Merge branch 'ppc-fixes' of git://git.bocc.de/dbox2 into for-2.6.24
[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
35 #include <asm/pgalloc.h>
36 #include <asm/prom.h>
37 #include <asm/io.h>
38 #include <asm/mmu_context.h>
39 #include <asm/pgtable.h>
40 #include <asm/mmu.h>
41 #include <asm/smp.h>
42 #include <asm/machdep.h>
43 #include <asm/btext.h>
44 #include <asm/tlb.h>
45 #include <asm/lmb.h>
46 #include <asm/sections.h>
47 #include <asm/vdso.h>
48
49 #include "mmu_decl.h"
50
51 #ifndef CPU_FTR_COHERENT_ICACHE
52 #define CPU_FTR_COHERENT_ICACHE 0       /* XXX for now */
53 #define CPU_FTR_NOEXECUTE       0
54 #endif
55
56 int init_bootmem_done;
57 int mem_init_done;
58 unsigned long memory_limit;
59
60 int page_is_ram(unsigned long pfn)
61 {
62         unsigned long paddr = (pfn << PAGE_SHIFT);
63
64 #ifndef CONFIG_PPC64    /* XXX for now */
65         return paddr < __pa(high_memory);
66 #else
67         int i;
68         for (i=0; i < lmb.memory.cnt; i++) {
69                 unsigned long base;
70
71                 base = lmb.memory.region[i].base;
72
73                 if ((paddr >= base) &&
74                         (paddr < (base + lmb.memory.region[i].size))) {
75                         return 1;
76                 }
77         }
78
79         return 0;
80 #endif
81 }
82
83 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
84                               unsigned long size, pgprot_t vma_prot)
85 {
86         if (ppc_md.phys_mem_access_prot)
87                 return ppc_md.phys_mem_access_prot(file, pfn, size, vma_prot);
88
89         if (!page_is_ram(pfn))
90                 vma_prot = __pgprot(pgprot_val(vma_prot)
91                                     | _PAGE_GUARDED | _PAGE_NO_CACHE);
92         return vma_prot;
93 }
94 EXPORT_SYMBOL(phys_mem_access_prot);
95
96 #ifdef CONFIG_MEMORY_HOTPLUG
97
98 void online_page(struct page *page)
99 {
100         ClearPageReserved(page);
101         init_page_count(page);
102         __free_page(page);
103         totalram_pages++;
104         num_physpages++;
105 }
106
107 #ifdef CONFIG_NUMA
108 int memory_add_physaddr_to_nid(u64 start)
109 {
110         return hot_add_scn_to_nid(start);
111 }
112 #endif
113
114 int __devinit arch_add_memory(int nid, u64 start, u64 size)
115 {
116         struct pglist_data *pgdata;
117         struct zone *zone;
118         unsigned long start_pfn = start >> PAGE_SHIFT;
119         unsigned long nr_pages = size >> PAGE_SHIFT;
120
121         pgdata = NODE_DATA(nid);
122
123         start = (unsigned long)__va(start);
124         create_section_mapping(start, start + size);
125
126         /* this should work for most non-highmem platforms */
127         zone = pgdata->node_zones;
128
129         return __add_pages(zone, start_pfn, nr_pages);
130 }
131
132 /*
133  * First pass at this code will check to determine if the remove
134  * request is within the RMO.  Do not allow removal within the RMO.
