Merge git://git.infradead.org/hdrinstall-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  *  Amiga/APUS changes by Jesper Skov (jskov@cygnus.co.uk).
9  *  PPC44x/36-bit changes by Matt Porter (mporter@mvista.com)
10  *
11  *  Derived from "arch/i386/mm/init.c"
12  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
13  *
14  *  This program is free software; you can redistribute it and/or
15  *  modify it under the terms of the GNU General Public License
16  *  as published by the Free Software Foundation; either version
17  *  2 of the License, or (at your option) any later version.
18  *
19  */
20
21 #include <linux/module.h>
22 #include <linux/sched.h>
23 #include <linux/kernel.h>
24 #include <linux/errno.h>
25 #include <linux/string.h>
26 #include <linux/types.h>
27 #include <linux/mm.h>
28 #include <linux/stddef.h>
29 #include <linux/init.h>
30 #include <linux/bootmem.h>
31 #include <linux/highmem.h>
32 #include <linux/initrd.h>
33 #include <linux/pagemap.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/prom.h>
46 #include <asm/lmb.h>
47 #include <asm/sections.h>
48 #include <asm/vdso.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 extern void hash_preload(struct mm_struct *mm, unsigned long ea,
62                          unsigned long access, unsigned long trap);
63
64 /*
65  * This is called by /dev/mem to know if a given address has to
66  * be mapped non-cacheable or not
67  */
68 int page_is_ram(unsigned long pfn)
69 {
70         unsigned long paddr = (pfn << PAGE_SHIFT);
71
72 #ifndef CONFIG_PPC64    /* XXX for now */
73         return paddr < __pa(high_memory);
74 #else
75         int i;
76         for (i=0; i < lmb.memory.cnt; i++) {
77                 unsigned long base;
78
79                 base = lmb.memory.region[i].base;
80
81                 if ((paddr >= base) &&
82                         (paddr < (base + lmb.memory.region[i].size))) {
83                         return 1;
84                 }
85         }
86
87         return 0;
88 #endif
89 }
90 EXPORT_SYMBOL(page_is_ram);
91
92 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
93                               unsigned long size, pgprot_t vma_prot)
94 {
95         if (ppc_md.phys_mem_access_prot)
96                 return ppc_md.phys_mem_access_prot(file, pfn, size, vma_prot);
97
98         if (!page_is_ram(pfn))
99                 vma_prot = __pgprot(pgprot_val(vma_prot)
100                                     | _PAGE_GUARDED | _PAGE_NO_CACHE);
101         return vma_prot;
102 }
103 EXPORT_SYMBOL(phys_mem_access_prot);
104
105 #ifdef CONFIG_MEMORY_HOTPLUG
106
107 void online_page(struct page *page)
108 {
109         ClearPageReserved(page);
110         init_page_count(page);
111         __free_page(page);
112         totalram_pages++;
113         num_physpages++;
114 }
115
116 #ifdef CONFIG_NUMA
117 int memory_add_physaddr_to_nid(u64 start)
118 {
119         return hot_add_scn_to_nid(start);
120 }
121 #endif
122
123 int __devinit arch_add_memory(int nid, u64 start, u64 size)
124 {
125         struct pglist_data *pgdata;
126         struct zone *zone;
127         unsigned long start_pfn = start >> PAGE_SHIFT;
128         unsigned long nr_pages = size >> PAGE_SHIFT;
129
130         pgdata = NODE_DATA(nid);
131
132         start = (unsigned long)__va(start);
133         create_section_mapping(start, start + size);
134
135         /* this should work for most non-highmem platforms */
136         zone = pgdata->node_zones;
137
138         return __add_pages(zone, start_pfn, nr_pages);
139
140         return 0;
141 }
142
143 /*
144  * First pass at this code will check to determine if the remove
145  * request is within the RMO.  Do not allow removal within the RMO.
