Pull ec into test branch
[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         extern char saved_command_line[];
81
82         /* We need this before __setup() functions are called */
83
84         limit = MAX_MEM;
85         for (cp = saved_command_line; *cp; ) {
86                 if (memcmp(cp, "mem=", 4) == 0) {
87                         cp += 4;
88                         limit = memparse(cp, &end);
89                         if (end != cp)
90                                 break;
91                         cp = end;
92                 } else {
93                         while (*cp != ' ' && *cp)
94                                 ++cp;
95                         while (*cp == ' ')
96                                 ++cp;
97                 }
98         }
99
100         if (limit < mem_limit)
101                 mem_limit = limit;
102 }
103
104 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
105
106 static void __init setup_bootmem(void)
107 {
108         unsigned long bootmap_size;
109         unsigned long mem_max;
110         unsigned long bootmap_pages;
111         unsigned long bootmap_start_pfn;
112         unsigned long bootmap_pfn;
113 #ifndef CONFIG_DISCONTIGMEM
114         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
115         int npmem_holes;
116 #endif
117         int i, sysram_resource_count;
118
119         disable_sr_hashing(); /* Turn off space register hashing */
120
121         /*
122          * Sort the ranges. Since the number of ranges is typically
123          * small, and performance is not an issue here, just do
124          * a simple insertion sort.
125          */
126
127         for (i = 1; i < npmem_ranges; i++) {
128                 int j;
129
130                 for (j = i; j > 0; j--) {
131                         unsigned long tmp;
132
133                         if (pmem_ranges[j-1].start_pfn <
134                             pmem_ranges[j].start_pfn) {
135
136                                 break;
137                         }
138                         tmp = pmem_ranges[j-1].start_pfn;
139                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
140                         pmem_ranges[j].start_pfn = tmp;
141                         tmp = pmem_ranges[j-1].pages;
142                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
143                         pmem_ranges[j].pages = tmp;
144                 }
145         }
146
147 #ifndef CONFIG_DISCONTIGMEM
148         /*
149          * Throw out ranges that are too far apart (controlled by
150          * MAX_GAP).
151          */
152
153         for (i = 1; i < npmem_ranges; i++) {
154                 if (pmem_ranges[i].start_pfn -
155                         (pmem_ranges[i-1].start_pfn +
156                          pmem_ranges[i-1].pages) > MAX_GAP) {
157                         npmem_ranges = i;
158                         printk("Large gap in memory detected (%ld pages). "
159                                "Consider turning on CONFIG_DISCONTIGMEM\n",
160                                pmem_ranges[i].start_pfn -
161                                (pmem_ranges[i-1].start_pfn +
162                                 pmem_ranges[i-1].pages));
163                         break;
164                 }
165         }
166 #endif
167
168         if (npmem_ranges > 1) {
169
170                 /* Print the memory ranges */
171
172                 printk(KERN_INFO "Memory Ranges:\n");
173
174                 for (i = 0; i < npmem_ranges; i++) {
175                         unsigned long start;
176                         unsigned long size;
177
178                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
179                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
180                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
181                                 i,start, start + (size - 1), size >> 20);
182                 }
183         }
184
185         sysram_resource_count = npmem_ranges;
186         for (i = 0; i < sysram_resource_count; i++) {
187                 struct resource *res = &sysram_resources[i];
188                 res->name = "System RAM";
189                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
190                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
191                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
192                 request_resource(&iomem_resource, res);
193         }
194
195         /*
196          * For 32 bit kernels we limit the amount of memory we can
197          * support, in order to preserve enough kernel address space
198          * for other purposes. For 64 bit kernels we don't normally
199          * limit the memory, but this mechanism can be used to
200          * artificially limit the amount of memory (and it is written
201          * to work with multiple memory ranges).
