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 #include <linux/memremap.h>
  48
  49 #include <asm/pgalloc.h>
  50 #include <asm/page.h>
  51 #include <asm/prom.h>
  52 #include <asm/rtas.h>
  53 #include <asm/io.h>
  54 #include <asm/mmu_context.h>
  55 #include <asm/pgtable.h>
  56 #include <asm/mmu.h>
  57 #include <linux/uaccess.h>
  58 #include <asm/smp.h>
  59 #include <asm/machdep.h>
  60 #include <asm/tlb.h>
  61 #include <asm/eeh.h>
  62 #include <asm/processor.h>
  63 #include <asm/mmzone.h>
  64 #include <asm/cputable.h>
  65 #include <asm/sections.h>
  66 #include <asm/iommu.h>
  67 #include <asm/vdso.h>
  68
  69 #include <mm/mmu_decl.h>
  70
  71 phys_addr_t memstart_addr = ~0;
  72 EXPORT_SYMBOL_GPL(memstart_addr);
  73 phys_addr_t kernstart_addr;
  74 EXPORT_SYMBOL_GPL(kernstart_addr);
  75
  76 #ifdef CONFIG_SPARSEMEM_VMEMMAP
  77 /*
  78  * Given an address within the vmemmap, determine the pfn of the page that
  79  * represents the start of the section it is within.  Note that we have to
  80  * do this by hand as the proffered address may not be correctly aligned.
  81  * Subtraction of non-aligned pointers produces undefined results.
  82  */
  83 static unsigned long __meminit vmemmap_section_start(unsigned long page)
  84 {
  85         unsigned long offset = page - ((unsigned long)(vmemmap));
  86
  87         /* Return the pfn of the start of the section. */
  88         return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
  89 }
  90
  91 /*
  92  * Check if this vmemmap page is already initialised.  If any section
  93  * which overlaps this vmemmap page is initialised then this page is
  94  * initialised already.
  95  */
  96 static int __meminit vmemmap_populated(unsigned long start, int page_size)
  97 {
  98         unsigned long end = start + page_size;
  99         start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
 100
 101         for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
 102                 if (pfn_valid(page_to_pfn((struct page *)start)))
 103                         return 1;
 104
 105         return 0;
 106 }
 107
 108 /*
 109  * vmemmap virtual address space management does not have a traditonal page
 110  * table to track which virtual struct pages are backed by physical mapping.
 111  * The virtual to physical mappings are tracked in a simple linked list
 112  * format. 'vmemmap_list' maintains the entire vmemmap physical mapping at
 113  * all times where as the 'next' list maintains the available
 114  * vmemmap_backing structures which have been deleted from the
 115  * 'vmemmap_global' list during system runtime (memory hotplug remove
 116  * operation). The freed 'vmemmap_backing' structures are reused later when
 117  * new requests come in without allocating fresh memory. This pointer also
 118  * tracks the allocated 'vmemmap_backing' structures as we allocate one
 119  * full page memory at a time when we dont have any.
 120  */
 121 struct vmemmap_backing *vmemmap_list;
 122 static struct vmemmap_backing *next;
 123
 124 /*
 125  * The same pointer 'next' tracks individual chunks inside the allocated
 126  * full page during the boot time and again tracks the freeed nodes during
 127  * runtime. It is racy but it does not happen as they are separated by the
 128  * boot process. Will create problem if some how we have memory hotplug
 129  * operation during boot !!
