xref: /openbmc/linux/arch/parisc/mm/init.c (revision 37744fee)
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 #include <asm/sparsemem.h>
36 
37 extern int  data_start;
38 extern void parisc_kernel_start(void);	/* Kernel entry point in head.S */
39 
40 #if CONFIG_PGTABLE_LEVELS == 3
41 /* NOTE: This layout exactly conforms to the hybrid L2/L3 page table layout
42  * with the first pmd adjacent to the pgd and below it. gcc doesn't actually
43  * guarantee that global objects will be laid out in memory in the same order
44  * as the order of declaration, so put these in different sections and use
45  * the linker script to order them. */
46 pmd_t pmd0[PTRS_PER_PMD] __attribute__ ((__section__ (".data..vm0.pmd"), aligned(PAGE_SIZE)));
47 #endif
48 
49 pgd_t swapper_pg_dir[PTRS_PER_PGD] __attribute__ ((__section__ (".data..vm0.pgd"), aligned(PAGE_SIZE)));
50 pte_t pg0[PT_INITIAL * PTRS_PER_PTE] __attribute__ ((__section__ (".data..vm0.pte"), aligned(PAGE_SIZE)));
51 
52 static struct resource data_resource = {
53 	.name	= "Kernel data",
54 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
55 };
56 
57 static struct resource code_resource = {
58 	.name	= "Kernel code",
59 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
60 };
61 
62 static struct resource pdcdata_resource = {
63 	.name	= "PDC data (Page Zero)",
64 	.start	= 0,
65 	.end	= 0x9ff,
66 	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM,
67 };
68 
69 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __ro_after_init;
70 
71 /* The following array is initialized from the firmware specific
72  * information retrieved in kernel/inventory.c.
73  */
74 
75 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __initdata;
76 int npmem_ranges __initdata;
77 
78 #ifdef CONFIG_64BIT
79 #define MAX_MEM         (1UL << MAX_PHYSMEM_BITS)
80 #else /* !CONFIG_64BIT */
81 #define MAX_MEM         (3584U*1024U*1024U)
82 #endif /* !CONFIG_64BIT */
83 
84 static unsigned long mem_limit __read_mostly = MAX_MEM;
85 
86 static void __init mem_limit_func(void)
87 {
88 	char *cp, *end;
89 	unsigned long limit;
90 
91 	/* We need this before __setup() functions are called */
92 
93 	limit = MAX_MEM;
94 	for (cp = boot_command_line; *cp; ) {
95 		if (memcmp(cp, "mem=", 4) == 0) {
96 			cp += 4;
97 			limit = memparse(cp, &end);
98 			if (end != cp)
99 				break;
100 			cp = end;
101 		} else {
102 			while (*cp != ' ' && *cp)
103 				++cp;
104 			while (*cp == ' ')
105 				++cp;
106 		}
107 	}
108 
109 	if (limit < mem_limit)
110 		mem_limit = limit;
111 }
112 
113 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
114 
115 static void __init setup_bootmem(void)
116 {
117 	unsigned long mem_max;
118 #ifndef CONFIG_SPARSEMEM
119 	physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
120 	int npmem_holes;
121 #endif
122 	int i, sysram_resource_count;
123 
124 	disable_sr_hashing(); /* Turn off space register hashing */
125 
126 	/*
127 	 * Sort the ranges. Since the number of ranges is typically
128 	 * small, and performance is not an issue here, just do
129 	 * a simple insertion sort.
130 	 */
131 
132 	for (i = 1; i < npmem_ranges; i++) {
133 		int j;
134 
135 		for (j = i; j > 0; j--) {
136 			physmem_range_t tmp;
137 
138 			if (pmem_ranges[j-1].start_pfn <
139 			    pmem_ranges[j].start_pfn) {
140 
141 				break;
142 			}
143 			tmp = pmem_ranges[j-1];
144 			pmem_ranges[j-1] = pmem_ranges[j];
145 			pmem_ranges[j] = tmp;
146 		}
147 	}
148 
149 #ifndef CONFIG_SPARSEMEM
150 	/*
151 	 * Throw out ranges that are too far apart (controlled by
152 	 * MAX_GAP).
