Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[linux-2.6] / arch / powerpc / mm / mem.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  *  PPC44x/36-bit changes by Matt Porter (mporter@mvista.com)
9  *
10  *  Derived from "arch/i386/mm/init.c"
11  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
12  *
13  *  This program is free software; you can redistribute it and/or
14  *  modify it under the terms of the GNU General Public License
15  *  as published by the Free Software Foundation; either version
16  *  2 of the License, or (at your option) any later version.
17  *
18  */
19
20 #include <linux/module.h>
21 #include <linux/sched.h>
22 #include <linux/kernel.h>
23 #include <linux/errno.h>
24 #include <linux/string.h>
25 #include <linux/types.h>
26 #include <linux/mm.h>
27 #include <linux/stddef.h>
28 #include <linux/init.h>
29 #include <linux/bootmem.h>
30 #include <linux/highmem.h>
31 #include <linux/initrd.h>
32 #include <linux/pagemap.h>
33 #include <linux/suspend.h>
34 #include <linux/lmb.h>
35
36 #include <asm/pgalloc.h>
37 #include <asm/prom.h>
38 #include <asm/io.h>
39 #include <asm/mmu_context.h>
40 #include <asm/pgtable.h>
41 #include <asm/mmu.h>
42 #include <asm/smp.h>
43 #include <asm/machdep.h>
44 #include <asm/btext.h>
45 #include <asm/tlb.h>
46 #include <asm/sections.h>
47 #include <asm/vdso.h>
48 #include <asm/fixmap.h>
49
50 #include "mmu_decl.h"
51
52 #ifndef CPU_FTR_COHERENT_ICACHE
53 #define CPU_FTR_COHERENT_ICACHE 0       /* XXX for now */
54 #define CPU_FTR_NOEXECUTE       0
55 #endif
56
57 int init_bootmem_done;
58 int mem_init_done;
59 unsigned long memory_limit;
60
61 #ifdef CONFIG_HIGHMEM
62 pte_t *kmap_pte;
63 pgprot_t kmap_prot;
64
65 EXPORT_SYMBOL(kmap_prot);
66 EXPORT_SYMBOL(kmap_pte);
67
68 static inline pte_t *virt_to_kpte(unsigned long vaddr)
69 {
70         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
71                         vaddr), vaddr), vaddr);
72 }
73 #endif
74
75 int page_is_ram(unsigned long pfn)
76 {
77         unsigned long paddr = (pfn << PAGE_SHIFT);
78
79 #ifndef CONFIG_PPC64    /* XXX for now */
80         return paddr < __pa(high_memory);
81 #else
82         int i;
83         for (i=0; i < lmb.memory.cnt; i++) {
84                 unsigned long base;
85
86                 base = lmb.memory.region[i].base;
87
88                 if ((paddr >= base) &&
89                         (paddr < (base + lmb.memory.region[i].size))) {
90                         return 1;
91                 }
92         }
93
94         return 0;
95 #endif
96 }
97
98 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
99                               unsigned long size, pgprot_t vma_prot)
100 {
101         if (ppc_md.phys_mem_access_prot)
102                 return ppc_md.phys_mem_access_prot(file, pfn, size, vma_prot);
103
104         if (!page_is_ram(pfn))
105                 vma_prot = __pgprot(pgprot_val(vma_prot)
106                                     | _PAGE_GUARDED | _PAGE_NO_CACHE);
107         return vma_prot;
108 }
109 EXPORT_SYMBOL(phys_mem_access_prot);
110
111 #ifdef CONFIG_MEMORY_HOTPLUG
112
113 #ifdef CONFIG_NUMA
114 int memory_add_physaddr_to_nid(u64 start)
115 {
116         return hot_add_scn_to_nid(start);
117 }
118 #endif
119
120 int arch_add_memory(int nid, u64 start, u64 size)
121 {
122         struct pglist_data *pgdata;
123         struct zone *zone;
124         unsigned long start_pfn = start >> PAGE_SHIFT;
125         unsigned long nr_pages = size >> PAGE_SHIFT;
126
127         pgdata = NODE_DATA(nid);
128
129         start = (unsigned long)__va(start);
130         create_section_mapping(start, start + size);
131
132         /* this should work for most non-highmem platforms */
133         zone = pgdata->node_zones;
134
135         return __add_pages(zone, start_pfn, nr_pages);
136 }
137
138 #ifdef CONFIG_MEMORY_HOTREMOVE
139 int remove_memory(u64 start, u64 size)
140 {
141         unsigned long start_pfn, end_pfn;
142         int ret;
143
144         start_pfn = start >> PAGE_SHIFT;
145         end_pfn = start_pfn + (size >> PAGE_SHIFT);
146         ret = offline_pages(start_pfn, end_pfn, 120 * HZ);
147         if (ret)
148                 goto out;
149         /* Arch-specific calls go here - next patch */
150 out:
151         return ret;
152 }
153 #endif /* CONFIG_MEMORY_HOTREMOVE */
154 #endif /* CONFIG_MEMORY_HOTPLUG */
155
156 /*
157  * walk_memory_resource() needs to make sure there is no holes in a given
158  * memory range.  PPC64 does not maintain the memory layout in /proc/iomem.
