xref: /kernel/linux/linux-5.10/arch/parisc/mm/init.c (revision 8c2ecf20)
1// SPDX-License-Identifier: GPL-2.0
2/*
3 *  linux/arch/parisc/mm/init.c
4 *
5 *  Copyright (C) 1995	Linus Torvalds
6 *  Copyright 1999 SuSE GmbH
7 *    changed by Philipp Rumpf
8 *  Copyright 1999 Philipp Rumpf (prumpf@tux.org)
9 *  Copyright 2004 Randolph Chung (tausq@debian.org)
10 *  Copyright 2006-2007 Helge Deller (deller@gmx.de)
11 *
12 */
13
14
15#include <linux/module.h>
16#include <linux/mm.h>
17#include <linux/memblock.h>
18#include <linux/gfp.h>
19#include <linux/delay.h>
20#include <linux/init.h>
21#include <linux/initrd.h>
22#include <linux/swap.h>
23#include <linux/unistd.h>
24#include <linux/nodemask.h>	/* for node_online_map */
25#include <linux/pagemap.h>	/* for release_pages */
26#include <linux/compat.h>
27
28#include <asm/pgalloc.h>
29#include <asm/tlb.h>
30#include <asm/pdc_chassis.h>
31#include <asm/mmzone.h>
32#include <asm/sections.h>
33#include <asm/msgbuf.h>
34#include <asm/sparsemem.h>
35
36extern int  data_start;
37extern void parisc_kernel_start(void);	/* Kernel entry point in head.S */
38
39#if CONFIG_PGTABLE_LEVELS == 3
40pmd_t pmd0[PTRS_PER_PMD] __section(".data..vm0.pmd") __attribute__ ((aligned(PAGE_SIZE)));
41#endif
42
43pgd_t swapper_pg_dir[PTRS_PER_PGD] __section(".data..vm0.pgd") __attribute__ ((aligned(PAGE_SIZE)));
44pte_t pg0[PT_INITIAL * PTRS_PER_PTE] __section(".data..vm0.pte") __attribute__ ((aligned(PAGE_SIZE)));
45
46static struct resource data_resource = {
47	.name	= "Kernel data",
48	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
49};
50
51static struct resource code_resource = {
52	.name	= "Kernel code",
53	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
54};
55
56static struct resource pdcdata_resource = {
57	.name	= "PDC data (Page Zero)",
58	.start	= 0,
59	.end	= 0x9ff,
60	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM,
61};
62
63static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __ro_after_init;
64
65/* The following array is initialized from the firmware specific
66 * information retrieved in kernel/inventory.c.
67 */
68
69physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __initdata;
70int npmem_ranges __initdata;
71
72#ifdef CONFIG_64BIT
73#define MAX_MEM         (1UL << MAX_PHYSMEM_BITS)
74#else /* !CONFIG_64BIT */
75#define MAX_MEM         (3584U*1024U*1024U)
76#endif /* !CONFIG_64BIT */
77
78static unsigned long mem_limit __read_mostly = MAX_MEM;
79
80static void __init mem_limit_func(void)
81{
82	char *cp, *end;
83	unsigned long limit;
84
85	/* We need this before __setup() functions are called */
86
87	limit = MAX_MEM;
88	for (cp = boot_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
109static void __init setup_bootmem(void)
110{
111	unsigned long mem_max;
112#ifndef CONFIG_SPARSEMEM
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			physmem_range_t 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];
138			pmem_ranges[j-1] = pmem_ranges[j];
139			pmem_ranges[j] = tmp;
140		}
141	}
142
143#ifndef CONFIG_SPARSEMEM
144	/*
145	 * Throw out ranges that are too far apart (controlled by
146	 * MAX_GAP).
