xref: /openbmc/linux/arch/parisc/mm/init.c (revision b8bb76713ec50df2f11efee386e16f93d51e1076)
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 unsigned char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
40 #endif
41 
42 static struct resource data_resource = {
43 	.name	= "Kernel data",
44 	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM,
45 };
46 
47 static struct resource code_resource = {
48 	.name	= "Kernel code",
49 	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM,
50 };
51 
52 static struct resource pdcdata_resource = {
53 	.name	= "PDC data (Page Zero)",
54 	.start	= 0,
55 	.end	= 0x9ff,
56 	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM,
57 };
58 
59 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
60 
61 /* The following array is initialized from the firmware specific
62  * information retrieved in kernel/inventory.c.
63  */
64 
65 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
66 int npmem_ranges __read_mostly;
67 
68 #ifdef CONFIG_64BIT
69 #define MAX_MEM         (~0UL)
70 #else /* !CONFIG_64BIT */
71 #define MAX_MEM         (3584U*1024U*1024U)
72 #endif /* !CONFIG_64BIT */
73 
74 static unsigned long mem_limit __read_mostly = MAX_MEM;
75 
76 static void __init mem_limit_func(void)
77 {
78 	char *cp, *end;
79 	unsigned long limit;
80 
81 	/* We need this before __setup() functions are called */
82 
83 	limit = MAX_MEM;
84 	for (cp = boot_command_line; *cp; ) {
85 		if (memcmp(cp, "mem=", 4) == 0) {
86 			cp += 4;
87 			limit = memparse(cp, &end);
88 			if (end != cp)
89 				break;
90 			cp = end;
91 		} else {
92 			while (*cp != ' ' && *cp)
93 				++cp;
94 			while (*cp == ' ')
95 				++cp;
96 		}
97 	}
98 
99 	if (limit < mem_limit)
100 		mem_limit = limit;
101 }
102 
103 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
104 
105 static void __init setup_bootmem(void)
106 {
107 	unsigned long bootmap_size;
108 	unsigned long mem_max;
109 	unsigned long bootmap_pages;
110 	unsigned long bootmap_start_pfn;
111 	unsigned long bootmap_pfn;
112 #ifndef CONFIG_DISCONTIGMEM
113 	physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
114 	int npmem_holes;
115 #endif
116 	int i, sysram_resource_count;
117 
118 	disable_sr_hashing(); /* Turn off space register hashing */
119 
120 	/*
121 	 * Sort the ranges. Since the number of ranges is typically
122 	 * small, and performance is not an issue here, just do
123 	 * a simple insertion sort.
124 	 */
125 
126 	for (i = 1; i < npmem_ranges; i++) {
127 		int j;
128 
129 		for (j = i; j > 0; j--) {
130 			unsigned long tmp;
131 
132 			if (pmem_ranges[j-1].start_pfn <
133 			    pmem_ranges[j].start_pfn) {
134 
135 				break;
136 			}
137 			tmp = pmem_ranges[j-1].start_pfn;
138 			pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
139 			pmem_ranges[j].start_pfn = tmp;
140 			tmp = pmem_ranges[j-1].pages;
141 			pmem_ranges[j-1].pages = pmem_ranges[j].pages;
142 			pmem_ranges[j].pages = tmp;
143 		}
144 	}
145 
146 #ifndef CONFIG_DISCONTIGMEM
147 	/*
148 	 * Throw out ranges that are too far apart (controlled by
149 	 * MAX_GAP).
