arch/powerpc/mm/init_64.c

   1 /*
   2  *  PowerPC version
   3  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
   4  *
   5  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
   6  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
   7  *    Copyright (C) 1996 Paul Mackerras
   8  *
   9  *  Derived from "arch/i386/mm/init.c"
  10  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  11  *
  12  *  Dave Engebretsen <engebret@us.ibm.com>
  13  *      Rework for PPC64 port.
  14  *
  15  *  This program is free software; you can redistribute it and/or
  16  *  modify it under the terms of the GNU General Public License
  17  *  as published by the Free Software Foundation; either version
  18  *  2 of the License, or (at your option) any later version.
  19  *
  20  */
  21
  22 #undef DEBUG
  23
  24 #include <linux/signal.h>
  25 #include <linux/sched.h>
  26 #include <linux/kernel.h>
  27 #include <linux/errno.h>
  28 #include <linux/string.h>
  29 #include <linux/types.h>
  30 #include <linux/mman.h>
  31 #include <linux/mm.h>
  32 #include <linux/swap.h>
  33 #include <linux/stddef.h>
  34 #include <linux/vmalloc.h>
  35 #include <linux/init.h>
  36 #include <linux/delay.h>
  37 #include <linux/highmem.h>
  38 #include <linux/idr.h>
  39 #include <linux/nodemask.h>
  40 #include <linux/module.h>
  41 #include <linux/poison.h>
  42 #include <linux/memblock.h>
  43 #include <linux/hugetlb.h>
  44 #include <linux/slab.h>
  45 #include <linux/of_fdt.h>
  46 #include <linux/libfdt.h>
  47
  48 #include <asm/pgalloc.h>
  49 #include <asm/page.h>
  50 #include <asm/prom.h>
  51 #include <asm/rtas.h>
  52 #include <asm/io.h>
  53 #include <asm/mmu_context.h>
  54 #include <asm/pgtable.h>
  55 #include <asm/mmu.h>
  56 #include <linux/uaccess.h>
  57 #include <asm/smp.h>
  58 #include <asm/machdep.h>
  59 #include <asm/tlb.h>
  60 #include <asm/eeh.h>
  61 #include <asm/processor.h>
  62 #include <asm/mmzone.h>
  63 #include <asm/cputable.h>
  64 #include <asm/sections.h>
  65 #include <asm/iommu.h>
  66 #include <asm/vdso.h>
  67
  68 #include "mmu_decl.h"
  69
  70 #ifdef CONFIG_PPC_STD_MMU_64
  71 #if H_PGTABLE_RANGE > USER_VSID_RANGE
  72 #warning Limited user VSID range means pagetable space is wasted
  73 #endif
  74
  75 #if (TASK_SIZE_USER64 < H_PGTABLE_RANGE) && (TASK_SIZE_USER64 < USER_VSID_RANGE)
  76 #warning TASK_SIZE is smaller than it needs to be.
  77 #endif
  78 #endif /* CONFIG_PPC_STD_MMU_64 */
  79
  80 phys_addr_t memstart_addr = ~0;
  81 EXPORT_SYMBOL_GPL(memstart_addr);
  82 phys_addr_t kernstart_addr;
  83 EXPORT_SYMBOL_GPL(kernstart_addr);
  84
  85 #ifdef CONFIG_SPARSEMEM_VMEMMAP
  86 /*
  87  * Given an address within the vmemmap, determine the pfn of the page that
  88  * represents the start of the section it is within.  Note that we have to
  89  * do this by hand as the proffered address may not be correctly aligned.
  90  * Subtraction of non-aligned pointers produces undefined results.
  91  */
  92 static unsigned long __meminit vmemmap_section_start(unsigned long page)
  93 {
  94         unsigned long offset = page - ((unsigned long)(vmemmap));
  95
  96         /* Return the pfn of the start of the section. */
  97         return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
  98 }
  99
 100 /*
 101  * Check if this vmemmap page is already initialised.  If any section
 102  * which overlaps this vmemmap page is initialised then this page is
 103  * initialised already.
