[PATCH] NUMA: Add zone_to_nid function
[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  *  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 char _text;      /* start of kernel code, defined by linker */
35 extern int  data_start;
36 extern char _end;       /* end of BSS, defined by linker */
37 extern char __init_begin, __init_end;
38
39 #ifdef CONFIG_DISCONTIGMEM
40 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
41 bootmem_data_t bmem_data[MAX_NUMNODES] __read_mostly;
42 unsigned char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
43 #endif
44
45 static struct resource data_resource = {
46         .name   = "Kernel data",
47         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
48 };
49
50 static struct resource code_resource = {
51         .name   = "Kernel code",
52         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
53 };
54
55 static struct resource pdcdata_resource = {
56         .name   = "PDC data (Page Zero)",
57         .start  = 0,
58         .end    = 0x9ff,
59         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM,
60 };
61
62 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
63
64 /* The following array is initialized from the firmware specific
65  * information retrieved in kernel/inventory.c.
66  */
67
68 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
69 int npmem_ranges __read_mostly;
70
71 #ifdef __LP64__
72 #define MAX_MEM         (~0UL)
73 #else /* !__LP64__ */
74 #define MAX_MEM         (3584U*1024U*1024U)
75 #endif /* !__LP64__ */
76
77 static unsigned long mem_limit __read_mostly = MAX_MEM;
78
79 static void __init mem_limit_func(void)
80 {
81         char *cp, *end;
82         unsigned long limit;
83         extern char saved_command_line[];
84
85         /* We need this before __setup() functions are called */
86
87         limit = MAX_MEM;
88         for (cp = saved_command_line; *cp; ) {
89                 if (memcmp(cp, "mem=", 4) == 0) {
90                         cp += 4;
91                         limit = memparse(cp, &end);
92                         if (end != cp)
93                                 break;
94                         cp = end;
95                 } else {
96                         while (*cp != ' ' && *cp)
97                                 ++cp;
98                         while (*cp == ' ')
99                                 ++cp;
100                 }
101         }
102
103         if (limit < mem_limit)
104                 mem_limit = limit;
105 }
106
107 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
108
109 static void __init setup_bootmem(void)
110 {
111         unsigned long bootmap_size;
112         unsigned long mem_max;
113         unsigned long bootmap_pages;
114         unsigned long bootmap_start_pfn;
115         unsigned long bootmap_pfn;
116 #ifndef CONFIG_DISCONTIGMEM
117         physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
118         int npmem_holes;
119 #endif
120         int i, sysram_resource_count;
121
122         disable_sr_hashing(); /* Turn off space register hashing */
123
124         /*
125          * Sort the ranges. Since the number of ranges is typically
126          * small, and performance is not an issue here, just do
127          * a simple insertion sort.
128          */
129
130         for (i = 1; i < npmem_ranges; i++) {
131                 int j;
132
133                 for (j = i; j > 0; j--) {
134                         unsigned long tmp;
135
136                         if (pmem_ranges[j-1].start_pfn <
137                             pmem_ranges[j].start_pfn) {
138
139                                 break;
140                         }
141                         tmp = pmem_ranges[j-1].start_pfn;
142                         pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
143                         pmem_ranges[j].start_pfn = tmp;
144                         tmp = pmem_ranges[j-1].pages;
145                         pmem_ranges[j-1].pages = pmem_ranges[j].pages;
146                         pmem_ranges[j].pages = tmp;
147                 }
148         }
149
150 #ifndef CONFIG_DISCONTIGMEM
151         /*
152          * Throw out ranges that are too far apart (controlled by
153          * MAX_GAP).