202          */
203
204         mem_limit_func();       /* check for "mem=" argument */
205
206         mem_max = 0;
207         num_physpages = 0;
208         for (i = 0; i < npmem_ranges; i++) {
209                 unsigned long rsize;
210
211                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
212                 if ((mem_max + rsize) > mem_limit) {
213                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
214                         if (mem_max == mem_limit)
215                                 npmem_ranges = i;
216                         else {
217                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
218                                                        - (mem_max >> PAGE_SHIFT);
219                                 npmem_ranges = i + 1;
220                                 mem_max = mem_limit;
221                         }
222                 num_physpages += pmem_ranges[i].pages;
223                         break;
224                 }
225             num_physpages += pmem_ranges[i].pages;
226                 mem_max += rsize;
227         }
228
229         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
230
231 #ifndef CONFIG_DISCONTIGMEM
232         /* Merge the ranges, keeping track of the holes */
233
234         {
235                 unsigned long end_pfn;
236                 unsigned long hole_pages;
237
238                 npmem_holes = 0;
239                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
240                 for (i = 1; i < npmem_ranges; i++) {
241
242                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
243                         if (hole_pages) {
244                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
245                                 pmem_holes[npmem_holes++].pages = hole_pages;
246                                 end_pfn += hole_pages;
247                         }
248                         end_pfn += pmem_ranges[i].pages;
249                 }
250
251                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
252                 npmem_ranges = 1;
253         }
254 #endif
255
256         bootmap_pages = 0;
257         for (i = 0; i < npmem_ranges; i++)
258                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
259
260         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
261
262 #ifdef CONFIG_DISCONTIGMEM
263         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
264                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
265                 NODE_DATA(i)->bdata = &bmem_data[i];
266         }
267         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
268
269         for (i = 0; i < npmem_ranges; i++)
270                 node_set_online(i);
271 #endif
272
273         /*
274          * Initialize and free the full range of memory in each range.
275          * Note that the only writing these routines do are to the bootmap,
276          * and we've made sure to locate the bootmap properly so that they
277          * won't be writing over anything important.
278          */
279
280         bootmap_pfn = bootmap_start_pfn;
281         max_pfn = 0;
282         for (i = 0; i < npmem_ranges; i++) {
283                 unsigned long start_pfn;
284                 unsigned long npages;
285
286                 start_pfn = pmem_ranges[i].start_pfn;
287                 npages = pmem_ranges[i].pages;
288
289                 bootmap_size = init_bootmem_node(NODE_DATA(i),
290                                                 bootmap_pfn,
291                                                 start_pfn,
292                                                 (start_pfn + npages) );
293                 free_bootmem_node(NODE_DATA(i),
294                                   (start_pfn << PAGE_SHIFT),
295                                   (npages << PAGE_SHIFT) );
296                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
297                 if ((start_pfn + npages) > max_pfn)
298                         max_pfn = start_pfn + npages;
299         }
300
301         /* IOMMU is always used to access "high mem" on those boxes
302          * that can support enough mem that a PCI device couldn't
303          * directly DMA to any physical addresses.
304          * ISA DMA support will need to revisit this.
305          */
306         max_low_pfn = max_pfn;
307
308         if ((bootmap_pfn - bootmap_start_pfn) != bootmap_pages) {
309                 printk(KERN_WARNING "WARNING! bootmap sizing is messed up!\n");
310                 BUG();
311         }
312
313         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
314
315 #define PDC_CONSOLE_IO_IODC_SIZE 32768
316
317         reserve_bootmem_node(NODE_DATA(0), 0UL,
318                         (unsigned long)(PAGE0->mem_free + PDC_CONSOLE_IO_IODC_SIZE));
319         reserve_bootmem_node(NODE_DATA(0), __pa((unsigned long)_text),
320                         (unsigned long)(_end - _text));
321         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
322                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT));
323
324 #ifndef CONFIG_DISCONTIGMEM
325
326         /* reserve the holes */
327
328         for (i = 0; i < npmem_holes; i++) {
329                 reserve_bootmem_node(NODE_DATA(0),
330                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
331                                 (pmem_holes[i].pages << PAGE_SHIFT));
332         }
333 #endif
334
335 #ifdef CONFIG_BLK_DEV_INITRD
336         if (initrd_start) {
337                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
338                 if (__pa(initrd_start) < mem_max) {
339                         unsigned long initrd_reserve;
340
341                         if (__pa(initrd_end) > mem_max) {
342                                 initrd_reserve = mem_max - __pa(initrd_start);
343                         } else {
344                                 initrd_reserve = initrd_end - initrd_start;
345                         }
346                         initrd_below_start_ok = 1;
347                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
348
349                         reserve_bootmem_node(NODE_DATA(0),__pa(initrd_start), initrd_reserve);
350                 }
351         }
352 #endif
353
354         data_resource.start =  virt_to_phys(&data_start);
355         data_resource.end = virt_to_phys(_end) - 1;
356         code_resource.start = virt_to_phys(_text);
357         code_resource.end = virt_to_phys(&data_start)-1;
358
359         /* We don't know which region the kernel will be in, so try
360          * all of them.