 104  */
 105 static int __meminit vmemmap_populated(unsigned long start, int page_size)
 106 {
 107         unsigned long end = start + page_size;
 108         start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
 109
 110         for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
 111                 if (pfn_valid(page_to_pfn((struct page *)start)))
 112                         return 1;
 113
 114         return 0;
 115 }
 116
 117 struct vmemmap_backing *vmemmap_list;
 118 static struct vmemmap_backing *next;
 119 static int num_left;
 120 static int num_freed;
 121
 122 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
 123 {
 124         struct vmemmap_backing *vmem_back;
 125         /* get from freed entries first */
 126         if (num_freed) {
 127                 num_freed--;
 128                 vmem_back = next;
 129                 next = next->list;
 130
 131                 return vmem_back;
 132         }
 133
 134         /* allocate a page when required and hand out chunks */
 135         if (!num_left) {
 136                 next = vmemmap_alloc_block(PAGE_SIZE, node);
 137                 if (unlikely(!next)) {
 138                         WARN_ON(1);
 139                         return NULL;
 140                 }
 141                 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
 142         }
 143
 144         num_left--;
 145
 146         return next++;
 147 }
 148
 149 static __meminit void vmemmap_list_populate(unsigned long phys,
 150                                             unsigned long start,
 151                                             int node)
 152 {
 153         struct vmemmap_backing *vmem_back;
 154
 155         vmem_back = vmemmap_list_alloc(node);
 156         if (unlikely(!vmem_back)) {
 157                 WARN_ON(1);
 158                 return;
 159         }
 160
 161         vmem_back->phys = phys;
 162         vmem_back->virt_addr = start;
 163         vmem_back->list = vmemmap_list;
 164
 165         vmemmap_list = vmem_back;
 166 }
 167
 168 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 169 {
 170         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 171
 172         /* Align to the page size of the linear mapping. */
 173         start = _ALIGN_DOWN(start, page_size);
 174
 175         pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
 176
 177         for (; start < end; start += page_size) {
 178                 void *p;
 179                 int rc;
 180
 181                 if (vmemmap_populated(start, page_size))
 182                         continue;
 183
 184                 p = vmemmap_alloc_block(page_size, node);
 185                 if (!p)
 186                         return -ENOMEM;
 187
 188                 vmemmap_list_populate(__pa(p), start, node);
 189
 190                 pr_debug("      * %016lx..%016lx allocated at %p\n",
 191                          start, start + page_size, p);
 192
 193                 rc = vmemmap_create_mapping(start, page_size, __pa(p));
 194                 if (rc < 0) {
 195                         pr_warning(
 196                                 "vmemmap_populate: Unable to create vmemmap mapping: %d\n",
 197                                 rc);
 198                         return -EFAULT;
 199                 }
 200         }
 201
 202         return 0;
 203 }
 204
 205 #ifdef CONFIG_MEMORY_HOTPLUG
 206 static unsigned long vmemmap_list_free(unsigned long start)
 207 {
 208         struct vmemmap_backing *vmem_back, *vmem_back_prev;
 209
 210         vmem_back_prev = vmem_back = vmemmap_list;
 211
 212         /* look for it with prev pointer recorded */
 213         for (; vmem_back; vmem_back = vmem_back->list) {
 214                 if (vmem_back->virt_addr == start)
 215                         break;
 216                 vmem_back_prev = vmem_back;
 217         }
 218
 219         if (unlikely(!vmem_back)) {
 220                 WARN_ON(1);
 221                 return 0;
 222         }
 223
 224         /* remove it from vmemmap_list */
 225         if (vmem_back == vmemmap_list) /* remove head */
 226                 vmemmap_list = vmem_back->list;
 227         else
 228                 vmem_back_prev->list = vmem_back->list;
 229
 230         /* next point to this freed entry */
 231         vmem_back->list = next;
 232         next = vmem_back;
 233         num_freed++;
 234
 235         return vmem_back->phys;
 236 }
 237
 238 void __ref vmemmap_free(unsigned long start, unsigned long end)
 239 {
 240         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 241
 242         start = _ALIGN_DOWN(start, page_size);
 243
 244         pr_debug("vmemmap_free %lx...%lx\n", start, end);
 245
 246         for (; start < end; start += page_size) {
 247                 unsigned long addr;
 248
 249                 /*
 250                  * the section has already be marked as invalid, so
 251                  * vmemmap_populated() true means some other sections still
 252                  * in this page, so skip it.
 253                  */
 254                 if (vmemmap_populated(start, page_size))
 255                         continue;
 256
 257                 addr = vmemmap_list_free(start);
 258                 if (addr) {
 259                         struct page *page = pfn_to_page(addr >> PAGE_SHIFT);
 260
 261                         if (PageReserved(page)) {
 262                                 /* allocated from bootmem */
 263                                 if (page_size < PAGE_SIZE) {
 264                                         /*
 265                                          * this shouldn't happen, but if it is
 266                                          * the case, leave the memory there
 267                                          */
 268                                         WARN_ON_ONCE(1);
 269                                 } else {
 270                                         unsigned int nr_pages =
 271                                                 1 << get_order(page_size);
 272                                         while (nr_pages--)
 273                                                 free_reserved_page(page++);
 274                                 }
 275                         } else
 276                                 free_pages((unsigned long)(__va(addr)),
 277                                                         get_order(page_size));
 278
 279                         vmemmap_remove_mapping(start, page_size);
 280                 }
 281         }
 282 }
 283 #endif
 284 void register_page_bootmem_memmap(unsigned long section_nr,
 285                                   struct page *start_page, unsigned long size)
 286 {
 287 }
 288
 289 /*
 290  * We do not have access to the sparsemem vmemmap, so we fallback to
 291  * walking the list of sparsemem blocks which we already maintain for
 292  * the sake of crashdump. In the long run, we might want to maintain
 293  * a tree if performance of that linear walk becomes a problem.
