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_INVALID,
  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_PROTECT : 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                                 _SEGMENT_ENTRY_YOUNG |
 122                                 (ro ? _SEGMENT_ENTRY_PROTECT : 0);
 123                         address += PMD_SIZE;
 124                         continue;
 125                 }
 126 #endif
 127                 if (pmd_none(*pm_dir)) {
 128                         pt_dir = vmem_pte_alloc(address);
 129                         if (!pt_dir)
 130                                 goto out;
 131                         pmd_populate(&init_mm, pm_dir, pt_dir);
 132                 }
 133
 134                 pt_dir = pte_offset_kernel(pm_dir, address);
 135                 pte_val(*pt_dir) = __pa(address) |
 136                         pgprot_val(ro ? PAGE_KERNEL_RO : PAGE_KERNEL);
 137                 address += PAGE_SIZE;
 138         }
 139         ret = 0;
 140 out:
 141         return ret;
 142 }
 143
 144 /*
 145  * Remove a physical memory range from the 1:1 mapping.
 146  * Currently only invalidates page table entries.
 147  */
 148 static void vmem_remove_range(unsigned long start, unsigned long size)
 149 {
 150         unsigned long end = start + size;
 151         unsigned long address = start;
 152         pgd_t *pg_dir;
 153         pud_t *pu_dir;
 154         pmd_t *pm_dir;
 155         pte_t *pt_dir;
 156         pte_t  pte;
 157
 158         pte_val(pte) = _PAGE_INVALID;
 159         while (address < end) {
 160                 pg_dir = pgd_offset_k(address);
 161                 if (pgd_none(*pg_dir)) {
 162                         address += PGDIR_SIZE;
 163                         continue;
 164                 }
 165                 pu_dir = pud_offset(pg_dir, address);
 166                 if (pud_none(*pu_dir)) {
 167                         address += PUD_SIZE;
 168                         continue;
 169                 }
 170                 if (pud_large(*pu_dir)) {
 171                         pud_clear(pu_dir);
 172                         address += PUD_SIZE;
 173                         continue;
 174                 }
 175                 pm_dir = pmd_offset(pu_dir, address);
 176                 if (pmd_none(*pm_dir)) {
 177                         address += PMD_SIZE;
 178                         continue;
 179                 }
 180                 if (pmd_large(*pm_dir)) {
 181                         pmd_clear(pm_dir);
 182                         address += PMD_SIZE;
 183                         continue;
 184                 }
 185                 pt_dir = pte_offset_kernel(pm_dir, address);
 186                 *pt_dir = pte;
 187                 address += PAGE_SIZE;
 188         }
 189         flush_tlb_kernel_range(start, end);
 190 }
 191
 192 /*
 193  * Add a backed mem_map array to the virtual mem_map array.
 194  */
 195 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 196 {
 197         unsigned long address = start;
 198         pgd_t *pg_dir;
 199         pud_t *pu_dir;
 200         pmd_t *pm_dir;
 201         pte_t *pt_dir;
 202         int ret = -ENOMEM;
 203
 204         for (address = start; address < end;) {
 205                 pg_dir = pgd_offset_k(address);
 206                 if (pgd_none(*pg_dir)) {
 207                         pu_dir = vmem_pud_alloc();
 208                         if (!pu_dir)
 209                                 goto out;
 210                         pgd_populate(&init_mm, pg_dir, pu_dir);
 211                 }
 212
 213                 pu_dir = pud_offset(pg_dir, address);
 214                 if (pud_none(*pu_dir)) {
 215                         pm_dir = vmem_pmd_alloc();
 216                         if (!pm_dir)
 217                                 goto out;
 218                         pud_populate(&init_mm, pu_dir, pm_dir);
 219                 }
 220
 221                 pm_dir = pmd_offset(pu_dir, address);
 222                 if (pmd_none(*pm_dir)) {
 223 #ifdef CONFIG_64BIT
 224                         /* Use 1MB frames for vmemmap if available. We always
 225                          * use large frames even if they are only partially
 226                          * used.
