[PARISC] Unbreak discontigmem mem_init()
[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/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 bootmem_data_t bmem_data[MAX_NUMNODES] __read_mostly;
40 unsigned char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
41 #endif
42
43 static struct resource data_resource = {
44         .name   = "Kernel data",
45         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
46 };
47
48 static struct resource code_resource = {
49         .name   = "Kernel code",
50         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
51 };
52
53 static struct resource pdcdata_resource = {
54         .name   = "PDC data (Page Zero)",
55         .start  = 0,
56         .end    = 0x9ff,
57         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
58 };
59
60 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
61
62 /* The following array is initialized from the firmware specific
63  * information retrieved in kernel/inventory.c.
64  */
65
66 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
67 int npmem_ranges __read_mostly;
68
69 #ifdef __LP64__
70 #define MAX_MEM         (~0UL)
71 #else /* !__LP64__ */
72 #define MAX_MEM         (3584U*1024U*1024U)
73 #endif /* !__LP64__ */
74
75 static unsigned long mem_limit __read_mostly = MAX_MEM;
76
77 static void __init mem_limit_func(void)
78 {
79         char *cp, *end;
80         unsigned long limit;
81         extern char saved_command_line[];
82
83         /* We need this before __setup() functions are called */
84
85         limit = MAX_MEM;
86         for (cp = saved_command_line; *cp; ) {
87                 if (memcmp(cp, "mem=", 4) == 0) {
88                         cp += 4;
89                         limit = memparse(cp, &end);
90                         if (end != cp)
91                                 break;
92                         cp = end;
93                 } else {
94                         while (*cp != ' ' && *cp)
95                                 ++cp;
96                         while (*cp == ' ')
97                                 ++cp;
98                 }
99         }
100
101         if (limit < mem_limit)
102                 mem_limit = limit;
103 }
104
105 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
106
107 static void __init setup_bootmem(void)
108 {
109         unsigned long bootmap_size;
110         unsigned long mem_max;
111         unsigned long bootmap_pages;
112         unsigned long bootmap_start_pfn;
113         unsigned long bootmap_pfn;
114 #ifndef CONFIG_DISCONTIGMEM
115         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
116         int npmem_holes;
117 #endif
118         int i, sysram_resource_count;
119
120         disable_sr_hashing(); /* Turn off space register hashing */
121
122         /*
123          * Sort the ranges. Since the number of ranges is typically
124          * small, and performance is not an issue here, just do
125          * a simple insertion sort.
126          */
127
128         for (i = 1; i < npmem_ranges; i++) {
129                 int j;
130
131                 for (j = i; j > 0; j--) {
132                         unsigned long tmp;
133
134                         if (pmem_ranges[j-1].start_pfn <
135                             pmem_ranges[j].start_pfn) {
136
137                                 break;
138                         }
139                         tmp = pmem_ranges[j-1].start_pfn;
140                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
141                         pmem_ranges[j].start_pfn = tmp;
142                         tmp = pmem_ranges[j-1].pages;
143                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
144                         pmem_ranges[j].pages = tmp;
145                 }
146         }
147
148 #ifndef CONFIG_DISCONTIGMEM
149         /*
150          * Throw out ranges that are too far apart (controlled by
151          * MAX_GAP).
