xref: /openbmc/linux/arch/parisc/mm/init.c (revision e868d61272caa648214046a096e5a6bfc068dc8c)
1 /*
2  *  linux/arch/parisc/mm/init.c
3  *
4  *  Copyright (C) 1995	Linus Torvalds
5  *  Copyright 1999 SuSE GmbH
6  *    changed by Philipp Rumpf
7  *  Copyright 1999 Philipp Rumpf (prumpf@tux.org)
8  *  Copyright 2004 Randolph Chung (tausq@debian.org)
9  *  Copyright 2006-2007 Helge Deller (deller@gmx.de)
10  *
11  */
12 
13 
14 #include <linux/module.h>
15 #include <linux/mm.h>
16 #include <linux/bootmem.h>
17 #include <linux/delay.h>
18 #include <linux/init.h>
19 #include <linux/pci.h>		/* for hppa_dma_ops and pcxl_dma_ops */
20 #include <linux/initrd.h>
21 #include <linux/swap.h>
22 #include <linux/unistd.h>
23 #include <linux/nodemask.h>	/* for node_online_map */
24 #include <linux/pagemap.h>	/* for release_pages and page_cache_release */
25 
26 #include <asm/pgalloc.h>
27 #include <asm/pgtable.h>
28 #include <asm/tlb.h>
29 #include <asm/pdc_chassis.h>
30 #include <asm/mmzone.h>
31 #include <asm/sections.h>
32 
33 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
34 
35 extern int  data_start;
36 
37 #ifdef CONFIG_DISCONTIGMEM
38 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
39 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 CONFIG_64BIT
70 #define MAX_MEM         (~0UL)
71 #else /* !CONFIG_64BIT */
72 #define MAX_MEM         (3584U*1024U*1024U)
73 #endif /* !CONFIG_64BIT */
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 
82 	/* We need this before __setup() functions are called */
83 
84 	limit = MAX_MEM;
85 	for (cp = boot_command_line; *cp; ) {
86 		if (memcmp(cp, "mem=", 4) == 0) {
87 			cp += 4;
88 			limit = memparse(cp, &end);
89 			if (end != cp)
90 				break;
91 			cp = end;
92 		} else {
93 			while (*cp != ' ' && *cp)
94 				++cp;
95 			while (*cp == ' ')
96 				++cp;
97 		}
98 	}
99 
100 	if (limit < mem_limit)
101 		mem_limit = limit;
102 }
103 
104 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
105 
106 static void __init setup_bootmem(void)
107 {
108 	unsigned long bootmap_size;
109 	unsigned long mem_max;
110 	unsigned long bootmap_pages;
111 	unsigned long bootmap_start_pfn;
112 	unsigned long bootmap_pfn;
113 #ifndef CONFIG_DISCONTIGMEM
114 	physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
115 	int npmem_holes;
116 #endif
117 	int i, sysram_resource_count;
118 
119 	disable_sr_hashing(); /* Turn off space register hashing */
120 
121 	/*
122 	 * Sort the ranges. Since the number of ranges is typically
123 	 * small, and performance is not an issue here, just do
124 	 * a simple insertion sort.
125 	 */
126 
127 	for (i = 1; i < npmem_ranges; i++) {
128 		int j;
129 
130 		for (j = i; j > 0; j--) {
131 			unsigned long tmp;
132 
133 			if (pmem_ranges[j-1].start_pfn <
134 			    pmem_ranges[j].start_pfn) {
135 
136 				break;
137 			}
138 			tmp = pmem_ranges[j-1].start_pfn;
139 			pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
140 			pmem_ranges[j].start_pfn = tmp;
141 			tmp = pmem_ranges[j-1].pages;
142 			pmem_ranges[j-1].pages = pmem_ranges[j].pages;
143 			pmem_ranges[j].pages = tmp;
144 		}
145 	}
146 
147 #ifndef CONFIG_DISCONTIGMEM
148 	/*
149 	 * Throw out ranges that are too far apart (controlled by
150 	 * MAX_GAP).