 227                          * Otherwise we would have also page tables since
 228                          * vmemmap_populate gets called for each section
 229                          * separately. */
 230                         if (MACHINE_HAS_EDAT1) {
 231                                 void *new_page;
 232
 233                                 new_page = vmemmap_alloc_block(PMD_SIZE, node);
 234                                 if (!new_page)
 235                                         goto out;
 236                                 pmd_val(*pm_dir) = __pa(new_page) |
 237                                         _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE |
 238                                         _SEGMENT_ENTRY_CO;
 239                                 address = (address + PMD_SIZE) & PMD_MASK;
 240                                 continue;
 241                         }
 242 #endif
 243                         pt_dir = vmem_pte_alloc(address);
 244                         if (!pt_dir)
 245                                 goto out;
 246                         pmd_populate(&init_mm, pm_dir, pt_dir);
 247                 } else if (pmd_large(*pm_dir)) {
 248                         address = (address + PMD_SIZE) & PMD_MASK;
 249                         continue;
 250                 }
 251
 252                 pt_dir = pte_offset_kernel(pm_dir, address);
 253                 if (pte_none(*pt_dir)) {
 254                         unsigned long new_page;
 255
 256                         new_page =__pa(vmem_alloc_pages(0));
 257                         if (!new_page)
 258                                 goto out;
 259                         pte_val(*pt_dir) =
 260                                 __pa(new_page) | pgprot_val(PAGE_KERNEL);
 261                 }
 262                 address += PAGE_SIZE;
 263         }
 264         memset((void *)start, 0, end - start);
 265         ret = 0;
 266 out:
 267         return ret;
 268 }
 269
 270 void vmemmap_free(unsigned long start, unsigned long end)
 271 {
 272 }
 273
 274 /*
 275  * Add memory segment to the segment list if it doesn't overlap with
 276  * an already present segment.
 277  */
 278 static int insert_memory_segment(struct memory_segment *seg)
 279 {
 280         struct memory_segment *tmp;
 281
 282         if (seg->start + seg->size > VMEM_MAX_PHYS ||
 283             seg->start + seg->size < seg->start)
 284                 return -ERANGE;
 285
 286         list_for_each_entry(tmp, &mem_segs, list) {
 287                 if (seg->start >= tmp->start + tmp->size)
 288                         continue;
 289                 if (seg->start + seg->size <= tmp->start)
 290                         continue;
 291                 return -ENOSPC;
 292         }
 293         list_add(&seg->list, &mem_segs);
 294         return 0;
 295 }
 296
 297 /*
 298  * Remove memory segment from the segment list.
 299  */
 300 static void remove_memory_segment(struct memory_segment *seg)
 301 {
 302         list_del(&seg->list);
 303 }
 304
 305 static void __remove_shared_memory(struct memory_segment *seg)
 306 {
 307         remove_memory_segment(seg);
 308         vmem_remove_range(seg->start, seg->size);
 309 }
 310
 311 int vmem_remove_mapping(unsigned long start, unsigned long size)
 312 {
 313         struct memory_segment *seg;
 314         int ret;
 315
 316         mutex_lock(&vmem_mutex);
 317
 318         ret = -ENOENT;
 319         list_for_each_entry(seg, &mem_segs, list) {
 320                 if (seg->start == start && seg->size == size)
 321                         break;
 322         }
 323
 324         if (seg->start != start || seg->size != size)
 325                 goto out;
 326
 327         ret = 0;
 328         __remove_shared_memory(seg);
 329         kfree(seg);
 330 out:
 331         mutex_unlock(&vmem_mutex);
 332         return ret;
 333 }
 334
 335 int vmem_add_mapping(unsigned long start, unsigned long size)
 336 {
 337         struct memory_segment *seg;
 338         int ret;
 339
 340         mutex_lock(&vmem_mutex);
 341         ret = -ENOMEM;
 342         seg = kzalloc(sizeof(*seg), GFP_KERNEL);
 343         if (!seg)
 344                 goto out;
 345         seg->start = start;
 346         seg->size = size;
 347
 348         ret = insert_memory_segment(seg);
 349         if (ret)
 350                 goto out_free;
 351
 352         ret = vmem_add_mem(start, size, 0);
 353         if (ret)
 354                 goto out_remove;
 355         goto out;
 356
 357 out_remove:
 358         __remove_shared_memory(seg);
 359 out_free:
 360         kfree(seg);
 361 out:
 362         mutex_unlock(&vmem_mutex);
 363         return ret;
 364 }
 365
 366 /*
 367  * map whole physical memory to virtual memory (identity mapping)
 368  * we reserve enough space in the vmalloc area for vmemmap to hotplug
 369  * additional memory segments.
 370  */
 371 void __init vmem_map_init(void)
 372 {
 373         unsigned long ro_start, ro_end;
 374         unsigned long start, end;
 375         int i;
 376
 377         ro_start = PFN_ALIGN((unsigned long)&_stext);
 378         ro_end = (unsigned long)&_eshared & PAGE_MASK;
 379         for (i = 0; i < MEMORY_CHUNKS; i++) {
 380                 if (!memory_chunk[i].size)
 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                 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
 416                 if (!seg)
 417                         panic("Out of memory...\n");
 418                 seg->start = memory_chunk[i].addr;
 419                 seg->size = memory_chunk[i].size;
 420                 insert_memory_segment(seg);
 421         }
 422         mutex_unlock(&vmem_mutex);
 423         return 0;
 424 }
 425
 426 core_initcall(vmem_convert_memory_chunk);