Merge branch 'upstream' into upstream-jgarzik
[linux-2.6] / arch / parisc / mm / init.c
1 /*
2  *  linux/arch/parisc/mm/init.c
3  *
4  *  Copyright (C) 1995  Linus Torvalds
5  *  Copyright 1999 SuSE GmbH
6  *    changed by Philipp Rumpf
7  *  Copyright 1999 Philipp Rumpf (prumpf@tux.org)
8  *  Copyright 2004 Randolph Chung (tausq@debian.org)
9  *  Copyright 2006 Helge Deller (deller@gmx.de)
10  *
11  */
12
13
14 #include <linux/module.h>
15 #include <linux/mm.h>
16 #include <linux/bootmem.h>
17 #include <linux/delay.h>
18 #include <linux/init.h>
19 #include <linux/pci.h>          /* for hppa_dma_ops and pcxl_dma_ops */
20 #include <linux/initrd.h>
21 #include <linux/swap.h>
22 #include <linux/unistd.h>
23 #include <linux/nodemask.h>     /* for node_online_map */
24 #include <linux/pagemap.h>      /* for release_pages and page_cache_release */
25
26 #include <asm/pgalloc.h>
27 #include <asm/tlb.h>
28 #include <asm/pdc_chassis.h>
29 #include <asm/mmzone.h>
30 #include <asm/sections.h>
31
32 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
33
34 extern int  data_start;
35
36 #ifdef CONFIG_DISCONTIGMEM
37 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
38 bootmem_data_t bmem_data[MAX_NUMNODES] __read_mostly;
39 unsigned char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
40 #endif
41
42 static struct resource data_resource = {
43         .name   = "Kernel data",
44         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
45 };
46
47 static struct resource code_resource = {
48         .name   = "Kernel code",
49         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
50 };
51
52 static struct resource pdcdata_resource = {
53         .name   = "PDC data (Page Zero)",
54         .start  = 0,
55         .end    = 0x9ff,
56         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
57 };
58
59 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
60
61 /* The following array is initialized from the firmware specific
62  * information retrieved in kernel/inventory.c.
63  */
64
65 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
66 int npmem_ranges __read_mostly;
67
68 #ifdef __LP64__
69 #define MAX_MEM         (~0UL)
70 #else /* !__LP64__ */
71 #define MAX_MEM         (3584U*1024U*1024U)
72 #endif /* !__LP64__ */
73
74 static unsigned long mem_limit __read_mostly = MAX_MEM;
75
76 static void __init mem_limit_func(void)
77 {
78         char *cp, *end;
79         unsigned long limit;
80
81         /* We need this before __setup() functions are called */
82
83         limit = MAX_MEM;
84         for (cp = boot_command_line; *cp; ) {
85                 if (memcmp(cp, "mem=", 4) == 0) {
86                         cp += 4;
87                         limit = memparse(cp, &end);
88                         if (end != cp)
89                                 break;
90                         cp = end;
91                 } else {
92                         while (*cp != ' ' && *cp)
93                                 ++cp;
94                         while (*cp == ' ')
95                                 ++cp;
96                 }
97         }
98
99         if (limit < mem_limit)
100                 mem_limit = limit;
101 }
102
103 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
104
105 static void __init setup_bootmem(void)
106 {
107         unsigned long bootmap_size;
108         unsigned long mem_max;
109         unsigned long bootmap_pages;
110         unsigned long bootmap_start_pfn;
111         unsigned long bootmap_pfn;
112 #ifndef CONFIG_DISCONTIGMEM
113         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
114         int npmem_holes;
115 #endif
116         int i, sysram_resource_count;
117
118         disable_sr_hashing(); /* Turn off space register hashing */
119
120         /*
121          * Sort the ranges. Since the number of ranges is typically
122          * small, and performance is not an issue here, just do
123          * a simple insertion sort.
124          */
125
126         for (i = 1; i < npmem_ranges; i++) {
127                 int j;
128
129                 for (j = i; j > 0; j--) {
130                         unsigned long tmp;
131
132                         if (pmem_ranges[j-1].start_pfn <
133                             pmem_ranges[j].start_pfn) {
134
135                                 break;
136                         }
137                         tmp = pmem_ranges[j-1].start_pfn;
138                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
139                         pmem_ranges[j].start_pfn = tmp;
140                         tmp = pmem_ranges[j-1].pages;
141                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
142                         pmem_ranges[j].pages = tmp;
143                 }
144         }
145
146 #ifndef CONFIG_DISCONTIGMEM
147         /*
148          * Throw out ranges that are too far apart (controlled by
149          * MAX_GAP).