 294  *
 295  * realmode_pfn_to_page functions can fail due to:
 296  * 1) As real sparsemem blocks do not lay in RAM continously (they
 297  * are in virtual address space which is not available in the real mode),
 298  * the requested page struct can be split between blocks so get_page/put_page
 299  * may fail.
 300  * 2) When huge pages are used, the get_page/put_page API will fail
 301  * in real mode as the linked addresses in the page struct are virtual
 302  * too.
 303  */
 304 struct page *realmode_pfn_to_page(unsigned long pfn)
 305 {
 306         struct vmemmap_backing *vmem_back;
 307         struct page *page;
 308         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 309         unsigned long pg_va = (unsigned long) pfn_to_page(pfn);
 310
 311         for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) {
 312                 if (pg_va < vmem_back->virt_addr)
 313                         continue;
 314
 315                 /* After vmemmap_list entry free is possible, need check all */
 316                 if ((pg_va + sizeof(struct page)) <=
 317                                 (vmem_back->virt_addr + page_size)) {
 318                         page = (struct page *) (vmem_back->phys + pg_va -
 319                                 vmem_back->virt_addr);
 320                         return page;
 321                 }
 322         }
 323
 324         /* Probably that page struct is split between real pages */
 325         return NULL;
 326 }
 327 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
 328
 329 #elif defined(CONFIG_FLATMEM)
 330
 331 struct page *realmode_pfn_to_page(unsigned long pfn)
 332 {
 333         struct page *page = pfn_to_page(pfn);
 334         return page;
 335 }
 336 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
 337
 338 #endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */
 339
 340 #ifdef CONFIG_PPC_STD_MMU_64
 341 static bool disable_radix;
 342 static int __init parse_disable_radix(char *p)
 343 {
 344         disable_radix = true;
 345         return 0;
 346 }
 347 early_param("disable_radix", parse_disable_radix);
 348
 349 /*
 350  * If we're running under a hypervisor, we need to check the contents of
 351  * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do
 352  * radix.  If not, we clear the radix feature bit so we fall back to hash.
 353  */
 354 static void early_check_vec5(void)
 355 {
 356         unsigned long root, chosen;
 357         int size;
 358         const u8 *vec5;
 359         u8 mmu_supported;
 360
 361         root = of_get_flat_dt_root();
 362         chosen = of_get_flat_dt_subnode_by_name(root, "chosen");
 363         if (chosen == -FDT_ERR_NOTFOUND) {
 364                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 365                 return;
 366         }
 367         vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size);
 368         if (!vec5) {
 369                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 370                 return;
 371         }
 372         if (size <= OV5_INDX(OV5_MMU_SUPPORT)) {
 373                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 374                 return;
 375         }
 376
 377         /* Check for supported configuration */
 378         mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] &
 379                         OV5_FEAT(OV5_MMU_SUPPORT);
 380         if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) {
 381                 /* Hypervisor only supports radix - check enabled && GTSE */
 382                 if (!early_radix_enabled()) {
 383                         pr_warn("WARNING: Ignoring cmdline option disable_radix\n");
 384                 }
 385                 if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] &
 386                                                 OV5_FEAT(OV5_RADIX_GTSE))) {
 387                         pr_warn("WARNING: Hypervisor doesn't support RADIX with GTSE\n");
 388                 }
 389                 /* Do radix anyway - the hypervisor said we had to */
 390                 cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX;
 391         } else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) {
 392                 /* Hypervisor only supports hash - disable radix */
 393                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 394         }
 395 }
 396
 397 void __init mmu_early_init_devtree(void)
 398 {
 399         /* Disable radix mode based on kernel command line. */
 400         if (disable_radix)
 401                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 402
 403         /*
 404          * Check /chosen/ibm,architecture-vec-5 if running as a guest.
 405          * When running bare-metal, we can use radix if we like
 406          * even though the ibm,architecture-vec-5 property created by
 407          * skiboot doesn't have the necessary bits set.
 408          */
 409         if (!(mfmsr() & MSR_HV))
 410                 early_check_vec5();
 411
 412         if (early_radix_enabled())
 413                 radix__early_init_devtree();
 414         else
 415                 hash__early_init_devtree();
 416 }
 417 #endif /* CONFIG_PPC_STD_MMU_64 */