 130  */
 131 static int num_left;
 132 static int num_freed;
 133
 134 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
 135 {
 136         struct vmemmap_backing *vmem_back;
 137         /* get from freed entries first */
 138         if (num_freed) {
 139                 num_freed--;
 140                 vmem_back = next;
 141                 next = next->list;
 142
 143                 return vmem_back;
 144         }
 145
 146         /* allocate a page when required and hand out chunks */
 147         if (!num_left) {
 148                 next = vmemmap_alloc_block(PAGE_SIZE, node);
 149                 if (unlikely(!next)) {
 150                         WARN_ON(1);
 151                         return NULL;
 152                 }
 153                 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
 154         }
 155
 156         num_left--;
 157
 158         return next++;
 159 }
 160
 161 static __meminit void vmemmap_list_populate(unsigned long phys,
 162                                             unsigned long start,
 163                                             int node)
 164 {
 165         struct vmemmap_backing *vmem_back;
 166
 167         vmem_back = vmemmap_list_alloc(node);
 168         if (unlikely(!vmem_back)) {
 169                 WARN_ON(1);
 170                 return;
 171         }
 172
 173         vmem_back->phys = phys;
 174         vmem_back->virt_addr = start;
 175         vmem_back->list = vmemmap_list;
 176
 177         vmemmap_list = vmem_back;
 178 }
 179
 180 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
 181                 struct vmem_altmap *altmap)
 182 {
 183         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 184
 185         /* Align to the page size of the linear mapping. */
 186         start = _ALIGN_DOWN(start, page_size);
 187
 188         pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
 189
 190         for (; start < end; start += page_size) {
 191                 void *p = NULL;
 192                 int rc;
 193
 194                 if (vmemmap_populated(start, page_size))
 195                         continue;
 196
 197                 /*
 198                  * Allocate from the altmap first if we have one. This may
 199                  * fail due to alignment issues when using 16MB hugepages, so
 200                  * fall back to system memory if the altmap allocation fail.
 201                  */
 202                 if (altmap)
 203                         p = altmap_alloc_block_buf(page_size, altmap);
 204                 if (!p)
 205                         p = vmemmap_alloc_block_buf(page_size, node);
 206                 if (!p)
 207                         return -ENOMEM;
 208
 209                 vmemmap_list_populate(__pa(p), start, node);
 210
 211                 pr_debug("      * %016lx..%016lx allocated at %p\n",
 212                          start, start + page_size, p);
 213
 214                 rc = vmemmap_create_mapping(start, page_size, __pa(p));
 215                 if (rc < 0) {
 216                         pr_warn("%s: Unable to create vmemmap mapping: %d\n",
 217                                 __func__, rc);
 218                         return -EFAULT;
 219                 }
 220         }
 221
 222         return 0;
 223 }
 224
 225 #ifdef CONFIG_MEMORY_HOTPLUG
 226 static unsigned long vmemmap_list_free(unsigned long start)
 227 {
 228         struct vmemmap_backing *vmem_back, *vmem_back_prev;
 229
 230         vmem_back_prev = vmem_back = vmemmap_list;
 231
 232         /* look for it with prev pointer recorded */
 233         for (; vmem_back; vmem_back = vmem_back->list) {
 234                 if (vmem_back->virt_addr == start)
 235                         break;
 236                 vmem_back_prev = vmem_back;
 237         }
 238
 239         if (unlikely(!vmem_back)) {
 240                 WARN_ON(1);
 241                 return 0;
 242         }
 243
 244         /* remove it from vmemmap_list */
 245         if (vmem_back == vmemmap_list) /* remove head */
 246                 vmemmap_list = vmem_back->list;
 247         else
 248                 vmem_back_prev->list = vmem_back->list;
 249
 250         /* next point to this freed entry */
 251         vmem_back->list = next;
 252         next = vmem_back;
 253         num_freed++;
 254
 255         return vmem_back->phys;
 256 }
 257
 258 void __ref vmemmap_free(unsigned long start, unsigned long end,
 259                 struct vmem_altmap *altmap)
 260 {
 261         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 262         unsigned long page_order = get_order(page_size);
 263         unsigned long alt_start = ~0, alt_end = ~0;
 264         unsigned long base_pfn;
 265
 266         start = _ALIGN_DOWN(start, page_size);
 267         if (altmap) {
 268                 alt_start = altmap->base_pfn;
 269                 alt_end = altmap->base_pfn + altmap->reserve +
 270                           altmap->free + altmap->alloc + altmap->align;
 271         }
 272
 273         pr_debug("vmemmap_free %lx...%lx\n", start, end);
 274
 275         for (; start < end; start += page_size) {
 276                 unsigned long nr_pages, addr;
 277                 struct page *page;
 278
 279                 /*
 280                  * the section has already be marked as invalid, so
 281                  * vmemmap_populated() true means some other sections still
 282                  * in this page, so skip it.