150          */
151
152         for (i = 1; i < npmem_ranges; i++) {
153                 if (pmem_ranges[i].start_pfn -
154                         (pmem_ranges[i-1].start_pfn +
155                          pmem_ranges[i-1].pages) > MAX_GAP) {
156                         npmem_ranges = i;
157                         printk("Large gap in memory detected (%ld pages). "
158                                "Consider turning on CONFIG_DISCONTIGMEM\n",
159                                pmem_ranges[i].start_pfn -
160                                (pmem_ranges[i-1].start_pfn +
161                                 pmem_ranges[i-1].pages));
162                         break;
163                 }
164         }
165 #endif
166
167         if (npmem_ranges > 1) {
168
169                 /* Print the memory ranges */
170
171                 printk(KERN_INFO "Memory Ranges:\n");
172
173                 for (i = 0; i < npmem_ranges; i++) {
174                         unsigned long start;
175                         unsigned long size;
176
177                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
178                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
179                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
180                                 i,start, start + (size - 1), size >> 20);
181                 }
182         }
183
184         sysram_resource_count = npmem_ranges;
185         for (i = 0; i < sysram_resource_count; i++) {
186                 struct resource *res = &sysram_resources[i];
187                 res->name = "System RAM";
188                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
189                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
190                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
191                 request_resource(&iomem_resource, res);
192         }
193
194         /*
195          * For 32 bit kernels we limit the amount of memory we can
196          * support, in order to preserve enough kernel address space
197          * for other purposes. For 64 bit kernels we don't normally
198          * limit the memory, but this mechanism can be used to
199          * artificially limit the amount of memory (and it is written
200          * to work with multiple memory ranges).
201          */
202
203         mem_limit_func();       /* check for "mem=" argument */
204
205         mem_max = 0;
206         num_physpages = 0;
207         for (i = 0; i < npmem_ranges; i++) {
208                 unsigned long rsize;
209
210                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
211                 if ((mem_max + rsize) > mem_limit) {
212                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
213                         if (mem_max == mem_limit)
214                                 npmem_ranges = i;
215                         else {
216                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
217                                                        - (mem_max >> PAGE_SHIFT);
218                                 npmem_ranges = i + 1;
219                                 mem_max = mem_limit;
220                         }
221                 num_physpages += pmem_ranges[i].pages;
222                         break;
223                 }
224             num_physpages += pmem_ranges[i].pages;
225                 mem_max += rsize;
226         }
227
228         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
229
230 #ifndef CONFIG_DISCONTIGMEM
231         /* Merge the ranges, keeping track of the holes */
232
233         {
234                 unsigned long end_pfn;
235                 unsigned long hole_pages;
236
237                 npmem_holes = 0;
238                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
239                 for (i = 1; i < npmem_ranges; i++) {
240
241                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
242                         if (hole_pages) {
243                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
244                                 pmem_holes[npmem_holes++].pages = hole_pages;
245                                 end_pfn += hole_pages;
246                         }
247                         end_pfn += pmem_ranges[i].pages;
248                 }
249
250                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
251                 npmem_ranges = 1;
252         }
253 #endif
254
255         bootmap_pages = 0;
256         for (i = 0; i < npmem_ranges; i++)
257                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
258
259         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
260
261 #ifdef CONFIG_DISCONTIGMEM
262         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
263                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
264                 NODE_DATA(i)->bdata = &bmem_data[i];
265         }
266         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
267
268         for (i = 0; i < npmem_ranges; i++)
269                 node_set_online(i);
270 #endif
271
272         /*
273          * Initialize and free the full range of memory in each range.
274          * Note that the only writing these routines do are to the bootmap,
275          * and we've made sure to locate the bootmap properly so that they
276          * won't be writing over anything important.