147	 */
148
149	for (i = 1; i < npmem_ranges; i++) {
150		if (pmem_ranges[i].start_pfn -
151			(pmem_ranges[i-1].start_pfn +
152			 pmem_ranges[i-1].pages) > MAX_GAP) {
153			npmem_ranges = i;
154			printk("Large gap in memory detected (%ld pages). "
155			       "Consider turning on CONFIG_SPARSEMEM\n",
156			       pmem_ranges[i].start_pfn -
157			       (pmem_ranges[i-1].start_pfn +
158			        pmem_ranges[i-1].pages));
159			break;
160		}
161	}
162#endif
163
164	/* Print the memory ranges */
165	pr_info("Memory Ranges:\n");
166
167	for (i = 0; i < npmem_ranges; i++) {
168		struct resource *res = &sysram_resources[i];
169		unsigned long start;
170		unsigned long size;
171
172		size = (pmem_ranges[i].pages << PAGE_SHIFT);
173		start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
174		pr_info("%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
175			i, start, start + (size - 1), size >> 20);
176
177		/* request memory resource */
178		res->name = "System RAM";
179		res->start = start;
180		res->end = start + size - 1;
181		res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
182		request_resource(&iomem_resource, res);
183	}
184
185	sysram_resource_count = npmem_ranges;
186
187	/*
188	 * For 32 bit kernels we limit the amount of memory we can
189	 * support, in order to preserve enough kernel address space
190	 * for other purposes. For 64 bit kernels we don't normally
191	 * limit the memory, but this mechanism can be used to
192	 * artificially limit the amount of memory (and it is written
193	 * to work with multiple memory ranges).
194	 */
195
196	mem_limit_func();       /* check for "mem=" argument */
197
198	mem_max = 0;
199	for (i = 0; i < npmem_ranges; i++) {
200		unsigned long rsize;
201
202		rsize = pmem_ranges[i].pages << PAGE_SHIFT;
203		if ((mem_max + rsize) > mem_limit) {
204			printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
205			if (mem_max == mem_limit)
206				npmem_ranges = i;
207			else {
208				pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
209						       - (mem_max >> PAGE_SHIFT);
210				npmem_ranges = i + 1;
211				mem_max = mem_limit;
212			}
213			break;
214		}
215		mem_max += rsize;
216	}
217
218	printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
219
220#ifndef CONFIG_SPARSEMEM
221	/* Merge the ranges, keeping track of the holes */
222	{
223		unsigned long end_pfn;
224		unsigned long hole_pages;
225
226		npmem_holes = 0;
227		end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
228		for (i = 1; i < npmem_ranges; i++) {
229
230			hole_pages = pmem_ranges[i].start_pfn - end_pfn;
231			if (hole_pages) {
232				pmem_holes[npmem_holes].start_pfn = end_pfn;
233				pmem_holes[npmem_holes++].pages = hole_pages;
234				end_pfn += hole_pages;
235			}
236			end_pfn += pmem_ranges[i].pages;
237		}
238
239		pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
240		npmem_ranges = 1;
241	}
242#endif
243
244	/*
245	 * Initialize and free the full range of memory in each range.
246	 */
247
248	max_pfn = 0;
249	for (i = 0; i < npmem_ranges; i++) {
250		unsigned long start_pfn;
251		unsigned long npages;
252		unsigned long start;
253		unsigned long size;
254
255		start_pfn = pmem_ranges[i].start_pfn;
256		npages = pmem_ranges[i].pages;
257
258		start = start_pfn << PAGE_SHIFT;
259		size = npages << PAGE_SHIFT;
260
261		/* add system RAM memblock */
262		memblock_add(start, size);
263
264		if ((start_pfn + npages) > max_pfn)
265			max_pfn = start_pfn + npages;
266	}
267
268	/*
269	 * We can't use memblock top-down allocations because we only
270	 * created the initial mapping up to KERNEL_INITIAL_SIZE in
271	 * the assembly bootup code.
272	 */
273	memblock_set_bottom_up(true);
274
275	/* IOMMU is always used to access "high mem" on those boxes
276	 * that can support enough mem that a PCI device couldn't
277	 * directly DMA to any physical addresses.