151 	 */
152 
153 	for (i = 1; i < npmem_ranges; i++) {
154 		if (pmem_ranges[i].start_pfn -
155 			(pmem_ranges[i-1].start_pfn +
156 			 pmem_ranges[i-1].pages) > MAX_GAP) {
157 			npmem_ranges = i;
158 			printk("Large gap in memory detected (%ld pages). "
159 			       "Consider turning on CONFIG_DISCONTIGMEM\n",
160 			       pmem_ranges[i].start_pfn -
161 			       (pmem_ranges[i-1].start_pfn +
162 			        pmem_ranges[i-1].pages));
163 			break;
164 		}
165 	}
166 #endif
167 
168 	if (npmem_ranges > 1) {
169 
170 		/* Print the memory ranges */
171 
172 		printk(KERN_INFO "Memory Ranges:\n");
173 
174 		for (i = 0; i < npmem_ranges; i++) {
175 			unsigned long start;
176 			unsigned long size;
177 
178 			size = (pmem_ranges[i].pages << PAGE_SHIFT);
179 			start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
180 			printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
181 				i,start, start + (size - 1), size >> 20);
182 		}
183 	}
184 
185 	sysram_resource_count = npmem_ranges;
186 	for (i = 0; i < sysram_resource_count; i++) {
187 		struct resource *res = &sysram_resources[i];
188 		res->name = "System RAM";
189 		res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
190 		res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
191 		res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
192 		request_resource(&iomem_resource, res);
193 	}
194 
195 	/*
196 	 * For 32 bit kernels we limit the amount of memory we can
197 	 * support, in order to preserve enough kernel address space
198 	 * for other purposes. For 64 bit kernels we don't normally
199 	 * limit the memory, but this mechanism can be used to
200 	 * artificially limit the amount of memory (and it is written
201 	 * to work with multiple memory ranges).
202 	 */
203 
204 	mem_limit_func();       /* check for "mem=" argument */
205 
206 	mem_max = 0;
207 	num_physpages = 0;
208 	for (i = 0; i < npmem_ranges; i++) {
209 		unsigned long rsize;
210 
211 		rsize = pmem_ranges[i].pages << PAGE_SHIFT;
212 		if ((mem_max + rsize) > mem_limit) {
213 			printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
214 			if (mem_max == mem_limit)
215 				npmem_ranges = i;
216 			else {
217 				pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
218 						       - (mem_max >> PAGE_SHIFT);
219 				npmem_ranges = i + 1;
220 				mem_max = mem_limit;
221 			}
222 	        num_physpages += pmem_ranges[i].pages;
223 			break;
224 		}
225 	    num_physpages += pmem_ranges[i].pages;
226 		mem_max += rsize;
227 	}
228 
229 	printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
230 
231 #ifndef CONFIG_DISCONTIGMEM
232 	/* Merge the ranges, keeping track of the holes */
233 
234 	{
235 		unsigned long end_pfn;
236 		unsigned long hole_pages;
237 
238 		npmem_holes = 0;
239 		end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
240 		for (i = 1; i < npmem_ranges; i++) {
241 
242 			hole_pages = pmem_ranges[i].start_pfn - end_pfn;
243 			if (hole_pages) {
244 				pmem_holes[npmem_holes].start_pfn = end_pfn;
245 				pmem_holes[npmem_holes++].pages = hole_pages;
246 				end_pfn += hole_pages;
247 			}
248 			end_pfn += pmem_ranges[i].pages;
249 		}
250 
251 		pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
252 		npmem_ranges = 1;
253 	}
254 #endif
255 
256 	bootmap_pages = 0;
257 	for (i = 0; i < npmem_ranges; i++)
258 		bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
259 
260 	bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
261 
262 #ifdef CONFIG_DISCONTIGMEM
263 	for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
264 		memset(NODE_DATA(i), 0, sizeof(pg_data_t));
265 		NODE_DATA(i)->bdata = &bmem_data[i];
266 	}
267 	memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
268 
269 	for (i = 0; i < npmem_ranges; i++)
270 		node_set_online(i);
271 #endif
272 
273 	/*
274 	 * Initialize and free the full range of memory in each range.
275 	 * Note that the only writing these routines do are to the bootmap,
276 	 * and we've made sure to locate the bootmap properly so that they
277 	 * won't be writing over anything important.