277          */
278
279         bootmap_pfn = bootmap_start_pfn;
280         max_pfn = 0;
281         for (i = 0; i < npmem_ranges; i++) {
282                 unsigned long start_pfn;
283                 unsigned long npages;
284
285                 start_pfn = pmem_ranges[i].start_pfn;
286                 npages = pmem_ranges[i].pages;
287
288                 bootmap_size = init_bootmem_node(NODE_DATA(i),
289                                                 bootmap_pfn,
290                                                 start_pfn,
291                                                 (start_pfn + npages) );
292                 free_bootmem_node(NODE_DATA(i),
293                                   (start_pfn << PAGE_SHIFT),
294                                   (npages << PAGE_SHIFT) );
295                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
296                 if ((start_pfn + npages) > max_pfn)
297                         max_pfn = start_pfn + npages;
298         }
299
300         /* IOMMU is always used to access "high mem" on those boxes
301          * that can support enough mem that a PCI device couldn't
302          * directly DMA to any physical addresses.
303          * ISA DMA support will need to revisit this.
304          */
305         max_low_pfn = max_pfn;
306
307         if ((bootmap_pfn - bootmap_start_pfn) != bootmap_pages) {
308                 printk(KERN_WARNING "WARNING! bootmap sizing is messed up!\n");
309                 BUG();
310         }
311
312         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
313
314 #define PDC_CONSOLE_IO_IODC_SIZE 32768
315
316         reserve_bootmem_node(NODE_DATA(0), 0UL,
317                         (unsigned long)(PAGE0->mem_free + PDC_CONSOLE_IO_IODC_SIZE));
318         reserve_bootmem_node(NODE_DATA(0), __pa((unsigned long)_text),
319                         (unsigned long)(_end - _text));
320         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
321                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT));
322
323 #ifndef CONFIG_DISCONTIGMEM
324
325         /* reserve the holes */
326
327         for (i = 0; i < npmem_holes; i++) {
328                 reserve_bootmem_node(NODE_DATA(0),
329                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
330                                 (pmem_holes[i].pages << PAGE_SHIFT));
331         }
332 #endif
333
334 #ifdef CONFIG_BLK_DEV_INITRD
335         if (initrd_start) {
336                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
337                 if (__pa(initrd_start) < mem_max) {
338                         unsigned long initrd_reserve;
339
340                         if (__pa(initrd_end) > mem_max) {
341                                 initrd_reserve = mem_max - __pa(initrd_start);
342                         } else {
343                                 initrd_reserve = initrd_end - initrd_start;
344                         }
345                         initrd_below_start_ok = 1;
346                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
347
348                         reserve_bootmem_node(NODE_DATA(0),__pa(initrd_start), initrd_reserve);
349                 }
350         }
351 #endif
352
353         data_resource.start =  virt_to_phys(&data_start);
354         data_resource.end = virt_to_phys(_end) - 1;
355         code_resource.start = virt_to_phys(_text);
356         code_resource.end = virt_to_phys(&data_start)-1;
357
358         /* We don't know which region the kernel will be in, so try
359          * all of them.
360          */
361         for (i = 0; i < sysram_resource_count; i++) {
362                 struct resource *res = &sysram_resources[i];
363                 request_resource(res, &code_resource);
364                 request_resource(res, &data_resource);
365         }
366         request_resource(&sysram_resources[0], &pdcdata_resource);
367 }
368
369 void free_initmem(void)
370 {
371         unsigned long addr, init_begin, init_end;
372
373         printk(KERN_INFO "Freeing unused kernel memory: ");
374
375 #ifdef CONFIG_DEBUG_KERNEL
376         /* Attempt to catch anyone trying to execute code here
377          * by filling the page with BRK insns.
378          * 
379          * If we disable interrupts for all CPUs, then IPI stops working.
380          * Kinda breaks the global cache flushing.
381          */
382         local_irq_disable();
383
384         memset(__init_begin, 0x00,
385                 (unsigned long)__init_end - (unsigned long)__init_begin);
386
387         flush_data_cache();
388         asm volatile("sync" : : );
389         flush_icache_range((unsigned long)__init_begin, (unsigned long)__init_end);
390         asm volatile("sync" : : );
391
392         local_irq_enable();
393 #endif
394         
395         /* align __init_begin and __init_end to page size,
396            ignoring linker script where we might have tried to save RAM */
397         init_begin = PAGE_ALIGN((unsigned long)(__init_begin));
398         init_end   = PAGE_ALIGN((unsigned long)(__init_end));
399         for (addr = init_begin; addr < init_end; addr += PAGE_SIZE) {
400                 ClearPageReserved(virt_to_page(addr));
401                 init_page_count(virt_to_page(addr));
402                 free_page(addr);
403                 num_physpages++;
404                 totalram_pages++;
405         }
406
407         /* set up a new led state on systems shipped LED State panel */
408         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
409         
410         printk("%luk freed\n", (init_end - init_begin) >> 10);
411 }
412
413
414 #ifdef CONFIG_DEBUG_RODATA
415 void mark_rodata_ro(void)
416 {
417         /* rodata memory was already mapped with KERNEL_RO access rights by
418            pagetable_init() and map_pages(). No need to do additional stuff here */
419         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
420                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
421 }
422 #endif
423
424
425 /*
426  * Just an arbitrary offset to serve as a "hole" between mapping areas
427  * (between top of physical memory and a potential pcxl dma mapping
428  * area, and below the vmalloc mapping area).
