arch/s390/mm/vmem.c

   1 /*
   2  *  arch/s390/mm/vmem.c
   3  *
   4  *    Copyright IBM Corp. 2006
   5  *    Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
   6  */
   7
   8 #include <linux/bootmem.h>
   9 #include <linux/pfn.h>
  10 #include <linux/mm.h>
  11 #include <linux/module.h>
  12 #include <linux/list.h>
  13 #include <linux/hugetlb.h>
  14 #include <asm/pgalloc.h>
  15 #include <asm/pgtable.h>
  16 #include <asm/setup.h>
  17 #include <asm/tlbflush.h>
  18 #include <asm/sections.h>
  19
  20 static DEFINE_MUTEX(vmem_mutex);
  21
  22 struct memory_segment {
  23         struct list_head list;
  24         unsigned long start;
  25         unsigned long size;
  26 };
  27
  28 static LIST_HEAD(mem_segs);
  29
  30 static void __ref *vmem_alloc_pages(unsigned int order)
  31 {
  32         if (slab_is_available())
  33                 return (void *)__get_free_pages(GFP_KERNEL, order);
  34         return alloc_bootmem_pages((1 << order) * PAGE_SIZE);
  35 }
  36
  37 static inline pud_t *vmem_pud_alloc(void)
  38 {
  39         pud_t *pud = NULL;
  40
  41 #ifdef CONFIG_64BIT
  42         pud = vmem_alloc_pages(2);
  43         if (!pud)
  44                 return NULL;
  45         clear_table((unsigned long *) pud, _REGION3_ENTRY_EMPTY, PAGE_SIZE * 4);
  46 #endif
  47         return pud;
  48 }
  49
  50 static inline pmd_t *vmem_pmd_alloc(void)
  51 {
  52         pmd_t *pmd = NULL;
  53
  54 #ifdef CONFIG_64BIT
  55         pmd = vmem_alloc_pages(2);
  56         if (!pmd)
  57                 return NULL;
  58         clear_table((unsigned long *) pmd, _SEGMENT_ENTRY_EMPTY, PAGE_SIZE * 4);
  59 #endif
  60         return pmd;
  61 }
  62
  63 static pte_t __ref *vmem_pte_alloc(void)
  64 {
  65         pte_t *pte;
  66
  67         if (slab_is_available())
  68                 pte = (pte_t *) page_table_alloc(&init_mm);
  69         else
  70                 pte = alloc_bootmem(PTRS_PER_PTE * sizeof(pte_t));
  71         if (!pte)
  72                 return NULL;
  73         if (MACHINE_HAS_HPAGE)
  74                 clear_table((unsigned long *) pte, _PAGE_TYPE_EMPTY | _PAGE_CO,
  75                             PTRS_PER_PTE * sizeof(pte_t));
  76         else
  77                 clear_table((unsigned long *) pte, _PAGE_TYPE_EMPTY,
  78                             PTRS_PER_PTE * sizeof(pte_t));
  79         return pte;
  80 }
  81
  82 /*
  83  * Add a physical memory range to the 1:1 mapping.
  84  */
  85 static int vmem_add_mem(unsigned long start, unsigned long size, int ro)
  86 {
  87         unsigned long address;
  88         pgd_t *pg_dir;
  89         pud_t *pu_dir;
  90         pmd_t *pm_dir;
  91         pte_t *pt_dir;
  92         pte_t  pte;
  93         int ret = -ENOMEM;
  94
  95         for (address = start; address < start + size; address += PAGE_SIZE) {
  96                 pg_dir = pgd_offset_k(address);
  97                 if (pgd_none(*pg_dir)) {
  98                         pu_dir = vmem_pud_alloc();
  99                         if (!pu_dir)
 100                                 goto out;
 101                         pgd_populate_kernel(&init_mm, pg_dir, pu_dir);
 102                 }
 103
 104                 pu_dir = pud_offset(pg_dir, address);
 105                 if (pud_none(*pu_dir)) {
 106                         pm_dir = vmem_pmd_alloc();
 107                         if (!pm_dir)
 108                                 goto out;
 109                         pud_populate_kernel(&init_mm, pu_dir, pm_dir);
 110                 }
 111
 112                 pte = mk_pte_phys(address, __pgprot(ro ? _PAGE_RO : 0));
 113                 pm_dir = pmd_offset(pu_dir, address);
 114
 115 #ifdef __s390x__
 116                 if (MACHINE_HAS_HPAGE && !(address & ~HPAGE_MASK) &&
 117                     (address + HPAGE_SIZE <= start + size) &&
 118                     (address >= HPAGE_SIZE)) {
 119                         pte_val(pte) |= _SEGMENT_ENTRY_LARGE |
 120                                         _SEGMENT_ENTRY_CO;
 121                         pmd_val(*pm_dir) = pte_val(pte);
 122                         address += HPAGE_SIZE - PAGE_SIZE;
 123                         continue;
 124                 }
 125 #endif
 126                 if (pmd_none(*pm_dir)) {
 127                         pt_dir = vmem_pte_alloc();
 128                         if (!pt_dir)
 129                                 goto out;
 130                         pmd_populate_kernel(&init_mm, pm_dir, pt_dir);
 131                 }
 132
 133                 pt_dir = pte_offset_kernel(pm_dir, address);
 134                 *pt_dir = pte;
 135         }
 136         ret = 0;
 137 out:
 138         flush_tlb_kernel_range(start, start + size);
 139         return ret;
 140 }
 141
 142 /*
 143  * Remove a physical memory range from the 1:1 mapping.
