arch/s390/mm/vmem.c

   1 /*
   2  *    Copyright IBM Corp. 2006
   3  *    Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
   4  */
   5
   6 #include <linux/bootmem.h>
   7 #include <linux/pfn.h>
   8 #include <linux/mm.h>
   9 #include <linux/module.h>
  10 #include <linux/list.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/slab.h>
  13 #include <asm/pgalloc.h>
  14 #include <asm/pgtable.h>
  15 #include <asm/setup.h>
  16 #include <asm/tlbflush.h>
  17 #include <asm/sections.h>
  18
  19 static DEFINE_MUTEX(vmem_mutex);
  20
  21 struct memory_segment {
  22         struct list_head list;
  23         unsigned long start;
  24         unsigned long size;
  25 };
  26
  27 static LIST_HEAD(mem_segs);
  28
  29 static void __ref *vmem_alloc_pages(unsigned int order)
  30 {
  31         if (slab_is_available())
  32                 return (void *)__get_free_pages(GFP_KERNEL, order);
  33         return alloc_bootmem_pages((1 << order) * PAGE_SIZE);
  34 }
  35
  36 static inline pud_t *vmem_pud_alloc(void)
  37 {
  38         pud_t *pud = NULL;
  39
  40 #ifdef CONFIG_64BIT
  41         pud = vmem_alloc_pages(2);
  42         if (!pud)
  43                 return NULL;
  44         clear_table((unsigned long *) pud, _REGION3_ENTRY_EMPTY, PAGE_SIZE * 4);
  45 #endif
  46         return pud;
  47 }
  48
  49 static inline pmd_t *vmem_pmd_alloc(void)
  50 {
  51         pmd_t *pmd = NULL;
  52
  53 #ifdef CONFIG_64BIT
  54         pmd = vmem_alloc_pages(2);
  55         if (!pmd)
  56                 return NULL;
  57         clear_table((unsigned long *) pmd, _SEGMENT_ENTRY_EMPTY, PAGE_SIZE * 4);
  58 #endif
  59         return pmd;
  60 }
  61
  62 static pte_t __ref *vmem_pte_alloc(unsigned long address)
  63 {
  64         pte_t *pte;
  65
  66         if (slab_is_available())
  67                 pte = (pte_t *) page_table_alloc(&init_mm, address);
  68         else
  69                 pte = alloc_bootmem(PTRS_PER_PTE * sizeof(pte_t));
  70         if (!pte)
  71                 return NULL;
  72         clear_table((unsigned long *) pte, _PAGE_TYPE_EMPTY,
  73                     PTRS_PER_PTE * sizeof(pte_t));
  74         return pte;
  75 }
  76
  77 /*
  78  * Add a physical memory range to the 1:1 mapping.
  79  */
  80 static int vmem_add_mem(unsigned long start, unsigned long size, int ro)
  81 {
  82         unsigned long end = start + size;
  83         unsigned long address = start;
  84         pgd_t *pg_dir;
  85         pud_t *pu_dir;
  86         pmd_t *pm_dir;
  87         pte_t *pt_dir;
  88         int ret = -ENOMEM;
  89
  90         while (address < end) {
  91                 pg_dir = pgd_offset_k(address);
  92                 if (pgd_none(*pg_dir)) {
  93                         pu_dir = vmem_pud_alloc();
  94                         if (!pu_dir)
  95                                 goto out;
  96                         pgd_populate(&init_mm, pg_dir, pu_dir);
  97                 }
  98                 pu_dir = pud_offset(pg_dir, address);
  99 #if defined(CONFIG_64BIT) && !defined(CONFIG_DEBUG_PAGEALLOC)
 100                 if (MACHINE_HAS_EDAT2 && pud_none(*pu_dir) && address &&
 101                     !(address & ~PUD_MASK) && (address + PUD_SIZE <= end)) {
 102                         pud_val(*pu_dir) = __pa(address) |
 103                                 _REGION_ENTRY_TYPE_R3 | _REGION3_ENTRY_LARGE |
 104                                 (ro ? _REGION_ENTRY_RO : 0);
 105                         address += PUD_SIZE;