150 	 */
151 
152 	for (i = 1; i < npmem_ranges; i++) {
153 		if (pmem_ranges[i].start_pfn -
154 			(pmem_ranges[i-1].start_pfn +
155 			 pmem_ranges[i-1].pages) > MAX_GAP) {
156 			npmem_ranges = i;
157 			printk("Large gap in memory detected (%ld pages). "
158 			       "Consider turning on CONFIG_DISCONTIGMEM\n",
159 			       pmem_ranges[i].start_pfn -
160 			       (pmem_ranges[i-1].start_pfn +
161 			        pmem_ranges[i-1].pages));
162 			break;
163 		}
164 	}
165 #endif
166 
167 	if (npmem_ranges > 1) {
168 
169 		/* Print the memory ranges */
170 
171 		printk(KERN_INFO "Memory Ranges:\n");
172 
173 		for (i = 0; i < npmem_ranges; i++) {
174 			unsigned long start;
175 			unsigned long size;
176 
177 			size = (pmem_ranges[i].pages << PAGE_SHIFT);
178 			start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
179 			printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
180 				i,start, start + (size - 1), size >> 20);
181 		}
182 	}
183 
184 	sysram_resource_count = npmem_ranges;
185 	for (i = 0; i < sysram_resource_count; i++) {
186 		struct resource *res = &sysram_resources[i];
187 		res->name = "System RAM";
188 		res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
189 		res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
190 		res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
191 		request_resource(&iomem_resource, res);
192 	}
193 
194 	/*
195 	 * For 32 bit kernels we limit the amount of memory we can
196 	 * support, in order to preserve enough kernel address space
197 	 * for other purposes. For 64 bit kernels we don't normally
198 	 * limit the memory, but this mechanism can be used to
199 	 * artificially limit the amount of memory (and it is written
200 	 * to work with multiple memory ranges).
201 	 */
202 
203 	mem_limit_func();       /* check for "mem=" argument */
204 
205 	mem_max = 0;
206 	num_physpages = 0;
207 	for (i = 0; i < npmem_ranges; i++) {
208 		unsigned long rsize;
209 
210 		rsize = pmem_ranges[i].pages << PAGE_SHIFT;
211 		if ((mem_max + rsize) > mem_limit) {
212 			printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
213 			if (mem_max == mem_limit)
214 				npmem_ranges = i;
215 			else {
216 				pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
217 						       - (mem_max >> PAGE_SHIFT);
218 				npmem_ranges = i + 1;
219 				mem_max = mem_limit;
220 			}
221 	        num_physpages += pmem_ranges[i].pages;
222 			break;
223 		}
224 	    num_physpages += pmem_ranges[i].pages;
225 		mem_max += rsize;
226 	}
227 
228 	printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
229 
230 #ifndef CONFIG_DISCONTIGMEM
231 	/* Merge the ranges, keeping track of the holes */
232 
233 	{
234 		unsigned long end_pfn;
235 		unsigned long hole_pages;
236 
237 		npmem_holes = 0;
238 		end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
239 		for (i = 1; i < npmem_ranges; i++) {
240 
241 			hole_pages = pmem_ranges[i].start_pfn - end_pfn;
242 			if (hole_pages) {
243 				pmem_holes[npmem_holes].start_pfn = end_pfn;
244 				pmem_holes[npmem_holes++].pages = hole_pages;
245 				end_pfn += hole_pages;
246 			}
247 			end_pfn += pmem_ranges[i].pages;
248 		}
249 
250 		pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
251 		npmem_ranges = 1;
252 	}
253 #endif
254 
255 	bootmap_pages = 0;
256 	for (i = 0; i < npmem_ranges; i++)
257 		bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
258 
259 	bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
260 
261 #ifdef CONFIG_DISCONTIGMEM
262 	for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
263 		memset(NODE_DATA(i), 0, sizeof(pg_data_t));
264 		NODE_DATA(i)->bdata = &bootmem_node_data[i];
265 	}
266 	memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
267 
268 	for (i = 0; i < npmem_ranges; i++)
269 		node_set_online(i);
270 #endif
271 
272 	/*
273 	 * Initialize and free the full range of memory in each range.
274 	 * Note that the only writing these routines do are to the bootmap,
275 	 * and we've made sure to locate the bootmap properly so that they
276 	 * won't be writing over anything important.