429  *
430  * The current 32K value just means that there will be a 32K "hole"
431  * between mapping areas. That means that  any out-of-bounds memory
432  * accesses will hopefully be caught. The vmalloc() routines leaves
433  * a hole of 4kB between each vmalloced area for the same reason.
434  */
435
436  /* Leave room for gateway page expansion */
437 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
438 #error KERNEL_MAP_START is in gateway reserved region
439 #endif
440 #define MAP_START (KERNEL_MAP_START)
441
442 #define VM_MAP_OFFSET  (32*1024)
443 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
444                                      & ~(VM_MAP_OFFSET-1)))
445
446 void *vmalloc_start __read_mostly;
447 EXPORT_SYMBOL(vmalloc_start);
448
449 #ifdef CONFIG_PA11
450 unsigned long pcxl_dma_start __read_mostly;
451 #endif
452
453 void __init mem_init(void)
454 {
455         high_memory = __va((max_pfn << PAGE_SHIFT));
456
457 #ifndef CONFIG_DISCONTIGMEM
458         max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
459         totalram_pages += free_all_bootmem();
460 #else
461         {
462                 int i;
463
464                 for (i = 0; i < npmem_ranges; i++)
465                         totalram_pages += free_all_bootmem_node(NODE_DATA(i));
466         }
467 #endif
468
469         printk(KERN_INFO "Memory: %luk available\n", num_physpages << (PAGE_SHIFT-10));
470
471 #ifdef CONFIG_PA11
472         if (hppa_dma_ops == &pcxl_dma_ops) {
473                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
474                 vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
475         } else {
476                 pcxl_dma_start = 0;
477                 vmalloc_start = SET_MAP_OFFSET(MAP_START);
478         }
479 #else
480         vmalloc_start = SET_MAP_OFFSET(MAP_START);
481 #endif
482
483 }
484
485 unsigned long *empty_zero_page __read_mostly;
486
487 void show_mem(void)
488 {
489         int i,free = 0,total = 0,reserved = 0;
490         int shared = 0, cached = 0;
491
492         printk(KERN_INFO "Mem-info:\n");
493         show_free_areas();
494         printk(KERN_INFO "Free swap:     %6ldkB\n",
495                                 nr_swap_pages<<(PAGE_SHIFT-10));
496 #ifndef CONFIG_DISCONTIGMEM
497         i = max_mapnr;
498         while (i-- > 0) {
499                 total++;
500                 if (PageReserved(mem_map+i))
501                         reserved++;
502                 else if (PageSwapCache(mem_map+i))
503                         cached++;
504                 else if (!page_count(&mem_map[i]))
505                         free++;
506                 else
507                         shared += page_count(&mem_map[i]) - 1;
508         }
509 #else
510         for (i = 0; i < npmem_ranges; i++) {
511                 int j;
512
513                 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
514                         struct page *p;
515                         unsigned long flags;
516
517                         pgdat_resize_lock(NODE_DATA(i), &flags);
518                         p = nid_page_nr(i, j) - node_start_pfn(i);
519
520                         total++;
521                         if (PageReserved(p))
522                                 reserved++;
523                         else if (PageSwapCache(p))
524                                 cached++;
525                         else if (!page_count(p))
526                                 free++;
527                         else
528                                 shared += page_count(p) - 1;
529                         pgdat_resize_unlock(NODE_DATA(i), &flags);
530                 }
531         }
532 #endif
533         printk(KERN_INFO "%d pages of RAM\n", total);
534         printk(KERN_INFO "%d reserved pages\n", reserved);
535         printk(KERN_INFO "%d pages shared\n", shared);
536         printk(KERN_INFO "%d pages swap cached\n", cached);
537
538
539 #ifdef CONFIG_DISCONTIGMEM
540         {
541                 struct zonelist *zl;
542                 int i, j, k;
543
544                 for (i = 0; i < npmem_ranges; i++) {
545                         for (j = 0; j < MAX_NR_ZONES; j++) {
546                                 zl = NODE_DATA(i)->node_zonelists + j;
547
548                                 printk("Zone list for zone %d on node %d: ", j, i);
549                                 for (k = 0; zl->zones[k] != NULL; k++) 
550                                         printk("[%d/%s] ", zone_to_nid(zl->zones[k]), zl->zones[k]->name);
