xref: /linux/arch/parisc/mm/init.c (revision be709d48329a500621d2a05835283150ae137b45)
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/pgtable.h>
30 #include <asm/tlb.h>
31 #include <asm/pdc_chassis.h>
32 #include <asm/mmzone.h>
33 #include <asm/sections.h>
34 #include <asm/msgbuf.h>
35 
36 extern int  data_start;
37 extern void parisc_kernel_start(void);	/* Kernel entry point in head.S */
38 
39 #if CONFIG_PGTABLE_LEVELS == 3
40 /* NOTE: This layout exactly conforms to the hybrid L2/L3 page table layout
41  * with the first pmd adjacent to the pgd and below it. gcc doesn't actually
42  * guarantee that global objects will be laid out in memory in the same order
43  * as the order of declaration, so put these in different sections and use
44  * the linker script to order them. */
45 pmd_t pmd0[PTRS_PER_PMD] __attribute__ ((__section__ (".data..vm0.pmd"), aligned(PAGE_SIZE)));
46 #endif
47 
48 pgd_t swapper_pg_dir[PTRS_PER_PGD] __attribute__ ((__section__ (".data..vm0.pgd"), aligned(PAGE_SIZE)));
49 pte_t pg0[PT_INITIAL * PTRS_PER_PTE] __attribute__ ((__section__ (".data..vm0.pte"), aligned(PAGE_SIZE)));
50 
51 #ifdef CONFIG_DISCONTIGMEM
52 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
53 signed char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
54 #endif
55 
56 static struct resource data_resource = {
57 	.name	= "Kernel data",
58 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
59 };
60 
61 static struct resource code_resource = {
62 	.name	= "Kernel code",
63 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
64 };
65 
66 static struct resource pdcdata_resource = {
67 	.name	= "PDC data (Page Zero)",
68 	.start	= 0,
69 	.end	= 0x9ff,
70 	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM,
71 };
72 
73 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
74 
75 /* The following array is initialized from the firmware specific
76  * information retrieved in kernel/inventory.c.
77  */
78 
79 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
80 int npmem_ranges __read_mostly;
81 
82 #ifdef CONFIG_64BIT
83 #define MAX_MEM         (~0UL)
84 #else /* !CONFIG_64BIT */
85 #define MAX_MEM         (3584U*1024U*1024U)
86 #endif /* !CONFIG_64BIT */
87 
88 static unsigned long mem_limit __read_mostly = MAX_MEM;
89 
90 static void __init mem_limit_func(void)
91 {
92 	char *cp, *end;
93 	unsigned long limit;
94 
95 	/* We need this before __setup() functions are called */
96 
97 	limit = MAX_MEM;
98 	for (cp = boot_command_line; *cp; ) {
99 		if (memcmp(cp, "mem=", 4) == 0) {
100 			cp += 4;
101 			limit = memparse(cp, &end);
102 			if (end != cp)
103 				break;
104 			cp = end;
105 		} else {
106 			while (*cp != ' ' && *cp)
107 				++cp;
108 			while (*cp == ' ')
109 				++cp;
110 		}
111 	}
112 
113 	if (limit < mem_limit)
114 		mem_limit = limit;
115 }
116 
117 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
118 
119 static void __init setup_bootmem(void)
120 {
121 	unsigned long mem_max;
122 #ifndef CONFIG_DISCONTIGMEM
123 	physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
124 	int npmem_holes;
125 #endif
126 	int i, sysram_resource_count;
127 
128 	disable_sr_hashing(); /* Turn off space register hashing */
129 
130 	/*
131 	 * Sort the ranges. Since the number of ranges is typically
132 	 * small, and performance is not an issue here, just do
133 	 * a simple insertion sort.
134 	 */
135 
136 	for (i = 1; i < npmem_ranges; i++) {
137 		int j;
138 
139 		for (j = i; j > 0; j--) {
140 			unsigned long tmp;
141 
142 			if (pmem_ranges[j-1].start_pfn <
143 			    pmem_ranges[j].start_pfn) {
144 
145 				break;
146 			}
147 			tmp = pmem_ranges[j-1].start_pfn;
148 			pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
149 			pmem_ranges[j].start_pfn = tmp;
150 			tmp = pmem_ranges[j-1].pages;
151 			pmem_ranges[j-1].pages = pmem_ranges[j].pages;
152 			pmem_ranges[j].pages = tmp;
153 		}
154 	}
155 
156 #ifndef CONFIG_DISCONTIGMEM
157 	/*
158 	 * Throw out ranges that are too far apart (controlled by
159 	 * MAX_GAP).
