Merge branch 'for-linus' of git://neil.brown.name/md
[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 2006-2007 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/pgtable.h>
28 #include <asm/tlb.h>
29 #include <asm/pdc_chassis.h>
30 #include <asm/mmzone.h>
31 #include <asm/sections.h>
32
33 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
34
35 extern int  data_start;
36
37 #ifdef CONFIG_DISCONTIGMEM
38 struct node_map_data node_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 CONFIG_64BIT
69 #define MAX_MEM         (~0UL)
70 #else /* !CONFIG_64BIT */
71 #define MAX_MEM         (3584U*1024U*1024U)
72 #endif /* !CONFIG_64BIT */
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 = &bootmem_node_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 +
318                                 PDC_CONSOLE_IO_IODC_SIZE), BOOTMEM_DEFAULT);
319         reserve_bootmem_node(NODE_DATA(0), __pa((unsigned long)_text),
320                         (unsigned long)(_end - _text), BOOTMEM_DEFAULT);
321         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
322                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT),
323                         BOOTMEM_DEFAULT);
324
325 #ifndef CONFIG_DISCONTIGMEM
326
327         /* reserve the holes */
328
329         for (i = 0; i < npmem_holes; i++) {
330                 reserve_bootmem_node(NODE_DATA(0),
331                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
332                                 (pmem_holes[i].pages << PAGE_SHIFT),
333                                 BOOTMEM_DEFAULT);
334         }
335 #endif
336
337 #ifdef CONFIG_BLK_DEV_INITRD
338         if (initrd_start) {
339                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
340                 if (__pa(initrd_start) < mem_max) {
341                         unsigned long initrd_reserve;
342
343                         if (__pa(initrd_end) > mem_max) {
344                                 initrd_reserve = mem_max - __pa(initrd_start);
345                         } else {
346                                 initrd_reserve = initrd_end - initrd_start;
347                         }
348                         initrd_below_start_ok = 1;
349                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
350
351                         reserve_bootmem_node(NODE_DATA(0), __pa(initrd_start),
352                                         initrd_reserve, BOOTMEM_DEFAULT);
353                 }
354         }
355 #endif
356
357         data_resource.start =  virt_to_phys(&data_start);
358         data_resource.end = virt_to_phys(_end) - 1;
359         code_resource.start = virt_to_phys(_text);
360         code_resource.end = virt_to_phys(&data_start)-1;
361
362         /* We don't know which region the kernel will be in, so try
363          * all of them.
364          */
365         for (i = 0; i < sysram_resource_count; i++) {
366                 struct resource *res = &sysram_resources[i];
367                 request_resource(res, &code_resource);
368                 request_resource(res, &data_resource);
369         }
370         request_resource(&sysram_resources[0], &pdcdata_resource);
371 }
372
373 void free_initmem(void)
374 {
375         unsigned long addr, init_begin, init_end;
376
377         printk(KERN_INFO "Freeing unused kernel memory: ");
378
379 #ifdef CONFIG_DEBUG_KERNEL
380         /* Attempt to catch anyone trying to execute code here
381          * by filling the page with BRK insns.
382          * 
383          * If we disable interrupts for all CPUs, then IPI stops working.
384          * Kinda breaks the global cache flushing.