153 	 */
154 
155 	for (i = 1; i < npmem_ranges; i++) {
156 		if (pmem_ranges[i].start_pfn -
157 			(pmem_ranges[i-1].start_pfn +
158 			 pmem_ranges[i-1].pages) > MAX_GAP) {
159 			npmem_ranges = i;
160 			printk("Large gap in memory detected (%ld pages). "
161 			       "Consider turning on CONFIG_SPARSEMEM\n",
162 			       pmem_ranges[i].start_pfn -
163 			       (pmem_ranges[i-1].start_pfn +
164 			        pmem_ranges[i-1].pages));
165 			break;
166 		}
167 	}
168 #endif
169 
170 	/* Print the memory ranges */
171 	pr_info("Memory Ranges:\n");
172 
173 	for (i = 0; i < npmem_ranges; i++) {
174 		struct resource *res = &sysram_resources[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 		pr_info("%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
181 			i, start, start + (size - 1), size >> 20);
182 
183 		/* request memory resource */
184 		res->name = "System RAM";
185 		res->start = start;
186 		res->end = start + size - 1;
187 		res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
188 		request_resource(&iomem_resource, res);
189 	}
190 
191 	sysram_resource_count = npmem_ranges;
192 
193 	/*
194 	 * For 32 bit kernels we limit the amount of memory we can
195 	 * support, in order to preserve enough kernel address space
196 	 * for other purposes. For 64 bit kernels we don't normally
197 	 * limit the memory, but this mechanism can be used to
198 	 * artificially limit the amount of memory (and it is written
199 	 * to work with multiple memory ranges).
200 	 */
201 
202 	mem_limit_func();       /* check for "mem=" argument */
203 
204 	mem_max = 0;
205 	for (i = 0; i < npmem_ranges; i++) {
206 		unsigned long rsize;
207 
208 		rsize = pmem_ranges[i].pages << PAGE_SHIFT;
209 		if ((mem_max + rsize) > mem_limit) {
210 			printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
211 			if (mem_max == mem_limit)
212 				npmem_ranges = i;
213 			else {
214 				pmem_ranges[i].pages =   (mem_limit >> PAGE_SHIFT)
215 						       - (mem_max >> PAGE_SHIFT);
216 				npmem_ranges = i + 1;
217 				mem_max = mem_limit;
218 			}
219 			break;
220 		}
221 		mem_max += rsize;
222 	}
223 
224 	printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
225 
226 #ifndef CONFIG_SPARSEMEM
227 	/* Merge the ranges, keeping track of the holes */
228 	{
229 		unsigned long end_pfn;
230 		unsigned long hole_pages;
231 
232 		npmem_holes = 0;
233 		end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
234 		for (i = 1; i < npmem_ranges; i++) {
235 
236 			hole_pages = pmem_ranges[i].start_pfn - end_pfn;
237 			if (hole_pages) {
238 				pmem_holes[npmem_holes].start_pfn = end_pfn;
239 				pmem_holes[npmem_holes++].pages = hole_pages;
240 				end_pfn += hole_pages;
241 			}
242 			end_pfn += pmem_ranges[i].pages;
243 		}
244 
245 		pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
246 		npmem_ranges = 1;
247 	}
248 #endif
249 
250 	/*
251 	 * Initialize and free the full range of memory in each range.
252 	 */
253 
254 	max_pfn = 0;
255 	for (i = 0; i < npmem_ranges; i++) {
256 		unsigned long start_pfn;
257 		unsigned long npages;
258 		unsigned long start;
259 		unsigned long size;
260 
261 		start_pfn = pmem_ranges[i].start_pfn;
262 		npages = pmem_ranges[i].pages;
263 
264 		start = start_pfn << PAGE_SHIFT;
265 		size = npages << PAGE_SHIFT;
266 
267 		/* add system RAM memblock */
268 		memblock_add(start, size);
269 
270 		if ((start_pfn + npages) > max_pfn)
271 			max_pfn = start_pfn + npages;
272 	}
273 
274 	/*
275 	 * We can't use memblock top-down allocations because we only
276 	 * created the initial mapping up to KERNEL_INITIAL_SIZE in
277 	 * the assembly bootup code.
278 	 */
279 	memblock_set_bottom_up(true);
280 
281 	/* IOMMU is always used to access "high mem" on those boxes
282 	 * that can support enough mem that a PCI device couldn't
283 	 * directly DMA to any physical addresses.
284 	 * ISA DMA support will need to revisit this.