277 	 */
278 
279 	bootmap_pfn = bootmap_start_pfn;
280 	max_pfn = 0;
281 	for (i = 0; i < npmem_ranges; i++) {
282 		unsigned long start_pfn;
283 		unsigned long npages;
284 
285 		start_pfn = pmem_ranges[i].start_pfn;
286 		npages = pmem_ranges[i].pages;
287 
288 		bootmap_size = init_bootmem_node(NODE_DATA(i),
289 						bootmap_pfn,
290 						start_pfn,
291 						(start_pfn + npages) );
292 		free_bootmem_node(NODE_DATA(i),
293 				  (start_pfn << PAGE_SHIFT),
294 				  (npages << PAGE_SHIFT) );
295 		bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
296 		if ((start_pfn + npages) > max_pfn)
297 			max_pfn = start_pfn + npages;
298 	}
299 
300 	/* IOMMU is always used to access "high mem" on those boxes
301 	 * that can support enough mem that a PCI device couldn't
302 	 * directly DMA to any physical addresses.
303 	 * ISA DMA support will need to revisit this.
304 	 */
305 	max_low_pfn = max_pfn;
306 
307 	/* bootmap sizing messed up? */
308 	BUG_ON((bootmap_pfn - bootmap_start_pfn) != bootmap_pages);
309 
310 	/* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
311 
312 #define PDC_CONSOLE_IO_IODC_SIZE 32768
313 
314 	reserve_bootmem_node(NODE_DATA(0), 0UL,
315 			(unsigned long)(PAGE0->mem_free +
316 				PDC_CONSOLE_IO_IODC_SIZE), BOOTMEM_DEFAULT);
317 	reserve_bootmem_node(NODE_DATA(0), __pa((unsigned long)_text),
318 			(unsigned long)(_end - _text), BOOTMEM_DEFAULT);
319 	reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
320 			((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT),
321 			BOOTMEM_DEFAULT);
322 
323 #ifndef CONFIG_DISCONTIGMEM
324 
325 	/* reserve the holes */
326 
327 	for (i = 0; i < npmem_holes; i++) {
328 		reserve_bootmem_node(NODE_DATA(0),
329 				(pmem_holes[i].start_pfn << PAGE_SHIFT),
330 				(pmem_holes[i].pages << PAGE_SHIFT),
331 				BOOTMEM_DEFAULT);
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),
350 					initrd_reserve, BOOTMEM_DEFAULT);
351 		}
352 	}
353 #endif
354 
355 	data_resource.start =  virt_to_phys(&data_start);
356 	data_resource.end = virt_to_phys(_end) - 1;
357 	code_resource.start = virt_to_phys(_text);
358 	code_resource.end = virt_to_phys(&data_start)-1;
359 
360 	/* We don't know which region the kernel will be in, so try
361 	 * all of them.
362 	 */
363 	for (i = 0; i < sysram_resource_count; i++) {
364 		struct resource *res = &sysram_resources[i];
365 		request_resource(res, &code_resource);
366 		request_resource(res, &data_resource);
367 	}
368 	request_resource(&sysram_resources[0], &pdcdata_resource);
369 }
370 
371 void free_initmem(void)
372 {
373 	unsigned long addr, init_begin, init_end;
374 
375 	printk(KERN_INFO "Freeing unused kernel memory: ");
376 
377 #ifdef CONFIG_DEBUG_KERNEL
378 	/* Attempt to catch anyone trying to execute code here
379 	 * by filling the page with BRK insns.
380 	 *
381 	 * If we disable interrupts for all CPUs, then IPI stops working.
382 	 * Kinda breaks the global cache flushing.
