share the pmd a similar
 203  * tactic would be needed. This is essentially codepath-based locking
 204  * against pageattr.c; it is the unique case in which a valid change
 205  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
 206  * vmalloc faults work because attached pagetables are never freed.
 207  * -- wli
 208  */
 209 static inline void pgd_list_add(pgd_t *pgd)
 210 {
 211         struct page *page = virt_to_page(pgd);
 212
 213         list_add(&page->lru, &pgd_list);
 214 }
 215
 216 static inline void pgd_list_del(pgd_t *pgd)
 217 {
 218         struct page *page = virt_to_page(pgd);
 219
 220         list_del(&page->lru);
 221 }
 222
 223 #define UNSHARED_PTRS_PER_PGD                           \
 224         (SHARED_KERNEL_PMD ? USER_PTRS_PER_PGD : PTRS_PER_PGD)
 225
 226 static void pgd_ctor(void *p)
 227 {
 228         pgd_t *pgd = p;
 229         unsigned long flags;
 230
 231         /* Clear usermode parts of PGD */
 232         memset(pgd, 0, USER_PTRS_PER_PGD*sizeof(pgd_t));
 233
 234         spin_lock_irqsave(&pgd_lock, flags);
 235
 236         /* If the pgd points to a shared pagetable level (either the
 237            ptes in non-PAE, or shared PMD in PAE), then just copy the
 238            references from swapper_pg_dir. */
 239         if (PAGETABLE_LEVELS == 2 ||
 240             (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD)) {
 241                 clone_pgd_range(pgd + USER_PTRS_PER_PGD,
 242                                 swapper_pg_dir + USER_PTRS_PER_PGD,
 243                                 KERNEL_PGD_PTRS);
 244                 paravirt_alloc_pd_clone(__pa(pgd) >> PAGE_SHIFT,
 245                                         __pa(swapper_pg_dir) >> PAGE_SHIFT,
 246                                         USER_PTRS_PER_PGD,
 247                                         KERNEL_PGD_PTRS);
 248         }
 249
 250         /* list required to sync kernel mapping updates */
 251         if (!SHARED_KERNEL_PMD)
 252                 pgd_list_add(pgd);
 253
 254         spin_unlock_irqrestore(&pgd_lock, flags);
 255 }
 256
 257 static void pgd_dtor(void *pgd)
 258 {
 259         unsigned long flags; /* can be called from interrupt context */
 260
 261         if (SHARED_KERNEL_PMD)
 262                 return;
 263
 264         spin_lock_irqsave(&pgd_lock, flags);
 265         pgd_list_del(pgd);
 266         spin_unlock_irqrestore(&pgd_lock, flags);
 267 }
 268
 269 #ifdef CONFIG_X86_PAE
 270 /*
 271  * Mop up any pmd pages which may still be attached to the pgd.
 272  * Normally they will be freed by munmap/exit_mmap, but any pmd we
 273  * preallocate which never got a corresponding vma will need to be
 274  * freed manually.
 275  */
 276 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
 277 {
 278         int i;
 279
 280         for(i = 0; i < UNSHARED_PTRS_PER_PGD; i++) {
 281                 pgd_t pgd = pgdp[i];
 282
 283                 if (pgd_val(pgd) != 0) {
 284                         pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
 285
 286                         pgdp[i] = native_make_pgd(0);
 287
 288                         paravirt_release_pd(pgd_val(pgd) >> PAGE_SHIFT);
 289                         pmd_free(mm, pmd);
 290                 }
 291         }
 292 }
 293
 294 /*
 295  * In PAE mode, we need to do a cr3 reload (=tlb flush) when
 296  * updating the top-level pagetable entries to guarantee the
 297  * processor notices the update.  Since this is expensive, and
 298  * all 4 top-level entries are used almost immediately in a
 299  * new process's life, we just pre-populate them here.
 300  *
 301  * Also, if we're in a paravirt environment where the kernel pmd is
 302  * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
 303  * and initialize the kernel pmds here.
 304  */
 305 static int pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd)
 306 {
 307         pud_t *pud;
 308         unsigned long addr;
 309         int i;
 310
 311         pud = pud_offset(pgd, 0);
 312         for (addr = i = 0; i < UNSHARED_PTRS_PER_PGD;
 313              i++, pud++, addr += PUD_SIZE) {
 314                 pmd_t *pmd = pmd_alloc_one(mm, addr);
 315
 316                 if (!pmd) {
 317                         pgd_mop_up_pmds(mm, pgd);
 318                         return 0;
 319                 }
 320
 321                 if (i >= USER_PTRS_PER_PGD)
 322                         memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
 323                                sizeof(pmd_t) * PTRS_PER_PMD);
 324
 325                 pud_populate(mm, pud, pmd);
 326         }
 327
 328         return 1;
 329 }
 330 #else  /* !CONFIG_X86_PAE */
 331 /* No need to prepopulate any pagetable entries in non-PAE modes. */
 332 static int pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd)
 333 {
 334         return 1;
 335 }
 336
 337 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
 338 {
 339 }
 340 #endif  /* CONFIG_X86_PAE */
 341
 342 pgd_t *pgd_alloc(struct mm_struct *mm)
 343 {
 344         pgd_t *pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
 345
 346         /* so that alloc_pd can use it */
 347         mm->pgd = pgd;
 348         if (pgd)
 349                 pgd_ctor(pgd);
 350
 351         if (pgd && !pgd_prepopulate_pmd(mm, pgd)) {
 352                 pgd_dtor(pgd);
 353                 free_page((unsigned long)pgd);
 354                 pgd = NULL;
 355         }
 356
 357         return pgd;
 358 }
 359
 360 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
 361 {
 362         pgd_mop_up_pmds(mm, pgd);
 363         pgd_dtor(pgd);
 364         free_page((unsigned long)pgd);
 365 }
 366
 367 void __pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
 368 {
 369         pgtable_page_dtor(pte);
 370         paravirt_release_pt(page_to_pfn(pte));
 371         tlb_remove_page(tlb, pte);
 372 }
 373
 374 #ifdef CONFIG_X86_PAE
 375
 376 void __pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
 377 {
 378         paravirt_release_pd(__pa(pmd) >> PAGE_SHIFT);
 379         tlb_remove_page(tlb, virt_to_page(pmd));
 380 }
 381
 382 #endif
 383
 384 int pmd_bad(pmd_t pmd)
 385 {
 386         WARN_ON_ONCE(pmd_bad_v1(pmd) != pmd_bad_v2(pmd));
 387
 388         return pmd_bad_v1(pmd);
 389 }