 106                         continue;
 107                 }
 108 #endif
 109                 if (pud_none(*pu_dir)) {
 110                         pm_dir = vmem_pmd_alloc();
 111                         if (!pm_dir)
 112                                 goto out;
 113                         pud_populate(&init_mm, pu_dir, pm_dir);
 114                 }
 115                 pm_dir = pmd_offset(pu_dir, address);
 116 #if defined(CONFIG_64BIT) && !defined(CONFIG_DEBUG_PAGEALLOC)
 117                 if (MACHINE_HAS_EDAT1 && pmd_none(*pm_dir) && address &&
 118                     !(address & ~PMD_MASK) && (address + PMD_SIZE <= end)) {
 119                         pmd_val(*pm_dir) = __pa(address) |
 120                                 _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE |
 121                                 (ro ? _SEGMENT_ENTRY_RO : 0);
 122                         address += PMD_SIZE;
 123                         continue;
 124                 }
 125 #endif
 126                 if (pmd_none(*pm_dir)) {
 127                         pt_dir = vmem_pte_alloc(address);
 128                         if (!pt_dir)
 129                                 goto out;
 130                         pmd_populate(&init_mm, pm_dir, pt_dir);
 131                 }
 132
 133                 pt_dir = pte_offset_kernel(pm_dir, address);
 134                 pte_val(*pt_dir) = __pa(address) | (ro ? _PAGE_RO : 0);
 135                 address += PAGE_SIZE;
 136         }
 137         ret = 0;
 138 out:
 139         flush_tlb_kernel_range(start, end);
 140         return ret;
 141 }
 142
 143 /*
 144  * Remove a physical memory range from the 1:1 mapping.
 145  * Currently only invalidates page table entries.
 146  */
 147 static void vmem_remove_range(unsigned long start, unsigned long size)
 148 {
 149         unsigned long end = start + size;
 150         unsigned long address = start;
 151         pgd_t *pg_dir;
 152         pud_t *pu_dir;
 153         pmd_t *pm_dir;
 154         pte_t *pt_dir;
 155         pte_t  pte;
 156
 157         pte_val(pte) = _PAGE_TYPE_EMPTY;
 158         while (address < end) {
 159                 pg_dir = pgd_offset_k(address);
 160                 if (pgd_none(*pg_dir)) {
 161                         address += PGDIR_SIZE;
 162                         continue;
 163                 }
 164                 pu_dir = pud_offset(pg_dir, address);
 165                 if (pud_none(*pu_dir)) {
 166                         address += PUD_SIZE;
 167                         continue;
 168                 }
 169                 if (pud_large(*pu_dir)) {
 170                         pud_clear(pu_dir);
 171                         address += PUD_SIZE;
 172                         continue;
 173                 }
 174                 pm_dir = pmd_offset(pu_dir, address);
 175                 if (pmd_none(*pm_dir)) {
 176                         address += PMD_SIZE;
 177                         continue;
 178                 }
 179                 if (pmd_large(*pm_dir)) {
 180                         pmd_clear(pm_dir);
 181                         address += PMD_SIZE;
 182                         continue;
 183                 }
 184                 pt_dir = pte_offset_kernel(pm_dir, address);
 185                 *pt_dir = pte;
 186                 address += PAGE_SIZE;
 187         }
 188         flush_tlb_kernel_range(start, end);
 189 }
 190
 191 /*
 192  * Add a backed mem_map array to the virtual mem_map array.