278	 * ISA DMA support will need to revisit this.
279	 */
280	max_low_pfn = max_pfn;
281
282	/* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
283
284#define PDC_CONSOLE_IO_IODC_SIZE 32768
285
286	memblock_reserve(0UL, (unsigned long)(PAGE0->mem_free +
287				PDC_CONSOLE_IO_IODC_SIZE));
288	memblock_reserve(__pa(KERNEL_BINARY_TEXT_START),
289			(unsigned long)(_end - KERNEL_BINARY_TEXT_START));
290
291#ifndef CONFIG_SPARSEMEM
292
293	/* reserve the holes */
294
295	for (i = 0; i < npmem_holes; i++) {
296		memblock_reserve((pmem_holes[i].start_pfn << PAGE_SHIFT),
297				(pmem_holes[i].pages << PAGE_SHIFT));
298	}
299#endif
300
301#ifdef CONFIG_BLK_DEV_INITRD
302	if (initrd_start) {
303		printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
304		if (__pa(initrd_start) < mem_max) {
305			unsigned long initrd_reserve;
306
307			if (__pa(initrd_end) > mem_max) {
308				initrd_reserve = mem_max - __pa(initrd_start);
309			} else {
310				initrd_reserve = initrd_end - initrd_start;
311			}
312			initrd_below_start_ok = 1;
313			printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
314
315			memblock_reserve(__pa(initrd_start), initrd_reserve);
316		}
317	}
318#endif
319
320	data_resource.start =  virt_to_phys(&data_start);
321	data_resource.end = virt_to_phys(_end) - 1;
322	code_resource.start = virt_to_phys(_text);
323	code_resource.end = virt_to_phys(&data_start)-1;
324
325	/* We don't know which region the kernel will be in, so try
326	 * all of them.
327	 */
328	for (i = 0; i < sysram_resource_count; i++) {
329		struct resource *res = &sysram_resources[i];
330		request_resource(res, &code_resource);
331		request_resource(res, &data_resource);
332	}
333	request_resource(&sysram_resources[0], &pdcdata_resource);
334
335	/* Initialize Page Deallocation Table (PDT) and check for bad memory. */
336	pdc_pdt_init();
337
338	memblock_allow_resize();
339	memblock_dump_all();
340}
341
342static bool kernel_set_to_readonly;
343
344static void __ref map_pages(unsigned long start_vaddr,
345			    unsigned long start_paddr, unsigned long size,
346			    pgprot_t pgprot, int force)
347{
348	pmd_t *pmd;
349	pte_t *pg_table;
350	unsigned long end_paddr;
351	unsigned long start_pmd;
352	unsigned long start_pte;
353	unsigned long tmp1;
354	unsigned long tmp2;
355	unsigned long address;
356	unsigned long vaddr;
357	unsigned long ro_start;
358	unsigned long ro_end;
359	unsigned long kernel_start, kernel_end;
360
361	ro_start = __pa((unsigned long)_text);
362	ro_end   = __pa((unsigned long)&data_start);
363	kernel_start = __pa((unsigned long)&__init_begin);
364	kernel_end  = __pa((unsigned long)&_end);
365
366	end_paddr = start_paddr + size;
367
368	/* for 2-level configuration PTRS_PER_PMD is 0 so start_pmd will be 0 */
369	start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
370	start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
371
372	address = start_paddr;
373	vaddr = start_vaddr;
374	while (address < end_paddr) {
375		pgd_t *pgd = pgd_offset_k(vaddr);
376		p4d_t *p4d = p4d_offset(pgd, vaddr);
377		pud_t *pud = pud_offset(p4d, vaddr);
378
379#if CONFIG_PGTABLE_LEVELS == 3
380		if (pud_none(*pud)) {
381			pmd = memblock_alloc(PAGE_SIZE << PMD_ORDER,
382					     PAGE_SIZE << PMD_ORDER);
383			if (!pmd)
384				panic("pmd allocation failed.\n");
385			pud_populate(NULL, pud, pmd);
386		}
387#endif
388
389		pmd = pmd_offset(pud, vaddr);
390		for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
391			if (pmd_none(*pmd)) {
392				pg_table = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
393				if (!pg_table)
394					panic("page table allocation failed\n");
395				pmd_populate_kernel(NULL, pmd, pg_table);
396			}
397
398			pg_table = pte_offset_kernel(pmd, vaddr);
399			for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
400				pte_t pte;
401				pgprot_t prot;
402				bool huge = false;
403
404				if (force) {
405					prot = pgprot;
406				} else if (address < kernel_start || address >= kernel_end) {
407					/* outside kernel memory */
408					prot = PAGE_KERNEL;
409				} else if (!kernel_set_to_readonly) {
410					/* still initializing, allow writing to RO memory */
411					prot = PAGE_KERNEL_RWX;
412					huge = true;
413				} else if (address >= ro_start) {
414					/* Code (ro) and Data areas */
415					prot = (address < ro_end) ?