160 	 */
161 
162 	for (i = 1; i < npmem_ranges; i++) {
163 		if (pmem_ranges[i].start_pfn -
164 			(pmem_ranges[i-1].start_pfn +
165 			 pmem_ranges[i-1].pages) > MAX_GAP) {
166 			npmem_ranges = i;
167 			printk("Large gap in memory detected (%ld pages). "
168 			       "Consider turning on CONFIG_DISCONTIGMEM\n",
169 			       pmem_ranges[i].start_pfn -
170 			       (pmem_ranges[i-1].start_pfn +
171 			        pmem_ranges[i-1].pages));
172 			break;
173 		}
174 	}
175 #endif
176 
177 	/* Print the memory ranges */
178 	pr_info("Memory Ranges:\n");
179 
180 	for (i = 0; i < npmem_ranges; i++) {
181 		struct resource *res = &sysram_resources[i];
182 		unsigned long start;
183 		unsigned long size;
184 
185 		size = (pmem_ranges[i].pages << PAGE_SHIFT);
186 		start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
187 		pr_info("%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
188 			i, start, start + (size - 1), size >> 20);
189 
190 		/* request memory resource */
191 		res->name = "System RAM";
192 		res->start = start;
193 		res->end = start + size - 1;
194 		res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
195 		request_resource(&iomem_resource, res);
196 	}
197 
198 	sysram_resource_count = npmem_ranges;
199 
200 	/*
201 	 * For 32 bit kernels we limit the amount of memory we can
202 	 * support, in order to preserve enough kernel address space
203 	 * for other purposes. For 64 bit kernels we don't normally
204 	 * limit the memory, but this mechanism can be used to
205 	 * artificially limit the amount of memory (and it is written
206 	 * to work with multiple memory ranges).
207 	 */
208 
209 	mem_limit_func();       /* check for "mem=" argument */
210 
211 	mem_max = 0;
212 	for (i = 0; i < npmem_ranges; i++) {
213 		unsigned long rsize;
214 
215 		rsize = pmem_ranges[i].pages << PAGE_SHIFT;
216 		if ((mem_max + rsize) > mem_limit) {
217 			printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
218 			if (mem_max == mem_limit)
219 				npmem_ranges = i;
220 			else {
221 				pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
222 						       - (mem_max >> PAGE_SHIFT);
223 				npmem_ranges = i + 1;
224 				mem_max = mem_limit;
225 			}
226 			break;
227 		}
228 		mem_max += rsize;
229 	}
230 
231 	printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
232 
233 #ifndef CONFIG_DISCONTIGMEM
234 	/* Merge the ranges, keeping track of the holes */
235 
236 	{
237 		unsigned long end_pfn;
238 		unsigned long hole_pages;
239 
240 		npmem_holes = 0;
241 		end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
242 		for (i = 1; i < npmem_ranges; i++) {
243 
244 			hole_pages = pmem_ranges[i].start_pfn - end_pfn;
245 			if (hole_pages) {
246 				pmem_holes[npmem_holes].start_pfn = end_pfn;
247 				pmem_holes[npmem_holes++].pages = hole_pages;
248 				end_pfn += hole_pages;
249 			}
250 			end_pfn += pmem_ranges[i].pages;
251 		}
252 
253 		pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
254 		npmem_ranges = 1;
255 	}
256 #endif
257 
258 #ifdef CONFIG_DISCONTIGMEM
259 	for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
260 		memset(NODE_DATA(i), 0, sizeof(pg_data_t));
261 	}
262 	memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
263 
264 	for (i = 0; i < npmem_ranges; i++) {
265 		node_set_state(i, N_NORMAL_MEMORY);
266 		node_set_online(i);
267 	}
268 #endif
269 
270 	/*
271 	 * Initialize and free the full range of memory in each range.