361          */
362         for (i = 0; i < sysram_resource_count; i++) {
363                 struct resource *res = &sysram_resources[i];
364                 request_resource(res, &code_resource);
365                 request_resource(res, &data_resource);
366         }
367         request_resource(&sysram_resources[0], &pdcdata_resource);
368 }
369
370 void free_initmem(void)
371 {
372         unsigned long addr, init_begin, init_end;
373
374         printk(KERN_INFO "Freeing unused kernel memory: ");
375
376 #ifdef CONFIG_DEBUG_KERNEL
377         /* Attempt to catch anyone trying to execute code here
378          * by filling the page with BRK insns.
379          * 
380          * If we disable interrupts for all CPUs, then IPI stops working.
381          * Kinda breaks the global cache flushing.
382          */
383         local_irq_disable();
384
385         memset(__init_begin, 0x00,
386                 (unsigned long)__init_end - (unsigned long)__init_begin);
387
388         flush_data_cache();
389         asm volatile("sync" : : );
390         flush_icache_range((unsigned long)__init_begin, (unsigned long)__init_end);
391         asm volatile("sync" : : );
392
393         local_irq_enable();
394 #endif
395         
396         /* align __init_begin and __init_end to page size,
397            ignoring linker script where we might have tried to save RAM */
398         init_begin = PAGE_ALIGN((unsigned long)(__init_begin));
399         init_end   = PAGE_ALIGN((unsigned long)(__init_end));
400         for (addr = init_begin; addr < init_end; addr += PAGE_SIZE) {
401                 ClearPageReserved(virt_to_page(addr));
402                 init_page_count(virt_to_page(addr));
403                 free_page(addr);
404                 num_physpages++;
405                 totalram_pages++;
406         }
407
408         /* set up a new led state on systems shipped LED State panel */
409         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
410         
411         printk("%luk freed\n", (init_end - init_begin) >> 10);
412 }
413
414
415 #ifdef CONFIG_DEBUG_RODATA
416 void mark_rodata_ro(void)
417 {
418         /* rodata memory was already mapped with KERNEL_RO access rights by
419            pagetable_init() and map_pages(). No need to do additional stuff here */
420         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
421                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
422 }
423 #endif
424
425
426 /*
427  * Just an arbitrary offset to serve as a "hole" between mapping areas
428  * (between top of physical memory and a potential pcxl dma mapping
429  * area, and below the vmalloc mapping area).
430  *
431  * The current 32K value just means that there will be a 32K "hole"
432  * between mapping areas. That means that  any out-of-bounds memory
433  * accesses will hopefully be caught. The vmalloc() routines leaves
434  * a hole of 4kB between each vmalloced area for the same reason.