154          */
155
156         for (i = 1; i < npmem_ranges; i++) {
157                 if (pmem_ranges[i].start_pfn -
158                         (pmem_ranges[i-1].start_pfn +
159                          pmem_ranges[i-1].pages) > MAX_GAP) {
160                         npmem_ranges = i;
161                         printk("Large gap in memory detected (%ld pages). "
162                                "Consider turning on CONFIG_DISCONTIGMEM\n",
163                                pmem_ranges[i].start_pfn -
164                                (pmem_ranges[i-1].start_pfn +
165                                 pmem_ranges[i-1].pages));
166                         break;
167                 }
168         }
169 #endif
170
171         if (npmem_ranges > 1) {
172
173                 /* Print the memory ranges */
174
175                 printk(KERN_INFO "Memory Ranges:\n");
176
177                 for (i = 0; i < npmem_ranges; i++) {
178                         unsigned long start;
179                         unsigned long size;
180
181                         size = (pmem_ranges[i].pages << PAGE_SHIFT);
182                         start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
183                         printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
184                                 i,start, start + (size - 1), size >> 20);
185                 }
186         }
187
188         sysram_resource_count = npmem_ranges;
189         for (i = 0; i < sysram_resource_count; i++) {
190                 struct resource *res = &sysram_resources[i];
191                 res->name = "System RAM";
192                 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
193                 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
194                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
195                 request_resource(&iomem_resource, res);
196         }
197
198         /*
199          * For 32 bit kernels we limit the amount of memory we can
200          * support, in order to preserve enough kernel address space
201          * for other purposes. For 64 bit kernels we don't normally
202          * limit the memory, but this mechanism can be used to
203          * artificially limit the amount of memory (and it is written
204          * to work with multiple memory ranges).
205          */
206
207         mem_limit_func();       /* check for "mem=" argument */
208
209         mem_max = 0;
210         num_physpages = 0;
211         for (i = 0; i < npmem_ranges; i++) {
212                 unsigned long rsize;
213
214                 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
215                 if ((mem_max + rsize) > mem_limit) {
216                         printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
217                         if (mem_max == mem_limit)
218                                 npmem_ranges = i;
219                         else {
220                                 pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
221                                                        - (mem_max >> PAGE_SHIFT);
222                                 npmem_ranges = i + 1;
223                                 mem_max = mem_limit;
224                         }
225                 num_physpages += pmem_ranges[i].pages;
226                         break;
227                 }
228             num_physpages += pmem_ranges[i].pages;
229                 mem_max += rsize;
230         }
231
232         printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
233
234 #ifndef CONFIG_DISCONTIGMEM
235         /* Merge the ranges, keeping track of the holes */
236
237         {
238                 unsigned long end_pfn;
239                 unsigned long hole_pages;
240
241                 npmem_holes = 0;
242                 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
243                 for (i = 1; i < npmem_ranges; i++) {
244
245                         hole_pages = pmem_ranges[i].start_pfn - end_pfn;
246                         if (hole_pages) {
247                                 pmem_holes[npmem_holes].start_pfn = end_pfn;
248                                 pmem_holes[npmem_holes++].pages = hole_pages;
249                                 end_pfn += hole_pages;
250                         }
251                         end_pfn += pmem_ranges[i].pages;
252                 }
253
254                 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
255                 npmem_ranges = 1;
256         }
257 #endif
258
259         bootmap_pages = 0;
260         for (i = 0; i < npmem_ranges; i++)
261                 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
262
263         bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
264
265 #ifdef CONFIG_DISCONTIGMEM
266         for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
267                 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
268                 NODE_DATA(i)->bdata = &bmem_data[i];
269         }
270         memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
271
272         for (i = 0; i < npmem_ranges; i++)
273                 node_set_online(i);
274 #endif
275
276         /*
277          * Initialize and free the full range of memory in each range.
278          * Note that the only writing these routines do are to the bootmap,
279          * and we've made sure to locate the bootmap properly so that they
280          * won't be writing over anything important.
281          */
282
283         bootmap_pfn = bootmap_start_pfn;
284         max_pfn = 0;
285         for (i = 0; i < npmem_ranges; i++) {
286                 unsigned long start_pfn;
287                 unsigned long npages;
288
289                 start_pfn = pmem_ranges[i].start_pfn;
290                 npages = pmem_ranges[i].pages;
291
292                 bootmap_size = init_bootmem_node(NODE_DATA(i),
293                                                 bootmap_pfn,
294                                                 start_pfn,
295                                                 (start_pfn + npages) );
296                 free_bootmem_node(NODE_DATA(i),
297                                   (start_pfn << PAGE_SHIFT),
298                                   (npages << PAGE_SHIFT) );
299                 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
300                 if ((start_pfn + npages) > max_pfn)
301                         max_pfn = start_pfn + npages;
302         }
303
304         /* IOMMU is always used to access "high mem" on those boxes
305          * that can support enough mem that a PCI device couldn't
306          * directly DMA to any physical addresses.