551                                 printk("\n");
552                         }
553                 }
554         }
555 #endif
556 }
557
558
559 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
560 {
561         pgd_t *pg_dir;
562         pmd_t *pmd;
563         pte_t *pg_table;
564         unsigned long end_paddr;
565         unsigned long start_pmd;
566         unsigned long start_pte;
567         unsigned long tmp1;
568         unsigned long tmp2;
569         unsigned long address;
570         unsigned long ro_start;
571         unsigned long ro_end;
572         unsigned long fv_addr;
573         unsigned long gw_addr;
574         extern const unsigned long fault_vector_20;
575         extern void * const linux_gateway_page;
576
577         ro_start = __pa((unsigned long)_text);
578         ro_end   = __pa((unsigned long)&data_start);
579         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
580         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
581
582         end_paddr = start_paddr + size;
583
584         pg_dir = pgd_offset_k(start_vaddr);
585
586 #if PTRS_PER_PMD == 1
587         start_pmd = 0;
588 #else
589         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
590 #endif
591         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
592
593         address = start_paddr;
594         while (address < end_paddr) {
595 #if PTRS_PER_PMD == 1
596                 pmd = (pmd_t *)__pa(pg_dir);
597 #else
598                 pmd = (pmd_t *)pgd_address(*pg_dir);
599
600                 /*
601                  * pmd is physical at this point
602                  */
603
604                 if (!pmd) {
605                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
606                         pmd = (pmd_t *) __pa(pmd);
607                 }
608
609                 pgd_populate(NULL, pg_dir, __va(pmd));
610 #endif
611                 pg_dir++;
612
613                 /* now change pmd to kernel virtual addresses */
614
615                 pmd = (pmd_t *)__va(pmd) + start_pmd;
616                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
617
618                         /*
619                          * pg_table is physical at this point
620                          */
621
622                         pg_table = (pte_t *)pmd_address(*pmd);
623                         if (!pg_table) {
624                                 pg_table = (pte_t *)
625                                         alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
626                                 pg_table = (pte_t *) __pa(pg_table);
627                         }
628
629                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
630
631                         /* now change pg_table to kernel virtual addresses */
632
633                         pg_table = (pte_t *) __va(pg_table) + start_pte;
634                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
635                                 pte_t pte;
636
637                                 /*
638                                  * Map the fault vector writable so we can
639                                  * write the HPMC checksum.
640                                  */
641 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
642                                 if (address >= ro_start && address < ro_end
643                                                         && address != fv_addr
644                                                         && address != gw_addr)
645                                     pte = __mk_pte(address, PAGE_KERNEL_RO);
646                                 else
647 #endif
648                                     pte = __mk_pte(address, pgprot);
649
650                                 if (address >= end_paddr)
651                                         pte_val(pte) = 0;
652
653                                 set_pte(pg_table, pte);
654
655                                 address += PAGE_SIZE;
656                         }
657                         start_pte = 0;
658
659                         if (address >= end_paddr)
660                             break;
661                 }
662                 start_pmd = 0;
663         }
664 }
665
666 /*
667  * pagetable_init() sets up the page tables
668  *
669  * Note that gateway_init() places the Linux gateway page at page 0.
670  * Since gateway pages cannot be dereferenced this has the desirable
671  * side effect of trapping those pesky NULL-reference errors in the
672  * kernel.