152          */
153
154         for (i = 1; i < npmem_ranges; i++) {
155                 if (pmem_ranges[i].start_pfn -
156                         (pmem_ranges[i-1].start_pfn +
157                          pmem_ranges[i-1].pages) > MAX_GAP) {
158                         npmem_ranges = i;
159                         printk("Large gap in memory detected (%ld pages). "
160                                "Consider turning on CONFIG_DISCONTIGMEM\n",
161                                pmem_ranges[i].start_pfn -
162                                (pmem_ranges[i-1].start_pfn +
163                                 pmem_ranges[i-1].pages));
164                         break;
165                 }
166         }
167 #endif
168
169         if (npmem_ranges > 1) {
170
171                 /* Print the memory ranges */
172
173                 printk(KERN_INFO "Memory Ranges:\n");
174
175                 for (i = 0; i < npmem_ranges; i++) {
176                         unsigned long start;
177                         unsigned long size;
178
179                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
180                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
181                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
182                                 i,start, start + (size - 1), size >> 20);
183                 }
184         }
185
186         sysram_resource_count = npmem_ranges;
187         for (i = 0; i < sysram_resource_count; i++) {
188                 struct resource *res = &sysram_resources[i];
189                 res->name = "System RAM";
190                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
191                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
192                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
193                 request_resource(&iomem_resource, res);
194         }
195
196         /*
197          * For 32 bit kernels we limit the amount of memory we can
198          * support, in order to preserve enough kernel address space
199          * for other purposes. For 64 bit kernels we don't normally
200          * limit the memory, but this mechanism can be used to
201          * artificially limit the amount of memory (and it is written
202          * to work with multiple memory ranges).
203          */
204
205         mem_limit_func();       /* check for "mem=" argument */
206
207         mem_max = 0;
208         num_physpages = 0;
209         for (i = 0; i < npmem_ranges; i++) {
210                 unsigned long rsize;
211
212                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
213                 if ((mem_max + rsize) > mem_limit) {
214                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
215                         if (mem_max == mem_limit)
216                                 npmem_ranges = i;
217                         else {
218                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
219                                                        - (mem_max >> PAGE_SHIFT);
220                                 npmem_ranges = i + 1;
221                                 mem_max = mem_limit;
222                         }
223                 num_physpages += pmem_ranges[i].pages;
224                         break;
225                 }
226             num_physpages += pmem_ranges[i].pages;
227                 mem_max += rsize;
228         }
229
230         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
231
232 #ifndef CONFIG_DISCONTIGMEM
233         /* Merge the ranges, keeping track of the holes */
234
235         {
236                 unsigned long end_pfn;
237                 unsigned long hole_pages;
238
239                 npmem_holes = 0;
240                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
241                 for (i = 1; i < npmem_ranges; i++) {
242
243                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
244                         if (hole_pages) {
245                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
246                                 pmem_holes[npmem_holes++].pages = hole_pages;
247                                 end_pfn += hole_pages;
248                         }
249                         end_pfn += pmem_ranges[i].pages;
250                 }
251
252                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
253                 npmem_ranges = 1;
254         }
255 #endif
256
257         bootmap_pages = 0;
258         for (i = 0; i < npmem_ranges; i++)
259                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
260
261         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
262
263 #ifdef CONFIG_DISCONTIGMEM
264         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
265                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
266                 NODE_DATA(i)->bdata = &bmem_data[i];
267         }
268         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
269
270         for (i = 0; i < npmem_ranges; i++)
271                 node_set_online(i);
272 #endif
273
274         /*
275          * Initialize and free the full range of memory in each range.
276          * Note that the only writing these routines do are to the bootmap,
277          * and we've made sure to locate the bootmap properly so that they
278          * won't be writing over anything important.
279          */
280
281         bootmap_pfn = bootmap_start_pfn;
282         max_pfn = 0;
283         for (i = 0; i < npmem_ranges; i++) {
284                 unsigned long start_pfn;
285                 unsigned long npages;
286
287                 start_pfn = pmem_ranges[i].start_pfn;
288                 npages = pmem_ranges[i].pages;
289
290                 bootmap_size = init_bootmem_node(NODE_DATA(i),
291                                                 bootmap_pfn,
292                                                 start_pfn,
293                                                 (start_pfn + npages) );
294                 free_bootmem_node(NODE_DATA(i),
295                                   (start_pfn << PAGE_SHIFT),
296                                   (npages << PAGE_SHIFT) );
297                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
298                 if ((start_pfn + npages) > max_pfn)
299                         max_pfn = start_pfn + npages;
300         }
301
302         /* IOMMU is always used to access "high mem" on those boxes
303          * that can support enough mem that a PCI device couldn't
304          * directly DMA to any physical addresses.