285 	 */
286 	max_low_pfn = max_pfn;
287 
288 	/* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
289 
290 #define PDC_CONSOLE_IO_IODC_SIZE 32768
291 
292 	memblock_reserve(0UL, (unsigned long)(PAGE0->mem_free +
293 				PDC_CONSOLE_IO_IODC_SIZE));
294 	memblock_reserve(__pa(KERNEL_BINARY_TEXT_START),
295 			(unsigned long)(_end - KERNEL_BINARY_TEXT_START));
296 
297 #ifndef CONFIG_SPARSEMEM
298 
299 	/* reserve the holes */
300 
301 	for (i = 0; i < npmem_holes; i++) {
302 		memblock_reserve((pmem_holes[i].start_pfn << PAGE_SHIFT),
303 				(pmem_holes[i].pages << PAGE_SHIFT));
304 	}
305 #endif
306 
307 #ifdef CONFIG_BLK_DEV_INITRD
308 	if (initrd_start) {
309 		printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
310 		if (__pa(initrd_start) < mem_max) {
311 			unsigned long initrd_reserve;
312 
313 			if (__pa(initrd_end) > mem_max) {
314 				initrd_reserve = mem_max - __pa(initrd_start);
315 			} else {
316 				initrd_reserve = initrd_end - initrd_start;
317 			}
318 			initrd_below_start_ok = 1;
319 			printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
320 
321 			memblock_reserve(__pa(initrd_start), initrd_reserve);
322 		}
323 	}
324 #endif
325 
326 	data_resource.start =  virt_to_phys(&data_start);
327 	data_resource.end = virt_to_phys(_end) - 1;
328 	code_resource.start = virt_to_phys(_text);
329 	code_resource.end = virt_to_phys(&data_start)-1;
330 
331 	/* We don't know which region the kernel will be in, so try
332 	 * all of them.
333 	 */
334 	for (i = 0; i < sysram_resource_count; i++) {
335 		struct resource *res = &sysram_resources[i];
336 		request_resource(res, &code_resource);
337 		request_resource(res, &data_resource);
338 	}
339 	request_resource(&sysram_resources[0], &pdcdata_resource);
340 
341 	/* Initialize Page Deallocation Table (PDT) and check for bad memory. */
342 	pdc_pdt_init();
343 
344 	memblock_allow_resize();
345 	memblock_dump_all();
346 }
347 
348 static bool kernel_set_to_readonly;
349 
350 static void __init map_pages(unsigned long start_vaddr,
351 			     unsigned long start_paddr, unsigned long size,
352 			     pgprot_t pgprot, int force)
353 {
354 	pmd_t *pmd;
355 	pte_t *pg_table;
356 	unsigned long end_paddr;
357 	unsigned long start_pmd;
358 	unsigned long start_pte;
359 	unsigned long tmp1;
360 	unsigned long tmp2;
361 	unsigned long address;
362 	unsigned long vaddr;
363 	unsigned long ro_start;
364 	unsigned long ro_end;
365 	unsigned long kernel_start, kernel_end;
366 
367 	ro_start = __pa((unsigned long)_text);
368 	ro_end   = __pa((unsigned long)&data_start);
369 	kernel_start = __pa((unsigned long)&__init_begin);
370 	kernel_end  = __pa((unsigned long)&_end);
371 
372 	end_paddr = start_paddr + size;
373 
374 	/* for 2-level configuration PTRS_PER_PMD is 0 so start_pmd will be 0 */
375 	start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
376 	start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
377 
378 	address = start_paddr;
379 	vaddr = start_vaddr;
380 	while (address < end_paddr) {
381 		pgd_t *pgd = pgd_offset_k(vaddr);
382 		p4d_t *p4d = p4d_offset(pgd, vaddr);
383 		pud_t *pud = pud_offset(p4d, vaddr);
384 
385 #if CONFIG_PGTABLE_LEVELS == 3
386 		if (pud_none(*pud)) {
387 			pmd = memblock_alloc(PAGE_SIZE << PMD_ORDER,
388 					     PAGE_SIZE << PMD_ORDER);
389 			if (!pmd)
390 				panic("pmd allocation failed.\n");
391 			pud_populate(NULL, pud, pmd);
392 		}
393 #endif
394 
395 		pmd = pmd_offset(pud, vaddr);
396 		for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
397 			if (pmd_none(*pmd)) {
398 				pg_table = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
399 				if (!pg_table)
400 					panic("page table allocation failed\n");
401 				pmd_populate_kernel(NULL, pmd, pg_table);
402 			}
403 
404 			pg_table = pte_offset_kernel(pmd, vaddr);
405 			for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
406 				pte_t pte;
407 				pgprot_t prot;
408 				bool huge = false;
409 
410 				if (force) {
411 					prot = pgprot;
412 				} else if (address < kernel_start || address >= kernel_end) {
413 					/* outside kernel memory */
414 					prot = PAGE_KERNEL;
415 				} else if (!kernel_set_to_readonly) {
416 					/* still initializing, allow writing to RO memory */
417 					prot = PAGE_KERNEL_RWX;
418 					huge = true;
419 				} else if (address >= ro_start) {
420 					/* Code (ro) and Data areas */
421 					prot = (address < ro_end) ?