146  */
147 int __devinit remove_memory(u64 start, u64 size)
148 {
149         struct zone *zone;
150         unsigned long start_pfn, end_pfn, nr_pages;
151
152         start_pfn = start >> PAGE_SHIFT;
153         nr_pages = size >> PAGE_SHIFT;
154         end_pfn = start_pfn + nr_pages;
155
156         printk("%s(): Attempting to remove memoy in range "
157                         "%lx to %lx\n", __func__, start, start+size);
158         /*
159          * check for range within RMO
160          */
161         zone = page_zone(pfn_to_page(start_pfn));
162
163         printk("%s(): memory will be removed from "
164                         "the %s zone\n", __func__, zone->name);
165
166         /*
167          * not handling removing memory ranges that
168          * overlap multiple zones yet
169          */
170         if (end_pfn > (zone->zone_start_pfn + zone->spanned_pages))
171                 goto overlap;
172
173         /* make sure it is NOT in RMO */
174         if ((start < lmb.rmo_size) || ((start+size) < lmb.rmo_size)) {
175                 printk("%s(): range to be removed must NOT be in RMO!\n",
176                         __func__);
177                 goto in_rmo;
178         }
179
180         return __remove_pages(zone, start_pfn, nr_pages);
181
182 overlap:
183         printk("%s(): memory range to be removed overlaps "
184                 "multiple zones!!!\n", __func__);
185 in_rmo:
186         return -1;
187 }
188 #endif /* CONFIG_MEMORY_HOTPLUG */
189
190 void show_mem(void)
191 {
192         unsigned long total = 0, reserved = 0;
193         unsigned long shared = 0, cached = 0;
194         unsigned long highmem = 0;
195         struct page *page;
196         pg_data_t *pgdat;
197         unsigned long i;
198
199         printk("Mem-info:\n");
200         show_free_areas();
201         printk("Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
202         for_each_online_pgdat(pgdat) {
203                 unsigned long flags;
204                 pgdat_resize_lock(pgdat, &flags);
205                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
206                         if (!pfn_valid(pgdat->node_start_pfn + i))
207                                 continue;
208                         page = pgdat_page_nr(pgdat, i);
209                         total++;
210                         if (PageHighMem(page))
211                                 highmem++;
212                         if (PageReserved(page))
213                                 reserved++;
214                         else if (PageSwapCache(page))
215                                 cached++;
216                         else if (page_count(page))
217                                 shared += page_count(page) - 1;
218                 }
219                 pgdat_resize_unlock(pgdat, &flags);
220         }
221         printk("%ld pages of RAM\n", total);
222 #ifdef CONFIG_HIGHMEM
223         printk("%ld pages of HIGHMEM\n", highmem);
224 #endif
225         printk("%ld reserved pages\n", reserved);
226         printk("%ld pages shared\n", shared);
227         printk("%ld pages swap cached\n", cached);
228 }
229
230 /*
231  * Initialize the bootmem system and give it all the memory we
232  * have available.  If we are using highmem, we only put the
233  * lowmem into the bootmem system.
234  */
235 #ifndef CONFIG_NEED_MULTIPLE_NODES
236 void __init do_init_bootmem(void)
237 {
238         unsigned long i;
239         unsigned long start, bootmap_pages;
240         unsigned long total_pages;
241         int boot_mapsize;
242
243         max_pfn = total_pages = lmb_end_of_DRAM() >> PAGE_SHIFT;
244 #ifdef CONFIG_HIGHMEM
245         total_pages = total_lowmem >> PAGE_SHIFT;
246 #endif
247
248         /*
249          * Find an area to use for the bootmem bitmap.  Calculate the size of
250          * bitmap required as (Total Memory) / PAGE_SIZE / BITS_PER_BYTE.
251          * Add 1 additional page in case the address isn't page-aligned.