416						PAGE_KERNEL_EXEC : PAGE_KERNEL;
417					huge = true;
418				} else {
419					prot = PAGE_KERNEL;
420				}
421
422				pte = __mk_pte(address, prot);
423				if (huge)
424					pte = pte_mkhuge(pte);
425
426				if (address >= end_paddr)
427					break;
428
429				set_pte(pg_table, pte);
430
431				address += PAGE_SIZE;
432				vaddr += PAGE_SIZE;
433			}
434			start_pte = 0;
435
436			if (address >= end_paddr)
437			    break;
438		}
439		start_pmd = 0;
440	}
441}
442
443void __init set_kernel_text_rw(int enable_read_write)
444{
445	unsigned long start = (unsigned long) __init_begin;
446	unsigned long end   = (unsigned long) &data_start;
447
448	map_pages(start, __pa(start), end-start,
449		PAGE_KERNEL_RWX, enable_read_write ? 1:0);
450
451	/* force the kernel to see the new page table entries */
452	flush_cache_all();
453	flush_tlb_all();
454}
455
456void free_initmem(void)
457{
458	unsigned long init_begin = (unsigned long)__init_begin;
459	unsigned long init_end = (unsigned long)__init_end;
460	unsigned long kernel_end  = (unsigned long)&_end;
461
462	/* Remap kernel text and data, but do not touch init section yet. */
463	kernel_set_to_readonly = true;
464	map_pages(init_end, __pa(init_end), kernel_end - init_end,
465		  PAGE_KERNEL, 0);
466
467	/* The init text pages are marked R-X.  We have to
468	 * flush the icache and mark them RW-
469	 *
470	 * Do a dummy remap of the data section first (the data
471	 * section is already PAGE_KERNEL) to pull in the TLB entries
472	 * for map_kernel */
473	map_pages(init_begin, __pa(init_begin), init_end - init_begin,
474		  PAGE_KERNEL_RWX, 1);
475	/* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
476	 * map_pages */
477	map_pages(init_begin, __pa(init_begin), init_end - init_begin,
478		  PAGE_KERNEL, 1);
479
480	/* force the kernel to see the new TLB entries */
481	__flush_tlb_range(0, init_begin, kernel_end);
482
483	/* finally dump all the instructions which were cached, since the
484	 * pages are no-longer executable */
485	flush_icache_range(init_begin, init_end);
486
487	free_initmem_default(POISON_FREE_INITMEM);
488
489	/* set up a new led state on systems shipped LED State panel */
490	pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
491}
492
493
494#ifdef CONFIG_STRICT_KERNEL_RWX
495void mark_rodata_ro(void)
496{
497	/* rodata memory was already mapped with KERNEL_RO access rights by
498           pagetable_init() and map_pages(). No need to do additional stuff here */
499	unsigned long roai_size = __end_ro_after_init - __start_ro_after_init;
500
501	pr_info("Write protected read-only-after-init data: %luk\n", roai_size >> 10);
502}
503#endif
504
505
506/*
507 * Just an arbitrary offset to serve as a "hole" between mapping areas
508 * (between top of physical memory and a potential pcxl dma mapping
509 * area, and below the vmalloc mapping area).