385          */
386         local_irq_disable();
387
388         memset(__init_begin, 0x00,
389                 (unsigned long)__init_end - (unsigned long)__init_begin);
390
391         flush_data_cache();
392         asm volatile("sync" : : );
393         flush_icache_range((unsigned long)__init_begin, (unsigned long)__init_end);
394         asm volatile("sync" : : );
395
396         local_irq_enable();
397 #endif
398         
399         /* align __init_begin and __init_end to page size,
400            ignoring linker script where we might have tried to save RAM */
401         init_begin = PAGE_ALIGN((unsigned long)(__init_begin));
402         init_end   = PAGE_ALIGN((unsigned long)(__init_end));
403         for (addr = init_begin; addr < init_end; addr += PAGE_SIZE) {
404                 ClearPageReserved(virt_to_page(addr));
405                 init_page_count(virt_to_page(addr));
406                 free_page(addr);
407                 num_physpages++;
408                 totalram_pages++;
409         }
410
411         /* set up a new led state on systems shipped LED State panel */
412         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
413         
414         printk("%luk freed\n", (init_end - init_begin) >> 10);
415 }
416
417
418 #ifdef CONFIG_DEBUG_RODATA
419 void mark_rodata_ro(void)
420 {
421         /* rodata memory was already mapped with KERNEL_RO access rights by
422            pagetable_init() and map_pages(). No need to do additional stuff here */
423         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
424                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
425 }
426 #endif
427
428
429 /*
430  * Just an arbitrary offset to serve as a "hole" between mapping areas
431  * (between top of physical memory and a potential pcxl dma mapping
432  * area, and below the vmalloc mapping area).
433  *
434  * The current 32K value just means that there will be a 32K "hole"
435  * between mapping areas. That means that  any out-of-bounds memory
436  * accesses will hopefully be caught. The vmalloc() routines leaves
437  * a hole of 4kB between each vmalloced area for the same reason.
438  */
439
440  /* Leave room for gateway page expansion */
441 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
442 #error KERNEL_MAP_START is in gateway reserved region
443 #endif
444 #define MAP_START (KERNEL_MAP_START)
445
446 #define VM_MAP_OFFSET  (32*1024)
447 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
448                                      & ~(VM_MAP_OFFSET-1)))
449
450 void *vmalloc_start __read_mostly;
451 EXPORT_SYMBOL(vmalloc_start);
452
453 #ifdef CONFIG_PA11
454 unsigned long pcxl_dma_start __read_mostly;
455 #endif
456
457 void __init mem_init(void)
458 {
459         int codesize, reservedpages, datasize, initsize;
460
461         high_memory = __va((max_pfn << PAGE_SHIFT));
462
463 #ifndef CONFIG_DISCONTIGMEM
464         max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
465         totalram_pages += free_all_bootmem();
466 #else
467         {
468                 int i;
469
470                 for (i = 0; i < npmem_ranges; i++)
471                         totalram_pages += free_all_bootmem_node(NODE_DATA(i));
472         }
473 #endif
474
475         codesize = (unsigned long)_etext - (unsigned long)_text;
476         datasize = (unsigned long)_edata - (unsigned long)_etext;
477         initsize = (unsigned long)__init_end - (unsigned long)__init_begin;
478
479         reservedpages = 0;
480 {
481         unsigned long pfn;
482 #ifdef CONFIG_DISCONTIGMEM
483         int i;
484
485         for (i = 0; i < npmem_ranges; i++) {
486                 for (pfn = node_start_pfn(i); pfn < node_end_pfn(i); pfn++) {
487                         if (PageReserved(pfn_to_page(pfn)))
488                                 reservedpages++;
489                 }
490         }
491 #else /* !CONFIG_DISCONTIGMEM */
492         for (pfn = 0; pfn < max_pfn; pfn++) {
493                 /*
494                  * Only count reserved RAM pages
495                  */
496                 if (PageReserved(pfn_to_page(pfn)))
497                         reservedpages++;
498         }
499 #endif
500 }
501
502 #ifdef CONFIG_PA11
503         if (hppa_dma_ops == &pcxl_dma_ops) {
504                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
505                 vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
506         } else {
507                 pcxl_dma_start = 0;
508                 vmalloc_start = SET_MAP_OFFSET(MAP_START);
509         }
510 #else
511         vmalloc_start = SET_MAP_OFFSET(MAP_START);
512 #endif
513
514         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, %dk reserved, %dk data, %dk init)\n",
515                 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
516                 num_physpages << (PAGE_SHIFT-10),
517                 codesize >> 10,
518                 reservedpages << (PAGE_SHIFT-10),
519                 datasize >> 10,
520                 initsize >> 10