435  */
436
437  /* Leave room for gateway page expansion */
438 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
439 #error KERNEL_MAP_START is in gateway reserved region
440 #endif
441 #define MAP_START (KERNEL_MAP_START)
442
443 #define VM_MAP_OFFSET  (32*1024)
444 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
445                                      & ~(VM_MAP_OFFSET-1)))
446
447 void *vmalloc_start __read_mostly;
448 EXPORT_SYMBOL(vmalloc_start);
449
450 #ifdef CONFIG_PA11
451 unsigned long pcxl_dma_start __read_mostly;
452 #endif
453
454 void __init mem_init(void)
455 {
456         high_memory = __va((max_pfn << PAGE_SHIFT));
457
458 #ifndef CONFIG_DISCONTIGMEM
459         max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
460         totalram_pages += free_all_bootmem();
461 #else
462         {
463                 int i;
464
465                 for (i = 0; i < npmem_ranges; i++)
466                         totalram_pages += free_all_bootmem_node(NODE_DATA(i));
467         }
468 #endif
469
470         printk(KERN_INFO "Memory: %luk available\n", num_physpages << (PAGE_SHIFT-10));
471
472 #ifdef CONFIG_PA11
473         if (hppa_dma_ops == &pcxl_dma_ops) {
474                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
475                 vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
476         } else {
477                 pcxl_dma_start = 0;
478                 vmalloc_start = SET_MAP_OFFSET(MAP_START);
479         }
480 #else
481         vmalloc_start = SET_MAP_OFFSET(MAP_START);
482 #endif
483
484 }
485
486 unsigned long *empty_zero_page __read_mostly;
487
488 void show_mem(void)
489 {
490         int i,free = 0,total = 0,reserved = 0;
491         int shared = 0, cached = 0;
492
493         printk(KERN_INFO "Mem-info:\n");
494         show_free_areas();
495         printk(KERN_INFO "Free swap:     %6ldkB\n",
496                                 nr_swap_pages<<(PAGE_SHIFT-10));
497 #ifndef CONFIG_DISCONTIGMEM
498         i = max_mapnr;
499         while (i-- > 0) {
500                 total++;
501                 if (PageReserved(mem_map+i))
502                         reserved++;
503                 else if (PageSwapCache(mem_map+i))
504                         cached++;
505                 else if (!page_count(&mem_map[i]))
506                         free++;
507                 else
508                         shared += page_count(&mem_map[i]) - 1;
509         }
510 #else
511         for (i = 0; i < npmem_ranges; i++) {
512                 int j;
513
514                 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
515                         struct page *p;
516                         unsigned long flags;
517
518                         pgdat_resize_lock(NODE_DATA(i), &flags);
519                         p = nid_page_nr(i, j) - node_start_pfn(i);
520
521                         total++;
522                         if (PageReserved(p))
523                                 reserved++;
524                         else if (PageSwapCache(p))
525                                 cached++;
526                         else if (!page_count(p))
527                                 free++;
528                         else
529                                 shared += page_count(p) - 1;
530                         pgdat_resize_unlock(NODE_DATA(i), &flags);
531                 }
532         }
533 #endif
534         printk(KERN_INFO "%d pages of RAM\n", total);
535         printk(KERN_INFO "%d reserved pages\n", reserved);
536         printk(KERN_INFO "%d pages shared\n", shared);
537         printk(KERN_INFO "%d pages swap cached\n", cached);
538
539
540 #ifdef CONFIG_DISCONTIGMEM
541         {
542                 struct zonelist *zl;
543                 int i, j, k;
544
545                 for (i = 0; i < npmem_ranges; i++) {
546                         for (j = 0; j < MAX_NR_ZONES; j++) {
547                                 zl = NODE_DATA(i)->node_zonelists + j;
548
549                                 printk("Zone list for zone %d on node %d: ", j, i);
550                                 for (k = 0; zl->zones[k] != NULL; k++) 
551                                         printk("[%d/%s] ", zone_to_nid(zl->zones[k]), zl->zones[k]->name);
552                                 printk("\n");
553                         }
554                 }
555         }
556 #endif
557 }
558
559
560 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
561 {
562         pgd_t *pg_dir;
563         pmd_t *pmd;
564         pte_t *pg_table;
565         unsigned long end_paddr;
566         unsigned long start_pmd;
567         unsigned long start_pte;
568         unsigned long tmp1;
569         unsigned long tmp2;
570         unsigned long address;
571         unsigned long ro_start;
572         unsigned long ro_end;
573         unsigned long fv_addr;
574         unsigned long gw_addr;
575         extern const unsigned long fault_vector_20;
576         extern void * const linux_gateway_page;
577
578         ro_start = __pa((unsigned long)_text);
579         ro_end   = __pa((unsigned long)&data_start);
580         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