278 	 */
279 
280 	bootmap_pfn = bootmap_start_pfn;
281 	max_pfn = 0;
282 	for (i = 0; i < npmem_ranges; i++) {
283 		unsigned long start_pfn;
284 		unsigned long npages;
285 
286 		start_pfn = pmem_ranges[i].start_pfn;
287 		npages = pmem_ranges[i].pages;
288 
289 		bootmap_size = init_bootmem_node(NODE_DATA(i),
290 						bootmap_pfn,
291 						start_pfn,
292 						(start_pfn + npages) );
293 		free_bootmem_node(NODE_DATA(i),
294 				  (start_pfn << PAGE_SHIFT),
295 				  (npages << PAGE_SHIFT) );
296 		bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
297 		if ((start_pfn + npages) > max_pfn)
298 			max_pfn = start_pfn + npages;
299 	}
300 
301 	/* IOMMU is always used to access "high mem" on those boxes
302 	 * that can support enough mem that a PCI device couldn't
303 	 * directly DMA to any physical addresses.
304 	 * ISA DMA support will need to revisit this.
305 	 */
306 	max_low_pfn = max_pfn;
307 
308 	if ((bootmap_pfn - bootmap_start_pfn) != bootmap_pages) {
309 		printk(KERN_WARNING "WARNING! bootmap sizing is messed up!\n");
310 		BUG();
311 	}
312 
313 	/* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
314 
315 #define PDC_CONSOLE_IO_IODC_SIZE 32768
316 
317 	reserve_bootmem_node(NODE_DATA(0), 0UL,
318 			(unsigned long)(PAGE0->mem_free + PDC_CONSOLE_IO_IODC_SIZE));
319 	reserve_bootmem_node(NODE_DATA(0), __pa((unsigned long)_text),
320 			(unsigned long)(_end - _text));
321 	reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
322 			((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT));
323 
324 #ifndef CONFIG_DISCONTIGMEM
325 
326 	/* reserve the holes */
327 
328 	for (i = 0; i < npmem_holes; i++) {
329 		reserve_bootmem_node(NODE_DATA(0),
330 				(pmem_holes[i].start_pfn << PAGE_SHIFT),
331 				(pmem_holes[i].pages << PAGE_SHIFT));
332 	}
333 #endif
334 
335 #ifdef CONFIG_BLK_DEV_INITRD
336 	if (initrd_start) {
337 		printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
338 		if (__pa(initrd_start) < mem_max) {
339 			unsigned long initrd_reserve;
340 
341 			if (__pa(initrd_end) > mem_max) {
342 				initrd_reserve = mem_max - __pa(initrd_start);
343 			} else {
344 				initrd_reserve = initrd_end - initrd_start;
345 			}
346 			initrd_below_start_ok = 1;
347 			printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
348 
349 			reserve_bootmem_node(NODE_DATA(0),__pa(initrd_start), initrd_reserve);
350 		}
351 	}
352 #endif
353 
354 	data_resource.start =  virt_to_phys(&data_start);
355 	data_resource.end = virt_to_phys(_end) - 1;
356 	code_resource.start = virt_to_phys(_text);
357 	code_resource.end = virt_to_phys(&data_start)-1;
358 
359 	/* We don't know which region the kernel will be in, so try
360 	 * all of them.
361 	 */
362 	for (i = 0; i < sysram_resource_count; i++) {
363 		struct resource *res = &sysram_resources[i];
364 		request_resource(res, &code_resource);
365 		request_resource(res, &data_resource);
366 	}
367 	request_resource(&sysram_resources[0], &pdcdata_resource);
368 }
369 
370 void free_initmem(void)
371 {
372 	unsigned long addr, init_begin, init_end;
373 
374 	printk(KERN_INFO "Freeing unused kernel memory: ");
375 
376 #ifdef CONFIG_DEBUG_KERNEL
377 	/* Attempt to catch anyone trying to execute code here
378 	 * by filling the page with BRK insns.
379 	 *
380 	 * If we disable interrupts for all CPUs, then IPI stops working.
381 	 * Kinda breaks the global cache flushing.