422 						PAGE_KERNEL_EXEC : PAGE_KERNEL;
423 					huge = true;
424 				} else {
425 					prot = PAGE_KERNEL;
426 				}
427 
428 				pte = __mk_pte(address, prot);
429 				if (huge)
430 					pte = pte_mkhuge(pte);
431 
432 				if (address >= end_paddr)
433 					break;
434 
435 				set_pte(pg_table, pte);
436 
437 				address += PAGE_SIZE;
438 				vaddr += PAGE_SIZE;
439 			}
440 			start_pte = 0;
441 
442 			if (address >= end_paddr)
443 			    break;
444 		}
445 		start_pmd = 0;
446 	}
447 }
448 
449 void __init set_kernel_text_rw(int enable_read_write)
450 {
451 	unsigned long start = (unsigned long) __init_begin;
452 	unsigned long end   = (unsigned long) &data_start;
453 
454 	map_pages(start, __pa(start), end-start,
455 		PAGE_KERNEL_RWX, enable_read_write ? 1:0);
456 
457 	/* force the kernel to see the new page table entries */
458 	flush_cache_all();
459 	flush_tlb_all();
460 }
461 
462 void __ref free_initmem(void)
463 {
464 	unsigned long init_begin = (unsigned long)__init_begin;
465 	unsigned long init_end = (unsigned long)__init_end;
466 	unsigned long kernel_end  = (unsigned long)&_end;
467 
468 	/* Remap kernel text and data, but do not touch init section yet. */
469 	kernel_set_to_readonly = true;
470 	map_pages(init_end, __pa(init_end), kernel_end - init_end,
471 		  PAGE_KERNEL, 0);
472 
473 	/* The init text pages are marked R-X.  We have to
474 	 * flush the icache and mark them RW-
475 	 *
476 	 * This is tricky, because map_pages is in the init section.
477 	 * Do a dummy remap of the data section first (the data
478 	 * section is already PAGE_KERNEL) to pull in the TLB entries
479 	 * for map_kernel */
480 	map_pages(init_begin, __pa(init_begin), init_end - init_begin,
481 		  PAGE_KERNEL_RWX, 1);
482 	/* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
483 	 * map_pages */
484 	map_pages(init_begin, __pa(init_begin), init_end - init_begin,
485 		  PAGE_KERNEL, 1);
486 
487 	/* force the kernel to see the new TLB entries */
488 	__flush_tlb_range(0, init_begin, kernel_end);
489 
490 	/* finally dump all the instructions which were cached, since the
491 	 * pages are no-longer executable */
492 	flush_icache_range(init_begin, init_end);
493 
494 	free_initmem_default(POISON_FREE_INITMEM);
495 
496 	/* set up a new led state on systems shipped LED State panel */
497 	pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
498 }
499 
500 
501 #ifdef CONFIG_STRICT_KERNEL_RWX
502 void mark_rodata_ro(void)
503 {
504 	/* rodata memory was already mapped with KERNEL_RO access rights by
505            pagetable_init() and map_pages(). No need to do additional stuff here */
506 	unsigned long roai_size = __end_ro_after_init - __start_ro_after_init;
507 
508 	pr_info("Write protected read-only-after-init data: %luk\n", roai_size >> 10);
509 }
510 #endif
511 
512 
513 /*
514  * Just an arbitrary offset to serve as a "hole" between mapping areas
515  * (between top of physical memory and a potential pcxl dma mapping
516  * area, and below the vmalloc mapping area).