383 	 */
384 	local_irq_disable();
385 
386 	memset(__init_begin, 0x00,
387 		(unsigned long)__init_end - (unsigned long)__init_begin);
388 
389 	flush_data_cache();
390 	asm volatile("sync" : : );
391 	flush_icache_range((unsigned long)__init_begin, (unsigned long)__init_end);
392 	asm volatile("sync" : : );
393 
394 	local_irq_enable();
395 #endif
396 
397 	/* align __init_begin and __init_end to page size,
398 	   ignoring linker script where we might have tried to save RAM */
399 	init_begin = PAGE_ALIGN((unsigned long)(__init_begin));
400 	init_end   = PAGE_ALIGN((unsigned long)(__init_end));
401 	for (addr = init_begin; addr < init_end; addr += PAGE_SIZE) {
402 		ClearPageReserved(virt_to_page(addr));
403 		init_page_count(virt_to_page(addr));
404 		free_page(addr);
405 		num_physpages++;
406 		totalram_pages++;
407 	}
408 
409 	/* set up a new led state on systems shipped LED State panel */
410 	pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
411 
412 	printk("%luk freed\n", (init_end - init_begin) >> 10);
413 }
414 
415 
416 #ifdef CONFIG_DEBUG_RODATA
417 void mark_rodata_ro(void)
418 {
419 	/* rodata memory was already mapped with KERNEL_RO access rights by
420            pagetable_init() and map_pages(). No need to do additional stuff here */
421 	printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
422 		(unsigned long)(__end_rodata - __start_rodata) >> 10);
423 }
424 #endif
425 
426 
427 /*
428  * Just an arbitrary offset to serve as a "hole" between mapping areas
429  * (between top of physical memory and a potential pcxl dma mapping
430  * area, and below the vmalloc mapping area).
431  *
432  * The current 32K value just means that there will be a 32K "hole"
433  * between mapping areas. That means that  any out-of-bounds memory
434  * accesses will hopefully be caught. The vmalloc() routines leaves
435  * a hole of 4kB between each vmalloced area for the same reason.
436  */
437 
438  /* Leave room for gateway page expansion */
439 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
440 #error KERNEL_MAP_START is in gateway reserved region
441 #endif
442 #define MAP_START (KERNEL_MAP_START)
443 
444 #define VM_MAP_OFFSET  (32*1024)
445 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
446 				     & ~(VM_MAP_OFFSET-1)))
447 
448 void *vmalloc_start __read_mostly;
449 EXPORT_SYMBOL(vmalloc_start);
450 
451 #ifdef CONFIG_PA11
452 unsigned long pcxl_dma_start __read_mostly;
453 #endif
454 
455 void __init mem_init(void)
456 {
457 	int codesize, reservedpages, datasize, initsize;
458 
459 	high_memory = __va((max_pfn << PAGE_SHIFT));
460 
461 #ifndef CONFIG_DISCONTIGMEM
462 	max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
463 	totalram_pages += free_all_bootmem();
464 #else
465 	{
466 		int i;
467 
468 		for (i = 0; i < npmem_ranges; i++)
469 			totalram_pages += free_all_bootmem_node(NODE_DATA(i));
470 	}
471 #endif
472 
473 	codesize = (unsigned long)_etext - (unsigned long)_text;
474 	datasize = (unsigned long)_edata - (unsigned long)_etext;
475 	initsize = (unsigned long)__init_end - (unsigned long)__init_begin;
476 
477 	reservedpages = 0;
478 {
479 	unsigned long pfn;
480 #ifdef CONFIG_DISCONTIGMEM
481 	int i;
482 
483 	for (i = 0; i < npmem_ranges; i++) {
484 		for (pfn = node_start_pfn(i); pfn < node_end_pfn(i); pfn++) {
485 			if (PageReserved(pfn_to_page(pfn)))
486 				reservedpages++;
487 		}
488 	}
489 #else /* !CONFIG_DISCONTIGMEM */
490 	for (pfn = 0; pfn < max_pfn; pfn++) {
491 		/*
492 		 * Only count reserved RAM pages
493 		 */
494 		if (PageReserved(pfn_to_page(pfn)))
495 			reservedpages++;
496 	}
497 #endif
498 }
499 
500 #ifdef CONFIG_PA11
501 	if (hppa_dma_ops == &pcxl_dma_ops) {
502 		pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
503 		vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
504 	} else {
505 		pcxl_dma_start = 0;
506 		vmalloc_start = SET_MAP_OFFSET(MAP_START);
507 	}
508 #else
509 	vmalloc_start = SET_MAP_OFFSET(MAP_START);
510 #endif
511 
512 	printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, %dk reserved, %dk data, %dk init)\n",
513 		(unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