 144  * Currently only invalidates page table entries.
 145  */
 146 static void vmem_remove_range(unsigned long start, unsigned long size)
 147 {
 148         unsigned long address;
 149         pgd_t *pg_dir;
 150         pud_t *pu_dir;
 151         pmd_t *pm_dir;
 152         pte_t *pt_dir;
 153         pte_t  pte;
 154
 155         pte_val(pte) = _PAGE_TYPE_EMPTY;
 156         for (address = start; address < start + size; address += PAGE_SIZE) {
 157                 pg_dir = pgd_offset_k(address);
 158                 pu_dir = pud_offset(pg_dir, address);
 159                 if (pud_none(*pu_dir))
 160                         continue;
 161                 pm_dir = pmd_offset(pu_dir, address);
 162                 if (pmd_none(*pm_dir))
 163                         continue;
 164
 165                 if (pmd_huge(*pm_dir)) {
 166                         pmd_clear_kernel(pm_dir);
 167                         address += HPAGE_SIZE - PAGE_SIZE;
 168                         continue;
 169                 }
 170
 171                 pt_dir = pte_offset_kernel(pm_dir, address);
 172                 *pt_dir = pte;
 173         }
 174         flush_tlb_kernel_range(start, start + size);
 175 }
 176
 177 /*
 178  * Add a backed mem_map array to the virtual mem_map array.
 179  */
 180 int __meminit vmemmap_populate(struct page *start, unsigned long nr, int node)
 181 {
 182         unsigned long address, start_addr, end_addr;
 183         pgd_t *pg_dir;
 184         pud_t *pu_dir;
 185         pmd_t *pm_dir;
 186         pte_t *pt_dir;
 187         pte_t  pte;
 188         int ret = -ENOMEM;
 189
 190         start_addr = (unsigned long) start;
 191         end_addr = (unsigned long) (start + nr);
 192
 193         for (address = start_addr; address < end_addr; address += PAGE_SIZE) {
 194                 pg_dir = pgd_offset_k(address);
 195                 if (pgd_none(*pg_dir)) {
 196                         pu_dir = vmem_pud_alloc();
 197                         if (!pu_dir)
 198                                 goto out;
 199                         pgd_populate_kernel(&init_mm, pg_dir, pu_dir);
 200                 }
 201
 202                 pu_dir = pud_offset(pg_dir, address);
 203                 if (pud_none(*pu_dir)) {
 204                         pm_dir = vmem_pmd_alloc();
 205                         if (!pm_dir)
 206                                 goto out;
 207                         pud_populate_kernel(&init_mm, pu_dir, pm_dir);
 208                 }
 209
 210                 pm_dir = pmd_offset(pu_dir, address);
 211                 if (pmd_none(*pm_dir)) {
 212                         pt_dir = vmem_pte_alloc();
 213                         if (!pt_dir)
 214                                 goto out;
 215                         pmd_populate_kernel(&init_mm, pm_dir, pt_dir);
 216                 }
 217
 218                 pt_dir = pte_offset_kernel(pm_dir, address);
 219                 if (pte_none(*pt_dir)) {
 220                         unsigned long new_page;
 221
 222                         new_page =__pa(vmem_alloc_pages(0));
 223                         if (!new_page)
 224                                 goto out;
 225                         pte = pfn_pte(new_page >> PAGE_SHIFT, PAGE_KERNEL);
 226                         *pt_dir = pte;
 227                 }
 228         }
 229         memset(start, 0, nr * sizeof(struct page));
 230         ret = 0;
 231 out:
 232         flush_tlb_kernel_range(start_addr, end_addr);
 233         return ret;
 234 }
 235
 236 /*
 237  * Add memory segment to the segment list if it doesn't overlap with
 238  * an already present segment.
 239  */
 240 static int insert_memory_segment(struct memory_segment *seg)
 241 {
 242         struct memory_segment *tmp;
 243
 244         if (seg->start + seg->size > VMEM_MAX_PHYS ||
 245             seg->start + seg->size < seg->start)
 246                 return -ERANGE;
 247
 248         list_for_each_entry(tmp, &mem_segs, list) {
 249                 if (seg->start >= tmp->start + tmp->size)
 250                         continue;
 251                 if (seg->start + seg->size <= tmp->start)
 252                         continue;
 253                 return -ENOSPC;
 254         }
 255         list_add(&seg->list, &mem_segs);
 256         return 0;
 257 }
 258
 259 /*
 260  * Remove memory segment from the segment list.