305          * ISA DMA support will need to revisit this.
306          */
307         max_low_pfn = max_pfn;
308
309         if ((bootmap_pfn - bootmap_start_pfn) != bootmap_pages) {
310                 printk(KERN_WARNING "WARNING! bootmap sizing is messed up!\n");
311                 BUG();
312         }
313
314         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
315
316 #define PDC_CONSOLE_IO_IODC_SIZE 32768
317
318         reserve_bootmem_node(NODE_DATA(0), 0UL,
319                         (unsigned long)(PAGE0->mem_free + PDC_CONSOLE_IO_IODC_SIZE));
320         reserve_bootmem_node(NODE_DATA(0), __pa((unsigned long)_text),
321                         (unsigned long)(_end - _text));
322         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
323                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT));
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         }
334 #endif
335
336 #ifdef CONFIG_BLK_DEV_INITRD
337         if (initrd_start) {
338                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
339                 if (__pa(initrd_start) < mem_max) {
340                         unsigned long initrd_reserve;
341
342                         if (__pa(initrd_end) > mem_max) {
343                                 initrd_reserve = mem_max - __pa(initrd_start);
344                         } else {
345                                 initrd_reserve = initrd_end - initrd_start;
346                         }
347                         initrd_below_start_ok = 1;
348                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
349
350                         reserve_bootmem_node(NODE_DATA(0),__pa(initrd_start), initrd_reserve);
351                 }
352         }
353 #endif
354
355         data_resource.start =  virt_to_phys(&data_start);
356         data_resource.end = virt_to_phys(_end) - 1;
357         code_resource.start = virt_to_phys(_text);
358         code_resource.end = virt_to_phys(&data_start)-1;
359
360         /* We don't know which region the kernel will be in, so try
361          * all of them.
362          */
363         for (i = 0; i < sysram_resource_count; i++) {
364                 struct resource *res = &sysram_resources[i];
365                 request_resource(res, &code_resource);
366                 request_resource(res, &data_resource);
367         }
368         request_resource(&sysram_resources[0], &pdcdata_resource);
369 }
370
371 void free_initmem(void)
372 {
373         unsigned long addr, init_begin, init_end;
374
375         printk(KERN_INFO "Freeing unused kernel memory: ");
376
377 #ifdef CONFIG_DEBUG_KERNEL
378         /* Attempt to catch anyone trying to execute code here
379          * by filling the page with BRK insns.
380          * 
381          * If we disable interrupts for all CPUs, then IPI stops working.
382          * Kinda breaks the global cache flushing.
383          */
384         local_irq_disable();
385
386         memset(__init_begin, 0x00,
387                 (unsigned long)__init_end - (unsigned long)__init_begin);
388
389         flush_data_cache();
390         asm volatile("sync" : : );
391         flush_icache_range((unsigned long)__init_begin, (unsigned long)__init_end);
392         asm volatile("sync" : : );
393
394         local_irq_enable();
395 #endif
396         
397         /* align __init_begin and __init_end to page size,
398            ignoring linker script where we might have tried to save RAM */
399         init_begin = PAGE_ALIGN((unsigned long)(__init_begin));
400         init_end   = PAGE_ALIGN((unsigned long)(__init_end));
401         for (addr = init_begin; addr < init_end; addr += PAGE_SIZE) {
402                 ClearPageReserved(virt_to_page(addr));
403                 init_page_count(virt_to_page(addr));
404                 free_page(addr);
405                 num_physpages++;
406                 totalram_pages++;
407         }
408
409         /* set up a new led state on systems shipped LED State panel */
410         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
411         
412         printk("%luk freed\n", (init_end - init_begin) >> 10);
413 }
414
415
416 #ifdef CONFIG_DEBUG_RODATA
417 void mark_rodata_ro(void)
418 {
419         /* rodata memory was already mapped with KERNEL_RO access rights by
420            pagetable_init() and map_pages(). No need to do additional stuff here */
421         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
422                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
423 }
424 #endif
425
426
427 /*
428  * Just an arbitrary offset to serve as a "hole" between mapping areas
429  * (between top of physical memory and a potential pcxl dma mapping
430  * area, and below the vmalloc mapping area).