 193  */
 194 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 195 {
 196         unsigned long address = start;
 197         pgd_t *pg_dir;
 198         pud_t *pu_dir;
 199         pmd_t *pm_dir;
 200         pte_t *pt_dir;
 201         int ret = -ENOMEM;
 202
 203         for (address = start; address < end;) {
 204                 pg_dir = pgd_offset_k(address);
 205                 if (pgd_none(*pg_dir)) {
 206                         pu_dir = vmem_pud_alloc();
 207                         if (!pu_dir)
 208                                 goto out;
 209                         pgd_populate(&init_mm, pg_dir, pu_dir);
 210                 }
 211
 212                 pu_dir = pud_offset(pg_dir, address);
 213                 if (pud_none(*pu_dir)) {
 214                         pm_dir = vmem_pmd_alloc();
 215                         if (!pm_dir)
 216                                 goto out;
 217                         pud_populate(&init_mm, pu_dir, pm_dir);
 218                 }
 219
 220                 pm_dir = pmd_offset(pu_dir, address);
 221                 if (pmd_none(*pm_dir)) {
 222 #ifdef CONFIG_64BIT
 223                         /* Use 1MB frames for vmemmap if available. We always
 224                          * use large frames even if they are only partially
 225                          * used.
 226                          * Otherwise we would have also page tables since
 227                          * vmemmap_populate gets called for each section
 228                          * separately. */
 229                         if (MACHINE_HAS_EDAT1) {
 230                                 void *new_page;
 231
 232                                 new_page = vmemmap_alloc_block(PMD_SIZE, node);
 233                                 if (!new_page)
 234                                         goto out;
 235                                 pmd_val(*pm_dir) = __pa(new_page) |
 236                                         _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE |
 237                                         _SEGMENT_ENTRY_CO;
 238                                 address = (address + PMD_SIZE) & PMD_MASK;
 239                                 continue;
 240                         }
 241 #endif
 242                         pt_dir = vmem_pte_alloc(address);
 243                         if (!pt_dir)
 244                                 goto out;
 245                         pmd_populate(&init_mm, pm_dir, pt_dir);
 246                 } else if (pmd_large(*pm_dir)) {
 247                         address = (address + PMD_SIZE) & PMD_MASK;
 248                         continue;
 249                 }
 250
 251                 pt_dir = pte_offset_kernel(pm_dir, address);
 252                 if (pte_none(*pt_dir)) {
 253                         unsigned long new_page;
 254
 255                         new_page =__pa(vmem_alloc_pages(0));
 256                         if (!new_page)
 257                                 goto out;
 258                         pte_val(*pt_dir) = __pa(new_page);
 259                 }
 260                 address += PAGE_SIZE;
 261         }
 262         memset((void *)start, 0, end - start);
 263         ret = 0;
 264 out:
 265         flush_tlb_kernel_range(start, end);
 266         return ret;
 267 }
 268
 269 void vmemmap_free(unsigned long start, unsigned long end)
 270 {
 271 }
 272
 273 /*
 274  * Add memory segment to the segment list if it doesn't overlap with
 275  * an already present segment.
 276  */
 277 static int insert_memory_segment(struct memory_segment *seg)
 278 {
 279         struct memory_segment *tmp;
 280
 281         if (seg->start + seg->size > VMEM_MAX_PHYS ||
 282             seg->start + seg->size < seg->start)
 283                 return -ERANGE;
 284
 285         list_for_each_entry(tmp, &mem_segs, list) {
 286                 if (seg->start >= tmp->start + tmp->size)
 287                         continue;
 288                 if (seg->start + seg->size <= tmp->start)
 289                         continue;
 290                 return -ENOSPC;
 291         }
 292         list_add(&seg->list, &mem_segs);
 293         return 0;
 294 }
 295
 296 /*
 297  * Remove memory segment from the segment list.