673  */
674 static void __init pagetable_init(void)
675 {
676         int range;
677
678         /* Map each physical memory range to its kernel vaddr */
679
680         for (range = 0; range < npmem_ranges; range++) {
681                 unsigned long start_paddr;
682                 unsigned long end_paddr;
683                 unsigned long size;
684
685                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
686                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
687                 size = pmem_ranges[range].pages << PAGE_SHIFT;
688
689                 map_pages((unsigned long)__va(start_paddr), start_paddr,
690                         size, PAGE_KERNEL);
691         }
692
693 #ifdef CONFIG_BLK_DEV_INITRD
694         if (initrd_end && initrd_end > mem_limit) {
695                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
696                 map_pages(initrd_start, __pa(initrd_start),
697                         initrd_end - initrd_start, PAGE_KERNEL);
698         }
699 #endif
700
701         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
702         memset(empty_zero_page, 0, PAGE_SIZE);
703 }
704
705 static void __init gateway_init(void)
706 {
707         unsigned long linux_gateway_page_addr;
708         /* FIXME: This is 'const' in order to trick the compiler
709            into not treating it as DP-relative data. */
710         extern void * const linux_gateway_page;
711
712         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
713
714         /*
715          * Setup Linux Gateway page.
716          *
717          * The Linux gateway page will reside in kernel space (on virtual
718          * page 0), so it doesn't need to be aliased into user space.
719          */
720
721         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
722                 PAGE_SIZE, PAGE_GATEWAY);
723 }
724
725 #ifdef CONFIG_HPUX
726 void
727 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
728 {
729         pgd_t *pg_dir;
730         pmd_t *pmd;
731         pte_t *pg_table;
732         unsigned long start_pmd;
733         unsigned long start_pte;
734         unsigned long address;
735         unsigned long hpux_gw_page_addr;
736         /* FIXME: This is 'const' in order to trick the compiler
737            into not treating it as DP-relative data. */
738         extern void * const hpux_gateway_page;
739
740         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
741
742         /*
743          * Setup HP-UX Gateway page.
744          *
745          * The HP-UX gateway page resides in the user address space,
746          * so it needs to be aliased into each process.
747          */
748
749         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
750
751 #if PTRS_PER_PMD == 1
752         start_pmd = 0;
753 #else
754         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
755 #endif
756         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
757
758         address = __pa(&hpux_gateway_page);
759 #if PTRS_PER_PMD == 1
760         pmd = (pmd_t *)__pa(pg_dir);
761 #else
762         pmd = (pmd_t *) pgd_address(*pg_dir);
763
764         /*
765          * pmd is physical at this point
766          */
767
768         if (!pmd) {
769                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
770                 pmd = (pmd_t *) __pa(pmd);
771         }
772
773         __pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
774 #endif
775         /* now change pmd to kernel virtual addresses */
776
777         pmd = (pmd_t *)__va(pmd) + start_pmd;
778
779         /*
780          * pg_table is physical at this point
781          */
782
783         pg_table = (pte_t *) pmd_address(*pmd);
784         if (!pg_table)
785                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
786
787         __pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
788
789         /* now change pg_table to kernel virtual addresses */
790
791         pg_table = (pte_t *) __va(pg_table) + start_pte;
792         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
793 }
794 EXPORT_SYMBOL(map_hpux_gateway_page);
795 #endif
796
797 void __init paging_init(void)
798 {
799         int i;
800
801         setup_bootmem();
802         pagetable_init();
803         gateway_init();
804         flush_cache_all_local(); /* start with known state */
805         flush_tlb_all_local(NULL);
806
807         for (i = 0; i < npmem_ranges; i++) {
808                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
809
810                 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
811
812 #ifdef CONFIG_DISCONTIGMEM
813                 /* Need to initialize the pfnnid_map before we can initialize
814                    the zone */
815                 {
816                     int j;
817                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
818                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
819                          j++) {
820                         pfnnid_map[j] = i;
821                     }
822                 }
823 #endif
824
825                 free_area_init_node(i, NODE_DATA(i), zones_size,
826                                 pmem_ranges[i].start_pfn, NULL);
827         }
828 }
829
830 #ifdef CONFIG_PA20
831
832 /*
833  * Currently, all PA20 chips have 18 bit protection id's, which is the
834  * limiting factor (space ids are 32 bits).
835  */
836
837 #define NR_SPACE_IDS 262144
838
839 #else
840
841 /*
842  * Currently we have a one-to-one relationship between space id's and
843  * protection id's. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
844  * support 15 bit protection id's, so that is the limiting factor.
845  * PCXT' has 18 bit protection id's, but only 16 bit spaceids, so it's
846  * probably not worth the effort for a special case here.