272 	 */
273 
274 	max_pfn = 0;
275 	for (i = 0; i < npmem_ranges; i++) {
276 		unsigned long start_pfn;
277 		unsigned long npages;
278 		unsigned long start;
279 		unsigned long size;
280 
281 		start_pfn = pmem_ranges[i].start_pfn;
282 		npages = pmem_ranges[i].pages;
283 
284 		start = start_pfn << PAGE_SHIFT;
285 		size = npages << PAGE_SHIFT;
286 
287 		/* add system RAM memblock */
288 		memblock_add(start, size);
289 
290 		if ((start_pfn + npages) > max_pfn)
291 			max_pfn = start_pfn + npages;
292 	}
293 
294 	/*
295 	 * We can't use memblock top-down allocations because we only
296 	 * created the initial mapping up to KERNEL_INITIAL_SIZE in
297 	 * the assembly bootup code.
298 	 */
299 	memblock_set_bottom_up(true);
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 	/* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
309 
310 #define PDC_CONSOLE_IO_IODC_SIZE 32768
311 
312 	memblock_reserve(0UL, (unsigned long)(PAGE0->mem_free +
313 				PDC_CONSOLE_IO_IODC_SIZE));
314 	memblock_reserve(__pa(KERNEL_BINARY_TEXT_START),
315 			(unsigned long)(_end - KERNEL_BINARY_TEXT_START));
316 
317 #ifndef CONFIG_DISCONTIGMEM
318 
319 	/* reserve the holes */
320 
321 	for (i = 0; i < npmem_holes; i++) {
322 		memblock_reserve((pmem_holes[i].start_pfn << PAGE_SHIFT),
323 				(pmem_holes[i].pages << PAGE_SHIFT));
324 	}
325 #endif
326 
327 #ifdef CONFIG_BLK_DEV_INITRD
328 	if (initrd_start) {
329 		printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
330 		if (__pa(initrd_start) < mem_max) {
331 			unsigned long initrd_reserve;
332 
333 			if (__pa(initrd_end) > mem_max) {
334 				initrd_reserve = mem_max - __pa(initrd_start);
335 			} else {
336 				initrd_reserve = initrd_end - initrd_start;
337 			}
338 			initrd_below_start_ok = 1;
339 			printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
340 
341 			memblock_reserve(__pa(initrd_start), initrd_reserve);
342 		}
343 	}
344 #endif
345 
346 	data_resource.start =  virt_to_phys(&data_start);
347 	data_resource.end = virt_to_phys(_end) - 1;
348 	code_resource.start = virt_to_phys(_text);
349 	code_resource.end = virt_to_phys(&data_start)-1;
350 
351 	/* We don't know which region the kernel will be in, so try
352 	 * all of them.
353 	 */
354 	for (i = 0; i < sysram_resource_count; i++) {
355 		struct resource *res = &sysram_resources[i];
356 		request_resource(res, &code_resource);
357 		request_resource(res, &data_resource);
358 	}
359 	request_resource(&sysram_resources[0], &pdcdata_resource);
360 
361 	/* Initialize Page Deallocation Table (PDT) and check for bad memory. */
362 	pdc_pdt_init();
363 }
364 
365 static int __init parisc_text_address(unsigned long vaddr)
366 {
367 	static unsigned long head_ptr __initdata;
368 
369 	if (!head_ptr)
370 		head_ptr = PAGE_MASK & (unsigned long)
371 			dereference_function_descriptor(&parisc_kernel_start);
372 
373 	return core_kernel_text(vaddr) || vaddr == head_ptr;
374 }
375 
376 static void __init map_pages(unsigned long start_vaddr,
377 			     unsigned long start_paddr, unsigned long size,
378 			     pgprot_t pgprot, int force)
379 {
380 	pgd_t *pg_dir;
381 	pmd_t *pmd;
382 	pte_t *pg_table;
383 	unsigned long end_paddr;
384 	unsigned long start_pmd;
385 	unsigned long start_pte;
386 	unsigned long tmp1;
387 	unsigned long tmp2;
388 	unsigned long address;
389 	unsigned long vaddr;
390 	unsigned long ro_start;
391 	unsigned long ro_end;
392 	unsigned long kernel_end;
393 