521         );
522
523 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
524         printk("virtual kernel memory layout:\n"
525                "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
526                "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
527                "      .init : 0x%p - 0x%p   (%4ld kB)\n"
528                "      .data : 0x%p - 0x%p   (%4ld kB)\n"
529                "      .text : 0x%p - 0x%p   (%4ld kB)\n",
530
531                (void*)VMALLOC_START, (void*)VMALLOC_END,
532                (VMALLOC_END - VMALLOC_START) >> 20,
533
534                __va(0), high_memory,
535                ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
536
537                __init_begin, __init_end,
538                ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
539
540                _etext, _edata,
541                ((unsigned long)_edata - (unsigned long)_etext) >> 10,
542
543                _text, _etext,
544                ((unsigned long)_etext - (unsigned long)_text) >> 10);
545 #endif
546 }
547
548 unsigned long *empty_zero_page __read_mostly;
549 EXPORT_SYMBOL(empty_zero_page);
550
551 void show_mem(void)
552 {
553         int i,free = 0,total = 0,reserved = 0;
554         int shared = 0, cached = 0;
555
556         printk(KERN_INFO "Mem-info:\n");
557         show_free_areas();
558 #ifndef CONFIG_DISCONTIGMEM
559         i = max_mapnr;
560         while (i-- > 0) {
561                 total++;
562                 if (PageReserved(mem_map+i))
563                         reserved++;
564                 else if (PageSwapCache(mem_map+i))
565                         cached++;
566                 else if (!page_count(&mem_map[i]))
567                         free++;
568                 else
569                         shared += page_count(&mem_map[i]) - 1;
570         }
571 #else
572         for (i = 0; i < npmem_ranges; i++) {
573                 int j;
574
575                 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
576                         struct page *p;
577                         unsigned long flags;
578
579                         pgdat_resize_lock(NODE_DATA(i), &flags);
580                         p = nid_page_nr(i, j) - node_start_pfn(i);
581
582                         total++;
583                         if (PageReserved(p))
584                                 reserved++;
585                         else if (PageSwapCache(p))
586                                 cached++;
587                         else if (!page_count(p))
588                                 free++;
589                         else
590                                 shared += page_count(p) - 1;
591                         pgdat_resize_unlock(NODE_DATA(i), &flags);
592                 }
593         }
594 #endif
595         printk(KERN_INFO "%d pages of RAM\n", total);
596         printk(KERN_INFO "%d reserved pages\n", reserved);
597         printk(KERN_INFO "%d pages shared\n", shared);
598         printk(KERN_INFO "%d pages swap cached\n", cached);
599
600
601 #ifdef CONFIG_DISCONTIGMEM
602         {
603                 struct zonelist *zl;
604                 int i, j;
605
606                 for (i = 0; i < npmem_ranges; i++) {
607                         zl = node_zonelist(i, 0);
608                         for (j = 0; j < MAX_NR_ZONES; j++) {
609                                 struct zoneref *z;
610                                 struct zone *zone;
611
612                                 printk("Zone list for zone %d on node %d: ", j, i);
613                                 for_each_zone_zonelist(zone, z, zl, j)
614                                         printk("[%d/%s] ", zone_to_nid(zone),
615                                                                 zone->name);
616                                 printk("\n");
617                         }
618                 }
619         }
620 #endif
621 }
622
623
624 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
625 {
626         pgd_t *pg_dir;
627         pmd_t *pmd;
628         pte_t *pg_table;
629         unsigned long end_paddr;
630         unsigned long start_pmd;
631         unsigned long start_pte;
632         unsigned long tmp1;
633         unsigned long tmp2;
634         unsigned long address;
635         unsigned long ro_start;
636         unsigned long ro_end;
637         unsigned long fv_addr;
638         unsigned long gw_addr;
639         extern const unsigned long fault_vector_20;
640         extern void * const linux_gateway_page;
641
642         ro_start = __pa((unsigned long)_text);
643         ro_end   = __pa((unsigned long)&data_start);
644         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
645         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
646
647         end_paddr = start_paddr + size;
648
649         pg_dir = pgd_offset_k(start_vaddr);
650
651 #if PTRS_PER_PMD == 1
652         start_pmd = 0;
653 #else
654         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
655 #endif