510 *
511 * The current 32K value just means that there will be a 32K "hole"
512 * between mapping areas. That means that  any out-of-bounds memory
513 * accesses will hopefully be caught. The vmalloc() routines leaves
514 * a hole of 4kB between each vmalloced area for the same reason.
515 */
516
517 /* Leave room for gateway page expansion */
518#if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
519#error KERNEL_MAP_START is in gateway reserved region
520#endif
521#define MAP_START (KERNEL_MAP_START)
522
523#define VM_MAP_OFFSET  (32*1024)
524#define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
525				     & ~(VM_MAP_OFFSET-1)))
526
527void *parisc_vmalloc_start __ro_after_init;
528EXPORT_SYMBOL(parisc_vmalloc_start);
529
530#ifdef CONFIG_PA11
531unsigned long pcxl_dma_start __ro_after_init;
532#endif
533
534void __init mem_init(void)
535{
536	/* Do sanity checks on IPC (compat) structures */
537	BUILD_BUG_ON(sizeof(struct ipc64_perm) != 48);
538#ifndef CONFIG_64BIT
539	BUILD_BUG_ON(sizeof(struct semid64_ds) != 80);
540	BUILD_BUG_ON(sizeof(struct msqid64_ds) != 104);
541	BUILD_BUG_ON(sizeof(struct shmid64_ds) != 104);
542#endif
543#ifdef CONFIG_COMPAT
544	BUILD_BUG_ON(sizeof(struct compat_ipc64_perm) != sizeof(struct ipc64_perm));
545	BUILD_BUG_ON(sizeof(struct compat_semid64_ds) != 80);
546	BUILD_BUG_ON(sizeof(struct compat_msqid64_ds) != 104);
547	BUILD_BUG_ON(sizeof(struct compat_shmid64_ds) != 104);
548#endif
549
550	/* Do sanity checks on page table constants */
551	BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
552	BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
553	BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
554	BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
555			> BITS_PER_LONG);
556#if CONFIG_PGTABLE_LEVELS == 3
557	BUILD_BUG_ON(PT_INITIAL > PTRS_PER_PMD);
558#else
559	BUILD_BUG_ON(PT_INITIAL > PTRS_PER_PGD);
560#endif
561
562	high_memory = __va((max_pfn << PAGE_SHIFT));
563	set_max_mapnr(max_low_pfn);
564	memblock_free_all();
565
566#ifdef CONFIG_PA11
567	if (boot_cpu_data.cpu_type == pcxl2 || boot_cpu_data.cpu_type == pcxl) {
568		pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
569		parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
570						+ PCXL_DMA_MAP_SIZE);
571	} else
572#endif
573		parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
574
575	mem_init_print_info(NULL);
576
577#if 0
578	/*
579	 * Do not expose the virtual kernel memory layout to userspace.
580	 * But keep code for debugging purposes.
581	 */
582	printk("virtual kernel memory layout:\n"
583	       "     vmalloc : 0x%px - 0x%px   (%4ld MB)\n"
584	       "     fixmap  : 0x%px - 0x%px   (%4ld kB)\n"
585	       "     memory  : 0x%px - 0x%px   (%4ld MB)\n"
586	       "       .init : 0x%px - 0x%px   (%4ld kB)\n"
587	       "       .data : 0x%px - 0x%px   (%4ld kB)\n"
588	       "       .text : 0x%px - 0x%px   (%4ld kB)\n",
589
590	       (void*)VMALLOC_START, (void*)VMALLOC_END,
591	       (VMALLOC_END - VMALLOC_START) >> 20,
592
593	       (void *)FIXMAP_START, (void *)(FIXMAP_START + FIXMAP_SIZE),
594	       (unsigned long)(FIXMAP_SIZE / 1024),
595
596	       __va(0), high_memory,
597	       ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
598
599	       __init_begin, __init_end,
600	       ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
601
602	       _etext, _edata,
603	       ((unsigned long)_edata - (unsigned long)_etext) >> 10,
604
605	       _text, _etext,
606	       ((unsigned long)_etext - (unsigned long)_text) >> 10);
607#endif
608}
609
610unsigned long *empty_zero_page __ro_after_init;
611EXPORT_SYMBOL(empty_zero_page);
612
613/*
614 * pagetable_init() sets up the page tables
615 *
616 * Note that gateway_init() places the Linux gateway page at page 0.