431  *
432  * The current 32K value just means that there will be a 32K "hole"
433  * between mapping areas. That means that  any out-of-bounds memory
434  * accesses will hopefully be caught. The vmalloc() routines leaves
435  * a hole of 4kB between each vmalloced area for the same reason.
436  */
437
438  /* Leave room for gateway page expansion */
439 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
440 #error KERNEL_MAP_START is in gateway reserved region
441 #endif
442 #define MAP_START (KERNEL_MAP_START)
443
444 #define VM_MAP_OFFSET  (32*1024)
445 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
446                                      & ~(VM_MAP_OFFSET-1)))
447
448 void *vmalloc_start __read_mostly;
449 EXPORT_SYMBOL(vmalloc_start);
450
451 #ifdef CONFIG_PA11
452 unsigned long pcxl_dma_start __read_mostly;
453 #endif
454
455 void __init mem_init(void)
456 {
457         int codesize, reservedpages, datasize, initsize;
458
459         high_memory = __va((max_pfn << PAGE_SHIFT));
460
461 #ifndef CONFIG_DISCONTIGMEM
462         max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
463         totalram_pages += free_all_bootmem();
464 #else
465         {
466                 int i;
467
468                 for (i = 0; i < npmem_ranges; i++)
469                         totalram_pages += free_all_bootmem_node(NODE_DATA(i));
470         }
471 #endif
472
473         codesize = (unsigned long)_etext - (unsigned long)_text;
474         datasize = (unsigned long)_edata - (unsigned long)_etext;
475         initsize = (unsigned long)__init_end - (unsigned long)__init_begin;
476
477         reservedpages = 0;
478 {
479         unsigned long pfn;
480 #ifdef CONFIG_DISCONTIGMEM
481         int i;
482
483         for (i = 0; i < npmem_ranges; i++) {
484                 for (pfn = node_start_pfn(i); pfn < node_end_pfn(i); pfn++) {
485                         if (PageReserved(pfn_to_page(pfn)))
486                                 reservedpages++;
487                 }
488         }
489 #else /* !CONFIG_DISCONTIGMEM */
490         for (pfn = 0; pfn < max_pfn; pfn++) {
491                 /*
492                  * Only count reserved RAM pages
493                  */
494                 if (PageReserved(pfn_to_page(pfn)))
495                         reservedpages++;
496         }
497 #endif
498 }
499
500 #ifdef CONFIG_PA11
501         if (hppa_dma_ops == &pcxl_dma_ops) {
502                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
503                 vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
504         } else {
505                 pcxl_dma_start = 0;
506                 vmalloc_start = SET_MAP_OFFSET(MAP_START);
507         }
508 #else
509         vmalloc_start = SET_MAP_OFFSET(MAP_START);
510 #endif
511
512         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, %dk reserved, %dk data, %dk init)\n",
513                 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
514                 num_physpages << (PAGE_SHIFT-10),
515                 codesize >> 10,
516                 reservedpages << (PAGE_SHIFT-10),
517                 datasize >> 10,
518                 initsize >> 10
519         );
520
521 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
522         printk("virtual kernel memory layout:\n"
523                "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
524                "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
525                "      .init : 0x%p - 0x%p   (%4ld kB)\n"
526                "      .data : 0x%p - 0x%p   (%4ld kB)\n"
527                "      .text : 0x%p - 0x%p   (%4ld kB)\n",
528
529                (void*)VMALLOC_START, (void*)VMALLOC_END,
530                (VMALLOC_END - VMALLOC_START) >> 20,
531
532                __va(0), high_memory,
533                ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
534
535                __init_begin, __init_end,
536                ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
537
538                _etext, _edata,
539                ((unsigned long)_edata - (unsigned long)_etext) >> 10,
540
541                _text, _etext,
542                ((unsigned long)_etext - (unsigned long)_text) >> 10);
543 #endif
544 }
545
546 unsigned long *empty_zero_page __read_mostly;
547
548 void show_mem(void)
549 {
550         int i,free = 0,total = 0,reserved = 0;
551         int shared = 0, cached = 0;