514 		num_physpages << (PAGE_SHIFT-10),
515 		codesize >> 10,
516 		reservedpages << (PAGE_SHIFT-10),
517 		datasize >> 10,
518 		initsize >> 10
519 	);
520 
521 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
522 	printk("virtual kernel memory layout:\n"
523 	       "    vmalloc : 0x%p - 0x%p   (%4ld MB)\n"
524 	       "    memory  : 0x%p - 0x%p   (%4ld MB)\n"
525 	       "      .init : 0x%p - 0x%p   (%4ld kB)\n"
526 	       "      .data : 0x%p - 0x%p   (%4ld kB)\n"
527 	       "      .text : 0x%p - 0x%p   (%4ld kB)\n",
528 
529 	       (void*)VMALLOC_START, (void*)VMALLOC_END,
530 	       (VMALLOC_END - VMALLOC_START) >> 20,
531 
532 	       __va(0), high_memory,
533 	       ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
534 
535 	       __init_begin, __init_end,
536 	       ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
537 
538 	       _etext, _edata,
539 	       ((unsigned long)_edata - (unsigned long)_etext) >> 10,
540 
541 	       _text, _etext,
542 	       ((unsigned long)_etext - (unsigned long)_text) >> 10);
543 #endif
544 }
545 
546 unsigned long *empty_zero_page __read_mostly;
547 EXPORT_SYMBOL(empty_zero_page);
548 
549 void show_mem(void)
550 {
551 	int i,free = 0,total = 0,reserved = 0;
552 	int shared = 0, cached = 0;
553 
554 	printk(KERN_INFO "Mem-info:\n");
555 	show_free_areas();
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;
603 
604 		for (i = 0; i < npmem_ranges; i++) {
605 			zl = node_zonelist(i, 0);
606 			for (j = 0; j < MAX_NR_ZONES; j++) {
607 				struct zoneref *z;
608 				struct zone *zone;
609 
610 				printk("Zone list for zone %d on node %d: ", j, i);
611 				for_each_zone_zonelist(zone, z, zl, j)
612 					printk("[%d/%s] ", zone_to_nid(zone),
613 								zone->name);
614 				printk("\n");
615 			}
616 		}
617 	}
618 #endif
619 }
620 
621 
622 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
623 {
624 	pgd_t *pg_dir;
625 	pmd_t *pmd;
626 	pte_t *pg_table;
627 	unsigned long end_paddr;
628 	unsigned long start_pmd;
629 	unsigned long start_pte;
630 	unsigned long tmp1;
631 	unsigned long tmp2;
632 	unsigned long address;
633 	unsigned long ro_start;
634 	unsigned long ro_end;
635 	unsigned long fv_addr;
636 	unsigned long gw_addr;
637 	extern const unsigned long fault_vector_20;
638 	extern void * const linux_gateway_page;
639 
640 	ro_start = __pa((unsigned long)_text);
641 	ro_end   = __pa((unsigned long)&data_start);
642 	fv_addr  = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
643 	gw_addr  = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
644 
645 	end_paddr = start_paddr + size;
646 
647 	pg_dir = pgd_offset_k(start_vaddr);
648 
649 #if PTRS_PER_PMD == 1
650 	start_pmd = 0;
651 #else
652 	start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
653 #endif
654 	start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
655 
656 	address = start_paddr;
657 	while (address < end_paddr) {
658 #if PTRS_PER_PMD == 1
659 		pmd = (pmd_t *)__pa(pg_dir);
660 #else
661 		pmd = (pmd_t *)pgd_address(*pg_dir);
662 
663 		/*
664 		 * pmd is physical at this point
665 		 */
666 
667 		if (!pmd) {
668 			pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
669 			pmd = (pmd_t *) __pa(pmd);
670 		}
671 
672 		pgd_populate(NULL, pg_dir, __va(pmd));
673 #endif
674 		pg_dir++;
675 
676 		/* now change pmd to kernel virtual addresses */
677 
678 		pmd = (pmd_t *)__va(pmd) + start_pmd;
679 		for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
680 
681 			/*
682 			 * pg_table is physical at this point
683 			 */
684 
685 			pg_table = (pte_t *)pmd_address(*pmd);
686 			if (!pg_table) {
687 				pg_table = (pte_t *)
688 					alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
689 				pg_table = (pte_t *) __pa(pg_table);
690 			}
691 
692 			pmd_populate_kernel(NULL, pmd, __va(pg_table));
693 
694 			/* now change pg_table to kernel virtual addresses */
695 
696 			pg_table = (pte_t *) __va(pg_table) + start_pte;
697 			for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
698 				pte_t pte;
699 
700 				/*
701 				 * Map the fault vector writable so we can
702 				 * write the HPMC checksum.