656         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
657
658         address = start_paddr;
659         while (address < end_paddr) {
660 #if PTRS_PER_PMD == 1
661                 pmd = (pmd_t *)__pa(pg_dir);
662 #else
663                 pmd = (pmd_t *)pgd_address(*pg_dir);
664
665                 /*
666                  * pmd is physical at this point
667                  */
668
669                 if (!pmd) {
670                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
671                         pmd = (pmd_t *) __pa(pmd);
672                 }
673
674                 pgd_populate(NULL, pg_dir, __va(pmd));
675 #endif
676                 pg_dir++;
677
678                 /* now change pmd to kernel virtual addresses */
679
680                 pmd = (pmd_t *)__va(pmd) + start_pmd;
681                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
682
683                         /*
684                          * pg_table is physical at this point
685                          */
686
687                         pg_table = (pte_t *)pmd_address(*pmd);
688                         if (!pg_table) {
689                                 pg_table = (pte_t *)
690                                         alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
691                                 pg_table = (pte_t *) __pa(pg_table);
692                         }
693
694                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
695
696                         /* now change pg_table to kernel virtual addresses */
697
698                         pg_table = (pte_t *) __va(pg_table) + start_pte;
699                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
700                                 pte_t pte;
701
702                                 /*
703                                  * Map the fault vector writable so we can
704                                  * write the HPMC checksum.
705                                  */
706 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
707                                 if (address >= ro_start && address < ro_end
708                                                         && address != fv_addr
709                                                         && address != gw_addr)
710                                     pte = __mk_pte(address, PAGE_KERNEL_RO);
711                                 else
712 #endif
713                                     pte = __mk_pte(address, pgprot);
714
715                                 if (address >= end_paddr)
716                                         pte_val(pte) = 0;
717
718                                 set_pte(pg_table, pte);
719
720                                 address += PAGE_SIZE;
721                         }
722                         start_pte = 0;
723
724                         if (address >= end_paddr)
725                             break;
726                 }
727                 start_pmd = 0;
728         }
729 }
730
731 /*
732  * pagetable_init() sets up the page tables
733  *
734  * Note that gateway_init() places the Linux gateway page at page 0.
735  * Since gateway pages cannot be dereferenced this has the desirable
736  * side effect of trapping those pesky NULL-reference errors in the
737  * kernel.
738  */
739 static void __init pagetable_init(void)
740 {
741         int range;
742
743         /* Map each physical memory range to its kernel vaddr */
744
745         for (range = 0; range < npmem_ranges; range++) {
746                 unsigned long start_paddr;
747                 unsigned long end_paddr;
748                 unsigned long size;
749
750                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
751                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
752                 size = pmem_ranges[range].pages << PAGE_SHIFT;
753
754                 map_pages((unsigned long)__va(start_paddr), start_paddr,
755                         size, PAGE_KERNEL);
756         }
757
758 #ifdef CONFIG_BLK_DEV_INITRD
759         if (initrd_end && initrd_end > mem_limit) {
760                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
761                 map_pages(initrd_start, __pa(initrd_start),
762                         initrd_end - initrd_start, PAGE_KERNEL);
763         }
764 #endif
765
766         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
767         memset(empty_zero_page, 0, PAGE_SIZE);
768 }
769
770 static void __init gateway_init(void)
771 {
772         unsigned long linux_gateway_page_addr;
773         /* FIXME: This is 'const' in order to trick the compiler
774            into not treating it as DP-relative data. */
775         extern void * const linux_gateway_page;
776
777         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
778
779         /*
780          * Setup Linux Gateway page.
781          *
782          * The Linux gateway page will reside in kernel space (on virtual
783          * page 0), so it doesn't need to be aliased into user space.