394 	ro_start = __pa((unsigned long)_text);
395 	ro_end   = __pa((unsigned long)&data_start);
396 	kernel_end  = __pa((unsigned long)&_end);
397 
398 	end_paddr = start_paddr + size;
399 
400 	pg_dir = pgd_offset_k(start_vaddr);
401 
402 #if PTRS_PER_PMD == 1
403 	start_pmd = 0;
404 #else
405 	start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
406 #endif
407 	start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
408 
409 	address = start_paddr;
410 	vaddr = start_vaddr;
411 	while (address < end_paddr) {
412 #if PTRS_PER_PMD == 1
413 		pmd = (pmd_t *)__pa(pg_dir);
414 #else
415 		pmd = (pmd_t *)pgd_address(*pg_dir);
416 
417 		/*
418 		 * pmd is physical at this point
419 		 */
420 
421 		if (!pmd) {
422 			pmd = memblock_alloc(PAGE_SIZE << PMD_ORDER,
423 					     PAGE_SIZE << PMD_ORDER);
424 			if (!pmd)
425 				panic("pmd allocation failed.\n");
426 			pmd = (pmd_t *) __pa(pmd);
427 		}
428 
429 		pgd_populate(NULL, pg_dir, __va(pmd));
430 #endif
431 		pg_dir++;
432 
433 		/* now change pmd to kernel virtual addresses */
434 
435 		pmd = (pmd_t *)__va(pmd) + start_pmd;
436 		for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
437 
438 			/*
439 			 * pg_table is physical at this point
440 			 */
441 
442 			pg_table = (pte_t *)pmd_address(*pmd);
443 			if (!pg_table) {
444 				pg_table = memblock_alloc(PAGE_SIZE,
445 							  PAGE_SIZE);
446 				if (!pg_table)
447 					panic("page table allocation failed\n");
448 				pg_table = (pte_t *) __pa(pg_table);
449 			}
450 
451 			pmd_populate_kernel(NULL, pmd, __va(pg_table));
452 
453 			/* now change pg_table to kernel virtual addresses */
454 
455 			pg_table = (pte_t *) __va(pg_table) + start_pte;
456 			for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
457 				pte_t pte;
458 
459 				if (force)
460 					pte =  __mk_pte(address, pgprot);
461 				else if (parisc_text_address(vaddr)) {
462 					pte = __mk_pte(address, PAGE_KERNEL_EXEC);
463 					if (address >= ro_start && address < kernel_end)
464 						pte = pte_mkhuge(pte);
465 				}
466 				else
467 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
468 				if (address >= ro_start && address < ro_end) {
469 					pte = __mk_pte(address, PAGE_KERNEL_EXEC);
470 					pte = pte_mkhuge(pte);
471 				} else
472 #endif
473 				{
474 					pte = __mk_pte(address, pgprot);
475 					if (address >= ro_start && address < kernel_end)
476 						pte = pte_mkhuge(pte);
477 				}
478 
479 				if (address >= end_paddr)
480 					break;
481 
482 				set_pte(pg_table, pte);
483 
484 				address += PAGE_SIZE;
485 				vaddr += PAGE_SIZE;
486 			}
487 			start_pte = 0;
488 
489 			if (address >= end_paddr)
490 			    break;
491 		}
492 		start_pmd = 0;
493 	}
494 }
495 
496 void __init set_kernel_text_rw(int enable_read_write)
497 {
498 	unsigned long start = (unsigned long) _text;
499 	unsigned long end   = (unsigned long) &data_start;
500 
501 	map_pages(start, __pa(start), end-start,
502 		PAGE_KERNEL_RWX, enable_read_write ? 1:0);
503 
504 	/* force the kernel to see the new page table entries */
505 	flush_cache_all();
506 	flush_tlb_all();
507 }
508 
509 void __ref free_initmem(void)
510 {
511 	unsigned long init_begin = (unsigned long)__init_begin;
512 	unsigned long init_end = (unsigned long)__init_end;
513 
514 	/* The init text pages are marked R-X.  We have to
515 	 * flush the icache and mark them RW-
516 	 *
517 	 * This is tricky, because map_pages is in the init section.