703 				 */
704 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
705 				if (address >= ro_start && address < ro_end
706 							&& address != fv_addr
707 							&& address != gw_addr)
708 				    pte = __mk_pte(address, PAGE_KERNEL_RO);
709 				else
710 #endif
711 				    pte = __mk_pte(address, pgprot);
712 
713 				if (address >= end_paddr)
714 					pte_val(pte) = 0;
715 
716 				set_pte(pg_table, pte);
717 
718 				address += PAGE_SIZE;
719 			}
720 			start_pte = 0;
721 
722 			if (address >= end_paddr)
723 			    break;
724 		}
725 		start_pmd = 0;
726 	}
727 }
728 
729 /*
730  * pagetable_init() sets up the page tables
731  *
732  * Note that gateway_init() places the Linux gateway page at page 0.
733  * Since gateway pages cannot be dereferenced this has the desirable
734  * side effect of trapping those pesky NULL-reference errors in the
735  * kernel.
736  */
737 static void __init pagetable_init(void)
738 {
739 	int range;
740 
741 	/* Map each physical memory range to its kernel vaddr */
742 
743 	for (range = 0; range < npmem_ranges; range++) {
744 		unsigned long start_paddr;
745 		unsigned long end_paddr;
746 		unsigned long size;
747 
748 		start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
749 		end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
750 		size = pmem_ranges[range].pages << PAGE_SHIFT;
751 
752 		map_pages((unsigned long)__va(start_paddr), start_paddr,
753 			size, PAGE_KERNEL);
754 	}
755 
756 #ifdef CONFIG_BLK_DEV_INITRD
757 	if (initrd_end && initrd_end > mem_limit) {
758 		printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
759 		map_pages(initrd_start, __pa(initrd_start),
760 			initrd_end - initrd_start, PAGE_KERNEL);
761 	}
762 #endif
763 
764 	empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
765 	memset(empty_zero_page, 0, PAGE_SIZE);
766 }
767 
768 static void __init gateway_init(void)
769 {
770 	unsigned long linux_gateway_page_addr;
771 	/* FIXME: This is 'const' in order to trick the compiler
772 	   into not treating it as DP-relative data. */
773 	extern void * const linux_gateway_page;
774 
775 	linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
776 
777 	/*
778 	 * Setup Linux Gateway page.
779 	 *
780 	 * The Linux gateway page will reside in kernel space (on virtual
781 	 * page 0), so it doesn't need to be aliased into user space.
782 	 */
783 
784 	map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
785 		PAGE_SIZE, PAGE_GATEWAY);
786 }
787 
788 #ifdef CONFIG_HPUX
789 void
790 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
791 {
792 	pgd_t *pg_dir;
793 	pmd_t *pmd;
794 	pte_t *pg_table;
795 	unsigned long start_pmd;
796 	unsigned long start_pte;
797 	unsigned long address;
798 	unsigned long hpux_gw_page_addr;
799 	/* FIXME: This is 'const' in order to trick the compiler
800 	   into not treating it as DP-relative data. */
801 	extern void * const hpux_gateway_page;
802 
803 	hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
804 
805 	/*
806 	 * Setup HP-UX Gateway page.
807 	 *
808 	 * The HP-UX gateway page resides in the user address space,
809 	 * so it needs to be aliased into each process.