517  *
518  * The current 32K value just means that there will be a 32K "hole"
519  * between mapping areas. That means that  any out-of-bounds memory
520  * accesses will hopefully be caught. The vmalloc() routines leaves
521  * a hole of 4kB between each vmalloced area for the same reason.
522  */
523 
524  /* Leave room for gateway page expansion */
525 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
526 #error KERNEL_MAP_START is in gateway reserved region
527 #endif
528 #define MAP_START (KERNEL_MAP_START)
529 
530 #define VM_MAP_OFFSET  (32*1024)
531 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
532 				     & ~(VM_MAP_OFFSET-1)))
533 
534 void *parisc_vmalloc_start __ro_after_init;
535 EXPORT_SYMBOL(parisc_vmalloc_start);
536 
537 #ifdef CONFIG_PA11
538 unsigned long pcxl_dma_start __ro_after_init;
539 #endif
540 
541 void __init mem_init(void)
542 {
543 	/* Do sanity checks on IPC (compat) structures */
544 	BUILD_BUG_ON(sizeof(struct ipc64_perm) != 48);
545 #ifndef CONFIG_64BIT
546 	BUILD_BUG_ON(sizeof(struct semid64_ds) != 80);
547 	BUILD_BUG_ON(sizeof(struct msqid64_ds) != 104);
548 	BUILD_BUG_ON(sizeof(struct shmid64_ds) != 104);
549 #endif
550 #ifdef CONFIG_COMPAT
551 	BUILD_BUG_ON(sizeof(struct compat_ipc64_perm) != sizeof(struct ipc64_perm));
552 	BUILD_BUG_ON(sizeof(struct compat_semid64_ds) != 80);
553 	BUILD_BUG_ON(sizeof(struct compat_msqid64_ds) != 104);
554 	BUILD_BUG_ON(sizeof(struct compat_shmid64_ds) != 104);
555 #endif
556 
557 	/* Do sanity checks on page table constants */
558 	BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
559 	BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
560 	BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
561 	BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
562 			> BITS_PER_LONG);
563 
564 	high_memory = __va((max_pfn << PAGE_SHIFT));
565 	set_max_mapnr(max_low_pfn);
566 	memblock_free_all();
567 
568 #ifdef CONFIG_PA11
569 	if (boot_cpu_data.cpu_type == pcxl2 || boot_cpu_data.cpu_type == pcxl) {
570 		pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
571 		parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
572 						+ PCXL_DMA_MAP_SIZE);
573 	} else
574 #endif
575 		parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
576 
577 	mem_init_print_info(NULL);
578 
579 #if 0
580 	/*
581 	 * Do not expose the virtual kernel memory layout to userspace.
582 	 * But keep code for debugging purposes.
583 	 */
584 	printk("virtual kernel memory layout:\n"
585 	       "     vmalloc : 0x%px - 0x%px   (%4ld MB)\n"
586 	       "     fixmap  : 0x%px - 0x%px   (%4ld kB)\n"
587 	       "     memory  : 0x%px - 0x%px   (%4ld MB)\n"
588 	       "       .init : 0x%px - 0x%px   (%4ld kB)\n"
589 	       "       .data : 0x%px - 0x%px   (%4ld kB)\n"
590 	       "       .text : 0x%px - 0x%px   (%4ld kB)\n",
591 
592 	       (void*)VMALLOC_START, (void*)VMALLOC_END,
593 	       (VMALLOC_END - VMALLOC_START) >> 20,
594 
595 	       (void *)FIXMAP_START, (void *)(FIXMAP_START + FIXMAP_SIZE),
596 	       (unsigned long)(FIXMAP_SIZE / 1024),
597 
598 	       __va(0), high_memory,
599 	       ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
600 
601 	       __init_begin, __init_end,
602 	       ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
603 
604 	       _etext, _edata,
605 	       ((unsigned long)_edata - (unsigned long)_etext) >> 10,
606 
607 	       _text, _etext,
608 	       ((unsigned long)_etext - (unsigned long)_text) >> 10);
609 #endif
610 }
611 
612 unsigned long *empty_zero_page __ro_after_init;
613 EXPORT_SYMBOL(empty_zero_page);
614 
615 /*
616  * pagetable_init() sets up the page tables
617  *
618  * Note that gateway_init() places the Linux gateway page at page 0.