307          * ISA DMA support will need to revisit this.
308          */
309         max_low_pfn = max_pfn;
310
311         if ((bootmap_pfn - bootmap_start_pfn) != bootmap_pages) {
312                 printk(KERN_WARNING "WARNING! bootmap sizing is messed up!\n");
313                 BUG();
314         }
315
316         /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
317
318 #define PDC_CONSOLE_IO_IODC_SIZE 32768
319
320         reserve_bootmem_node(NODE_DATA(0), 0UL,
321                         (unsigned long)(PAGE0->mem_free + PDC_CONSOLE_IO_IODC_SIZE));
322         reserve_bootmem_node(NODE_DATA(0),__pa((unsigned long)&_text),
323                         (unsigned long)(&_end - &_text));
324         reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
325                         ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT));
326
327 #ifndef CONFIG_DISCONTIGMEM
328
329         /* reserve the holes */
330
331         for (i = 0; i < npmem_holes; i++) {
332                 reserve_bootmem_node(NODE_DATA(0),
333                                 (pmem_holes[i].start_pfn << PAGE_SHIFT),
334                                 (pmem_holes[i].pages << PAGE_SHIFT));
335         }
336 #endif
337
338 #ifdef CONFIG_BLK_DEV_INITRD
339         if (initrd_start) {
340                 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
341                 if (__pa(initrd_start) < mem_max) {
342                         unsigned long initrd_reserve;
343
344                         if (__pa(initrd_end) > mem_max) {
345                                 initrd_reserve = mem_max - __pa(initrd_start);
346                         } else {
347                                 initrd_reserve = initrd_end - initrd_start;
348                         }
349                         initrd_below_start_ok = 1;
350                         printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
351
352                         reserve_bootmem_node(NODE_DATA(0),__pa(initrd_start), initrd_reserve);
353                 }
354         }
355 #endif
356
357         data_resource.start =  virt_to_phys(&data_start);
358         data_resource.end = virt_to_phys(&_end)-1;
359         code_resource.start = virt_to_phys(&_text);
360         code_resource.end = virt_to_phys(&data_start)-1;
361
362         /* We don't know which region the kernel will be in, so try
363          * all of them.
364          */
365         for (i = 0; i < sysram_resource_count; i++) {
366                 struct resource *res = &sysram_resources[i];
367                 request_resource(res, &code_resource);
368                 request_resource(res, &data_resource);
369         }
370         request_resource(&sysram_resources[0], &pdcdata_resource);
371 }
372
373 void free_initmem(void)
374 {
375         unsigned long addr, init_begin, init_end;
376
377         printk(KERN_INFO "Freeing unused kernel memory: ");
378
379 #ifdef CONFIG_DEBUG_KERNEL
380         /* Attempt to catch anyone trying to execute code here
381          * by filling the page with BRK insns.
382          * 
383          * If we disable interrupts for all CPUs, then IPI stops working.
384          * Kinda breaks the global cache flushing.
385          */
386         local_irq_disable();
387
388         memset(&__init_begin, 0x00, 
389                 (unsigned long)&__init_end - (unsigned long)&__init_begin);
390
391         flush_data_cache();
392         asm volatile("sync" : : );
393         flush_icache_range((unsigned long)&__init_begin, (unsigned long)&__init_end);
394         asm volatile("sync" : : );
395
396         local_irq_enable();
397 #endif
398         
399         /* align __init_begin and __init_end to page size,
400            ignoring linker script where we might have tried to save RAM */
401         init_begin = PAGE_ALIGN((unsigned long)(&__init_begin));
402         init_end   = PAGE_ALIGN((unsigned long)(&__init_end));
403         for (addr = init_begin; addr < init_end; addr += PAGE_SIZE) {
404                 ClearPageReserved(virt_to_page(addr));
405                 init_page_count(virt_to_page(addr));
406                 free_page(addr);
407                 num_physpages++;
408                 totalram_pages++;
409         }
410
411         /* set up a new led state on systems shipped LED State panel */
412         pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
413         
414         printk("%luk freed\n", (init_end - init_begin) >> 10);
415 }
416
417
418 #ifdef CONFIG_DEBUG_RODATA
419 void mark_rodata_ro(void)
420 {
421         /* rodata memory was already mapped with KERNEL_RO access rights by
422            pagetable_init() and map_pages(). No need to do additional stuff here */
423         printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
424                 (unsigned long)(__end_rodata - __start_rodata) >> 10);
425 }
426 #endif
427
428
429 /*
430  * Just an arbitrary offset to serve as a "hole" between mapping areas
431  * (between top of physical memory and a potential pcxl dma mapping
432  * area, and below the vmalloc mapping area).