382 	 */
383 	local_irq_disable();
384 
385 	memset(__init_begin, 0x00,
386 		(unsigned long)__init_end - (unsigned long)__init_begin);
387 
388 	flush_data_cache();
389 	asm volatile("sync" : : );
390 	flush_icache_range((unsigned long)__init_begin, (unsigned long)__init_end);
391 	asm volatile("sync" : : );
392 
393 	local_irq_enable();
394 #endif
395 
396 	/* align __init_begin and __init_end to page size,
397 	   ignoring linker script where we might have tried to save RAM */
398 	init_begin = PAGE_ALIGN((unsigned long)(__init_begin));
399 	init_end   = PAGE_ALIGN((unsigned long)(__init_end));
400 	for (addr = init_begin; addr < init_end; addr += PAGE_SIZE) {
401 		ClearPageReserved(virt_to_page(addr));
402 		init_page_count(virt_to_page(addr));
403 		free_page(addr);
404 		num_physpages++;
405 		totalram_pages++;
406 	}
407 
408 	/* set up a new led state on systems shipped LED State panel */
409 	pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
410 
411 	printk("%luk freed\n", (init_end - init_begin) >> 10);
412 }
413 
414 
415 #ifdef CONFIG_DEBUG_RODATA
416 void mark_rodata_ro(void)
417 {
418 	/* rodata memory was already mapped with KERNEL_RO access rights by
419            pagetable_init() and map_pages(). No need to do additional stuff here */
420 	printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
421 		(unsigned long)(__end_rodata - __start_rodata) >> 10);
422 }
423 #endif
424 
425 
426 /*
427  * Just an arbitrary offset to serve as a "hole" between mapping areas
428  * (between top of physical memory and a potential pcxl dma mapping
429  * area, and below the vmalloc mapping area).
430  *
431  * The current 32K value just means that there will be a 32K "hole"
432  * between mapping areas. That means that  any out-of-bounds memory
433  * accesses will hopefully be caught. The vmalloc() routines leaves
434  * a hole of 4kB between each vmalloced area for the same reason.
435  */
436 
437  /* Leave room for gateway page expansion */
438 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
439 #error KERNEL_MAP_START is in gateway reserved region
440 #endif
441 #define MAP_START (KERNEL_MAP_START)
442 
443 #define VM_MAP_OFFSET  (32*1024)
444 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
445 				     & ~(VM_MAP_OFFSET-1)))
446 
447 void *vmalloc_start __read_mostly;
448 EXPORT_SYMBOL(vmalloc_start);
449 
450 #ifdef CONFIG_PA11
451 unsigned long pcxl_dma_start __read_mostly;
452 #endif
453 
454 void __init mem_init(void)
455 {
456 	int codesize, reservedpages, datasize, initsize;
457 
458 	high_memory = __va((max_pfn << PAGE_SHIFT));
459 
460 #ifndef CONFIG_DISCONTIGMEM
461 	max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
462 	totalram_pages += free_all_bootmem();
463 #else
464 	{
465 		int i;
466 
467 		for (i = 0; i < npmem_ranges; i++)
468 			totalram_pages += free_all_bootmem_node(NODE_DATA(i));
469 	}
470 #endif
471 
472 	codesize = (unsigned long)_etext - (unsigned long)_text;
473 	datasize = (unsigned long)_edata - (unsigned long)_etext;
474 	initsize = (unsigned long)__init_end - (unsigned long)__init_begin;
475 
476 	reservedpages = 0;
477 {
478 	unsigned long pfn;
479 #ifdef CONFIG_DISCONTIGMEM
480 	int i;
481 
482 	for (i = 0; i < npmem_ranges; i++) {
483 		for (pfn = node_start_pfn(i); pfn < node_end_pfn(i); pfn++) {
484 			if (PageReserved(pfn_to_page(pfn)))
485 				reservedpages++;
486 		}
487 	}
488 #else /* !CONFIG_DISCONTIGMEM */
489 	for (pfn = 0; pfn < max_pfn; pfn++) {
490 		/*
491 		 * Only count reserved RAM pages
492 		 */
493 		if (PageReserved(pfn_to_page(pfn)))
494 			reservedpages++;
495 	}
496 #endif
497 }
498 
499 #ifdef CONFIG_PA11
500 	if (hppa_dma_ops == &pcxl_dma_ops) {
501 		pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
502 		vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
503 	} else {
504 		pcxl_dma_start = 0;
505 		vmalloc_start = SET_MAP_OFFSET(MAP_START);
506 	}
507 #else