810 	 */
811 
812 	pg_dir = pgd_offset(mm,hpux_gw_page_addr);
813 
814 #if PTRS_PER_PMD == 1
815 	start_pmd = 0;
816 #else
817 	start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
818 #endif
819 	start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
820 
821 	address = __pa(&hpux_gateway_page);
822 #if PTRS_PER_PMD == 1
823 	pmd = (pmd_t *)__pa(pg_dir);
824 #else
825 	pmd = (pmd_t *) pgd_address(*pg_dir);
826 
827 	/*
828 	 * pmd is physical at this point
829 	 */
830 
831 	if (!pmd) {
832 		pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
833 		pmd = (pmd_t *) __pa(pmd);
834 	}
835 
836 	__pgd_val_set(*pg_dir, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pmd);
837 #endif
838 	/* now change pmd to kernel virtual addresses */
839 
840 	pmd = (pmd_t *)__va(pmd) + start_pmd;
841 
842 	/*
843 	 * pg_table is physical at this point
844 	 */
845 
846 	pg_table = (pte_t *) pmd_address(*pmd);
847 	if (!pg_table)
848 		pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
849 
850 	__pmd_val_set(*pmd, PxD_FLAG_PRESENT | PxD_FLAG_VALID | (unsigned long) pg_table);
851 
852 	/* now change pg_table to kernel virtual addresses */
853 
854 	pg_table = (pte_t *) __va(pg_table) + start_pte;
855 	set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
856 }
857 EXPORT_SYMBOL(map_hpux_gateway_page);
858 #endif
859 
860 void __init paging_init(void)
861 {
862 	int i;
863 
864 	setup_bootmem();
865 	pagetable_init();
866 	gateway_init();
867 	flush_cache_all_local(); /* start with known state */
868 	flush_tlb_all_local(NULL);
869 
870 	for (i = 0; i < npmem_ranges; i++) {
871 		unsigned long zones_size[MAX_NR_ZONES] = { 0, };
872 
873 		zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
874 
875 #ifdef CONFIG_DISCONTIGMEM
876 		/* Need to initialize the pfnnid_map before we can initialize
877 		   the zone */
878 		{
879 		    int j;
880 		    for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
881 			 j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
882 			 j++) {
883 			pfnnid_map[j] = i;
884 		    }
885 		}
886 #endif
887 
888 		free_area_init_node(i, zones_size,
889 				pmem_ranges[i].start_pfn, NULL);
890 	}
891 }
892 
893 #ifdef CONFIG_PA20
894 
895 /*
896  * Currently, all PA20 chips have 18 bit protection IDs, which is the
897  * limiting factor (space ids are 32 bits).
898  */
899 
900 #define NR_SPACE_IDS 262144
901 
902 #else
903 
904 /*
905  * Currently we have a one-to-one relationship between space IDs and
906  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
907  * support 15 bit protection IDs, so that is the limiting factor.
908  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
909  * probably not worth the effort for a special case here.
