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