 298  */
 299 static void remove_memory_segment(struct memory_segment *seg)
 300 {
 301         list_del(&seg->list);
 302 }
 303
 304 static void __remove_shared_memory(struct memory_segment *seg)
 305 {
 306         remove_memory_segment(seg);
 307         vmem_remove_range(seg->start, seg->size);
 308 }
 309
 310 int vmem_remove_mapping(unsigned long start, unsigned long size)
 311 {
 312         struct memory_segment *seg;
 313         int ret;
 314
 315         mutex_lock(&vmem_mutex);
 316
 317         ret = -ENOENT;
 318         list_for_each_entry(seg, &mem_segs, list) {
 319                 if (seg->start == start && seg->size == size)
 320                         break;
 321         }
 322
 323         if (seg->start != start || seg->size != size)
 324                 goto out;
 325
 326         ret = 0;
 327         __remove_shared_memory(seg);
 328         kfree(seg);
 329 out:
 330         mutex_unlock(&vmem_mutex);
 331         return ret;
 332 }
 333
 334 int vmem_add_mapping(unsigned long start, unsigned long size)
 335 {
 336         struct memory_segment *seg;
 337         int ret;
 338
 339         mutex_lock(&vmem_mutex);
 340         ret = -ENOMEM;
 341         seg = kzalloc(sizeof(*seg), GFP_KERNEL);
 342         if (!seg)
 343                 goto out;
 344         seg->start = start;
 345         seg->size = size;
 346
 347         ret = insert_memory_segment(seg);
 348         if (ret)
 349                 goto out_free;
 350
 351         ret = vmem_add_mem(start, size, 0);
 352         if (ret)
 353                 goto out_remove;
 354         goto out;
 355
 356 out_remove:
 357         __remove_shared_memory(seg);
 358 out_free:
 359         kfree(seg);
 360 out:
 361         mutex_unlock(&vmem_mutex);
 362         return ret;
 363 }
 364
 365 /*
 366  * map whole physical memory to virtual memory (identity mapping)
 367  * we reserve enough space in the vmalloc area for vmemmap to hotplug
 368  * additional memory segments.
 369  */
 370 void __init vmem_map_init(void)
 371 {
 372         unsigned long ro_start, ro_end;
 373         unsigned long start, end;
 374         int i;
 375
 376         ro_start = PFN_ALIGN((unsigned long)&_stext);
 377         ro_end = (unsigned long)&_eshared & PAGE_MASK;
 378         for (i = 0; i < MEMORY_CHUNKS && memory_chunk[i].size > 0; i++) {
 379                 if (memory_chunk[i].type == CHUNK_CRASHK ||
 380                     memory_chunk[i].type == CHUNK_OLDMEM)
 381                         continue;
 382                 start = memory_chunk[i].addr;
 383                 end = memory_chunk[i].addr + memory_chunk[i].size;
 384                 if (start >= ro_end || end <= ro_start)
 385                         vmem_add_mem(start, end - start, 0);
 386                 else if (start >= ro_start && end <= ro_end)
 387                         vmem_add_mem(start, end - start, 1);
 388                 else if (start >= ro_start) {
 389                         vmem_add_mem(start, ro_end - start, 1);
 390                         vmem_add_mem(ro_end, end - ro_end, 0);
 391                 } else if (end < ro_end) {
 392                         vmem_add_mem(start, ro_start - start, 0);
 393                         vmem_add_mem(ro_start, end - ro_start, 1);
 394                 } else {
 395                         vmem_add_mem(start, ro_start - start, 0);
 396                         vmem_add_mem(ro_start, ro_end - ro_start, 1);
 397                         vmem_add_mem(ro_end, end - ro_end, 0);
 398                 }
 399         }
 400 }
 401
 402 /*
 403  * Convert memory chunk array to a memory segment list so there is a single
 404  * list that contains both r/w memory and shared memory segments.
 405  */
 406 static int __init vmem_convert_memory_chunk(void)
 407 {
 408         struct memory_segment *seg;
 409         int i;
 410
 411         mutex_lock(&vmem_mutex);
 412         for (i = 0; i < MEMORY_CHUNKS; i++) {
 413                 if (!memory_chunk[i].size)
 414                         continue;
 415                 if (memory_chunk[i].type == CHUNK_CRASHK ||
 416                     memory_chunk[i].type == CHUNK_OLDMEM)
 417                         continue;
 418                 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
 419                 if (!seg)
 420                         panic("Out of memory...\n");
 421                 seg->start = memory_chunk[i].addr;
 422                 seg->size = memory_chunk[i].size;
 423                 insert_memory_segment(seg);
 424         }
 425         mutex_unlock(&vmem_mutex);
 426         return 0;
 427 }
 428
 429 core_initcall(vmem_convert_memory_chunk);