433  *
434  * The current 32K value just means that there will be a 32K "hole"
435  * between mapping areas. That means that  any out-of-bounds memory
436  * accesses will hopefully be caught. The vmalloc() routines leaves
437  * a hole of 4kB between each vmalloced area for the same reason.
438  */
439
440  /* Leave room for gateway page expansion */
441 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
442 #error KERNEL_MAP_START is in gateway reserved region
443 #endif
444 #define MAP_START (KERNEL_MAP_START)
445
446 #define VM_MAP_OFFSET  (32*1024)
447 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
448                                      & ~(VM_MAP_OFFSET-1)))
449
450 void *vmalloc_start __read_mostly;
451 EXPORT_SYMBOL(vmalloc_start);
452
453 #ifdef CONFIG_PA11
454 unsigned long pcxl_dma_start __read_mostly;
455 #endif
456
457 void __init mem_init(void)
458 {
459         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         printk(KERN_INFO "Memory: %luk available\n", num_physpages << (PAGE_SHIFT-10));
474
475 #ifdef CONFIG_PA11
476         if (hppa_dma_ops == &pcxl_dma_ops) {
477                 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
478                 vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
479         } else {
480                 pcxl_dma_start = 0;
481                 vmalloc_start = SET_MAP_OFFSET(MAP_START);
482         }
483 #else
484         vmalloc_start = SET_MAP_OFFSET(MAP_START);
485 #endif
486
487 }
488
489 unsigned long *empty_zero_page __read_mostly;
490
491 void show_mem(void)
492 {
493         int i,free = 0,total = 0,reserved = 0;
494         int shared = 0, cached = 0;
495
496         printk(KERN_INFO "Mem-info:\n");
497         show_free_areas();
498         printk(KERN_INFO "Free swap:     %6ldkB\n",
499                                 nr_swap_pages<<(PAGE_SHIFT-10));
500 #ifndef CONFIG_DISCONTIGMEM
501         i = max_mapnr;
502         while (i-- > 0) {
503                 total++;
504                 if (PageReserved(mem_map+i))
505                         reserved++;
506                 else if (PageSwapCache(mem_map+i))
507                         cached++;
508                 else if (!page_count(&mem_map[i]))
509                         free++;
510                 else
511                         shared += page_count(&mem_map[i]) - 1;
512         }
513 #else
514         for (i = 0; i < npmem_ranges; i++) {
515                 int j;
516
517                 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
518                         struct page *p;
519                         unsigned long flags;
520
521                         pgdat_resize_lock(NODE_DATA(i), &flags);
522                         p = nid_page_nr(i, j) - node_start_pfn(i);
523
524                         total++;
525                         if (PageReserved(p))
526                                 reserved++;
527                         else if (PageSwapCache(p))
528                                 cached++;
529                         else if (!page_count(p))
530                                 free++;
531                         else
532                                 shared += page_count(p) - 1;
533                         pgdat_resize_unlock(NODE_DATA(i), &flags);
534                 }
535         }
536 #endif
537         printk(KERN_INFO "%d pages of RAM\n", total);
538         printk(KERN_INFO "%d reserved pages\n", reserved);
539         printk(KERN_INFO "%d pages shared\n", shared);
540         printk(KERN_INFO "%d pages swap cached\n", cached);
541
542
543 #ifdef CONFIG_DISCONTIGMEM
544         {
545                 struct zonelist *zl;
546                 int i, j, k;
547
548                 for (i = 0; i < npmem_ranges; i++) {
549                         for (j = 0; j < MAX_NR_ZONES; j++) {
550                                 zl = NODE_DATA(i)->node_zonelists + j;
551
552                                 printk("Zone list for zone %d on node %d: ", j, i);
553                                 for (k = 0; zl->zones[k] != NULL; k++) 
554                                         printk("[%d/%s] ", zone_to_nid(zl->zones[k]), zl->zones[k]->name);
555                                 printk("\n");
556                         }
557                 }
558         }
559 #endif
560 }
561
562
563 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
564 {
565         pgd_t *pg_dir;
566         pmd_t *pmd;
567         pte_t *pg_table;
568         unsigned long end_paddr;
569         unsigned long start_pmd;
570         unsigned long start_pte;
571         unsigned long tmp1;
572         unsigned long tmp2;
573         unsigned long address;
574         unsigned long ro_start;
575         unsigned long ro_end;
576         unsigned long fv_addr;
577         unsigned long gw_addr;
578         extern const unsigned long fault_vector_20;
579         extern void * const linux_gateway_page;
580
581         ro_start = __pa((unsigned long)&_text);
582         ro_end   = __pa((unsigned long)&data_start);
583         fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
584         gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