581         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
582
583         end_paddr = start_paddr + size;
584
585         pg_dir = pgd_offset_k(start_vaddr);
586
587 #if PTRS_PER_PMD == 1
588         start_pmd = 0;
589 #else
590         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
591 #endif
592         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
593
594         address = start_paddr;
595         while (address < end_paddr) {
596 #if PTRS_PER_PMD == 1
597                 pmd = (pmd_t *)__pa(pg_dir);
598 #else
599                 pmd = (pmd_t *)pgd_address(*pg_dir);
600
601                 /*
602                  * pmd is physical at this point
603                  */
604
605                 if (!pmd) {
606                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
607                         pmd = (pmd_t *) __pa(pmd);
608                 }
609
610                 pgd_populate(NULL, pg_dir, __va(pmd));
611 #endif
612                 pg_dir++;
613
614                 /* now change pmd to kernel virtual addresses */
615
616                 pmd = (pmd_t *)__va(pmd) + start_pmd;
617                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
618
619                         /*
620                          * pg_table is physical at this point
621                          */
622
623                         pg_table = (pte_t *)pmd_address(*pmd);
624                         if (!pg_table) {
625                                 pg_table = (pte_t *)
626                                         alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
627                                 pg_table = (pte_t *) __pa(pg_table);
628                         }
629
630                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
631
632                         /* now change pg_table to kernel virtual addresses */
633
634                         pg_table = (pte_t *) __va(pg_table) + start_pte;
635                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
636                                 pte_t pte;
637
638                                 /*
639                                  * Map the fault vector writable so we can
640                                  * write the HPMC checksum.
641                                  */
642 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
643                                 if (address >= ro_start && address < ro_end
644                                                         && address != fv_addr
645                                                         && address != gw_addr)
646                                     pte = __mk_pte(address, PAGE_KERNEL_RO);
647                                 else
648 #endif
649                                     pte = __mk_pte(address, pgprot);
650
651                                 if (address >= end_paddr)
652                                         pte_val(pte) = 0;
653
654                                 set_pte(pg_table, pte);
655
656                                 address += PAGE_SIZE;
657                         }
658                         start_pte = 0;
659
660                         if (address >= end_paddr)
661                             break;
662                 }
663                 start_pmd = 0;
664         }
665 }
666
667 /*
668  * pagetable_init() sets up the page tables
669  *
670  * Note that gateway_init() places the Linux gateway page at page 0.
671  * Since gateway pages cannot be dereferenced this has the desirable
672  * side effect of trapping those pesky NULL-reference errors in the
673  * kernel.
674  */
675 static void __init pagetable_init(void)
676 {
677         int range;
678
679         /* Map each physical memory range to its kernel vaddr */
680
681         for (range = 0; range < npmem_ranges; range++) {
682                 unsigned long start_paddr;
683                 unsigned long end_paddr;
684                 unsigned long size;
685
686                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
687                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
688                 size = pmem_ranges[range].pages << PAGE_SHIFT;
689
690                 map_pages((unsigned long)__va(start_paddr), start_paddr,
691                         size, PAGE_KERNEL);
692         }
693
694 #ifdef CONFIG_BLK_DEV_INITRD
695         if (initrd_end && initrd_end > mem_limit) {
696                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
697                 map_pages(initrd_start, __pa(initrd_start),
698                         initrd_end - initrd_start, PAGE_KERNEL);
699         }
700 #endif
701
702         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
703         memset(empty_zero_page, 0, PAGE_SIZE);
704 }
705
706 static void __init gateway_init(void)
707 {
708         unsigned long linux_gateway_page_addr;
709         /* FIXME: This is 'const' in order to trick the compiler
710            into not treating it as DP-relative data. */
711         extern void * const linux_gateway_page;
712
713         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
714
715         /*
716          * Setup Linux Gateway page.