847  */
848
849 #define NR_SPACE_IDS 32768
850
851 #endif  /* !CONFIG_PA20 */
852
853 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
854 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
855
856 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
857 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
858 static unsigned long space_id_index;
859 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
860 static unsigned long dirty_space_ids = 0;
861
862 static DEFINE_SPINLOCK(sid_lock);
863
864 unsigned long alloc_sid(void)
865 {
866         unsigned long index;
867
868         spin_lock(&sid_lock);
869
870         if (free_space_ids == 0) {
871                 if (dirty_space_ids != 0) {
872                         spin_unlock(&sid_lock);
873                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
874                         spin_lock(&sid_lock);
875                 }
876                 BUG_ON(free_space_ids == 0);
877         }
878
879         free_space_ids--;
880
881         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
882         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
883         space_id_index = index;
884
885         spin_unlock(&sid_lock);
886
887         return index << SPACEID_SHIFT;
888 }
889
890 void free_sid(unsigned long spaceid)
891 {
892         unsigned long index = spaceid >> SPACEID_SHIFT;
893         unsigned long *dirty_space_offset;
894
895         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
896         index &= (BITS_PER_LONG - 1);
897
898         spin_lock(&sid_lock);
899
900         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
901
902         *dirty_space_offset |= (1L << index);
903         dirty_space_ids++;
904
905         spin_unlock(&sid_lock);
906 }
907
908
909 #ifdef CONFIG_SMP
910 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
911 {
912         int i;
913
914         /* NOTE: sid_lock must be held upon entry */
915
916         *ndirtyptr = dirty_space_ids;
917         if (dirty_space_ids != 0) {
918             for (i = 0; i < SID_ARRAY_SIZE; i++) {
919                 dirty_array[i] = dirty_space_id[i];
920                 dirty_space_id[i] = 0;
921             }
922             dirty_space_ids = 0;
923         }
924
925         return;
926 }
927
928 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
929 {
930         int i;
931
932         /* NOTE: sid_lock must be held upon entry */
933
934         if (ndirty != 0) {
935                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
936                         space_id[i] ^= dirty_array[i];
937                 }
938
939                 free_space_ids += ndirty;
940                 space_id_index = 0;
941         }
942 }
943
944 #else /* CONFIG_SMP */
945
946 static void recycle_sids(void)
947 {
948         int i;
949
950         /* NOTE: sid_lock must be held upon entry */
951
952         if (dirty_space_ids != 0) {
953                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
954                         space_id[i] ^= dirty_space_id[i];
955                         dirty_space_id[i] = 0;
956                 }
957
958                 free_space_ids += dirty_space_ids;
959                 dirty_space_ids = 0;
960                 space_id_index = 0;
961         }
962 }
963 #endif
964
965 /*
966  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
967  * purged, we can safely reuse the space ids that were released but
968  * not flushed from the tlb.
969  */
970
971 #ifdef CONFIG_SMP
972
973 static unsigned long recycle_ndirty;
974 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
975 static unsigned int recycle_inuse;
976
977 void flush_tlb_all(void)
978 {
979         int do_recycle;
980
981         do_recycle = 0;
982         spin_lock(&sid_lock);
983         if (dirty_space_ids > RECYCLE_THRESHOLD) {
984             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
985             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
986             recycle_inuse++;
987             do_recycle++;
988         }
989         spin_unlock(&sid_lock);
990         on_each_cpu(flush_tlb_all_local, NULL, 1, 1);
991         if (do_recycle) {
992             spin_lock(&sid_lock);
993             recycle_sids(recycle_ndirty,recycle_dirty_array);
994             recycle_inuse = 0;
995             spin_unlock(&sid_lock);
996         }
997 }
998 #else
999 void flush_tlb_all(void)
1000 {
1001         spin_lock(&sid_lock);
1002         flush_tlb_all_local(NULL);
1003         recycle_sids();
1004         spin_unlock(&sid_lock);
1005 }
1006 #endif
1007
1008 #ifdef CONFIG_BLK_DEV_INITRD
1009 void free_initrd_mem(unsigned long start, unsigned long end)
1010 {
1011         if (start >= end)
1012                 return;
1013         printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1014         for (; start < end; start += PAGE_SIZE) {
1015                 ClearPageReserved(virt_to_page(start));
1016                 init_page_count(virt_to_page(start));
1017                 free_page(start);
1018                 num_physpages++;
1019                 totalram_pages++;
1020         }
1021 }
1022 #endif