 283                  */
 284                 if (vmemmap_populated(start, page_size))
 285                         continue;
 286
 287                 addr = vmemmap_list_free(start);
 288                 if (!addr)
 289                         continue;
 290
 291                 page = pfn_to_page(addr >> PAGE_SHIFT);
 292                 nr_pages = 1 << page_order;
 293                 base_pfn = PHYS_PFN(addr);
 294
 295                 if (base_pfn >= alt_start && base_pfn < alt_end) {
 296                         vmem_altmap_free(altmap, nr_pages);
 297                 } else if (PageReserved(page)) {
 298                         /* allocated from bootmem */
 299                         if (page_size < PAGE_SIZE) {
 300                                 /*
 301                                  * this shouldn't happen, but if it is
 302                                  * the case, leave the memory there
 303                                  */
 304                                 WARN_ON_ONCE(1);
 305                         } else {
 306                                 while (nr_pages--)
 307                                         free_reserved_page(page++);
 308                         }
 309                 } else {
 310                         free_pages((unsigned long)(__va(addr)), page_order);
 311                 }
 312
 313                 vmemmap_remove_mapping(start, page_size);
 314         }
 315 }
 316 #endif
 317 void register_page_bootmem_memmap(unsigned long section_nr,
 318                                   struct page *start_page, unsigned long size)
 319 {
 320 }
 321
 322 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 323
 324 #ifdef CONFIG_PPC_BOOK3S_64
 325 static bool disable_radix = !IS_ENABLED(CONFIG_PPC_RADIX_MMU_DEFAULT);
 326
 327 static int __init parse_disable_radix(char *p)
 328 {
 329         bool val;
 330
 331         if (!p)
 332                 val = true;
 333         else if (kstrtobool(p, &val))
 334                 return -EINVAL;
 335
 336         disable_radix = val;
 337
 338         return 0;
 339 }
 340 early_param("disable_radix", parse_disable_radix);
 341
 342 /*
 343  * If we're running under a hypervisor, we need to check the contents of
 344  * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do
 345  * radix.  If not, we clear the radix feature bit so we fall back to hash.
 346  */
 347 static void __init early_check_vec5(void)
 348 {
 349         unsigned long root, chosen;
 350         int size;
 351         const u8 *vec5;
 352         u8 mmu_supported;
 353
 354         root = of_get_flat_dt_root();
 355         chosen = of_get_flat_dt_subnode_by_name(root, "chosen");
 356         if (chosen == -FDT_ERR_NOTFOUND) {
 357                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 358                 return;
 359         }
 360         vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size);
 361         if (!vec5) {
 362                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 363                 return;
 364         }
 365         if (size <= OV5_INDX(OV5_MMU_SUPPORT)) {
 366                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 367                 return;
 368         }
 369
 370         /* Check for supported configuration */
 371         mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] &
 372                         OV5_FEAT(OV5_MMU_SUPPORT);
 373         if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) {
 374                 /* Hypervisor only supports radix - check enabled && GTSE */
 375                 if (!early_radix_enabled()) {
 376                         pr_warn("WARNING: Ignoring cmdline option disable_radix\n");
 377                 }
 378                 if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] &
 379                                                 OV5_FEAT(OV5_RADIX_GTSE))) {
 380                         pr_warn("WARNING: Hypervisor doesn't support RADIX with GTSE\n");
 381                 }
 382                 /* Do radix anyway - the hypervisor said we had to */
 383                 cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX;
 384         } else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) {
 385                 /* Hypervisor only supports hash - disable radix */
 386                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 387         }
 388 }
 389
 390 void __init mmu_early_init_devtree(void)
 391 {
 392         /* Disable radix mode based on kernel command line. */
 393         if (disable_radix)
 394                 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
 395
 396         /*
 397          * Check /chosen/ibm,architecture-vec-5 if running as a guest.
 398          * When running bare-metal, we can use radix if we like
 399          * even though the ibm,architecture-vec-5 property created by
 400          * skiboot doesn't have the necessary bits set.
 401          */
 402         if (!(mfmsr() & MSR_HV))
 403                 early_check_vec5();
 404
 405         if (early_radix_enabled())
 406                 radix__early_init_devtree();
 407         else
 408                 hash__early_init_devtree();
 409 }
 410 #endif /* CONFIG_PPC_BOOK3S_64 */