159  * Instead it maintains it in lmb.memory structures.  Walk through the
160  * memory regions, find holes and callback for contiguous regions.
161  */
162 int
163 walk_memory_resource(unsigned long start_pfn, unsigned long nr_pages, void *arg,
164                         int (*func)(unsigned long, unsigned long, void *))
165 {
166         struct lmb_property res;
167         unsigned long pfn, len;
168         u64 end;
169         int ret = -1;
170
171         res.base = (u64) start_pfn << PAGE_SHIFT;
172         res.size = (u64) nr_pages << PAGE_SHIFT;
173
174         end = res.base + res.size - 1;
175         while ((res.base < end) && (lmb_find(&res) >= 0)) {
176                 pfn = (unsigned long)(res.base >> PAGE_SHIFT);
177                 len = (unsigned long)(res.size >> PAGE_SHIFT);
178                 ret = (*func)(pfn, len, arg);
179                 if (ret)
180                         break;
181                 res.base += (res.size + 1);
182                 res.size = (end - res.base + 1);
183         }
184         return ret;
185 }
186 EXPORT_SYMBOL_GPL(walk_memory_resource);
187
188 void show_mem(void)
189 {
190         unsigned long total = 0, reserved = 0;
191         unsigned long shared = 0, cached = 0;
192         unsigned long highmem = 0;
193         struct page *page;
194         pg_data_t *pgdat;
195         unsigned long i;
196
197         printk("Mem-info:\n");
198         show_free_areas();
199         for_each_online_pgdat(pgdat) {
200                 unsigned long flags;
201                 pgdat_resize_lock(pgdat, &flags);
202                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
203                         if (!pfn_valid(pgdat->node_start_pfn + i))
204                                 continue;
205                         page = pgdat_page_nr(pgdat, i);
206                         total++;
207                         if (PageHighMem(page))
208                                 highmem++;
209                         if (PageReserved(page))
210                                 reserved++;
211                         else if (PageSwapCache(page))
212                                 cached++;
213                         else if (page_count(page))
214                                 shared += page_count(page) - 1;
215                 }
216                 pgdat_resize_unlock(pgdat, &flags);
217         }
218         printk("%ld pages of RAM\n", total);
219 #ifdef CONFIG_HIGHMEM
220         printk("%ld pages of HIGHMEM\n", highmem);
221 #endif
222         printk("%ld reserved pages\n", reserved);
223         printk("%ld pages shared\n", shared);
224         printk("%ld pages swap cached\n", cached);
225 }
226
227 /*
228  * Initialize the bootmem system and give it all the memory we
229  * have available.  If we are using highmem, we only put the
230  * lowmem into the bootmem system.
231  */
232 #ifndef CONFIG_NEED_MULTIPLE_NODES
233 void __init do_init_bootmem(void)
234 {
235         unsigned long i;
236         unsigned long start, bootmap_pages;
237         unsigned long total_pages;
238         int boot_mapsize;
239
240         max_low_pfn = max_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
241         total_pages = (lmb_end_of_DRAM() - memstart_addr) >> PAGE_SHIFT;
242 #ifdef CONFIG_HIGHMEM
243         total_pages = total_lowmem >> PAGE_SHIFT;
244         max_low_pfn = lowmem_end_addr >> PAGE_SHIFT;
245 #endif
246
247         /*
248          * Find an area to use for the bootmem bitmap.  Calculate the size of
249          * bitmap required as (Total Memory) / PAGE_SIZE / BITS_PER_BYTE.
250          * Add 1 additional page in case the address isn't page-aligned.