784          */
785
786         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
787                 PAGE_SIZE, PAGE_GATEWAY);
788 }
789
790 #ifdef CONFIG_HPUX
791 void
792 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
793 {
794         pgd_t *pg_dir;
795         pmd_t *pmd;
796         pte_t *pg_table;
797         unsigned long start_pmd;
798         unsigned long start_pte;
799         unsigned long address;
800         unsigned long hpux_gw_page_addr;
801         /* FIXME: This is 'const' in order to trick the compiler
802            into not treating it as DP-relative data. */
803         extern void * const hpux_gateway_page;
804
805         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
806
807         /*
808          * Setup HP-UX Gateway page.
809          *
810          * The HP-UX gateway page resides in the user address space,
811          * so it needs to be aliased into each process.
812          */
813
814         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
815
816 #if PTRS_PER_PMD == 1
817         start_pmd = 0;
818 #else
819         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
820 #endif
821         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
822
823         address = __pa(&hpux_gateway_page);
824 #if PTRS_PER_PMD == 1
825         pmd = (pmd_t *)__pa(pg_dir);
826 #else
827         pmd = (pmd_t *) pgd_address(*pg_dir);
828
829         /*
830          * pmd is physical at this point
831          */
832
833         if (!pmd) {
834                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
835                 pmd = (pmd_t *) __pa(pmd);
836         }
837
838         __pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
839 #endif
840         /* now change pmd to kernel virtual addresses */
841
842         pmd = (pmd_t *)__va(pmd) + start_pmd;
843
844         /*
845          * pg_table is physical at this point
846          */
847
848         pg_table = (pte_t *) pmd_address(*pmd);
849         if (!pg_table)
850                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
851
852         __pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
853
854         /* now change pg_table to kernel virtual addresses */
855
856         pg_table = (pte_t *) __va(pg_table) + start_pte;
857         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
858 }
859 EXPORT_SYMBOL(map_hpux_gateway_page);
860 #endif
861
862 void __init paging_init(void)
863 {
864         int i;
865
866         setup_bootmem();
867         pagetable_init();
868         gateway_init();
869         flush_cache_all_local(); /* start with known state */
870         flush_tlb_all_local(NULL);
871
872         for (i = 0; i < npmem_ranges; i++) {
873                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
874
875                 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
876
877 #ifdef CONFIG_DISCONTIGMEM
878                 /* Need to initialize the pfnnid_map before we can initialize
879                    the zone */
880                 {
881                     int j;
882                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
883                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
884                          j++) {
885                         pfnnid_map[j] = i;
886                     }
887                 }
888 #endif
889
890                 free_area_init_node(i, zones_size,
891                                 pmem_ranges[i].start_pfn, NULL);
892         }
893 }
894
895 #ifdef CONFIG_PA20
896
897 /*
898  * Currently, all PA20 chips have 18 bit protection IDs, which is the
899  * limiting factor (space ids are 32 bits).
900  */
901
902 #define NR_SPACE_IDS 262144
903
904 #else
905
906 /*
907  * Currently we have a one-to-one relationship between space IDs and
908  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
909  * support 15 bit protection IDs, so that is the limiting factor.
910  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
911  * probably not worth the effort for a special case here.