135  */
136 int __devinit remove_memory(u64 start, u64 size)
137 {
138         struct zone *zone;
139         unsigned long start_pfn, end_pfn, nr_pages;
140
141         start_pfn = start >> PAGE_SHIFT;
142         nr_pages = size >> PAGE_SHIFT;
143         end_pfn = start_pfn + nr_pages;
144
145         printk("%s(): Attempting to remove memoy in range "
146                         "%lx to %lx\n", __func__, start, start+size);
147         /*
148          * check for range within RMO
149          */
150         zone = page_zone(pfn_to_page(start_pfn));
151
152         printk("%s(): memory will be removed from "
153                         "the %s zone\n", __func__, zone->name);
154
155         /*
156          * not handling removing memory ranges that
157          * overlap multiple zones yet
158          */
159         if (end_pfn > (zone->zone_start_pfn + zone->spanned_pages))
160                 goto overlap;
161
162         /* make sure it is NOT in RMO */
163         if ((start < lmb.rmo_size) || ((start+size) < lmb.rmo_size)) {
164                 printk("%s(): range to be removed must NOT be in RMO!\n",
165                         __func__);
166                 goto in_rmo;
167         }
168
169         return __remove_pages(zone, start_pfn, nr_pages);
170
171 overlap:
172         printk("%s(): memory range to be removed overlaps "
173                 "multiple zones!!!\n", __func__);
174 in_rmo:
175         return -1;
176 }
177 #endif /* CONFIG_MEMORY_HOTPLUG */
178
179 void show_mem(void)
180 {
181         unsigned long total = 0, reserved = 0;
182         unsigned long shared = 0, cached = 0;
183         unsigned long highmem = 0;
184         struct page *page;
185         pg_data_t *pgdat;
186         unsigned long i;
187
188         printk("Mem-info:\n");
189         show_free_areas();
190         printk("Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
191         for_each_online_pgdat(pgdat) {
192                 unsigned long flags;
193                 pgdat_resize_lock(pgdat, &flags);
194                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
195                         if (!pfn_valid(pgdat->node_start_pfn + i))
196                                 continue;
197                         page = pgdat_page_nr(pgdat, i);
198                         total++;
199                         if (PageHighMem(page))
200                                 highmem++;
201                         if (PageReserved(page))
202                                 reserved++;
203                         else if (PageSwapCache(page))
204                                 cached++;
205                         else if (page_count(page))
206                                 shared += page_count(page) - 1;
207                 }
208                 pgdat_resize_unlock(pgdat, &flags);
209         }
210         printk("%ld pages of RAM\n", total);
211 #ifdef CONFIG_HIGHMEM
212         printk("%ld pages of HIGHMEM\n", highmem);
213 #endif
214         printk("%ld reserved pages\n", reserved);
215         printk("%ld pages shared\n", shared);
216         printk("%ld pages swap cached\n", cached);
217 }
218
219 /*
220  * Initialize the bootmem system and give it all the memory we
221  * have available.  If we are using highmem, we only put the
222  * lowmem into the bootmem system.
223  */
224 #ifndef CONFIG_NEED_MULTIPLE_NODES
225 void __init do_init_bootmem(void)
226 {
227         unsigned long i;
228         unsigned long start, bootmap_pages;
229         unsigned long total_pages;
230         int boot_mapsize;
231
232         max_pfn = total_pages = lmb_end_of_DRAM() >> PAGE_SHIFT;
233 #ifdef CONFIG_HIGHMEM
234         total_pages = total_lowmem >> PAGE_SHIFT;
235 #endif
236
237         /*
238          * Find an area to use for the bootmem bitmap.  Calculate the size of
239          * bitmap required as (Total Memory) / PAGE_SIZE / BITS_PER_BYTE.
240          * Add 1 additional page in case the address isn't page-aligned.
241          */
242         bootmap_pages = bootmem_bootmap_pages(total_pages);
243
244         start = lmb_alloc(bootmap_pages << PAGE_SHIFT, PAGE_SIZE);
245
246         boot_mapsize = init_bootmem(start >> PAGE_SHIFT, total_pages);
247
248         /* Add active regions with valid PFNs */
249         for (i = 0; i < lmb.memory.cnt; i++) {
250                 unsigned long start_pfn, end_pfn;
251                 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
252                 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
253                 add_active_range(0, start_pfn, end_pfn);
254         }
255
256         /* Add all physical memory to the bootmem map, mark each area
257          * present.
258          */
259 #ifdef CONFIG_HIGHMEM
260         free_bootmem_with_active_regions(0, total_lowmem >> PAGE_SHIFT);
261 #else
262         free_bootmem_with_active_regions(0, max_pfn);
263 #endif
264
265         /* reserve the sections we're already using */
266         for (i = 0; i < lmb.reserved.cnt; i++)
267                 reserve_bootmem(lmb.reserved.region[i].base,
268                                 lmb_size_bytes(&lmb.reserved, i));
269
270         /* XXX need to clip this if using highmem? */
271         sparse_memory_present_with_active_regions(0);
272
273         init_bootmem_done = 1;
274 }
275
276 /* mark pages that don't exist as nosave */
277 static int __init mark_nonram_nosave(void)
278 {
279         unsigned long lmb_next_region_start_pfn,
280                       lmb_region_max_pfn;
281         int i;
282
283         for (i = 0; i < lmb.memory.cnt - 1; i++) {
284                 lmb_region_max_pfn =
285                         (lmb.memory.region[i].base >> PAGE_SHIFT) +
286                         (lmb.memory.region[i].size >> PAGE_SHIFT);
287                 lmb_next_region_start_pfn =
288                         lmb.memory.region[i+1].base >> PAGE_SHIFT;
289
290                 if (lmb_region_max_pfn < lmb_next_region_start_pfn)
291                         register_nosave_region(lmb_region_max_pfn,
292                                                lmb_next_region_start_pfn);
293         }
294
295         return 0;
296 }
297
298 /*
299  * paging_init() sets up the page tables - in fact we've already done this.