552
553         printk(KERN_INFO "Mem-info:\n");
554         show_free_areas();
555         printk(KERN_INFO "Free swap:     %6ldkB\n",
556                                 nr_swap_pages<<(PAGE_SHIFT-10));
557 #ifndef CONFIG_DISCONTIGMEM
558         i = max_mapnr;
559         while (i-- > 0) {
560                 total++;
561                 if (PageReserved(mem_map+i))
562                         reserved++;
563                 else if (PageSwapCache(mem_map+i))
564                         cached++;
565                 else if (!page_count(&mem_map[i]))
566                         free++;
567                 else
568                         shared += page_count(&mem_map[i]) - 1;
569         }
570 #else
571         for (i = 0; i < npmem_ranges; i++) {
572                 int j;
573
574                 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
575                         struct page *p;
576                         unsigned long flags;
577
578                         pgdat_resize_lock(NODE_DATA(i), &flags);
579                         p = nid_page_nr(i, j) - node_start_pfn(i);
580
581                         total++;
582                         if (PageReserved(p))
583                                 reserved++;
584                         else if (PageSwapCache(p))
585                                 cached++;
586                         else if (!page_count(p))
587                                 free++;
588                         else
589                                 shared += page_count(p) - 1;
590                         pgdat_resize_unlock(NODE_DATA(i), &flags);
591                 }
592         }
593 #endif
594         printk(KERN_INFO "%d pages of RAM\n", total);
595         printk(KERN_INFO "%d reserved pages\n", reserved);
596         printk(KERN_INFO "%d pages shared\n", shared);
597         printk(KERN_INFO "%d pages swap cached\n", cached);
598
599
600 #ifdef CONFIG_DISCONTIGMEM
601         {
602                 struct zonelist *zl;
603                 int i, j, k;
604
605                 for (i = 0; i < npmem_ranges; i++) {
606                         for (j = 0; j < MAX_NR_ZONES; j++) {
607                                 zl = NODE_DATA(i)->node_zonelists + j;
608
609                                 printk("Zone list for zone %d on node %d: ", j, i);
610                                 for (k = 0; zl->zones[k] != NULL; k++) 
611                                         printk("[%d/%s] ", zone_to_nid(zl->zones[k]), zl->zones[k]->name);
612                                 printk("\n");
613                         }
614                 }
615         }
616 #endif
617 }
618
619
620 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
621 {
622         pgd_t *pg_dir;
623         pmd_t *pmd;
624         pte_t *pg_table;
625         unsigned long end_paddr;
626         unsigned long start_pmd;
627         unsigned long start_pte;
628         unsigned long tmp1;
629         unsigned long tmp2;
630         unsigned long address;
631         unsigned long ro_start;
632         unsigned long ro_end;
633         unsigned long fv_addr;
634         unsigned long gw_addr;
635         extern const unsigned long fault_vector_20;
636         extern void * const linux_gateway_page;
637
638         ro_start = __pa((unsigned long)_text);
639         ro_end   = __pa((unsigned long)&data_start);
640         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
641         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
642
643         end_paddr = start_paddr + size;
644
645         pg_dir = pgd_offset_k(start_vaddr);
646
647 #if PTRS_PER_PMD == 1
648         start_pmd = 0;
649 #else
650         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
651 #endif
652         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
653
654         address = start_paddr;
655         while (address < end_paddr) {
656 #if PTRS_PER_PMD == 1
657                 pmd = (pmd_t *)__pa(pg_dir);
658 #else
659                 pmd = (pmd_t *)pgd_address(*pg_dir);
660
661                 /*
662                  * pmd is physical at this point
663                  */
664
665                 if (!pmd) {
666                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
667                         pmd = (pmd_t *) __pa(pmd);
668                 }
669
670                 pgd_populate(NULL, pg_dir, __va(pmd));