717          *
718          * The Linux gateway page will reside in kernel space (on virtual
719          * page 0), so it doesn't need to be aliased into user space.
720          */
721
722         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
723                 PAGE_SIZE, PAGE_GATEWAY);
724 }
725
726 #ifdef CONFIG_HPUX
727 void
728 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
729 {
730         pgd_t *pg_dir;
731         pmd_t *pmd;
732         pte_t *pg_table;
733         unsigned long start_pmd;
734         unsigned long start_pte;
735         unsigned long address;
736         unsigned long hpux_gw_page_addr;
737         /* FIXME: This is 'const' in order to trick the compiler
738            into not treating it as DP-relative data. */
739         extern void * const hpux_gateway_page;
740
741         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
742
743         /*
744          * Setup HP-UX Gateway page.
745          *
746          * The HP-UX gateway page resides in the user address space,
747          * so it needs to be aliased into each process.
748          */
749
750         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
751
752 #if PTRS_PER_PMD == 1
753         start_pmd = 0;
754 #else
755         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
756 #endif
757         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
758
759         address = __pa(&hpux_gateway_page);
760 #if PTRS_PER_PMD == 1
761         pmd = (pmd_t *)__pa(pg_dir);
762 #else
763         pmd = (pmd_t *) pgd_address(*pg_dir);
764
765         /*
766          * pmd is physical at this point
767          */
768
769         if (!pmd) {
770                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
771                 pmd = (pmd_t *) __pa(pmd);
772         }
773
774         __pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
775 #endif
776         /* now change pmd to kernel virtual addresses */
777
778         pmd = (pmd_t *)__va(pmd) + start_pmd;
779
780         /*
781          * pg_table is physical at this point
782          */
783
784         pg_table = (pte_t *) pmd_address(*pmd);
785         if (!pg_table)
786                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
787
788         __pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
789
790         /* now change pg_table to kernel virtual addresses */
791
792         pg_table = (pte_t *) __va(pg_table) + start_pte;
793         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
794 }
795 EXPORT_SYMBOL(map_hpux_gateway_page);
796 #endif
797
798 void __init paging_init(void)
799 {
800         int i;
801
802         setup_bootmem();
803         pagetable_init();
804         gateway_init();
805         flush_cache_all_local(); /* start with known state */
806         flush_tlb_all_local(NULL);
807
808         for (i = 0; i < npmem_ranges; i++) {
809                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
810
811                 /* We have an IOMMU, so all memory can go into a single
812                    ZONE_DMA zone. */
813                 zones_size[ZONE_DMA] = pmem_ranges[i].pages;
814
815 #ifdef CONFIG_DISCONTIGMEM
816                 /* Need to initialize the pfnnid_map before we can initialize
817                    the zone */
818                 {
819                     int j;
820                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
821                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
822                          j++) {
823                         pfnnid_map[j] = i;
824                     }
825                 }
826 #endif
827
828                 free_area_init_node(i, NODE_DATA(i), zones_size,
829                                 pmem_ranges[i].start_pfn, NULL);
830         }
831 }
832
833 #ifdef CONFIG_PA20
834
835 /*
836  * Currently, all PA20 chips have 18 bit protection id's, which is the
837  * limiting factor (space ids are 32 bits).
838  */
839
840 #define NR_SPACE_IDS 262144
841
842 #else
843
844 /*
845  * Currently we have a one-to-one relationship between space id's and
846  * protection id's. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
847  * support 15 bit protection id's, so that is the limiting factor.
848  * PCXT' has 18 bit protection id's, but only 16 bit spaceids, so it's
849  * probably not worth the effort for a special case here.