617 * Since gateway pages cannot be dereferenced this has the desirable
618 * side effect of trapping those pesky NULL-reference errors in the
619 * kernel.
620 */
621static void __init pagetable_init(void)
622{
623	int range;
624
625	/* Map each physical memory range to its kernel vaddr */
626
627	for (range = 0; range < npmem_ranges; range++) {
628		unsigned long start_paddr;
629		unsigned long end_paddr;
630		unsigned long size;
631
632		start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
633		size = pmem_ranges[range].pages << PAGE_SHIFT;
634		end_paddr = start_paddr + size;
635
636		map_pages((unsigned long)__va(start_paddr), start_paddr,
637			  size, PAGE_KERNEL, 0);
638	}
639
640#ifdef CONFIG_BLK_DEV_INITRD
641	if (initrd_end && initrd_end > mem_limit) {
642		printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
643		map_pages(initrd_start, __pa(initrd_start),
644			  initrd_end - initrd_start, PAGE_KERNEL, 0);
645	}
646#endif
647
648	empty_zero_page = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
649	if (!empty_zero_page)
650		panic("zero page allocation failed.\n");
651
652}
653
654static void __init gateway_init(void)
655{
656	unsigned long linux_gateway_page_addr;
657	/* FIXME: This is 'const' in order to trick the compiler
658	   into not treating it as DP-relative data. */
659	extern void * const linux_gateway_page;
660
661	linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
662
663	/*
664	 * Setup Linux Gateway page.
665	 *
666	 * The Linux gateway page will reside in kernel space (on virtual
667	 * page 0), so it doesn't need to be aliased into user space.
668	 */
669
670	map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
671		  PAGE_SIZE, PAGE_GATEWAY, 1);
672}
673
674static void __init parisc_bootmem_free(void)
675{
676	unsigned long max_zone_pfn[MAX_NR_ZONES] = { 0, };
677
678	max_zone_pfn[0] = memblock_end_of_DRAM();
679
680	free_area_init(max_zone_pfn);
681}
682
683void __init paging_init(void)
684{
685	setup_bootmem();
686	pagetable_init();
687	gateway_init();
688	flush_cache_all_local(); /* start with known state */
689	flush_tlb_all_local(NULL);
690
691	sparse_init();
692	parisc_bootmem_free();
693}
694
695#ifdef CONFIG_PA20
696
697/*
698 * Currently, all PA20 chips have 18 bit protection IDs, which is the
699 * limiting factor (space ids are 32 bits).
700 */
701
702#define NR_SPACE_IDS 262144
703
704#else
705
706/*
707 * Currently we have a one-to-one relationship between space IDs and
708 * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
709 * support 15 bit protection IDs, so that is the limiting factor.
710 * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
711 * probably not worth the effort for a special case here.