252          */
253         bootmap_pages = bootmem_bootmap_pages(total_pages);
254
255         start = lmb_alloc(bootmap_pages << PAGE_SHIFT, PAGE_SIZE);
256
257         boot_mapsize = init_bootmem(start >> PAGE_SHIFT, total_pages);
258
259         /* Add all physical memory to the bootmem map, mark each area
260          * present.
261          */
262         for (i = 0; i < lmb.memory.cnt; i++) {
263                 unsigned long base = lmb.memory.region[i].base;
264                 unsigned long size = lmb_size_bytes(&lmb.memory, i);
265 #ifdef CONFIG_HIGHMEM
266                 if (base >= total_lowmem)
267                         continue;
268                 if (base + size > total_lowmem)
269                         size = total_lowmem - base;
270 #endif
271                 free_bootmem(base, size);
272         }
273
274         /* reserve the sections we're already using */
275         for (i = 0; i < lmb.reserved.cnt; i++)
276                 reserve_bootmem(lmb.reserved.region[i].base,
277                                 lmb_size_bytes(&lmb.reserved, i));
278
279         /* XXX need to clip this if using highmem? */
280         for (i = 0; i < lmb.memory.cnt; i++)
281                 memory_present(0, lmb_start_pfn(&lmb.memory, i),
282                                lmb_end_pfn(&lmb.memory, i));
283         init_bootmem_done = 1;
284 }
285
286 /*
287  * paging_init() sets up the page tables - in fact we've already done this.
288  */
289 void __init paging_init(void)
290 {
291         unsigned long zones_size[MAX_NR_ZONES];
292         unsigned long zholes_size[MAX_NR_ZONES];
293         unsigned long total_ram = lmb_phys_mem_size();
294         unsigned long top_of_ram = lmb_end_of_DRAM();
295
296 #ifdef CONFIG_HIGHMEM
297         map_page(PKMAP_BASE, 0, 0);     /* XXX gross */
298         pkmap_page_table = pte_offset_kernel(pmd_offset(pgd_offset_k
299                         (PKMAP_BASE), PKMAP_BASE), PKMAP_BASE);
300         map_page(KMAP_FIX_BEGIN, 0, 0); /* XXX gross */
301         kmap_pte = pte_offset_kernel(pmd_offset(pgd_offset_k
302                         (KMAP_FIX_BEGIN), KMAP_FIX_BEGIN), KMAP_FIX_BEGIN);
303         kmap_prot = PAGE_KERNEL;
304 #endif /* CONFIG_HIGHMEM */
305
306         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
307                top_of_ram, total_ram);
308         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
309                (top_of_ram - total_ram) >> 20);
310         /*
311          * All pages are DMA-able so we put them all in the DMA zone.
312          */
313         memset(zones_size, 0, sizeof(zones_size));
314         memset(zholes_size, 0, sizeof(zholes_size));
315
316         zones_size[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
317         zholes_size[ZONE_DMA] = (top_of_ram - total_ram) >> PAGE_SHIFT;
318
319 #ifdef CONFIG_HIGHMEM
320         zones_size[ZONE_DMA] = total_lowmem >> PAGE_SHIFT;
321         zones_size[ZONE_HIGHMEM] = (total_memory - total_lowmem) >> PAGE_SHIFT;
322         zholes_size[ZONE_HIGHMEM] = (top_of_ram - total_ram) >> PAGE_SHIFT;
323 #else
324         zones_size[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
325         zholes_size[ZONE_DMA] = (top_of_ram - total_ram) >> PAGE_SHIFT;
326 #endif /* CONFIG_HIGHMEM */
327
328         free_area_init_node(0, NODE_DATA(0), zones_size,
329                             __pa(PAGE_OFFSET) >> PAGE_SHIFT, zholes_size);
330 }
331 #endif /* ! CONFIG_NEED_MULTIPLE_NODES */
332
333 void __init mem_init(void)
334 {
335 #ifdef CONFIG_NEED_MULTIPLE_NODES
336         int nid;
337 #endif
338         pg_data_t *pgdat;
339         unsigned long i;
340         struct page *page;
341         unsigned long reservedpages = 0, codesize, initsize, datasize, bsssize;
342
343         num_physpages = lmb.memory.size >> PAGE_SHIFT;
344         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE);
345
346 #ifdef CONFIG_NEED_MULTIPLE_NODES
347         for_each_online_node(nid) {
348                 if (NODE_DATA(nid)->node_spanned_pages != 0) {
349                         printk("freeing bootmem node %d\n", nid);