251          */
252         bootmap_pages = bootmem_bootmap_pages(total_pages);
253
254         start = lmb_alloc(bootmap_pages << PAGE_SHIFT, PAGE_SIZE);
255
256         min_low_pfn = MEMORY_START >> PAGE_SHIFT;
257         boot_mapsize = init_bootmem_node(NODE_DATA(0), start >> PAGE_SHIFT, min_low_pfn, max_low_pfn);
258
259         /* Add active regions with valid PFNs */
260         for (i = 0; i < lmb.memory.cnt; i++) {
261                 unsigned long start_pfn, end_pfn;
262                 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
263                 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
264                 add_active_range(0, start_pfn, end_pfn);
265         }
266
267         /* Add all physical memory to the bootmem map, mark each area
268          * present.
269          */
270 #ifdef CONFIG_HIGHMEM
271         free_bootmem_with_active_regions(0, lowmem_end_addr >> PAGE_SHIFT);
272
273         /* reserve the sections we're already using */
274         for (i = 0; i < lmb.reserved.cnt; i++) {
275                 unsigned long addr = lmb.reserved.region[i].base +
276                                      lmb_size_bytes(&lmb.reserved, i) - 1;
277                 if (addr < lowmem_end_addr)
278                         reserve_bootmem(lmb.reserved.region[i].base,
279                                         lmb_size_bytes(&lmb.reserved, i),
280                                         BOOTMEM_DEFAULT);
281                 else if (lmb.reserved.region[i].base < lowmem_end_addr) {
282                         unsigned long adjusted_size = lowmem_end_addr -
283                                       lmb.reserved.region[i].base;
284                         reserve_bootmem(lmb.reserved.region[i].base,
285                                         adjusted_size, BOOTMEM_DEFAULT);
286                 }
287         }
288 #else
289         free_bootmem_with_active_regions(0, max_pfn);
290
291         /* reserve the sections we're already using */
292         for (i = 0; i < lmb.reserved.cnt; i++)
293                 reserve_bootmem(lmb.reserved.region[i].base,
294                                 lmb_size_bytes(&lmb.reserved, i),
295                                 BOOTMEM_DEFAULT);
296
297 #endif
298         /* XXX need to clip this if using highmem? */
299         sparse_memory_present_with_active_regions(0);
300
301         init_bootmem_done = 1;
302 }
303
304 /* mark pages that don't exist as nosave */
305 static int __init mark_nonram_nosave(void)
306 {
307         unsigned long lmb_next_region_start_pfn,
308                       lmb_region_max_pfn;
309         int i;
310
311         for (i = 0; i < lmb.memory.cnt - 1; i++) {
312                 lmb_region_max_pfn =
313                         (lmb.memory.region[i].base >> PAGE_SHIFT) +
314                         (lmb.memory.region[i].size >> PAGE_SHIFT);
315                 lmb_next_region_start_pfn =
316                         lmb.memory.region[i+1].base >> PAGE_SHIFT;
317
318                 if (lmb_region_max_pfn < lmb_next_region_start_pfn)
319                         register_nosave_region(lmb_region_max_pfn,
320                                                lmb_next_region_start_pfn);
321         }
322
323         return 0;
324 }
325
326 /*
327  * paging_init() sets up the page tables - in fact we've already done this.