585
586         end_paddr = start_paddr + size;
587
588         pg_dir = pgd_offset_k(start_vaddr);
589
590 #if PTRS_PER_PMD == 1
591         start_pmd = 0;
592 #else
593         start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
594 #endif
595         start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
596
597         address = start_paddr;
598         while (address < end_paddr) {
599 #if PTRS_PER_PMD == 1
600                 pmd = (pmd_t *)__pa(pg_dir);
601 #else
602                 pmd = (pmd_t *)pgd_address(*pg_dir);
603
604                 /*
605                  * pmd is physical at this point
606                  */
607
608                 if (!pmd) {
609                         pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
610                         pmd = (pmd_t *) __pa(pmd);
611                 }
612
613                 pgd_populate(NULL, pg_dir, __va(pmd));
614 #endif
615                 pg_dir++;
616
617                 /* now change pmd to kernel virtual addresses */
618
619                 pmd = (pmd_t *)__va(pmd) + start_pmd;
620                 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
621
622                         /*
623                          * pg_table is physical at this point
624                          */
625
626                         pg_table = (pte_t *)pmd_address(*pmd);
627                         if (!pg_table) {
628                                 pg_table = (pte_t *)
629                                         alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
630                                 pg_table = (pte_t *) __pa(pg_table);
631                         }
632
633                         pmd_populate_kernel(NULL, pmd, __va(pg_table));
634
635                         /* now change pg_table to kernel virtual addresses */
636
637                         pg_table = (pte_t *) __va(pg_table) + start_pte;
638                         for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
639                                 pte_t pte;
640
641                                 /*
642                                  * Map the fault vector writable so we can
643                                  * write the HPMC checksum.
644                                  */
645 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
646                                 if (address >= ro_start && address < ro_end
647                                                         && address != fv_addr
648                                                         && address != gw_addr)
649                                     pte = __mk_pte(address, PAGE_KERNEL_RO);
650                                 else
651 #endif
652                                     pte = __mk_pte(address, pgprot);
653
654                                 if (address >= end_paddr)
655                                         pte_val(pte) = 0;
656
657                                 set_pte(pg_table, pte);
658
659                                 address += PAGE_SIZE;
660                         }
661                         start_pte = 0;
662
663                         if (address >= end_paddr)
664                             break;
665                 }
666                 start_pmd = 0;
667         }
668 }
669
670 /*
671  * pagetable_init() sets up the page tables
672  *
673  * Note that gateway_init() places the Linux gateway page at page 0.
674  * Since gateway pages cannot be dereferenced this has the desirable
675  * side effect of trapping those pesky NULL-reference errors in the
676  * kernel.
677  */
678 static void __init pagetable_init(void)
679 {
680         int range;
681
682         /* Map each physical memory range to its kernel vaddr */
683
684         for (range = 0; range < npmem_ranges; range++) {
685                 unsigned long start_paddr;
686                 unsigned long end_paddr;
687                 unsigned long size;
688
689                 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
690                 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
691                 size = pmem_ranges[range].pages << PAGE_SHIFT;
692
693                 map_pages((unsigned long)__va(start_paddr), start_paddr,
694                         size, PAGE_KERNEL);
695         }
696
697 #ifdef CONFIG_BLK_DEV_INITRD
698         if (initrd_end && initrd_end > mem_limit) {
699                 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
700                 map_pages(initrd_start, __pa(initrd_start),
701                         initrd_end - initrd_start, PAGE_KERNEL);
702         }
703 #endif
704
705         empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
706         memset(empty_zero_page, 0, PAGE_SIZE);
707 }
708
709 static void __init gateway_init(void)
710 {
711         unsigned long linux_gateway_page_addr;
712         /* FIXME: This is 'const' in order to trick the compiler
713            into not treating it as DP-relative data. */
714         extern void * const linux_gateway_page;
715
716         linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
717
718         /*
719          * Setup Linux Gateway page.