910  */
911 
912 #define NR_SPACE_IDS 32768
913 
914 #endif  /* !CONFIG_PA20 */
915 
916 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
917 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
918 
919 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
920 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
921 static unsigned long space_id_index;
922 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
923 static unsigned long dirty_space_ids = 0;
924 
925 static DEFINE_SPINLOCK(sid_lock);
926 
927 unsigned long alloc_sid(void)
928 {
929 	unsigned long index;
930 
931 	spin_lock(&sid_lock);
932 
933 	if (free_space_ids == 0) {
934 		if (dirty_space_ids != 0) {
935 			spin_unlock(&sid_lock);
936 			flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
937 			spin_lock(&sid_lock);
938 		}
939 		BUG_ON(free_space_ids == 0);
940 	}
941 
942 	free_space_ids--;
943 
944 	index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
945 	space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
946 	space_id_index = index;
947 
948 	spin_unlock(&sid_lock);
949 
950 	return index << SPACEID_SHIFT;
951 }
952 
953 void free_sid(unsigned long spaceid)
954 {
955 	unsigned long index = spaceid >> SPACEID_SHIFT;
956 	unsigned long *dirty_space_offset;
957 
958 	dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
959 	index &= (BITS_PER_LONG - 1);
960 
961 	spin_lock(&sid_lock);
962 
963 	BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
964 
965 	*dirty_space_offset |= (1L << index);
966 	dirty_space_ids++;
967 
968 	spin_unlock(&sid_lock);
969 }
970 
971 
972 #ifdef CONFIG_SMP
973 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
974 {
975 	int i;
976 
977 	/* NOTE: sid_lock must be held upon entry */
978 
979 	*ndirtyptr = dirty_space_ids;
980 	if (dirty_space_ids != 0) {
981 	    for (i = 0; i < SID_ARRAY_SIZE; i++) {
982 		dirty_array[i] = dirty_space_id[i];
983 		dirty_space_id[i] = 0;
984 	    }
985 	    dirty_space_ids = 0;
986 	}
987 
988 	return;
989 }
990 
991 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
992 {
993 	int i;
994 
995 	/* NOTE: sid_lock must be held upon entry */
996 
997 	if (ndirty != 0) {
998 		for (i = 0; i < SID_ARRAY_SIZE; i++) {
999 			space_id[i] ^= dirty_array[i];
1000 		}
1001 
1002 		free_space_ids += ndirty;
1003 		space_id_index = 0;
1004 	}
1005 }
1006 
1007 #else /* CONFIG_SMP */
1008 
1009 static void recycle_sids(void)
1010 {
1011 	int i;
1012 
1013 	/* NOTE: sid_lock must be held upon entry */
1014 
1015 	if (dirty_space_ids != 0) {
1016 		for (i = 0; i < SID_ARRAY_SIZE; i++) {
1017 			space_id[i] ^= dirty_space_id[i];
1018 			dirty_space_id[i] = 0;
1019 		}
1020 
1021 		free_space_ids += dirty_space_ids;
1022 		dirty_space_ids = 0;
1023 		space_id_index = 0;
1024 	}
1025 }
1026 #endif
1027 
1028 /*
1029  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
1030  * purged, we can safely reuse the space ids that were released but
1031  * not flushed from the tlb.
1032  */
1033 
1034 #ifdef CONFIG_SMP
1035 
1036 static unsigned long recycle_ndirty;
1037 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
1038 static unsigned int recycle_inuse;
1039 
1040 void flush_tlb_all(void)
1041 {
1042 	int do_recycle;
1043 
1044 	do_recycle = 0;
1045 	spin_lock(&sid_lock);
1046 	if (dirty_space_ids > RECYCLE_THRESHOLD) {
1047 	    BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
1048 	    get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
1049 	    recycle_inuse++;
1050 	    do_recycle++;
1051 	}
1052 	spin_unlock(&sid_lock);
1053 	on_each_cpu(flush_tlb_all_local, NULL, 1);
1054 	if (do_recycle) {
1055 	    spin_lock(&sid_lock);
1056 	    recycle_sids(recycle_ndirty,recycle_dirty_array);
1057 	    recycle_inuse = 0;
1058 	    spin_unlock(&sid_lock);
1059 	}
1060 }
1061 #else
1062 void flush_tlb_all(void)
1063 {
1064 	spin_lock(&sid_lock);
1065 	flush_tlb_all_local(NULL);
1066 	recycle_sids();
1067 	spin_unlock(&sid_lock);
1068 }
1069 #endif
1070 
1071 #ifdef CONFIG_BLK_DEV_INITRD
1072 void free_initrd_mem(unsigned long start, unsigned long end)
1073 {
1074 	if (start >= end)
1075 		return;
1076 	printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1077 	for (; start < end; start += PAGE_SIZE) {
1078 		ClearPageReserved(virt_to_page(start));
1079 		init_page_count(virt_to_page(start));
1080 		free_page(start);
1081 		num_physpages++;
1082 		totalram_pages++;
1083 	}
1084 }
1085 #endif
1086