518 	 * Do a dummy remap of the data section first (the data
519 	 * section is already PAGE_KERNEL) to pull in the TLB entries
520 	 * for map_kernel */
521 	map_pages(init_begin, __pa(init_begin), init_end - init_begin,
522 		  PAGE_KERNEL_RWX, 1);
523 	/* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
524 	 * map_pages */
525 	map_pages(init_begin, __pa(init_begin), init_end - init_begin,
526 		  PAGE_KERNEL, 1);
527 
528 	/* force the kernel to see the new TLB entries */
529 	__flush_tlb_range(0, init_begin, init_end);
530 
531 	/* finally dump all the instructions which were cached, since the
532 	 * pages are no-longer executable */
533 	flush_icache_range(init_begin, init_end);
534 
535 	free_initmem_default(POISON_FREE_INITMEM);
536 
537 	/* set up a new led state on systems shipped LED State panel */
538 	pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
539 }
540 
541 
542 #ifdef CONFIG_STRICT_KERNEL_RWX
543 void mark_rodata_ro(void)
544 {
545 	/* rodata memory was already mapped with KERNEL_RO access rights by
546            pagetable_init() and map_pages(). No need to do additional stuff here */
547 	printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
548 		(unsigned long)(__end_rodata - __start_rodata) >> 10);
549 }
550 #endif
551 
552 
553 /*
554  * Just an arbitrary offset to serve as a "hole" between mapping areas
555  * (between top of physical memory and a potential pcxl dma mapping
556  * area, and below the vmalloc mapping area).
557  *
558  * The current 32K value just means that there will be a 32K "hole"
559  * between mapping areas. That means that  any out-of-bounds memory
560  * accesses will hopefully be caught. The vmalloc() routines leaves
561  * a hole of 4kB between each vmalloced area for the same reason.
562  */
563 
564  /* Leave room for gateway page expansion */
565 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
566 #error KERNEL_MAP_START is in gateway reserved region
567 #endif
568 #define MAP_START (KERNEL_MAP_START)
569 
570 #define VM_MAP_OFFSET  (32*1024)
571 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
572 				     & ~(VM_MAP_OFFSET-1)))
573 
574 void *parisc_vmalloc_start __read_mostly;
575 EXPORT_SYMBOL(parisc_vmalloc_start);
576 
577 #ifdef CONFIG_PA11
578 unsigned long pcxl_dma_start __read_mostly;
579 #endif
580 
581 void __init mem_init(void)
582 {
583 	/* Do sanity checks on IPC (compat) structures */
584 	BUILD_BUG_ON(sizeof(struct ipc64_perm) != 48);
585 #ifndef CONFIG_64BIT
586 	BUILD_BUG_ON(sizeof(struct semid64_ds) != 80);
587 	BUILD_BUG_ON(sizeof(struct msqid64_ds) != 104);
588 	BUILD_BUG_ON(sizeof(struct shmid64_ds) != 104);
589 #endif
590 #ifdef CONFIG_COMPAT
591 	BUILD_BUG_ON(sizeof(struct compat_ipc64_perm) != sizeof(struct ipc64_perm));
592 	BUILD_BUG_ON(sizeof(struct compat_semid64_ds) != 80);
593 	BUILD_BUG_ON(sizeof(struct compat_msqid64_ds) != 104);
594 	BUILD_BUG_ON(sizeof(struct compat_shmid64_ds) != 104);
595 #endif
596 
597 	/* Do sanity checks on page table constants */
598 	BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
599 	BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
600 	BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
601 	BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
602 			> BITS_PER_LONG);
603 
604 	high_memory = __va((max_pfn << PAGE_SHIFT));
605 	set_max_mapnr(page_to_pfn(virt_to_page(high_memory - 1)) + 1);
606 	memblock_free_all();
607 
608 #ifdef CONFIG_PA11
609 	if (boot_cpu_data.cpu_type == pcxl2 || boot_cpu_data.cpu_type == pcxl) {
610 		pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
611 		parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
612 						+ PCXL_DMA_MAP_SIZE);
613 	} else
614 #endif
615 		parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
616 
617 	mem_init_print_info(NULL);
618 
619 #if 0
620 	/*
621 	 * Do not expose the virtual kernel memory layout to userspace.
622 	 * But keep code for debugging purposes.
