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 #endif /* CONFIG_PPC_STD_MMU_64 */
  75
  76 phys_addr_t memstart_addr = ~0;
  77 EXPORT_SYMBOL_GPL(memstart_addr);
  78 phys_addr_t kernstart_addr;
  79 EXPORT_SYMBOL_GPL(kernstart_addr);
  80
  81 #ifdef CONFIG_SPARSEMEM_VMEMMAP
  82 /*
  83  * Given an address within the vmemmap, determine the pfn of the page that
  84  * represents the start of the section it is within.  Note that we have to
  85  * do this by hand as the proffered address may not be correctly aligned.
  86  * Subtraction of non-aligned pointers produces undefined results.
  87  */
  88 static unsigned long __meminit vmemmap_section_start(unsigned long page)
  89 {
  90         unsigned long offset = page - ((unsigned long)(vmemmap));
  91
  92         /* Return the pfn of the start of the section. */
  93         return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
  94 }
  95
  96 /*
  97  * Check if this vmemmap page is already initialised.  If any section
  98  * which overlaps this vmemmap page is initialised then this page is
  99  * initialised already.
 100  */
 101 static int __meminit vmemmap_populated(unsigned long start, int page_size)
 102 {
 103         unsigned long end = start + page_size;
 104         start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
 105
 106         for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
 107                 if (pfn_valid(page_to_pfn((struct page *)start)))
 108                         return 1;
 109
 110         return 0;
 111 }
 112
 113 struct vmemmap_backing *vmemmap_list;
 114 static struct vmemmap_backing *next;
 115 static int num_left;
 116 static int num_freed;
 117
 118 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
 119 {
 120         struct vmemmap_backing *vmem_back;
 121         /* get from freed entries first */
 122         if (num_freed) {
 123                 num_freed--;
 124                 vmem_back = next;
 125                 next = next->list;
 126
 127                 return vmem_back;
 128         }
 129
 130         /* allocate a page when required and hand out chunks */
 131         if (!num_left) {
 132                 next = vmemmap_alloc_block(PAGE_SIZE, node);
 133                 if (unlikely(!next)) {
 134                         WARN_ON(1);
 135                         return NULL;
 136                 }
 137                 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
 138         }
 139
 140         num_left--;
 141
 142         return next++;
 143 }
 144
 145 static __meminit void vmemmap_list_populate(unsigned long phys,
 146                                             unsigned long start,
 147                                             int node)
 148 {
 149         struct vmemmap_backing *vmem_back;
 150
 151         vmem_back = vmemmap_list_alloc(node);
 152         if (unlikely(!vmem_back)) {
 153                 WARN_ON(1);
 154                 return;
 155         }
 156
 157         vmem_back->phys = phys;
 158         vmem_back->virt_addr = start;
 159         vmem_back->list = vmemmap_list;
 160
 161         vmemmap_list = vmem_back;
 162 }
 163
 164 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 165 {
 166         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 167
 168         /* Align to the page size of the linear mapping. */
 169         start = _ALIGN_DOWN(start, page_size);
 170
 171         pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
 172
 173         for (; start < end; start += page_size) {
 174                 void *p;
 175                 int rc;
 176
 177                 if (vmemmap_populated(start, page_size))
 178                         continue;
 179
 180                 p = vmemmap_alloc_block(page_size, node);
 181                 if (!p)
 182                         return -ENOMEM;
 183
 184                 vmemmap_list_populate(__pa(p), start, node);
 185
 186                 pr_debug("      * %016lx..%016lx allocated at %p\n",
 187                          start, start + page_size, p);
 188
 189                 rc = vmemmap_create_mapping(start, page_size, __pa(p));
 190                 if (rc < 0) {
 191                         pr_warning(
 192                                 "vmemmap_populate: Unable to create vmemmap mapping: %d\n",
 193                                 rc);
 194                         return -EFAULT;
 195                 }
 196         }
 197
 198         return 0;
 199 }
 200
 201 #ifdef CONFIG_MEMORY_HOTPLUG
 202 static unsigned long vmemmap_list_free(unsigned long start)
 203 {
 204         struct vmemmap_backing *vmem_back, *vmem_back_prev;
 205
 206         vmem_back_prev = vmem_back = vmemmap_list;
 207
 208         /* look for it with prev pointer recorded */
 209         for (; vmem_back; vmem_back = vmem_back->list) {
 210                 if (vmem_back->virt_addr == start)
 211                         break;
 212                 vmem_back_prev = vmem_back;
 213         }
 214
 215         if (unlikely(!vmem_back)) {
 216                 WARN_ON(1);
 217                 return 0;
 218         }
 219
 220         /* remove it from vmemmap_list */
 221         if (vmem_back == vmemmap_list) /* remove head */
 222                 vmemmap_list = vmem_back->list;
 223         else
 224                 vmem_back_prev->list = vmem_back->list;
 225
 226         /* next point to this freed entry */
 227         vmem_back->list = next;
 228         next = vmem_back;
 229         num_freed++;
 230
 231         return vmem_back->phys;
 232 }
 233
 234 void __ref vmemmap_free(unsigned long start, unsigned long end)
 235 {
 236         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 237
 238         start = _ALIGN_DOWN(start, page_size);
 239
 240         pr_debug("vmemmap_free %lx...%lx\n", start, end);
 241
 242         for (; start < end; start += page_size) {
 243                 unsigned long addr;
 244
 245                 /*
 246                  * the section has already be marked as invalid, so
 247                  * vmemmap_populated() true means some other sections still
 248                  * in this page, so skip it.