720          *
721          * The Linux gateway page will reside in kernel space (on virtual
722          * page 0), so it doesn't need to be aliased into user space.
723          */
724
725         map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
726                 PAGE_SIZE, PAGE_GATEWAY);
727 }
728
729 #ifdef CONFIG_HPUX
730 void
731 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
732 {
733         pgd_t *pg_dir;
734         pmd_t *pmd;
735         pte_t *pg_table;
736         unsigned long start_pmd;
737         unsigned long start_pte;
738         unsigned long address;
739         unsigned long hpux_gw_page_addr;
740         /* FIXME: This is 'const' in order to trick the compiler
741            into not treating it as DP-relative data. */
742         extern void * const hpux_gateway_page;
743
744         hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
745
746         /*
747          * Setup HP-UX Gateway page.
748          *
749          * The HP-UX gateway page resides in the user address space,
750          * so it needs to be aliased into each process.
751          */
752
753         pg_dir = pgd_offset(mm,hpux_gw_page_addr);
754
755 #if PTRS_PER_PMD == 1
756         start_pmd = 0;
757 #else
758         start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
759 #endif
760         start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
761
762         address = __pa(&hpux_gateway_page);
763 #if PTRS_PER_PMD == 1
764         pmd = (pmd_t *)__pa(pg_dir);
765 #else
766         pmd = (pmd_t *) pgd_address(*pg_dir);
767
768         /*
769          * pmd is physical at this point
770          */
771
772         if (!pmd) {
773                 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
774                 pmd = (pmd_t *) __pa(pmd);
775         }
776
777         __pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
778 #endif
779         /* now change pmd to kernel virtual addresses */
780
781         pmd = (pmd_t *)__va(pmd) + start_pmd;
782
783         /*
784          * pg_table is physical at this point
785          */
786
787         pg_table = (pte_t *) pmd_address(*pmd);
788         if (!pg_table)
789                 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
790
791         __pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
792
793         /* now change pg_table to kernel virtual addresses */
794
795         pg_table = (pte_t *) __va(pg_table) + start_pte;
796         set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
797 }
798 EXPORT_SYMBOL(map_hpux_gateway_page);
799 #endif
800
801 void __init paging_init(void)
802 {
803         int i;
804
805         setup_bootmem();
806         pagetable_init();
807         gateway_init();
808         flush_cache_all_local(); /* start with known state */
809         flush_tlb_all_local(NULL);
810
811         for (i = 0; i < npmem_ranges; i++) {
812                 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
813
814                 /* We have an IOMMU, so all memory can go into a single
815                    ZONE_DMA zone. */
816                 zones_size[ZONE_DMA] = pmem_ranges[i].pages;
817
818 #ifdef CONFIG_DISCONTIGMEM
819                 /* Need to initialize the pfnnid_map before we can initialize
820                    the zone */
821                 {
822                     int j;
823                     for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
824                          j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
825                          j++) {
826                         pfnnid_map[j] = i;
827                     }
828                 }
829 #endif
830
831                 free_area_init_node(i, NODE_DATA(i), zones_size,
832                                 pmem_ranges[i].start_pfn, NULL);
833         }
834 }
835
836 #ifdef CONFIG_PA20
837
838 /*
839  * Currently, all PA20 chips have 18 bit protection id's, which is the
840  * limiting factor (space ids are 32 bits).
841  */
842
843 #define NR_SPACE_IDS 262144
844
845 #else
846
847 /*
848  * Currently we have a one-to-one relationship between space id's and
849  * protection id's. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
850  * support 15 bit protection id's, so that is the limiting factor.
851  * PCXT' has 18 bit protection id's, but only 16 bit spaceids, so it's
852  * probably not worth the effort for a special case here.