328  */
329 void __init paging_init(void)
330 {
331         unsigned long total_ram = lmb_phys_mem_size();
332         unsigned long top_of_ram = lmb_end_of_DRAM();
333         unsigned long max_zone_pfns[MAX_NR_ZONES];
334
335 #ifdef CONFIG_PPC32
336         unsigned long v = __fix_to_virt(__end_of_fixed_addresses - 1);
337         unsigned long end = __fix_to_virt(FIX_HOLE);
338
339         for (; v < end; v += PAGE_SIZE)
340                 map_page(v, 0, 0); /* XXX gross */
341 #endif
342
343 #ifdef CONFIG_HIGHMEM
344         map_page(PKMAP_BASE, 0, 0);     /* XXX gross */
345         pkmap_page_table = virt_to_kpte(PKMAP_BASE);
346
347         kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
348         kmap_prot = PAGE_KERNEL;
349 #endif /* CONFIG_HIGHMEM */
350
351         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
352                top_of_ram, total_ram);
353         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
354                (top_of_ram - total_ram) >> 20);
355         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
356 #ifdef CONFIG_HIGHMEM
357         max_zone_pfns[ZONE_DMA] = lowmem_end_addr >> PAGE_SHIFT;
358         max_zone_pfns[ZONE_HIGHMEM] = top_of_ram >> PAGE_SHIFT;
359 #else
360         max_zone_pfns[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
361 #endif
362         free_area_init_nodes(max_zone_pfns);
363
364         mark_nonram_nosave();
365 }
366 #endif /* ! CONFIG_NEED_MULTIPLE_NODES */
367
368 void __init mem_init(void)
369 {
370 #ifdef CONFIG_NEED_MULTIPLE_NODES
371         int nid;
372 #endif
373         pg_data_t *pgdat;
374         unsigned long i;
375         struct page *page;
376         unsigned long reservedpages = 0, codesize, initsize, datasize, bsssize;
377
378         num_physpages = lmb.memory.size >> PAGE_SHIFT;
379         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE);
380
381 #ifdef CONFIG_NEED_MULTIPLE_NODES
382         for_each_online_node(nid) {
383                 if (NODE_DATA(nid)->node_spanned_pages != 0) {
384                         printk("freeing bootmem node %d\n", nid);
385                         totalram_pages +=
386                                 free_all_bootmem_node(NODE_DATA(nid));
387                 }
388         }
389 #else
390         max_mapnr = max_pfn;
391         totalram_pages += free_all_bootmem();
392 #endif
393         for_each_online_pgdat(pgdat) {
394                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
395                         if (!pfn_valid(pgdat->node_start_pfn + i))
396                                 continue;
397                         page = pgdat_page_nr(pgdat, i);
398                         if (PageReserved(page))
399                                 reservedpages++;
400                 }
401         }
402
403         codesize = (unsigned long)&_sdata - (unsigned long)&_stext;
404         datasize = (unsigned long)&_edata - (unsigned long)&_sdata;
405         initsize = (unsigned long)&__init_end - (unsigned long)&__init_begin;
406         bsssize = (unsigned long)&__bss_stop - (unsigned long)&__bss_start;
407
408 #ifdef CONFIG_HIGHMEM
409         {
410                 unsigned long pfn, highmem_mapnr;
411
412                 highmem_mapnr = lowmem_end_addr >> PAGE_SHIFT;
413                 for (pfn = highmem_mapnr; pfn < max_mapnr; ++pfn) {
414                         struct page *page = pfn_to_page(pfn);
415                         if (lmb_is_reserved(pfn << PAGE_SHIFT))
416                                 continue;
417                         ClearPageReserved(page);
418                         init_page_count(page);
419                         __free_page(page);
420                         totalhigh_pages++;
421                         reservedpages--;
422                 }
423                 totalram_pages += totalhigh_pages;
424                 printk(KERN_DEBUG "High memory: %luk\n",
425                        totalhigh_pages << (PAGE_SHIFT-10));
426         }
427 #endif /* CONFIG_HIGHMEM */
428
429         printk(KERN_INFO "Memory: %luk/%luk available (%luk kernel code, "
430                "%luk reserved, %luk data, %luk bss, %luk init)\n",
431                 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
432                 num_physpages << (PAGE_SHIFT-10),
433                 codesize >> 10,
434                 reservedpages << (PAGE_SHIFT-10),
435                 datasize >> 10,
436                 bsssize >> 10,
437                 initsize >> 10);
438
439         mem_init_done = 1;
440 }
441
442 /*
443  * This is called when a page has been modified by the kernel.
444  * It just marks the page as not i-cache clean.  We do the i-cache
445  * flush later when the page is given to a user process, if necessary.