 249                  */
 250                 if (vmemmap_populated(start, page_size))
 251                         continue;
 252
 253                 addr = vmemmap_list_free(start);
 254                 if (addr) {
 255                         struct page *page = pfn_to_page(addr >> PAGE_SHIFT);
 256
 257                         if (PageReserved(page)) {
 258                                 /* allocated from bootmem */
 259                                 if (page_size < PAGE_SIZE) {
 260                                         /*
 261                                          * this shouldn't happen, but if it is
 262                                          * the case, leave the memory there
 263                                          */
 264                                         WARN_ON_ONCE(1);
 265                                 } else {
 266                                         unsigned int nr_pages =
 267                                                 1 << get_order(page_size);
 268                                         while (nr_pages--)
 269                                                 free_reserved_page(page++);
 270                                 }
 271                         } else
 272                                 free_pages((unsigned long)(__va(addr)),
 273                                                         get_order(page_size));
 274
 275                         vmemmap_remove_mapping(start, page_size);
 276                 }
 277         }
 278 }
 279 #endif
 280 void register_page_bootmem_memmap(unsigned long section_nr,
 281                                   struct page *start_page, unsigned long size)
 282 {
 283 }
 284
 285 /*
 286  * We do not have access to the sparsemem vmemmap, so we fallback to
 287  * walking the list of sparsemem blocks which we already maintain for
 288  * the sake of crashdump. In the long run, we might want to maintain
 289  * a tree if performance of that linear walk becomes a problem.
 290  *
 291  * realmode_pfn_to_page functions can fail due to:
 292  * 1) As real sparsemem blocks do not lay in RAM continously (they
 293  * are in virtual address space which is not available in the real mode),
 294  * the requested page struct can be split between blocks so get_page/put_page
 295  * may fail.
 296  * 2) When huge pages are used, the get_page/put_page API will fail
 297  * in real mode as the linked addresses in the page struct are virtual
 298  * too.
 299  */
 300 struct page *realmode_pfn_to_page(unsigned long pfn)
 301 {
 302         struct vmemmap_backing *vmem_back;
 303         struct page *page;
 304         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 305         unsigned long pg_va = (unsigned long) pfn_to_page(pfn);
 306
 307         for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) {
 308                 if (pg_va < vmem_back->virt_addr)
 309                         continue;
 310
 311                 /* After vmemmap_list entry free is possible, need check all */
 312                 if ((pg_va + sizeof(struct page)) <=
 313                                 (vmem_back->virt_addr + page_size)) {
 314                         page = (struct page *) (vmem_back->phys + pg_va -
 315                                 vmem_back->virt_addr);
 316                         return page;
 317                 }
 318         }
 319
 320         /* Probably that page struct is split between real pages */
 321         return NULL;
 322 }
 323 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
 324
 325 #elif defined(CONFIG_FLATMEM)
 326
 327 struct page *realmode_pfn_to_page(unsigned long pfn)
 328 {
 329         struct page *page = pfn_to_page(pfn);
 330         return page;
 331 }
 332 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
 333
 334 #endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */
 335
 336 #ifdef CONFIG_PPC_STD_MMU_64
 337 static bool disable_radix;
 338 static int __init parse_disable_radix(char *p)
 339 {
 340         disable_radix = true;
 341         return 0;
 342 }
 343 early_param("disable_radix", parse_disable_radix);
 344
 345 /*
 346  * If we're running under a hypervisor, we need to check the contents of
 347  * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do
 348  * radix.  If not, we clear the radix feature bit so we fall back to hash.
 349  */
 350 static void early_check_vec5(void)
 351 {
 352         unsigned long root, chosen;
 353         int size;
 354         const u8 *vec5;
 355         u8 mmu_supported;
 356
 357         root = of_get_flat_dt_root();
 358         chosen = of_get_flat_dt_subnode_by_name(root, "chosen");
 359         if (chosen == -FDT_ERR_NOTFOUND) {
 360                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 361                 return;
 362         }
 363         vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size);
 364         if (!vec5) {
 365                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 366                 return;
 367         }
 368         if (size <= OV5_INDX(OV5_MMU_SUPPORT)) {
 369                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 370                 return;
 371         }
 372
 373         /* Check for supported configuration */
 374         mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] &
 375                         OV5_FEAT(OV5_MMU_SUPPORT);
 376         if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) {
 377                 /* Hypervisor only supports radix - check enabled && GTSE */
 378                 if (!early_radix_enabled()) {
 379                         pr_warn("WARNING: Ignoring cmdline option disable_radix\n");
 380                 }
 381                 if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] &
 382                                                 OV5_FEAT(OV5_RADIX_GTSE))) {
 383                         pr_warn("WARNING: Hypervisor doesn't support RADIX with GTSE\n");
 384                 }
 385                 /* Do radix anyway - the hypervisor said we had to */
 386                 cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX;
 387         } else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) {
 388                 /* Hypervisor only supports hash - disable radix */
 389                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 390         }
 391 }
 392
 393 void __init mmu_early_init_devtree(void)
 394 {
 395         /* Disable radix mode based on kernel command line. */
 396         if (disable_radix)
 397                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 398
 399         /*
 400          * Check /chosen/ibm,architecture-vec-5 if running as a guest.
 401          * When running bare-metal, we can use radix if we like
 402          * even though the ibm,architecture-vec-5 property created by
 403          * skiboot doesn't have the necessary bits set.
 404          */
 405         if (!(mfmsr() & MSR_HV))
 406                 early_check_vec5();
 407
 408         if (early_radix_enabled())
 409                 radix__early_init_devtree();
 410         else
 411                 hash__early_init_devtree();
 412 }
 413 #endif /* CONFIG_PPC_STD_MMU_64 */