671 #endif
672                 pg_dir++;
673
674                 /* now change pmd to kernel virtual addresses */
675
676                 pmd = (pmd_t *)__va(pmd) + start_pmd;
677                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
678
679                         /*
680                          * pg_table is physical at this point
681                          */
682
683                         pg_table = (pte_t *)pmd_address(*pmd);
684                         if (!pg_table) {
685                                 pg_table = (pte_t *)
686                                         alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
687                                 pg_table = (pte_t *) __pa(pg_table);
688                         }
689
690                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
691
692                         /* now change pg_table to kernel virtual addresses */
693
694                         pg_table = (pte_t *) __va(pg_table) + start_pte;
695                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
696                                 pte_t pte;
697
698                                 /*
699                                  * Map the fault vector writable so we can
700                                  * write the HPMC checksum.
701                                  */
702 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
703                                 if (address >= ro_start && address < ro_end
704                                                         && address != fv_addr
705                                                         && address != gw_addr)
706                                     pte = __mk_pte(address, PAGE_KERNEL_RO);
707                                 else
708 #endif
709                                     pte = __mk_pte(address, pgprot);
710
711                                 if (address >= end_paddr)
712                                         pte_val(pte) = 0;
713
714                                 set_pte(pg_table, pte);
715
716                                 address += PAGE_SIZE;
717                         }
718                         start_pte = 0;
719
720                         if (address >= end_paddr)
721                             break;
722                 }
723                 start_pmd = 0;
724         }
725 }
726
727 /*
728  * pagetable_init() sets up the page tables
729  *
730  * Note that gateway_init() places the Linux gateway page at page 0.
731  * Since gateway pages cannot be dereferenced this has the desirable
732  * side effect of trapping those pesky NULL-reference errors in the
733  * kernel.
734  */
735 static void __init pagetable_init(void)
736 {
737         int range;
738
739         /* Map each physical memory range to its kernel vaddr */
740
741         for (range = 0; range < npmem_ranges; range++) {
742                 unsigned long start_paddr;
743                 unsigned long end_paddr;
744                 unsigned long size;
745
746                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
747                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
748                 size = pmem_ranges[range].pages << PAGE_SHIFT;
749
750                 map_pages((unsigned long)__va(start_paddr), start_paddr,
751                         size, PAGE_KERNEL);
752         }
753
754 #ifdef CONFIG_BLK_DEV_INITRD
755         if (initrd_end && initrd_end > mem_limit) {
756                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
757                 map_pages(initrd_start, __pa(initrd_start),
758                         initrd_end - initrd_start, PAGE_KERNEL);
759         }
760 #endif
761
762         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
763         memset(empty_zero_page, 0, PAGE_SIZE);
764 }
765
766 static void __init gateway_init(void)
767 {
768         unsigned long linux_gateway_page_addr;
769         /* FIXME: This is 'const' in order to trick the compiler
770            into not treating it as DP-relative data. */
771         extern void * const linux_gateway_page;
772
773         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
774
775         /*
776          * Setup Linux Gateway page.
777          *
778          * The Linux gateway page will reside in kernel space (on virtual
779          * page 0), so it doesn't need to be aliased into user space.