 261  */
 262 static void remove_memory_segment(struct memory_segment *seg)
 263 {
 264         list_del(&seg->list);
 265 }
 266
 267 static void __remove_shared_memory(struct memory_segment *seg)
 268 {
 269         remove_memory_segment(seg);
 270         vmem_remove_range(seg->start, seg->size);
 271 }
 272
 273 int vmem_remove_mapping(unsigned long start, unsigned long size)
 274 {
 275         struct memory_segment *seg;
 276         int ret;
 277
 278         mutex_lock(&vmem_mutex);
 279
 280         ret = -ENOENT;
 281         list_for_each_entry(seg, &mem_segs, list) {
 282                 if (seg->start == start && seg->size == size)
 283                         break;
 284         }
 285
 286         if (seg->start != start || seg->size != size)
 287                 goto out;
 288
 289         ret = 0;
 290         __remove_shared_memory(seg);
 291         kfree(seg);
 292 out:
 293         mutex_unlock(&vmem_mutex);
 294         return ret;
 295 }
 296
 297 int vmem_add_mapping(unsigned long start, unsigned long size)
 298 {
 299         struct memory_segment *seg;
 300         int ret;
 301
 302         mutex_lock(&vmem_mutex);
 303         ret = -ENOMEM;
 304         seg = kzalloc(sizeof(*seg), GFP_KERNEL);
 305         if (!seg)
 306                 goto out;
 307         seg->start = start;
 308         seg->size = size;
 309
 310         ret = insert_memory_segment(seg);
 311         if (ret)
 312                 goto out_free;
 313
 314         ret = vmem_add_mem(start, size, 0);
 315         if (ret)
 316                 goto out_remove;
 317         goto out;
 318
 319 out_remove:
 320         __remove_shared_memory(seg);
 321 out_free:
 322         kfree(seg);
 323 out:
 324         mutex_unlock(&vmem_mutex);
 325         return ret;
 326 }
 327
 328 /*
 329  * map whole physical memory to virtual memory (identity mapping)
 330  * we reserve enough space in the vmalloc area for vmemmap to hotplug
 331  * additional memory segments.
 332  */
 333 void __init vmem_map_init(void)
 334 {
 335         unsigned long ro_start, ro_end;
 336         unsigned long start, end;
 337         int i;
 338
 339         spin_lock_init(&init_mm.context.list_lock);
 340         INIT_LIST_HEAD(&init_mm.context.crst_list);
 341         INIT_LIST_HEAD(&init_mm.context.pgtable_list);
 342         init_mm.context.noexec = 0;
 343         ro_start = ((unsigned long)&_stext) & PAGE_MASK;
 344         ro_end = PFN_ALIGN((unsigned long)&_eshared);
 345         for (i = 0; i < MEMORY_CHUNKS && memory_chunk[i].size > 0; i++) {
 346                 start = memory_chunk[i].addr;
 347                 end = memory_chunk[i].addr + memory_chunk[i].size;
 348                 if (start >= ro_end || end <= ro_start)
 349                         vmem_add_mem(start, end - start, 0);
 350                 else if (start >= ro_start && end <= ro_end)
 351                         vmem_add_mem(start, end - start, 1);
 352                 else if (start >= ro_start) {
 353                         vmem_add_mem(start, ro_end - start, 1);
 354                         vmem_add_mem(ro_end, end - ro_end, 0);
 355                 } else if (end < ro_end) {
 356                         vmem_add_mem(start, ro_start - start, 0);
 357                         vmem_add_mem(ro_start, end - ro_start, 1);
 358                 } else {
 359                         vmem_add_mem(start, ro_start - start, 0);
 360                         vmem_add_mem(ro_start, ro_end - ro_start, 1);
 361                         vmem_add_mem(ro_end, end - ro_end, 0);
 362                 }
 363         }
 364 }
 365
 366 /*
 367  * Convert memory chunk array to a memory segment list so there is a single
 368  * list that contains both r/w memory and shared memory segments.
 369  */
 370 static int __init vmem_convert_memory_chunk(void)
 371 {
 372         struct memory_segment *seg;
 373         int i;
 374
 375         mutex_lock(&vmem_mutex);
 376         for (i = 0; i < MEMORY_CHUNKS; i++) {
 377                 if (!memory_chunk[i].size)
 378                         continue;
 379                 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
 380                 if (!seg)
 381                         panic("Out of memory...\n");
 382                 seg->start = memory_chunk[i].addr;
 383                 seg->size = memory_chunk[i].size;
 384                 insert_memory_segment(seg);
 385         }
 386         mutex_unlock(&vmem_mutex);
 387         return 0;
 388 }
 389
 390 core_initcall(vmem_convert_memory_chunk);