Merge branch 'master'
[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/config.h>
22 #include <linux/module.h>
23 #include <linux/sched.h>
24 #include <linux/kernel.h>
25 #include <linux/errno.h>
26 #include <linux/string.h>
27 #include <linux/types.h>
28 #include <linux/mm.h>
29 #include <linux/stddef.h>
30 #include <linux/init.h>
31 #include <linux/bootmem.h>
32 #include <linux/highmem.h>
33 #include <linux/initrd.h>
34 #include <linux/pagemap.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/prom.h>
47 #include <asm/lmb.h>
48 #include <asm/sections.h>
49 #include <asm/vdso.h>
50
51 #include "mmu_decl.h"
52
53 #ifndef CPU_FTR_COHERENT_ICACHE
54 #define CPU_FTR_COHERENT_ICACHE 0       /* XXX for now */
55 #define CPU_FTR_NOEXECUTE       0
56 #endif
57
58 int init_bootmem_done;
59 int mem_init_done;
60 unsigned long memory_limit;
61
62 extern void hash_preload(struct mm_struct *mm, unsigned long ea,
63                          unsigned long access, unsigned long trap);
64
65 /*
66  * This is called by /dev/mem to know if a given address has to
67  * be mapped non-cacheable or not
68  */
69 int page_is_ram(unsigned long pfn)
70 {
71         unsigned long paddr = (pfn << PAGE_SHIFT);
72
73 #ifndef CONFIG_PPC64    /* XXX for now */
74         return paddr < __pa(high_memory);
75 #else
76         int i;
77         for (i=0; i < lmb.memory.cnt; i++) {
78                 unsigned long base;
79
80                 base = lmb.memory.region[i].base;
81
82                 if ((paddr >= base) &&
83                         (paddr < (base + lmb.memory.region[i].size))) {
84                         return 1;
85                 }
86         }
87
88         return 0;
89 #endif
90 }
91 EXPORT_SYMBOL(page_is_ram);
92
93 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
94                               unsigned long size, pgprot_t vma_prot)
95 {
96         if (ppc_md.phys_mem_access_prot)
97                 return ppc_md.phys_mem_access_prot(file, pfn, size, vma_prot);
98
99         if (!page_is_ram(pfn))
100                 vma_prot = __pgprot(pgprot_val(vma_prot)
101                                     | _PAGE_GUARDED | _PAGE_NO_CACHE);
102         return vma_prot;
103 }
104 EXPORT_SYMBOL(phys_mem_access_prot);
105
106 #ifdef CONFIG_MEMORY_HOTPLUG
107
108 void online_page(struct page *page)
109 {
110         ClearPageReserved(page);
111         set_page_count(page, 0);
112         free_cold_page(page);
113         totalram_pages++;
114         num_physpages++;
115 }
116
117 /*
118  * This works only for the non-NUMA case.  Later, we'll need a lookup
119  * to convert from real physical addresses to nid, that doesn't use
120  * pfn_to_nid().
121  */
122 int __devinit add_memory(u64 start, u64 size)
123 {
124         struct pglist_data *pgdata = NODE_DATA(0);
125         struct zone *zone;
126         unsigned long start_pfn = start >> PAGE_SHIFT;
127         unsigned long nr_pages = size >> PAGE_SHIFT;
128
129         start += KERNELBASE;
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         return 0;
138 }
139
140 /*
141  * First pass at this code will check to determine if the remove
142  * request is within the RMO.  Do not allow removal within the RMO.