508 	vmalloc_start = SET_MAP_OFFSET(MAP_START);
509 #endif
510 
511 	printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, %dk reserved, %dk data, %dk init)\n",
512 		(unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
513 		num_physpages << (PAGE_SHIFT-10),
514 		codesize >> 10,
515 		reservedpages << (PAGE_SHIFT-10),
516 		datasize >> 10,
517 		initsize >> 10
518 	);
519 
520 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
521 	printk("virtual kernel memory layout:\n"
522 	       "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
523 	       "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
524 	       "      .init : 0x%p - 0x%p   (%4ld kB)\n"
525 	       "      .data : 0x%p - 0x%p   (%4ld kB)\n"
526 	       "      .text : 0x%p - 0x%p   (%4ld kB)\n",
527 
528 	       (void*)VMALLOC_START, (void*)VMALLOC_END,
529 	       (VMALLOC_END - VMALLOC_START) >> 20,
530 
531 	       __va(0), high_memory,
532 	       ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
533 
534 	       __init_begin, __init_end,
535 	       ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
536 
537 	       _etext, _edata,
538 	       ((unsigned long)_edata - (unsigned long)_etext) >> 10,
539 
540 	       _text, _etext,
541 	       ((unsigned long)_etext - (unsigned long)_text) >> 10);
542 #endif
543 }
544 
545 unsigned long *empty_zero_page __read_mostly;
546 
547 void show_mem(void)
548 {
549 	int i,free = 0,total = 0,reserved = 0;
550 	int shared = 0, cached = 0;
551 
552 	printk(KERN_INFO "Mem-info:\n");
553 	show_free_areas();
554 	printk(KERN_INFO "Free swap:	 %6ldkB\n",
555 				nr_swap_pages<<(PAGE_SHIFT-10));
556 #ifndef CONFIG_DISCONTIGMEM
557 	i = max_mapnr;
558 	while (i-- > 0) {
559 		total++;
560 		if (PageReserved(mem_map+i))
561 			reserved++;
562 		else if (PageSwapCache(mem_map+i))
563 			cached++;
564 		else if (!page_count(&mem_map[i]))
565 			free++;
566 		else
567 			shared += page_count(&mem_map[i]) - 1;
568 	}
569 #else
570 	for (i = 0; i < npmem_ranges; i++) {
571 		int j;
572 
573 		for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
574 			struct page *p;
575 			unsigned long flags;
576 
577 			pgdat_resize_lock(NODE_DATA(i), &flags);
578 			p = nid_page_nr(i, j) - node_start_pfn(i);
579 
580 			total++;
581 			if (PageReserved(p))
582 				reserved++;
583 			else if (PageSwapCache(p))
584 				cached++;
585 			else if (!page_count(p))
586 				free++;
587 			else
588 				shared += page_count(p) - 1;
589 			pgdat_resize_unlock(NODE_DATA(i), &flags);
590         	}
591 	}
592 #endif
593 	printk(KERN_INFO "%d pages of RAM\n", total);
594 	printk(KERN_INFO "%d reserved pages\n", reserved);
595 	printk(KERN_INFO "%d pages shared\n", shared);
596 	printk(KERN_INFO "%d pages swap cached\n", cached);
597 
598 
599 #ifdef CONFIG_DISCONTIGMEM
600 	{
601 		struct zonelist *zl;
602 		int i, j, k;
603 
604 		for (i = 0; i < npmem_ranges; i++) {
605 			for (j = 0; j < MAX_NR_ZONES; j++) {
606 				zl = NODE_DATA(i)->node_zonelists + j;
607 
608 				printk("Zone list for zone %d on node %d: ", j, i);
609 				for (k = 0; zl->zones[k] != NULL; k++)
610 					printk("[%ld/%s] ", zone_to_nid(zl->zones[k]), zl->zones[k]->name);
611 				printk("\n");
612 			}
613 		}
614 	}
615 #endif
616 }
617 
618 
619 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
620 {
621 	pgd_t *pg_dir;
622 	pmd_t *pmd;
623 	pte_t *pg_table;
624 	unsigned long end_paddr;
625 	unsigned long start_pmd;
626 	unsigned long start_pte;
627 	unsigned long tmp1;
628 	unsigned long tmp2;
629 	unsigned long address;
630 	unsigned long ro_start;
631 	unsigned long ro_end;
632 	unsigned long fv_addr;
633 	unsigned long gw_addr;
634 	extern const unsigned long fault_vector_20;
635 	extern void * const linux_gateway_page;
636 
637 	ro_start = __pa((unsigned long)_text);
638 	ro_end   = __pa((unsigned long)&data_start);
639 	fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
640 	gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