850  */
851
852 #define NR_SPACE_IDS 32768
853
854 #endif  /* !CONFIG_PA20 */
855
856 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
857 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
858
859 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
860 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
861 static unsigned long space_id_index;
862 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
863 static unsigned long dirty_space_ids = 0;
864
865 static DEFINE_SPINLOCK(sid_lock);
866
867 unsigned long alloc_sid(void)
868 {
869         unsigned long index;
870
871         spin_lock(&sid_lock);
872
873         if (free_space_ids == 0) {
874                 if (dirty_space_ids != 0) {
875                         spin_unlock(&sid_lock);
876                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
877                         spin_lock(&sid_lock);
878                 }
879                 BUG_ON(free_space_ids == 0);
880         }
881
882         free_space_ids--;
883
884         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
885         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
886         space_id_index = index;
887
888         spin_unlock(&sid_lock);
889
890         return index << SPACEID_SHIFT;
891 }
892
893 void free_sid(unsigned long spaceid)
894 {
895         unsigned long index = spaceid >> SPACEID_SHIFT;
896         unsigned long *dirty_space_offset;
897
898         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
899         index &= (BITS_PER_LONG - 1);
900
901         spin_lock(&sid_lock);
902
903         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
904
905         *dirty_space_offset |= (1L << index);
906         dirty_space_ids++;
907
908         spin_unlock(&sid_lock);
909 }
910
911
912 #ifdef CONFIG_SMP
913 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
914 {
915         int i;
916
917         /* NOTE: sid_lock must be held upon entry */
918
919         *ndirtyptr = dirty_space_ids;
920         if (dirty_space_ids != 0) {
921             for (i = 0; i < SID_ARRAY_SIZE; i++) {
922                 dirty_array[i] = dirty_space_id[i];
923                 dirty_space_id[i] = 0;
924             }
925             dirty_space_ids = 0;
926         }
927
928         return;
929 }
930
931 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
932 {
933         int i;
934
935         /* NOTE: sid_lock must be held upon entry */
936
937         if (ndirty != 0) {
938                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
939                         space_id[i] ^= dirty_array[i];
940                 }
941
942                 free_space_ids += ndirty;
943                 space_id_index = 0;
944         }
945 }
946
947 #else /* CONFIG_SMP */
948
949 static void recycle_sids(void)
950 {
951         int i;
952
953         /* NOTE: sid_lock must be held upon entry */
954
955         if (dirty_space_ids != 0) {
956                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
957                         space_id[i] ^= dirty_space_id[i];
958                         dirty_space_id[i] = 0;
959                 }
960
961                 free_space_ids += dirty_space_ids;
962                 dirty_space_ids = 0;
963                 space_id_index = 0;
964         }
965 }
966 #endif
967
968 /*
969  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
970  * purged, we can safely reuse the space ids that were released but
971  * not flushed from the tlb.
972  */
973
974 #ifdef CONFIG_SMP
975
976 static unsigned long recycle_ndirty;
977 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
978 static unsigned int recycle_inuse;
979
980 void flush_tlb_all(void)
981 {
982         int do_recycle;
983
984         do_recycle = 0;
985         spin_lock(&sid_lock);
986         if (dirty_space_ids > RECYCLE_THRESHOLD) {
987             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
988             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
989             recycle_inuse++;
990             do_recycle++;
991         }
992         spin_unlock(&sid_lock);
993         on_each_cpu(flush_tlb_all_local, NULL, 1, 1);
994         if (do_recycle) {
995             spin_lock(&sid_lock);
996             recycle_sids(recycle_ndirty,recycle_dirty_array);
997             recycle_inuse = 0;
998             spin_unlock(&sid_lock);
999         }
1000 }
1001 #else
1002 void flush_tlb_all(void)
1003 {
1004         spin_lock(&sid_lock);
1005         flush_tlb_all_local(NULL);
1006         recycle_sids();
1007         spin_unlock(&sid_lock);
1008 }
1009 #endif
1010
1011 #ifdef CONFIG_BLK_DEV_INITRD
1012 void free_initrd_mem(unsigned long start, unsigned long end)
1013 {
1014         if (start >= end)
1015                 return;
1016         printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1017         for (; start < end; start += PAGE_SIZE) {
1018                 ClearPageReserved(virt_to_page(start));
1019                 init_page_count(virt_to_page(start));
1020                 free_page(start);
1021                 num_physpages++;
1022                 totalram_pages++;
1023         }
1024 }
1025 #endif