143  */
144 int __devinit remove_memory(u64 start, u64 size)
145 {
146         struct zone *zone;
147         unsigned long start_pfn, end_pfn, nr_pages;
148
149         start_pfn = start >> PAGE_SHIFT;
150         nr_pages = size >> PAGE_SHIFT;
151         end_pfn = start_pfn + nr_pages;
152
153         printk("%s(): Attempting to remove memoy in range "
154                         "%lx to %lx\n", __func__, start, start+size);
155         /*
156          * check for range within RMO
157          */
158         zone = page_zone(pfn_to_page(start_pfn));
159
160         printk("%s(): memory will be removed from "
161                         "the %s zone\n", __func__, zone->name);
162
163         /*
164          * not handling removing memory ranges that
165          * overlap multiple zones yet
166          */
167         if (end_pfn > (zone->zone_start_pfn + zone->spanned_pages))
168                 goto overlap;
169
170         /* make sure it is NOT in RMO */
171         if ((start < lmb.rmo_size) || ((start+size) < lmb.rmo_size)) {
172                 printk("%s(): range to be removed must NOT be in RMO!\n",
173                         __func__);
174                 goto in_rmo;
175         }
176
177         return __remove_pages(zone, start_pfn, nr_pages);
178
179 overlap:
180         printk("%s(): memory range to be removed overlaps "
181                 "multiple zones!!!\n", __func__);
182 in_rmo:
183         return -1;
184 }
185 #endif /* CONFIG_MEMORY_HOTPLUG */
186
187 void show_mem(void)
188 {
189         unsigned long total = 0, reserved = 0;
190         unsigned long shared = 0, cached = 0;
191         unsigned long highmem = 0;
192         struct page *page;
193         pg_data_t *pgdat;
194         unsigned long i;
195
196         printk("Mem-info:\n");
197         show_free_areas();
198         printk("Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
199         for_each_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_pfn = total_pages = lmb_end_of_DRAM() >> PAGE_SHIFT;
241 #ifdef CONFIG_HIGHMEM
242         total_pages = total_lowmem >> PAGE_SHIFT;
243 #endif
244
245         /*
246          * Find an area to use for the bootmem bitmap.  Calculate the size of
247          * bitmap required as (Total Memory) / PAGE_SIZE / BITS_PER_BYTE.
248          * Add 1 additional page in case the address isn't page-aligned.
249          */
250         bootmap_pages = bootmem_bootmap_pages(total_pages);
251
252         start = lmb_alloc(bootmap_pages << PAGE_SHIFT, PAGE_SIZE);
253         BUG_ON(!start);
254
255         boot_mapsize = init_bootmem(start >> PAGE_SHIFT, total_pages);
256
257         /* Add all physical memory to the bootmem map, mark each area
258          * present.
259          */
260         for (i = 0; i < lmb.memory.cnt; i++) {
261                 unsigned long base = lmb.memory.region[i].base;
262                 unsigned long size = lmb_size_bytes(&lmb.memory, i);
263 #ifdef CONFIG_HIGHMEM
264                 if (base >= total_lowmem)
265                         continue;
266                 if (base + size > total_lowmem)
267                         size = total_lowmem - base;
268 #endif
269                 free_bootmem(base, size);
270         }
271
272         /* reserve the sections we're already using */
273         for (i = 0; i < lmb.reserved.cnt; i++)
274                 reserve_bootmem(lmb.reserved.region[i].base,
275                                 lmb_size_bytes(&lmb.reserved, i));
276
277         /* XXX need to clip this if using highmem? */
278         for (i = 0; i < lmb.memory.cnt; i++)
279                 memory_present(0, lmb_start_pfn(&lmb.memory, i),
280                                lmb_end_pfn(&lmb.memory, i));
281         init_bootmem_done = 1;
282 }
283
284 /*
285  * paging_init() sets up the page tables - in fact we've already done this.
286  */
287 void __init paging_init(void)
288 {
289         unsigned long zones_size[MAX_NR_ZONES];
290         unsigned long zholes_size[MAX_NR_ZONES];
291         unsigned long total_ram = lmb_phys_mem_size();
292         unsigned long top_of_ram = lmb_end_of_DRAM();
293
294 #ifdef CONFIG_HIGHMEM
295         map_page(PKMAP_BASE, 0, 0);     /* XXX gross */
296         pkmap_page_table = pte_offset_kernel(pmd_offset(pgd_offset_k
297                         (PKMAP_BASE), PKMAP_BASE), PKMAP_BASE);
298         map_page(KMAP_FIX_BEGIN, 0, 0); /* XXX gross */
299         kmap_pte = pte_offset_kernel(pmd_offset(pgd_offset_k
300                         (KMAP_FIX_BEGIN), KMAP_FIX_BEGIN), KMAP_FIX_BEGIN);
301         kmap_prot = PAGE_KERNEL;
302 #endif /* CONFIG_HIGHMEM */
303
304         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
305                top_of_ram, total_ram);
306         printk(KERN_INFO "Memory hole size: %ldMB\n",
307                (top_of_ram - total_ram) >> 20);
308         /*
309          * All pages are DMA-able so we put them all in the DMA zone.