853  */
854
855 #define NR_SPACE_IDS 32768
856
857 #endif  /* !CONFIG_PA20 */
858
859 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
860 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
861
862 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
863 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
864 static unsigned long space_id_index;
865 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
866 static unsigned long dirty_space_ids = 0;
867
868 static DEFINE_SPINLOCK(sid_lock);
869
870 unsigned long alloc_sid(void)
871 {
872         unsigned long index;
873
874         spin_lock(&sid_lock);
875
876         if (free_space_ids == 0) {
877                 if (dirty_space_ids != 0) {
878                         spin_unlock(&sid_lock);
879                         flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
880                         spin_lock(&sid_lock);
881                 }
882                 BUG_ON(free_space_ids == 0);
883         }
884
885         free_space_ids--;
886
887         index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
888         space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
889         space_id_index = index;
890
891         spin_unlock(&sid_lock);
892
893         return index << SPACEID_SHIFT;
894 }
895
896 void free_sid(unsigned long spaceid)
897 {
898         unsigned long index = spaceid >> SPACEID_SHIFT;
899         unsigned long *dirty_space_offset;
900
901         dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
902         index &= (BITS_PER_LONG - 1);
903
904         spin_lock(&sid_lock);
905
906         BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
907
908         *dirty_space_offset |= (1L << index);
909         dirty_space_ids++;
910
911         spin_unlock(&sid_lock);
912 }
913
914
915 #ifdef CONFIG_SMP
916 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
917 {
918         int i;
919
920         /* NOTE: sid_lock must be held upon entry */
921
922         *ndirtyptr = dirty_space_ids;
923         if (dirty_space_ids != 0) {
924             for (i = 0; i < SID_ARRAY_SIZE; i++) {
925                 dirty_array[i] = dirty_space_id[i];
926                 dirty_space_id[i] = 0;
927             }
928             dirty_space_ids = 0;
929         }
930
931         return;
932 }
933
934 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
935 {
936         int i;
937
938         /* NOTE: sid_lock must be held upon entry */
939
940         if (ndirty != 0) {
941                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
942                         space_id[i] ^= dirty_array[i];
943                 }
944
945                 free_space_ids += ndirty;
946                 space_id_index = 0;
947         }
948 }
949
950 #else /* CONFIG_SMP */
951
952 static void recycle_sids(void)
953 {
954         int i;
955
956         /* NOTE: sid_lock must be held upon entry */
957
958         if (dirty_space_ids != 0) {
959                 for (i = 0; i < SID_ARRAY_SIZE; i++) {
960                         space_id[i] ^= dirty_space_id[i];
961                         dirty_space_id[i] = 0;
962                 }
963
964                 free_space_ids += dirty_space_ids;
965                 dirty_space_ids = 0;
966                 space_id_index = 0;
967         }
968 }
969 #endif
970
971 /*
972  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
973  * purged, we can safely reuse the space ids that were released but
974  * not flushed from the tlb.
975  */
976
977 #ifdef CONFIG_SMP
978
979 static unsigned long recycle_ndirty;
980 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
981 static unsigned int recycle_inuse;
982
983 void flush_tlb_all(void)
984 {
985         int do_recycle;
986
987         do_recycle = 0;
988         spin_lock(&sid_lock);
989         if (dirty_space_ids > RECYCLE_THRESHOLD) {
990             BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
991             get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
992             recycle_inuse++;
993             do_recycle++;
994         }
995         spin_unlock(&sid_lock);
996         on_each_cpu(flush_tlb_all_local, NULL, 1, 1);
997         if (do_recycle) {
998             spin_lock(&sid_lock);
999             recycle_sids(recycle_ndirty,recycle_dirty_array);
1000             recycle_inuse = 0;
1001             spin_unlock(&sid_lock);
1002         }
1003 }
1004 #else
1005 void flush_tlb_all(void)
1006 {
1007         spin_lock(&sid_lock);
1008         flush_tlb_all_local(NULL);
1009         recycle_sids();
1010         spin_unlock(&sid_lock);
1011 }
1012 #endif
1013
1014 #ifdef CONFIG_BLK_DEV_INITRD
1015 void free_initrd_mem(unsigned long start, unsigned long end)
1016 {
1017         if (start >= end)
1018                 return;
1019         printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1020         for (; start < end; start += PAGE_SIZE) {
1021                 ClearPageReserved(virt_to_page(start));
1022                 init_page_count(virt_to_page(start));
1023                 free_page(start);
1024                 num_physpages++;
1025                 totalram_pages++;
1026         }
1027 }
1028 #endif