641 
642 	end_paddr = start_paddr + size;
643 
644 	pg_dir = pgd_offset_k(start_vaddr);
645 
646 #if PTRS_PER_PMD == 1
647 	start_pmd = 0;
648 #else
649 	start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
650 #endif
651 	start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
652 
653 	address = start_paddr;
654 	while (address < end_paddr) {
655 #if PTRS_PER_PMD == 1
656 		pmd = (pmd_t *)__pa(pg_dir);
657 #else
658 		pmd = (pmd_t *)pgd_address(*pg_dir);
659 
660 		/*
661 		 * pmd is physical at this point
662 		 */
663 
664 		if (!pmd) {
665 			pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
666 			pmd = (pmd_t *) __pa(pmd);
667 		}
668 
669 		pgd_populate(NULL, pg_dir, __va(pmd));
670 #endif
671 		pg_dir++;
672 
673 		/* now change pmd to kernel virtual addresses */
674 
675 		pmd = (pmd_t *)__va(pmd) + start_pmd;
676 		for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
677 
678 			/*
679 			 * pg_table is physical at this point
680 			 */
681 
682 			pg_table = (pte_t *)pmd_address(*pmd);
683 			if (!pg_table) {
684 				pg_table = (pte_t *)
685 					alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
686 				pg_table = (pte_t *) __pa(pg_table);
687 			}
688 
689 			pmd_populate_kernel(NULL, pmd, __va(pg_table));
690 
691 			/* now change pg_table to kernel virtual addresses */
692 
693 			pg_table = (pte_t *) __va(pg_table) + start_pte;
694 			for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
695 				pte_t pte;
696 
697 				/*
698 				 * Map the fault vector writable so we can
699 				 * write the HPMC checksum.
700 				 */
701 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
702 				if (address >= ro_start && address < ro_end
703 							&& address != fv_addr
704 							&& address != gw_addr)
705 				    pte = __mk_pte(address, PAGE_KERNEL_RO);
706 				else
707 #endif
708 				    pte = __mk_pte(address, pgprot);
709 
710 				if (address >= end_paddr)
711 					pte_val(pte) = 0;
712 
713 				set_pte(pg_table, pte);
714 
715 				address += PAGE_SIZE;
716 			}
717 			start_pte = 0;
718 
719 			if (address >= end_paddr)
720 			    break;
721 		}
722 		start_pmd = 0;
723 	}
724 }
725 
726 /*
727  * pagetable_init() sets up the page tables
728  *
729  * Note that gateway_init() places the Linux gateway page at page 0.
730  * Since gateway pages cannot be dereferenced this has the desirable
731  * side effect of trapping those pesky NULL-reference errors in the
732  * kernel.
733  */
734 static void __init pagetable_init(void)
735 {
736 	int range;
737 
738 	/* Map each physical memory range to its kernel vaddr */
739 
740 	for (range = 0; range < npmem_ranges; range++) {
741 		unsigned long start_paddr;
742 		unsigned long end_paddr;
743 		unsigned long size;
744 
745 		start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
746 		end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
747 		size = pmem_ranges[range].pages << PAGE_SHIFT;
748 
749 		map_pages((unsigned long)__va(start_paddr), start_paddr,
750 			size, PAGE_KERNEL);
751 	}
752 
753 #ifdef CONFIG_BLK_DEV_INITRD
754 	if (initrd_end && initrd_end > mem_limit) {
755 		printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
756 		map_pages(initrd_start, __pa(initrd_start),
757 			initrd_end - initrd_start, PAGE_KERNEL);
758 	}
759 #endif
760 
761 	empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
762 	memset(empty_zero_page, 0, PAGE_SIZE);
763 }
764 
765 static void __init gateway_init(void)
766 {
767 	unsigned long linux_gateway_page_addr;
768 	/* FIXME: This is 'const' in order to trick the compiler
769 	   into not treating it as DP-relative data. */
770 	extern void * const linux_gateway_page;
771 
772 	linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
773 
774 	/*
775 	 * Setup Linux Gateway page.
776 	 *
777 	 * The Linux gateway page will reside in kernel space (on virtual
778 	 * page 0), so it doesn't need to be aliased into user space.