310          */
311         memset(zones_size, 0, sizeof(zones_size));
312         memset(zholes_size, 0, sizeof(zholes_size));
313
314         zones_size[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
315         zholes_size[ZONE_DMA] = (top_of_ram - total_ram) >> PAGE_SHIFT;
316
317 #ifdef CONFIG_HIGHMEM
318         zones_size[ZONE_DMA] = total_lowmem >> PAGE_SHIFT;
319         zones_size[ZONE_HIGHMEM] = (total_memory - total_lowmem) >> PAGE_SHIFT;
320         zholes_size[ZONE_HIGHMEM] = (top_of_ram - total_ram) >> PAGE_SHIFT;
321 #else
322         zones_size[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
323         zholes_size[ZONE_DMA] = (top_of_ram - total_ram) >> PAGE_SHIFT;
324 #endif /* CONFIG_HIGHMEM */
325
326         free_area_init_node(0, NODE_DATA(0), zones_size,
327                             __pa(PAGE_OFFSET) >> PAGE_SHIFT, zholes_size);
328 }
329 #endif /* ! CONFIG_NEED_MULTIPLE_NODES */
330
331 void __init mem_init(void)
332 {
333 #ifdef CONFIG_NEED_MULTIPLE_NODES
334         int nid;
335 #endif
336         pg_data_t *pgdat;
337         unsigned long i;
338         struct page *page;
339         unsigned long reservedpages = 0, codesize, initsize, datasize, bsssize;
340
341         num_physpages = lmb.memory.size >> PAGE_SHIFT;
342         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE);
343
344 #ifdef CONFIG_NEED_MULTIPLE_NODES
345         for_each_online_node(nid) {
346                 if (NODE_DATA(nid)->node_spanned_pages != 0) {
347                         printk("freeing bootmem node %x\n", nid);
348                         totalram_pages +=
349                                 free_all_bootmem_node(NODE_DATA(nid));
350                 }
351         }
352 #else
353         max_mapnr = max_pfn;
354         totalram_pages += free_all_bootmem();
355 #endif
356         for_each_pgdat(pgdat) {
357                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
358                         if (!pfn_valid(pgdat->node_start_pfn + i))
359                                 continue;
360                         page = pgdat_page_nr(pgdat, i);
361                         if (PageReserved(page))
362                                 reservedpages++;
363                 }
364         }
365
366         codesize = (unsigned long)&_sdata - (unsigned long)&_stext;
367         datasize = (unsigned long)&_edata - (unsigned long)&_sdata;
368         initsize = (unsigned long)&__init_end - (unsigned long)&__init_begin;
369         bsssize = (unsigned long)&__bss_stop - (unsigned long)&__bss_start;
370
371 #ifdef CONFIG_HIGHMEM
372         {
373                 unsigned long pfn, highmem_mapnr;
374
375                 highmem_mapnr = total_lowmem >> PAGE_SHIFT;
376                 for (pfn = highmem_mapnr; pfn < max_mapnr; ++pfn) {
377                         struct page *page = pfn_to_page(pfn);
378
379                         ClearPageReserved(page);
380                         set_page_count(page, 1);
381                         __free_page(page);
382                         totalhigh_pages++;
383                 }
384                 totalram_pages += totalhigh_pages;
385                 printk(KERN_INFO "High memory: %luk\n",
386                        totalhigh_pages << (PAGE_SHIFT-10));
387         }
388 #endif /* CONFIG_HIGHMEM */
389
390         printk(KERN_INFO "Memory: %luk/%luk available (%luk kernel code, "
391                "%luk reserved, %luk data, %luk bss, %luk init)\n",
392                 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
393                 num_physpages << (PAGE_SHIFT-10),
394                 codesize >> 10,
395                 reservedpages << (PAGE_SHIFT-10),
396                 datasize >> 10,
397                 bsssize >> 10,
398                 initsize >> 10);
399
400         mem_init_done = 1;
401
402         /* Initialize the vDSO */
403         vdso_init();
404 }
405
406 /*
407  * This is called when a page has been modified by the kernel.
408  * It just marks the page as not i-cache clean.  We do the i-cache
409  * flush later when the page is given to a user process, if necessary.