300  */
301 void __init paging_init(void)
302 {
303         unsigned long total_ram = lmb_phys_mem_size();
304         unsigned long top_of_ram = lmb_end_of_DRAM();
305         unsigned long max_zone_pfns[MAX_NR_ZONES];
306
307 #ifdef CONFIG_HIGHMEM
308         map_page(PKMAP_BASE, 0, 0);     /* XXX gross */
309         pkmap_page_table = pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k
310                         (PKMAP_BASE), PKMAP_BASE), PKMAP_BASE), PKMAP_BASE);
311         map_page(KMAP_FIX_BEGIN, 0, 0); /* XXX gross */
312         kmap_pte = pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k
313                         (KMAP_FIX_BEGIN), KMAP_FIX_BEGIN), KMAP_FIX_BEGIN),
314                          KMAP_FIX_BEGIN);
315         kmap_prot = PAGE_KERNEL;
316 #endif /* CONFIG_HIGHMEM */
317
318         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
319                top_of_ram, total_ram);
320         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
321                (top_of_ram - total_ram) >> 20);
322         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
323 #ifdef CONFIG_HIGHMEM
324         max_zone_pfns[ZONE_DMA] = total_lowmem >> PAGE_SHIFT;
325         max_zone_pfns[ZONE_HIGHMEM] = top_of_ram >> PAGE_SHIFT;
326 #else
327         max_zone_pfns[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
328 #endif
329         free_area_init_nodes(max_zone_pfns);
330
331         mark_nonram_nosave();
332 }
333 #endif /* ! CONFIG_NEED_MULTIPLE_NODES */
334
335 void __init mem_init(void)
336 {
337 #ifdef CONFIG_NEED_MULTIPLE_NODES
338         int nid;
339 #endif
340         pg_data_t *pgdat;
341         unsigned long i;
342         struct page *page;
343         unsigned long reservedpages = 0, codesize, initsize, datasize, bsssize;
344
345         num_physpages = lmb.memory.size >> PAGE_SHIFT;
346         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE);
347
348 #ifdef CONFIG_NEED_MULTIPLE_NODES
349         for_each_online_node(nid) {
350                 if (NODE_DATA(nid)->node_spanned_pages != 0) {
351                         printk("freeing bootmem node %d\n", nid);
352                         totalram_pages +=
353                                 free_all_bootmem_node(NODE_DATA(nid));
354                 }
355         }
356 #else
357         max_mapnr = max_pfn;
358         totalram_pages += free_all_bootmem();
359 #endif
360         for_each_online_pgdat(pgdat) {
361                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
362                         if (!pfn_valid(pgdat->node_start_pfn + i))
363                                 continue;
364                         page = pgdat_page_nr(pgdat, i);
365                         if (PageReserved(page))
366                                 reservedpages++;
367                 }
368         }
369
370         codesize = (unsigned long)&_sdata - (unsigned long)&_stext;
371         datasize = (unsigned long)&_edata - (unsigned long)&_sdata;
372         initsize = (unsigned long)&__init_end - (unsigned long)&__init_begin;
373         bsssize = (unsigned long)&__bss_stop - (unsigned long)&__bss_start;
374
375 #ifdef CONFIG_HIGHMEM
376         {
377                 unsigned long pfn, highmem_mapnr;
378
379                 highmem_mapnr = total_lowmem >> PAGE_SHIFT;
380                 for (pfn = highmem_mapnr; pfn < max_mapnr; ++pfn) {
381                         struct page *page = pfn_to_page(pfn);
382
383                         ClearPageReserved(page);
384                         init_page_count(page);
385                         __free_page(page);
386                         totalhigh_pages++;
387                 }
388                 totalram_pages += totalhigh_pages;
389                 printk(KERN_DEBUG "High memory: %luk\n",
390                        totalhigh_pages << (PAGE_SHIFT-10));
391         }
392 #endif /* CONFIG_HIGHMEM */
393
394         printk(KERN_INFO "Memory: %luk/%luk available (%luk kernel code, "
395                "%luk reserved, %luk data, %luk bss, %luk init)\n",
396                 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
397                 num_physpages << (PAGE_SHIFT-10),
398                 codesize >> 10,
399                 reservedpages << (PAGE_SHIFT-10),
400                 datasize >> 10,
401                 bsssize >> 10,
402                 initsize >> 10);
403
404         mem_init_done = 1;
405 }
406
407 /*
408  * This is called when a page has been modified by the kernel.
409  * It just marks the page as not i-cache clean.  We do the i-cache
410  * flush later when the page is given to a user process, if necessary.