779 	 */
780 
781 	map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
782 		PAGE_SIZE, PAGE_GATEWAY);
783 }
784 
785 #ifdef CONFIG_HPUX
786 void
787 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
788 {
789 	pgd_t *pg_dir;
790 	pmd_t *pmd;
791 	pte_t *pg_table;
792 	unsigned long start_pmd;
793 	unsigned long start_pte;
794 	unsigned long address;
795 	unsigned long hpux_gw_page_addr;
796 	/* FIXME: This is 'const' in order to trick the compiler
797 	   into not treating it as DP-relative data. */
798 	extern void * const hpux_gateway_page;
799 
800 	hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
801 
802 	/*
803 	 * Setup HP-UX Gateway page.
804 	 *
805 	 * The HP-UX gateway page resides in the user address space,
806 	 * so it needs to be aliased into each process.
807 	 */
808 
809 	pg_dir = pgd_offset(mm,hpux_gw_page_addr);
810 
811 #if PTRS_PER_PMD == 1
812 	start_pmd = 0;
813 #else
814 	start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
815 #endif
816 	start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
817 
818 	address = __pa(&hpux_gateway_page);
819 #if PTRS_PER_PMD == 1
820 	pmd = (pmd_t *)__pa(pg_dir);
821 #else
822 	pmd = (pmd_t *) pgd_address(*pg_dir);
823 
824 	/*
825 	 * pmd is physical at this point
826 	 */
827 
828 	if (!pmd) {
829 		pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
830 		pmd = (pmd_t *) __pa(pmd);
831 	}
832 
833 	__pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
834 #endif
835 	/* now change pmd to kernel virtual addresses */
836 
837 	pmd = (pmd_t *)__va(pmd) + start_pmd;
838 
839 	/*
840 	 * pg_table is physical at this point
841 	 */
842 
843 	pg_table = (pte_t *) pmd_address(*pmd);
844 	if (!pg_table)
845 		pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
846 
847 	__pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
848 
849 	/* now change pg_table to kernel virtual addresses */
850 
851 	pg_table = (pte_t *) __va(pg_table) + start_pte;
852 	set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
853 }
854 EXPORT_SYMBOL(map_hpux_gateway_page);
855 #endif
856 
857 void __init paging_init(void)
858 {
859 	int i;
860 
861 	setup_bootmem();
862 	pagetable_init();
863 	gateway_init();
864 	flush_cache_all_local(); /* start with known state */
865 	flush_tlb_all_local(NULL);
866 
867 	for (i = 0; i < npmem_ranges; i++) {
868 		unsigned long zones_size[MAX_NR_ZONES] = { 0, };
869 
870 		zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
871 
872 #ifdef CONFIG_DISCONTIGMEM
873 		/* Need to initialize the pfnnid_map before we can initialize
874 		   the zone */
875 		{
876 		    int j;
877 		    for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
878 			 j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
879 			 j++) {
880 			pfnnid_map[j] = i;
881 		    }
882 		}
883 #endif
884 
885 		free_area_init_node(i, NODE_DATA(i), zones_size,
886 				pmem_ranges[i].start_pfn, NULL);
887 	}
888 }
889 
890 #ifdef CONFIG_PA20
891 
892 /*
893  * Currently, all PA20 chips have 18 bit protection id's, which is the
894  * limiting factor (space ids are 32 bits).
895  */
896 
897 #define NR_SPACE_IDS 262144
898 
899 #else
900 
901 /*
902  * Currently we have a one-to-one relationship between space id's and
903  * protection id's. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
904  * support 15 bit protection id's, so that is the limiting factor.
905  * PCXT' has 18 bit protection id's, but only 16 bit spaceids, so it's
906  * probably not worth the effort for a special case here.