410  */
411 void flush_dcache_page(struct page *page)
412 {
413         if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
414                 return;
415         /* avoid an atomic op if possible */
416         if (test_bit(PG_arch_1, &page->flags))
417                 clear_bit(PG_arch_1, &page->flags);
418 }
419 EXPORT_SYMBOL(flush_dcache_page);
420
421 void flush_dcache_icache_page(struct page *page)
422 {
423 #ifdef CONFIG_BOOKE
424         void *start = kmap_atomic(page, KM_PPC_SYNC_ICACHE);
425         __flush_dcache_icache(start);
426         kunmap_atomic(start, KM_PPC_SYNC_ICACHE);
427 #elif defined(CONFIG_8xx) || defined(CONFIG_PPC64)
428         /* On 8xx there is no need to kmap since highmem is not supported */
429         __flush_dcache_icache(page_address(page)); 
430 #else
431         __flush_dcache_icache_phys(page_to_pfn(page) << PAGE_SHIFT);
432 #endif
433
434 }
435 void clear_user_page(void *page, unsigned long vaddr, struct page *pg)
436 {
437         clear_page(page);
438
439         if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
440                 return;
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
447         /* avoid an atomic op if possible */
448         if (test_bit(PG_arch_1, &pg->flags))
449                 clear_bit(PG_arch_1, &pg->flags);
450 }
451 EXPORT_SYMBOL(clear_user_page);
452
453 void copy_user_page(void *vto, void *vfrom, unsigned long vaddr,
454                     struct page *pg)
455 {
456         copy_page(vto, vfrom);
457
458         /*
459          * We should be able to use the following optimisation, however
460          * there are two problems.
461          * Firstly a bug in some versions of binutils meant PLT sections
462          * were not marked executable.
463          * Secondly the first word in the GOT section is blrl, used
464          * to establish the GOT address. Until recently the GOT was
465          * not marked executable.
466          * - Anton
467          */
468 #if 0
469         if (!vma->vm_file && ((vma->vm_flags & VM_EXEC) == 0))
470                 return;
471 #endif
472
473         if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
474                 return;
475
476         /* avoid an atomic op if possible */
477         if (test_bit(PG_arch_1, &pg->flags))
478                 clear_bit(PG_arch_1, &pg->flags);
479 }
480
481 void flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
482                              unsigned long addr, int len)
483 {
484         unsigned long maddr;
485
486         maddr = (unsigned long) kmap(page) + (addr & ~PAGE_MASK);
487         flush_icache_range(maddr, maddr + len);
488         kunmap(page);
489 }
490 EXPORT_SYMBOL(flush_icache_user_range);
491
492 /*
493  * This is called at the end of handling a user page fault, when the
494  * fault has been handled by updating a PTE in the linux page tables.
495  * We use it to preload an HPTE into the hash table corresponding to
496  * the updated linux PTE.
497  * 
498  * This must always be called with the pte lock held.
499  */
500 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
501                       pte_t pte)
502 {
503 #ifdef CONFIG_PPC_STD_MMU
504         unsigned long access = 0, trap;
505 #endif
506         unsigned long pfn = pte_pfn(pte);
507
508         /* handle i-cache coherency */
509         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE) &&
510             !cpu_has_feature(CPU_FTR_NOEXECUTE) &&
511             pfn_valid(pfn)) {
512                 struct page *page = pfn_to_page(pfn);
513                 if (!PageReserved(page)
514                     && !test_bit(PG_arch_1, &page->flags)) {
515                         if (vma->vm_mm == current->active_mm) {
516 #ifdef CONFIG_8xx
517                         /* On 8xx, cache control instructions (particularly 
518                          * "dcbst" from flush_dcache_icache) fault as write 
519                          * operation if there is an unpopulated TLB entry 
520                          * for the address in question. To workaround that, 
521                          * we invalidate the TLB here, thus avoiding dcbst 
522                          * misbehaviour.
523                          */
524                                 _tlbie(address);
525 #endif
526                                 __flush_dcache_icache((void *) address);
527                         } else
528                                 flush_dcache_icache_page(page);
529                         set_bit(PG_arch_1, &page->flags);
530                 }
531         }
532
533 #ifdef CONFIG_PPC_STD_MMU
534         /* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
535         if (!pte_young(pte) || address >= TASK_SIZE)
536                 return;
537
538         /* We try to figure out if we are coming from an instruction
539          * access fault and pass that down to __hash_page so we avoid
540          * double-faulting on execution of fresh text. We have to test
541          * for regs NULL since init will get here first thing at boot
542          *
543          * We also avoid filling the hash if not coming from a fault
544          */
545         if (current->thread.regs == NULL)
546                 return;
547         trap = TRAP(current->thread.regs);
548         if (trap == 0x400)
549                 access |= _PAGE_EXEC;
550         else if (trap != 0x300)
551                 return;
552         hash_preload(vma->vm_mm, address, access, trap);
553 #endif /* CONFIG_PPC_STD_MMU */
554 }