350                         totalram_pages +=
351                                 free_all_bootmem_node(NODE_DATA(nid));
352                 }
353         }
354 #else
355         max_mapnr = max_pfn;
356         totalram_pages += free_all_bootmem();
357 #endif
358         for_each_online_pgdat(pgdat) {
359                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
360                         if (!pfn_valid(pgdat->node_start_pfn + i))
361                                 continue;
362                         page = pgdat_page_nr(pgdat, i);
363                         if (PageReserved(page))
364                                 reservedpages++;
365                 }
366         }
367
368         codesize = (unsigned long)&_sdata - (unsigned long)&_stext;
369         datasize = (unsigned long)&_edata - (unsigned long)&_sdata;
370         initsize = (unsigned long)&__init_end - (unsigned long)&__init_begin;
371         bsssize = (unsigned long)&__bss_stop - (unsigned long)&__bss_start;
372
373 #ifdef CONFIG_HIGHMEM
374         {
375                 unsigned long pfn, highmem_mapnr;
376
377                 highmem_mapnr = total_lowmem >> PAGE_SHIFT;
378                 for (pfn = highmem_mapnr; pfn < max_mapnr; ++pfn) {
379                         struct page *page = pfn_to_page(pfn);
380
381                         ClearPageReserved(page);
382                         init_page_count(page);
383                         __free_page(page);
384                         totalhigh_pages++;
385                 }
386                 totalram_pages += totalhigh_pages;
387                 printk(KERN_DEBUG "High memory: %luk\n",
388                        totalhigh_pages << (PAGE_SHIFT-10));
389         }
390 #endif /* CONFIG_HIGHMEM */
391
392         printk(KERN_INFO "Memory: %luk/%luk available (%luk kernel code, "
393                "%luk reserved, %luk data, %luk bss, %luk init)\n",
394                 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
395                 num_physpages << (PAGE_SHIFT-10),
396                 codesize >> 10,
397                 reservedpages << (PAGE_SHIFT-10),
398                 datasize >> 10,
399                 bsssize >> 10,
400                 initsize >> 10);
401
402         mem_init_done = 1;
403
404         /* Initialize the vDSO */
405         vdso_init();
406 }
407
408 /*
409  * This is called when a page has been modified by the kernel.
410  * It just marks the page as not i-cache clean.  We do the i-cache
411  * flush later when the page is given to a user process, if necessary.
412  */
413 void flush_dcache_page(struct page *page)
414 {
415         if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
416                 return;
417         /* avoid an atomic op if possible */
418         if (test_bit(PG_arch_1, &page->flags))
419                 clear_bit(PG_arch_1, &page->flags);
420 }
421 EXPORT_SYMBOL(flush_dcache_page);
422
423 void flush_dcache_icache_page(struct page *page)
424 {
425 #ifdef CONFIG_BOOKE
426         void *start = kmap_atomic(page, KM_PPC_SYNC_ICACHE);
427         __flush_dcache_icache(start);
428         kunmap_atomic(start, KM_PPC_SYNC_ICACHE);
429 #elif defined(CONFIG_8xx) || defined(CONFIG_PPC64)
430         /* On 8xx there is no need to kmap since highmem is not supported */
431         __flush_dcache_icache(page_address(page)); 
432 #else
433         __flush_dcache_icache_phys(page_to_pfn(page) << PAGE_SHIFT);
434 #endif
435
436 }
437 void clear_user_page(void *page, unsigned long vaddr, struct page *pg)
438 {
439         clear_page(page);
440
441         /*
442          * We shouldnt have to do this, but some versions of glibc
443          * require it (ld.so assumes zero filled pages are icache clean)
444          * - Anton
445          */
446         flush_dcache_page(pg);
447 }
448 EXPORT_SYMBOL(clear_user_page);
449
450 void copy_user_page(void *vto, void *vfrom, unsigned long vaddr,
451                     struct page *pg)
452 {
453         copy_page(vto, vfrom);
454
455         /*
456          * We should be able to use the following optimisation, however
457          * there are two problems.
