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