446  */
447 void flush_dcache_page(struct page *page)
448 {
449         if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
450                 return;
451         /* avoid an atomic op if possible */
452         if (test_bit(PG_arch_1, &page->flags))
453                 clear_bit(PG_arch_1, &page->flags);
454 }
455 EXPORT_SYMBOL(flush_dcache_page);
456
457 void flush_dcache_icache_page(struct page *page)
458 {
459 #ifdef CONFIG_BOOKE
460         void *start = kmap_atomic(page, KM_PPC_SYNC_ICACHE);
461         __flush_dcache_icache(start);
462         kunmap_atomic(start, KM_PPC_SYNC_ICACHE);
463 #elif defined(CONFIG_8xx) || defined(CONFIG_PPC64)
464         /* On 8xx there is no need to kmap since highmem is not supported */
465         __flush_dcache_icache(page_address(page)); 
466 #else
467         __flush_dcache_icache_phys(page_to_pfn(page) << PAGE_SHIFT);
468 #endif
469
470 }
471 void clear_user_page(void *page, unsigned long vaddr, struct page *pg)
472 {
473         clear_page(page);
474
475         /*
476          * We shouldnt have to do this, but some versions of glibc
477          * require it (ld.so assumes zero filled pages are icache clean)
478          * - Anton
479          */
480         flush_dcache_page(pg);
481 }
482 EXPORT_SYMBOL(clear_user_page);
483
484 void copy_user_page(void *vto, void *vfrom, unsigned long vaddr,
485                     struct page *pg)
486 {
487         copy_page(vto, vfrom);
488
489         /*
490          * We should be able to use the following optimisation, however
491          * there are two problems.
492          * Firstly a bug in some versions of binutils meant PLT sections
493          * were not marked executable.
494          * Secondly the first word in the GOT section is blrl, used
495          * to establish the GOT address. Until recently the GOT was
496          * not marked executable.
497          * - Anton
498          */
499 #if 0
500         if (!vma->vm_file && ((vma->vm_flags & VM_EXEC) == 0))
501                 return;
502 #endif
503
504         flush_dcache_page(pg);
505 }
506
507 void flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
508                              unsigned long addr, int len)
509 {
510         unsigned long maddr;
511
512         maddr = (unsigned long) kmap(page) + (addr & ~PAGE_MASK);
513         flush_icache_range(maddr, maddr + len);
514         kunmap(page);
515 }
516 EXPORT_SYMBOL(flush_icache_user_range);
517
518 /*
519  * This is called at the end of handling a user page fault, when the
520  * fault has been handled by updating a PTE in the linux page tables.
521  * We use it to preload an HPTE into the hash table corresponding to
522  * the updated linux PTE.
523  * 
524  * This must always be called with the pte lock held.
525  */
526 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
527                       pte_t pte)
528 {
529 #ifdef CONFIG_PPC_STD_MMU
530         unsigned long access = 0, trap;
531 #endif
532         unsigned long pfn = pte_pfn(pte);
533
534         /* handle i-cache coherency */
535         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE) &&
536             !cpu_has_feature(CPU_FTR_NOEXECUTE) &&
537             pfn_valid(pfn)) {
538                 struct page *page = pfn_to_page(pfn);
539 #ifdef CONFIG_8xx
540                 /* On 8xx, cache control instructions (particularly
541                  * "dcbst" from flush_dcache_icache) fault as write
542                  * operation if there is an unpopulated TLB entry
543                  * for the address in question. To workaround that,
544                  * we invalidate the TLB here, thus avoiding dcbst
545                  * misbehaviour.
546                  */
547                 _tlbie(address, 0 /* 8xx doesn't care about PID */);
548 #endif
549                 /* The _PAGE_USER test should really be _PAGE_EXEC, but
550                  * older glibc versions execute some code from no-exec
551                  * pages, which for now we are supporting.  If exec-only
552                  * pages are ever implemented, this will have to change.
553                  */
554                 if (!PageReserved(page) && (pte_val(pte) & _PAGE_USER)
555                     && !test_bit(PG_arch_1, &page->flags)) {
556                         if (vma->vm_mm == current->active_mm) {
557                                 __flush_dcache_icache((void *) address);
558                         } else
559                                 flush_dcache_icache_page(page);
560                         set_bit(PG_arch_1, &page->flags);
561                 }
562         }
563
564 #ifdef CONFIG_PPC_STD_MMU
565         /* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
566         if (!pte_young(pte) || address >= TASK_SIZE)
567                 return;
568
569         /* We try to figure out if we are coming from an instruction
570          * access fault and pass that down to __hash_page so we avoid
571          * double-faulting on execution of fresh text. We have to test
572          * for regs NULL since init will get here first thing at boot
573          *
574          * We also avoid filling the hash if not coming from a fault
575          */
576         if (current->thread.regs == NULL)
577                 return;
578         trap = TRAP(current->thread.regs);
579         if (trap == 0x400)
580                 access |= _PAGE_EXEC;
581         else if (trap != 0x300)
582                 return;
583         hash_preload(vma->vm_mm, address, access, trap);
584 #endif /* CONFIG_PPC_STD_MMU */
585 }