411  */
412 void flush_dcache_page(struct page *page)
413 {
414         if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
415                 return;
416         /* avoid an atomic op if possible */
417         if (test_bit(PG_arch_1, &page->flags))
418                 clear_bit(PG_arch_1, &page->flags);
419 }
420 EXPORT_SYMBOL(flush_dcache_page);
421
422 void flush_dcache_icache_page(struct page *page)
423 {
424 #ifdef CONFIG_BOOKE
425         void *start = kmap_atomic(page, KM_PPC_SYNC_ICACHE);
426         __flush_dcache_icache(start);
427         kunmap_atomic(start, KM_PPC_SYNC_ICACHE);
428 #elif defined(CONFIG_8xx) || defined(CONFIG_PPC64)
429         /* On 8xx there is no need to kmap since highmem is not supported */
430         __flush_dcache_icache(page_address(page)); 
431 #else
432         __flush_dcache_icache_phys(page_to_pfn(page) << PAGE_SHIFT);
433 #endif
434
435 }
436 void clear_user_page(void *page, unsigned long vaddr, struct page *pg)
437 {
438         clear_page(page);
439
440         /*
441          * We shouldnt have to do this, but some versions of glibc
442          * require it (ld.so assumes zero filled pages are icache clean)
443          * - Anton
444          */
445         flush_dcache_page(pg);
446 }
447 EXPORT_SYMBOL(clear_user_page);
448
449 void copy_user_page(void *vto, void *vfrom, unsigned long vaddr,
450                     struct page *pg)
451 {
452         copy_page(vto, vfrom);
453
454         /*
455          * We should be able to use the following optimisation, however
456          * there are two problems.
457          * Firstly a bug in some versions of binutils meant PLT sections
458          * were not marked executable.
459          * Secondly the first word in the GOT section is blrl, used
460          * to establish the GOT address. Until recently the GOT was
461          * not marked executable.
462          * - Anton
463          */
464 #if 0
465         if (!vma->vm_file && ((vma->vm_flags & VM_EXEC) == 0))
466                 return;
467 #endif
468
469         flush_dcache_page(pg);
470 }
471
472 void flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
473                              unsigned long addr, int len)
474 {
475         unsigned long maddr;
476
477         maddr = (unsigned long) kmap(page) + (addr & ~PAGE_MASK);
478         flush_icache_range(maddr, maddr + len);
479         kunmap(page);
480 }
481 EXPORT_SYMBOL(flush_icache_user_range);
482
483 /*
484  * This is called at the end of handling a user page fault, when the
485  * fault has been handled by updating a PTE in the linux page tables.
486  * We use it to preload an HPTE into the hash table corresponding to
487  * the updated linux PTE.
488  * 
489  * This must always be called with the pte lock held.
490  */
491 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
492                       pte_t pte)
493 {
494 #ifdef CONFIG_PPC_STD_MMU
495         unsigned long access = 0, trap;
496 #endif
497         unsigned long pfn = pte_pfn(pte);
498
499         /* handle i-cache coherency */
500         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE) &&
501             !cpu_has_feature(CPU_FTR_NOEXECUTE) &&
502             pfn_valid(pfn)) {
503                 struct page *page = pfn_to_page(pfn);
504 #ifdef CONFIG_8xx
505                 /* On 8xx, cache control instructions (particularly
506                  * "dcbst" from flush_dcache_icache) fault as write
507                  * operation if there is an unpopulated TLB entry
508                  * for the address in question. To workaround that,
509                  * we invalidate the TLB here, thus avoiding dcbst
510                  * misbehaviour.
511                  */
512                 _tlbie(address);
513 #endif
514                 if (!PageReserved(page)
515                     && !test_bit(PG_arch_1, &page->flags)) {
516                         if (vma->vm_mm == current->active_mm) {
517                                 __flush_dcache_icache((void *) address);
518                         } else
519                                 flush_dcache_icache_page(page);
520                         set_bit(PG_arch_1, &page->flags);
521                 }
522         }
523
524 #ifdef CONFIG_PPC_STD_MMU
525         /* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
526         if (!pte_young(pte) || address >= TASK_SIZE)
527                 return;
528
529         /* We try to figure out if we are coming from an instruction
530          * access fault and pass that down to __hash_page so we avoid
531          * double-faulting on execution of fresh text. We have to test
532          * for regs NULL since init will get here first thing at boot
533          *
534          * We also avoid filling the hash if not coming from a fault
535          */
536         if (current->thread.regs == NULL)
537                 return;
538         trap = TRAP(current->thread.regs);
539         if (trap == 0x400)
540                 access |= _PAGE_EXEC;
541         else if (trap != 0x300)
542                 return;
543         hash_preload(vma->vm_mm, address, access, trap);
544 #endif /* CONFIG_PPC_STD_MMU */
545 }