623 	 */
624 	printk("virtual kernel memory layout:\n"
625 	       "    vmalloc : 0x%px - 0x%px   (%4ld MB)\n"
626 	       "    memory  : 0x%px - 0x%px   (%4ld MB)\n"
627 	       "      .init : 0x%px - 0x%px   (%4ld kB)\n"
628 	       "      .data : 0x%px - 0x%px   (%4ld kB)\n"
629 	       "      .text : 0x%px - 0x%px   (%4ld kB)\n",
630 
631 	       (void*)VMALLOC_START, (void*)VMALLOC_END,
632 	       (VMALLOC_END - VMALLOC_START) >> 20,
633 
634 	       __va(0), high_memory,
635 	       ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
636 
637 	       __init_begin, __init_end,
638 	       ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
639 
640 	       _etext, _edata,
641 	       ((unsigned long)_edata - (unsigned long)_etext) >> 10,
642 
643 	       _text, _etext,
644 	       ((unsigned long)_etext - (unsigned long)_text) >> 10);
645 #endif
646 }
647 
648 unsigned long *empty_zero_page __read_mostly;
649 EXPORT_SYMBOL(empty_zero_page);
650 
651 /*
652  * pagetable_init() sets up the page tables
653  *
654  * Note that gateway_init() places the Linux gateway page at page 0.
655  * Since gateway pages cannot be dereferenced this has the desirable
656  * side effect of trapping those pesky NULL-reference errors in the
657  * kernel.
658  */
659 static void __init pagetable_init(void)
660 {
661 	int range;
662 
663 	/* Map each physical memory range to its kernel vaddr */
664 
665 	for (range = 0; range < npmem_ranges; range++) {
666 		unsigned long start_paddr;
667 		unsigned long end_paddr;
668 		unsigned long size;
669 
670 		start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
671 		size = pmem_ranges[range].pages << PAGE_SHIFT;
672 		end_paddr = start_paddr + size;
673 
674 		map_pages((unsigned long)__va(start_paddr), start_paddr,
675 			  size, PAGE_KERNEL, 0);
676 	}
677 
678 #ifdef CONFIG_BLK_DEV_INITRD
679 	if (initrd_end && initrd_end > mem_limit) {
680 		printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
681 		map_pages(initrd_start, __pa(initrd_start),
682 			  initrd_end - initrd_start, PAGE_KERNEL, 0);
683 	}
684 #endif
685 
686 	empty_zero_page = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
687 	if (!empty_zero_page)
688 		panic("zero page allocation failed.\n");
689 
690 }
691 
692 static void __init gateway_init(void)
693 {
694 	unsigned long linux_gateway_page_addr;
695 	/* FIXME: This is 'const' in order to trick the compiler
696 	   into not treating it as DP-relative data. */
697 	extern void * const linux_gateway_page;
698 
699 	linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
700 
701 	/*
702 	 * Setup Linux Gateway page.
703 	 *
704 	 * The Linux gateway page will reside in kernel space (on virtual
705 	 * page 0), so it doesn't need to be aliased into user space.
706 	 */
707 
708 	map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
709 		  PAGE_SIZE, PAGE_GATEWAY, 1);
710 }
711 
712 void __init paging_init(void)
713 {
714 	int i;
715 
716 	setup_bootmem();
717 	pagetable_init();
718 	gateway_init();
719 	flush_cache_all_local(); /* start with known state */
720 	flush_tlb_all_local(NULL);
721 
722 	for (i = 0; i < npmem_ranges; i++) {
723 		unsigned long zones_size[MAX_NR_ZONES] = { 0, };
724 
725 		zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
726 
727 #ifdef CONFIG_DISCONTIGMEM
728 		/* Need to initialize the pfnnid_map before we can initialize
729 		   the zone */
730 		{
731 		    int j;
732 		    for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
733 			 j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
734 			 j++) {
735 			pfnnid_map[j] = i;
736 		    }
737 		}
738 #endif
739 
740 		free_area_init_node(i, zones_size,
741 				pmem_ranges[i].start_pfn, NULL);
742 	}
743 }
744 
745 #ifdef CONFIG_PA20
746 
747 /*
748  * Currently, all PA20 chips have 18 bit protection IDs, which is the
749  * limiting factor (space ids are 32 bits).
750  */
751 
752 #define NR_SPACE_IDS 262144
753 
754 #else
755 
756 /*
757  * Currently we have a one-to-one relationship between space IDs and
758  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
759  * support 15 bit protection IDs, so that is the limiting factor.
760  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
761  * probably not worth the effort for a special case here.