780          */
781
782         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
783                 PAGE_SIZE, PAGE_GATEWAY);
784 }
785
786 #ifdef CONFIG_HPUX
787 void
788 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
789 {
790         pgd_t *pg_dir;
791         pmd_t *pmd;
792         pte_t *pg_table;
793         unsigned long start_pmd;
794         unsigned long start_pte;
795         unsigned long address;
796         unsigned long hpux_gw_page_addr;
797         /* FIXME: This is 'const' in order to trick the compiler
798            into not treating it as DP-relative data. */
799         extern void * const hpux_gateway_page;
800
801         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
802
803         /*
804          * Setup HP-UX Gateway page.
805          *
806          * The HP-UX gateway page resides in the user address space,
807          * so it needs to be aliased into each process.
808          */
809
810         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
811
812 #if PTRS_PER_PMD == 1
813         start_pmd = 0;
814 #else
815         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
816 #endif
817         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
818
819         address = __pa(&hpux_gateway_page);
820 #if PTRS_PER_PMD == 1
821         pmd = (pmd_t *)__pa(pg_dir);
822 #else
823         pmd = (pmd_t *) pgd_address(*pg_dir);
824
825         /*
826          * pmd is physical at this point
827          */
828
829         if (!pmd) {
830                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
831                 pmd = (pmd_t *) __pa(pmd);
832         }
833
834         __pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
835 #endif
836         /* now change pmd to kernel virtual addresses */
837
838         pmd = (pmd_t *)__va(pmd) + start_pmd;
839
840         /*
841          * pg_table is physical at this point
842          */
843
844         pg_table = (pte_t *) pmd_address(*pmd);
845         if (!pg_table)
846                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
847
848         __pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
849
850         /* now change pg_table to kernel virtual addresses */
851
852         pg_table = (pte_t *) __va(pg_table) + start_pte;
853         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
854 }
855 EXPORT_SYMBOL(map_hpux_gateway_page);
856 #endif
857
858 void __init paging_init(void)
859 {
860         int i;
861
862         setup_bootmem();
863         pagetable_init();
864         gateway_init();
865         flush_cache_all_local(); /* start with known state */
866         flush_tlb_all_local(NULL);
867
868         for (i = 0; i < npmem_ranges; i++) {
869                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
870
871                 /* We have an IOMMU, so all memory can go into a single
872                    ZONE_DMA zone. */
873                 zones_size[ZONE_DMA] = pmem_ranges[i].pages;
874
875 #ifdef CONFIG_DISCONTIGMEM
876                 /* Need to initialize the pfnnid_map before we can initialize
877                    the zone */
878                 {
879                     int j;
880                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
881                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
882                          j++) {
883                         pfnnid_map[j] = i;
884                     }
885                 }
886 #endif
887
888                 free_area_init_node(i, NODE_DATA(i), zones_size,
889                                 pmem_ranges[i].start_pfn, NULL);
890         }
891 }
892
893 #ifdef CONFIG_PA20
894
895 /*
896  * Currently, all PA20 chips have 18 bit protection id's, which is the
897  * limiting factor (space ids are 32 bits).
898  */
899
900 #define NR_SPACE_IDS 262144
901
902 #else
903
904 /*
905  * Currently we have a one-to-one relationship between space id's and
906  * protection id's. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
907  * support 15 bit protection id's, so that is the limiting factor.
908  * PCXT' has 18 bit protection id's, but only 16 bit spaceids, so it's
909  * probably not worth the effort for a special case here.