912  */
913
914 #define NR_SPACE_IDS 32768
915
916 #endif  /* !CONFIG_PA20 */
917
918 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
919 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
920
921 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
922 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
923 static unsigned long space_id_index;
924 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
925 static unsigned long dirty_space_ids = 0;
926
927 static DEFINE_SPINLOCK(sid_lock);
928
929 unsigned long alloc_sid(void)
930 {
931         unsigned long index;
932
933         spin_lock(&sid_lock);
934
935         if (free_space_ids == 0) {
936                 if (dirty_space_ids != 0) {
937                         spin_unlock(&sid_lock);
938                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
939                         spin_lock(&sid_lock);
940                 }
941                 BUG_ON(free_space_ids == 0);
942         }
943
944         free_space_ids--;
945
946         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
947         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
948         space_id_index = index;
949
950         spin_unlock(&sid_lock);
951
952         return index << SPACEID_SHIFT;
953 }
954
955 void free_sid(unsigned long spaceid)
956 {
957         unsigned long index = spaceid >> SPACEID_SHIFT;
958         unsigned long *dirty_space_offset;
959
960         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
961         index &= (BITS_PER_LONG - 1);
962
963         spin_lock(&sid_lock);
964
965         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
966
967         *dirty_space_offset |= (1L << index);
968         dirty_space_ids++;
969
970         spin_unlock(&sid_lock);
971 }
972
973
974 #ifdef CONFIG_SMP
975 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
976 {
977         int i;
978
979         /* NOTE: sid_lock must be held upon entry */
980
981         *ndirtyptr = dirty_space_ids;
982         if (dirty_space_ids != 0) {
983             for (i = 0; i < SID_ARRAY_SIZE; i++) {
984                 dirty_array[i] = dirty_space_id[i];
985                 dirty_space_id[i] = 0;
986             }
987             dirty_space_ids = 0;
988         }
989
990         return;
991 }
992
993 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
994 {
995         int i;
996
997         /* NOTE: sid_lock must be held upon entry */
998
999         if (ndirty != 0) {
1000                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
1001                         space_id[i] ^= dirty_array[i];
1002                 }
1003
1004                 free_space_ids += ndirty;
1005                 space_id_index = 0;
1006         }
1007 }
1008
1009 #else /* CONFIG_SMP */
1010
1011 static void recycle_sids(void)
1012 {
1013         int i;
1014
1015         /* NOTE: sid_lock must be held upon entry */
1016
1017         if (dirty_space_ids != 0) {
1018                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
1019                         space_id[i] ^= dirty_space_id[i];
1020                         dirty_space_id[i] = 0;
1021                 }
1022
1023                 free_space_ids += dirty_space_ids;
1024                 dirty_space_ids = 0;
1025                 space_id_index = 0;
1026         }
1027 }
1028 #endif
1029
1030 /*
1031  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
1032  * purged, we can safely reuse the space ids that were released but
1033  * not flushed from the tlb.
1034  */
1035
1036 #ifdef CONFIG_SMP
1037
1038 static unsigned long recycle_ndirty;
1039 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
1040 static unsigned int recycle_inuse;
1041
1042 void flush_tlb_all(void)
1043 {
1044         int do_recycle;
1045
1046         do_recycle = 0;
1047         spin_lock(&sid_lock);
1048         if (dirty_space_ids > RECYCLE_THRESHOLD) {
1049             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
1050             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
1051             recycle_inuse++;
1052             do_recycle++;
1053         }
1054         spin_unlock(&sid_lock);
1055         on_each_cpu(flush_tlb_all_local, NULL, 1);
1056         if (do_recycle) {
1057             spin_lock(&sid_lock);
1058             recycle_sids(recycle_ndirty,recycle_dirty_array);
1059             recycle_inuse = 0;
1060             spin_unlock(&sid_lock);
1061         }
1062 }
1063 #else
1064 void flush_tlb_all(void)
1065 {
1066         spin_lock(&sid_lock);
1067         flush_tlb_all_local(NULL);
1068         recycle_sids();
1069         spin_unlock(&sid_lock);
1070 }
1071 #endif
1072
1073 #ifdef CONFIG_BLK_DEV_INITRD
1074 void free_initrd_mem(unsigned long start, unsigned long end)
1075 {
1076         if (start >= end)
1077                 return;
1078         printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1079         for (; start < end; start += PAGE_SIZE) {
1080                 ClearPageReserved(virt_to_page(start));
1081                 init_page_count(virt_to_page(start));
1082                 free_page(start);
1083                 num_physpages++;
1084                 totalram_pages++;
1085         }
1086 }
1087 #endif