762  */
763 
764 #define NR_SPACE_IDS 32768
765 
766 #endif  /* !CONFIG_PA20 */
767 
768 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
769 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
770 
771 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
772 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
773 static unsigned long space_id_index;
774 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
775 static unsigned long dirty_space_ids = 0;
776 
777 static DEFINE_SPINLOCK(sid_lock);
778 
779 unsigned long alloc_sid(void)
780 {
781 	unsigned long index;
782 
783 	spin_lock(&sid_lock);
784 
785 	if (free_space_ids == 0) {
786 		if (dirty_space_ids != 0) {
787 			spin_unlock(&sid_lock);
788 			flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
789 			spin_lock(&sid_lock);
790 		}
791 		BUG_ON(free_space_ids == 0);
792 	}
793 
794 	free_space_ids--;
795 
796 	index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
797 	space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
798 	space_id_index = index;
799 
800 	spin_unlock(&sid_lock);
801 
802 	return index << SPACEID_SHIFT;
803 }
804 
805 void free_sid(unsigned long spaceid)
806 {
807 	unsigned long index = spaceid >> SPACEID_SHIFT;
808 	unsigned long *dirty_space_offset;
809 
810 	dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
811 	index &= (BITS_PER_LONG - 1);
812 
813 	spin_lock(&sid_lock);
814 
815 	BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
816 
817 	*dirty_space_offset |= (1L << index);
818 	dirty_space_ids++;
819 
820 	spin_unlock(&sid_lock);
821 }
822 
823 
824 #ifdef CONFIG_SMP
825 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
826 {
827 	int i;
828 
829 	/* NOTE: sid_lock must be held upon entry */
830 
831 	*ndirtyptr = dirty_space_ids;
832 	if (dirty_space_ids != 0) {
833 	    for (i = 0; i < SID_ARRAY_SIZE; i++) {
834 		dirty_array[i] = dirty_space_id[i];
835 		dirty_space_id[i] = 0;
836 	    }
837 	    dirty_space_ids = 0;
838 	}
839 
840 	return;
841 }
842 
843 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
844 {
845 	int i;
846 
847 	/* NOTE: sid_lock must be held upon entry */
848 
849 	if (ndirty != 0) {
850 		for (i = 0; i < SID_ARRAY_SIZE; i++) {
851 			space_id[i] ^= dirty_array[i];
852 		}
853 
854 		free_space_ids += ndirty;
855 		space_id_index = 0;
856 	}
857 }
858 
859 #else /* CONFIG_SMP */
860 
861 static void recycle_sids(void)
862 {
863 	int i;
864 
865 	/* NOTE: sid_lock must be held upon entry */
866 
867 	if (dirty_space_ids != 0) {
868 		for (i = 0; i < SID_ARRAY_SIZE; i++) {
869 			space_id[i] ^= dirty_space_id[i];
870 			dirty_space_id[i] = 0;
871 		}
872 
873 		free_space_ids += dirty_space_ids;
874 		dirty_space_ids = 0;
875 		space_id_index = 0;
876 	}
877 }
878 #endif
879 
880 /*
881  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
882  * purged, we can safely reuse the space ids that were released but
883  * not flushed from the tlb.
884  */
885 
886 #ifdef CONFIG_SMP
887 
888 static unsigned long recycle_ndirty;
889 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
890 static unsigned int recycle_inuse;
891 
892 void flush_tlb_all(void)
893 {
894 	int do_recycle;
895 
896 	__inc_irq_stat(irq_tlb_count);
897 	do_recycle = 0;
898 	spin_lock(&sid_lock);
899 	if (dirty_space_ids > RECYCLE_THRESHOLD) {
900 	    BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
901 	    get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
902 	    recycle_inuse++;
903 	    do_recycle++;
904 	}
905 	spin_unlock(&sid_lock);
906 	on_each_cpu(flush_tlb_all_local, NULL, 1);
907 	if (do_recycle) {
908 	    spin_lock(&sid_lock);
909 	    recycle_sids(recycle_ndirty,recycle_dirty_array);
910 	    recycle_inuse = 0;
911 	    spin_unlock(&sid_lock);
912 	}
913 }
914 #else
915 void flush_tlb_all(void)
916 {
917 	__inc_irq_stat(irq_tlb_count);
918 	spin_lock(&sid_lock);
919 	flush_tlb_all_local(NULL);
920 	recycle_sids();
921 	spin_unlock(&sid_lock);
922 }
923 #endif
924 
925 #ifdef CONFIG_BLK_DEV_INITRD
926 void free_initrd_mem(unsigned long start, unsigned long end)
927 {
928 	free_reserved_area((void *)start, (void *)end, -1, "initrd");
929 }
930 #endif
931