712 */
713
714#define NR_SPACE_IDS 32768
715
716#endif  /* !CONFIG_PA20 */
717
718#define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
719#define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
720
721static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
722static unsigned long dirty_space_id[SID_ARRAY_SIZE];
723static unsigned long space_id_index;
724static unsigned long free_space_ids = NR_SPACE_IDS - 1;
725static unsigned long dirty_space_ids = 0;
726
727static DEFINE_SPINLOCK(sid_lock);
728
729unsigned long alloc_sid(void)
730{
731	unsigned long index;
732
733	spin_lock(&sid_lock);
734
735	if (free_space_ids == 0) {
736		if (dirty_space_ids != 0) {
737			spin_unlock(&sid_lock);
738			flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
739			spin_lock(&sid_lock);
740		}
741		BUG_ON(free_space_ids == 0);
742	}
743
744	free_space_ids--;
745
746	index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
747	space_id[BIT_WORD(index)] |= BIT_MASK(index);
748	space_id_index = index;
749
750	spin_unlock(&sid_lock);
751
752	return index << SPACEID_SHIFT;
753}
754
755void free_sid(unsigned long spaceid)
756{
757	unsigned long index = spaceid >> SPACEID_SHIFT;
758	unsigned long *dirty_space_offset, mask;
759
760	dirty_space_offset = &dirty_space_id[BIT_WORD(index)];
761	mask = BIT_MASK(index);
762
763	spin_lock(&sid_lock);
764
765	BUG_ON(*dirty_space_offset & mask); /* attempt to free space id twice */
766
767	*dirty_space_offset |= mask;
768	dirty_space_ids++;
769
770	spin_unlock(&sid_lock);
771}
772
773
774#ifdef CONFIG_SMP
775static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
776{
777	int i;
778
779	/* NOTE: sid_lock must be held upon entry */
780
781	*ndirtyptr = dirty_space_ids;
782	if (dirty_space_ids != 0) {
783	    for (i = 0; i < SID_ARRAY_SIZE; i++) {
784		dirty_array[i] = dirty_space_id[i];
785		dirty_space_id[i] = 0;
786	    }
787	    dirty_space_ids = 0;
788	}
789
790	return;
791}
792
793static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
794{
795	int i;
796
797	/* NOTE: sid_lock must be held upon entry */
798
799	if (ndirty != 0) {
800		for (i = 0; i < SID_ARRAY_SIZE; i++) {
801			space_id[i] ^= dirty_array[i];
802		}
803
804		free_space_ids += ndirty;
805		space_id_index = 0;
806	}
807}
808
809#else /* CONFIG_SMP */
810
811static void recycle_sids(void)
812{
813	int i;
814
815	/* NOTE: sid_lock must be held upon entry */
816
817	if (dirty_space_ids != 0) {
818		for (i = 0; i < SID_ARRAY_SIZE; i++) {
819			space_id[i] ^= dirty_space_id[i];
820			dirty_space_id[i] = 0;
821		}
822
823		free_space_ids += dirty_space_ids;
824		dirty_space_ids = 0;
825		space_id_index = 0;
826	}
827}
828#endif
829
830/*
831 * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
832 * purged, we can safely reuse the space ids that were released but
833 * not flushed from the tlb.
834 */
835
836#ifdef CONFIG_SMP
837
838static unsigned long recycle_ndirty;
839static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
840static unsigned int recycle_inuse;
841
842void flush_tlb_all(void)
843{
844	int do_recycle;
845
846	do_recycle = 0;
847	spin_lock(&sid_lock);
848	__inc_irq_stat(irq_tlb_count);
849	if (dirty_space_ids > RECYCLE_THRESHOLD) {
850	    BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
851	    get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
852	    recycle_inuse++;
853	    do_recycle++;
854	}
855	spin_unlock(&sid_lock);
856	on_each_cpu(flush_tlb_all_local, NULL, 1);
857	if (do_recycle) {
858	    spin_lock(&sid_lock);
859	    recycle_sids(recycle_ndirty,recycle_dirty_array);
860	    recycle_inuse = 0;
861	    spin_unlock(&sid_lock);
862	}
863}
864#else
865void flush_tlb_all(void)
866{
867	spin_lock(&sid_lock);
868	__inc_irq_stat(irq_tlb_count);
869	flush_tlb_all_local(NULL);
870	recycle_sids();
871	spin_unlock(&sid_lock);
872}
873#endif
874