910  */
911
912 #define NR_SPACE_IDS 32768
913
914 #endif  /* !CONFIG_PA20 */
915
916 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
917 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
918
919 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
920 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
921 static unsigned long space_id_index;
922 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
923 static unsigned long dirty_space_ids = 0;
924
925 static DEFINE_SPINLOCK(sid_lock);
926
927 unsigned long alloc_sid(void)
928 {
929         unsigned long index;
930
931         spin_lock(&sid_lock);
932
933         if (free_space_ids == 0) {
934                 if (dirty_space_ids != 0) {
935                         spin_unlock(&sid_lock);
936                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
937                         spin_lock(&sid_lock);
938                 }
939                 BUG_ON(free_space_ids == 0);
940         }
941
942         free_space_ids--;
943
944         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
945         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
946         space_id_index = index;
947
948         spin_unlock(&sid_lock);
949
950         return index << SPACEID_SHIFT;
951 }
952
953 void free_sid(unsigned long spaceid)
954 {
955         unsigned long index = spaceid >> SPACEID_SHIFT;
956         unsigned long *dirty_space_offset;
957
958         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
959         index &= (BITS_PER_LONG - 1);
960
961         spin_lock(&sid_lock);
962
963         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
964
965         *dirty_space_offset |= (1L << index);
966         dirty_space_ids++;
967
968         spin_unlock(&sid_lock);
969 }
970
971
972 #ifdef CONFIG_SMP
973 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
974 {
975         int i;
976
977         /* NOTE: sid_lock must be held upon entry */
978
979         *ndirtyptr = dirty_space_ids;
980         if (dirty_space_ids != 0) {
981             for (i = 0; i < SID_ARRAY_SIZE; i++) {
982                 dirty_array[i] = dirty_space_id[i];
983                 dirty_space_id[i] = 0;
984             }
985             dirty_space_ids = 0;
986         }
987
988         return;
989 }
990
991 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
992 {
993         int i;
994
995         /* NOTE: sid_lock must be held upon entry */
996
997         if (ndirty != 0) {
998                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
999                         space_id[i] ^= dirty_array[i];
1000                 }
1001
1002                 free_space_ids += ndirty;
1003                 space_id_index = 0;
1004         }
1005 }
1006
1007 #else /* CONFIG_SMP */
1008
1009 static void recycle_sids(void)
1010 {
1011         int i;
1012
1013         /* NOTE: sid_lock must be held upon entry */
1014
1015         if (dirty_space_ids != 0) {
1016                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
1017                         space_id[i] ^= dirty_space_id[i];
1018                         dirty_space_id[i] = 0;
1019                 }
1020
1021                 free_space_ids += dirty_space_ids;
1022                 dirty_space_ids = 0;
1023                 space_id_index = 0;
1024         }
1025 }
1026 #endif
1027
1028 /*
1029  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
1030  * purged, we can safely reuse the space ids that were released but
1031  * not flushed from the tlb.
1032  */
1033
1034 #ifdef CONFIG_SMP
1035
1036 static unsigned long recycle_ndirty;
1037 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
1038 static unsigned int recycle_inuse;
1039
1040 void flush_tlb_all(void)
1041 {
1042         int do_recycle;
1043
1044         do_recycle = 0;
1045         spin_lock(&sid_lock);
1046         if (dirty_space_ids > RECYCLE_THRESHOLD) {
1047             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
1048             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
1049             recycle_inuse++;
1050             do_recycle++;
1051         }
1052         spin_unlock(&sid_lock);
1053         on_each_cpu(flush_tlb_all_local, NULL, 1, 1);
1054         if (do_recycle) {
1055             spin_lock(&sid_lock);
1056             recycle_sids(recycle_ndirty,recycle_dirty_array);
1057             recycle_inuse = 0;
1058             spin_unlock(&sid_lock);
1059         }
1060 }
1061 #else
1062 void flush_tlb_all(void)
1063 {
1064         spin_lock(&sid_lock);
1065         flush_tlb_all_local(NULL);
1066         recycle_sids();
1067         spin_unlock(&sid_lock);
1068 }
1069 #endif
1070
1071 #ifdef CONFIG_BLK_DEV_INITRD
1072 void free_initrd_mem(unsigned long start, unsigned long end)
1073 {
1074         if (start >= end)
1075                 return;
1076         printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1077         for (; start < end; start += PAGE_SIZE) {
1078                 ClearPageReserved(virt_to_page(start));
1079                 init_page_count(virt_to_page(start));
1080                 free_page(start);
1081                 num_physpages++;
1082                 totalram_pages++;
1083         }
1084 }
1085 #endif