907  */
908 
909 #define NR_SPACE_IDS 32768
910 
911 #endif  /* !CONFIG_PA20 */
912 
913 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
914 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
915 
916 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
917 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
918 static unsigned long space_id_index;
919 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
920 static unsigned long dirty_space_ids = 0;
921 
922 static DEFINE_SPINLOCK(sid_lock);
923 
924 unsigned long alloc_sid(void)
925 {
926 	unsigned long index;
927 
928 	spin_lock(&sid_lock);
929 
930 	if (free_space_ids == 0) {
931 		if (dirty_space_ids != 0) {
932 			spin_unlock(&sid_lock);
933 			flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
934 			spin_lock(&sid_lock);
935 		}
936 		BUG_ON(free_space_ids == 0);
937 	}
938 
939 	free_space_ids--;
940 
941 	index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
942 	space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
943 	space_id_index = index;
944 
945 	spin_unlock(&sid_lock);
946 
947 	return index << SPACEID_SHIFT;
948 }
949 
950 void free_sid(unsigned long spaceid)
951 {
952 	unsigned long index = spaceid >> SPACEID_SHIFT;
953 	unsigned long *dirty_space_offset;
954 
955 	dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
956 	index &= (BITS_PER_LONG - 1);
957 
958 	spin_lock(&sid_lock);
959 
960 	BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
961 
962 	*dirty_space_offset |= (1L << index);
963 	dirty_space_ids++;
964 
965 	spin_unlock(&sid_lock);
966 }
967 
968 
969 #ifdef CONFIG_SMP
970 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
971 {
972 	int i;
973 
974 	/* NOTE: sid_lock must be held upon entry */
975 
976 	*ndirtyptr = dirty_space_ids;
977 	if (dirty_space_ids != 0) {
978 	    for (i = 0; i < SID_ARRAY_SIZE; i++) {
979 		dirty_array[i] = dirty_space_id[i];
980 		dirty_space_id[i] = 0;
981 	    }
982 	    dirty_space_ids = 0;
983 	}
984 
985 	return;
986 }
987 
988 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
989 {
990 	int i;
991 
992 	/* NOTE: sid_lock must be held upon entry */
993 
994 	if (ndirty != 0) {
995 		for (i = 0; i < SID_ARRAY_SIZE; i++) {
996 			space_id[i] ^= dirty_array[i];
997 		}
998 
999 		free_space_ids += ndirty;
1000 		space_id_index = 0;
1001 	}
1002 }
1003 
1004 #else /* CONFIG_SMP */
1005 
1006 static void recycle_sids(void)
1007 {
1008 	int i;
1009 
1010 	/* NOTE: sid_lock must be held upon entry */
1011 
1012 	if (dirty_space_ids != 0) {
1013 		for (i = 0; i < SID_ARRAY_SIZE; i++) {
1014 			space_id[i] ^= dirty_space_id[i];
1015 			dirty_space_id[i] = 0;
1016 		}
1017 
1018 		free_space_ids += dirty_space_ids;
1019 		dirty_space_ids = 0;
1020 		space_id_index = 0;
1021 	}
1022 }
1023 #endif
1024 
1025 /*
1026  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
1027  * purged, we can safely reuse the space ids that were released but
1028  * not flushed from the tlb.
1029  */
1030 
1031 #ifdef CONFIG_SMP
1032 
1033 static unsigned long recycle_ndirty;
1034 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
1035 static unsigned int recycle_inuse;
1036 
1037 void flush_tlb_all(void)
1038 {
1039 	int do_recycle;
1040 
1041 	do_recycle = 0;
1042 	spin_lock(&sid_lock);
1043 	if (dirty_space_ids > RECYCLE_THRESHOLD) {
1044 	    BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
1045 	    get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
1046 	    recycle_inuse++;
1047 	    do_recycle++;
1048 	}
1049 	spin_unlock(&sid_lock);
1050 	on_each_cpu(flush_tlb_all_local, NULL, 1, 1);
1051 	if (do_recycle) {
1052 	    spin_lock(&sid_lock);
1053 	    recycle_sids(recycle_ndirty,recycle_dirty_array);
1054 	    recycle_inuse = 0;
1055 	    spin_unlock(&sid_lock);
1056 	}
1057 }
1058 #else
1059 void flush_tlb_all(void)
1060 {
1061 	spin_lock(&sid_lock);
1062 	flush_tlb_all_local(NULL);
1063 	recycle_sids();
1064 	spin_unlock(&sid_lock);
1065 }
1066 #endif
1067 
1068 #ifdef CONFIG_BLK_DEV_INITRD
1069 void free_initrd_mem(unsigned long start, unsigned long end)
1070 {
1071 	if (start >= end)
1072 		return;
1073 	printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1074 	for (; start < end; start += PAGE_SIZE) {
1075 		ClearPageReserved(virt_to_page(start));
1076 		init_page_count(virt_to_page(start));
1077 		free_page(start);
1078 		num_physpages++;
1079 		totalram_pages++;
1080 	}
1081 }
1082 #endif
1083