458          * Firstly a bug in some versions of binutils meant PLT sections
459          * were not marked executable.
460          * Secondly the first word in the GOT section is blrl, used
461          * to establish the GOT address. Until recently the GOT was
462          * not marked executable.
463          * - Anton
464          */
465 #if 0
466         if (!vma->vm_file && ((vma->vm_flags & VM_EXEC) == 0))
467                 return;
468 #endif
469
470         flush_dcache_page(pg);
471 }
472
473 void flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
474                              unsigned long addr, int len)
475 {
476         unsigned long maddr;
477
478         maddr = (unsigned long) kmap(page) + (addr & ~PAGE_MASK);
479         flush_icache_range(maddr, maddr + len);
480         kunmap(page);
481 }
482 EXPORT_SYMBOL(flush_icache_user_range);
483
484 /*
485  * This is called at the end of handling a user page fault, when the
486  * fault has been handled by updating a PTE in the linux page tables.
487  * We use it to preload an HPTE into the hash table corresponding to
488  * the updated linux PTE.
489  * 
490  * This must always be called with the pte lock held.
491  */
492 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
493                       pte_t pte)
494 {
495 #ifdef CONFIG_PPC_STD_MMU
496         unsigned long access = 0, trap;
497 #endif
498         unsigned long pfn = pte_pfn(pte);
499
500         /* handle i-cache coherency */
501         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE) &&
502             !cpu_has_feature(CPU_FTR_NOEXECUTE) &&
503             pfn_valid(pfn)) {
504                 struct page *page = pfn_to_page(pfn);
505                 if (!PageReserved(page)
506                     && !test_bit(PG_arch_1, &page->flags)) {
507                         if (vma->vm_mm == current->active_mm) {
508 #ifdef CONFIG_8xx
509                         /* On 8xx, cache control instructions (particularly 
510                          * "dcbst" from flush_dcache_icache) fault as write 
511                          * operation if there is an unpopulated TLB entry 
512                          * for the address in question. To workaround that, 
513                          * we invalidate the TLB here, thus avoiding dcbst 
514                          * misbehaviour.
515                          */
516                                 _tlbie(address);
517 #endif
518                                 __flush_dcache_icache((void *) address);
519                         } else
520                                 flush_dcache_icache_page(page);
521                         set_bit(PG_arch_1, &page->flags);
522                 }
523         }
524
525 #ifdef CONFIG_PPC_STD_MMU
526         /* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
527         if (!pte_young(pte) || address >= TASK_SIZE)
528                 return;
529
530         /* We try to figure out if we are coming from an instruction
531          * access fault and pass that down to __hash_page so we avoid
532          * double-faulting on execution of fresh text. We have to test
533          * for regs NULL since init will get here first thing at boot
534          *
535          * We also avoid filling the hash if not coming from a fault
536          */
537         if (current->thread.regs == NULL)
538                 return;
539         trap = TRAP(current->thread.regs);
540         if (trap == 0x400)
541                 access |= _PAGE_EXEC;
542         else if (trap != 0x300)
543                 return;
544         hash_preload(vma->vm_mm, address, access, trap);
545 #endif /* CONFIG_PPC_STD_MMU */
546 }