619  * Since gateway pages cannot be dereferenced this has the desirable
620  * side effect of trapping those pesky NULL-reference errors in the
621  * kernel.
622  */
623 static void __init pagetable_init(void)
624 {
625 	int range;
626 
627 	/* Map each physical memory range to its kernel vaddr */
628 
629 	for (range = 0; range < npmem_ranges; range++) {
630 		unsigned long start_paddr;
631 		unsigned long end_paddr;
632 		unsigned long size;
633 
634 		start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
635 		size = pmem_ranges[range].pages << PAGE_SHIFT;
636 		end_paddr = start_paddr + size;
637 
638 		map_pages((unsigned long)__va(start_paddr), start_paddr,
639 			  size, PAGE_KERNEL, 0);
640 	}
641 
642 #ifdef CONFIG_BLK_DEV_INITRD
643 	if (initrd_end && initrd_end > mem_limit) {
644 		printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
645 		map_pages(initrd_start, __pa(initrd_start),
646 			  initrd_end - initrd_start, PAGE_KERNEL, 0);
647 	}
648 #endif
649 
650 	empty_zero_page = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
651 	if (!empty_zero_page)
652 		panic("zero page allocation failed.\n");
653 
654 }
655 
656 static void __init gateway_init(void)
657 {
658 	unsigned long linux_gateway_page_addr;
659 	/* FIXME: This is 'const' in order to trick the compiler
660 	   into not treating it as DP-relative data. */
661 	extern void * const linux_gateway_page;
662 
663 	linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
664 
665 	/*
666 	 * Setup Linux Gateway page.
667 	 *
668 	 * The Linux gateway page will reside in kernel space (on virtual
669 	 * page 0), so it doesn't need to be aliased into user space.
670 	 */
671 
672 	map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
673 		  PAGE_SIZE, PAGE_GATEWAY, 1);
674 }
675 
676 static void __init parisc_bootmem_free(void)
677 {
678 	unsigned long zones_size[MAX_NR_ZONES] = { 0, };
679 	unsigned long holes_size[MAX_NR_ZONES] = { 0, };
680 	unsigned long mem_start_pfn = ~0UL, mem_end_pfn = 0, mem_size_pfn = 0;
681 	int i;
682 
683 	for (i = 0; i < npmem_ranges; i++) {
684 		unsigned long start = pmem_ranges[i].start_pfn;
685 		unsigned long size = pmem_ranges[i].pages;
686 		unsigned long end = start + size;
687 
688 		if (mem_start_pfn > start)
689 			mem_start_pfn = start;
690 		if (mem_end_pfn < end)
691 			mem_end_pfn = end;
692 		mem_size_pfn += size;
693 	}
694 
695 	zones_size[0] = mem_end_pfn - mem_start_pfn;
696 	holes_size[0] = zones_size[0] - mem_size_pfn;
697 
698 	free_area_init_node(0, zones_size, mem_start_pfn, holes_size);
699 }
700 
701 void __init paging_init(void)
702 {
703 	setup_bootmem();
704 	pagetable_init();
705 	gateway_init();
706 	flush_cache_all_local(); /* start with known state */
707 	flush_tlb_all_local(NULL);
708 
709 	/*
710 	 * Mark all memblocks as present for sparsemem using
711 	 * memory_present() and then initialize sparsemem.
712 	 */
713 	memblocks_present();
714 	sparse_init();
715 	parisc_bootmem_free();
716 }
717 
718 #ifdef CONFIG_PA20
719 
720 /*
721  * Currently, all PA20 chips have 18 bit protection IDs, which is the
722  * limiting factor (space ids are 32 bits).
723  */
724 
725 #define NR_SPACE_IDS 262144
726 
727 #else
728 
729 /*
730  * Currently we have a one-to-one relationship between space IDs and
731  * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
732  * support 15 bit protection IDs, so that is the limiting factor.
733  * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
734  * probably not worth the effort for a special case here.
735  */
736 
737 #define NR_SPACE_IDS 32768
738 
739 #endif  /* !CONFIG_PA20 */
740 
741 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
742 #define SID_ARRAY_SIZE  (NR_SPACE_IDS / (8 * sizeof(long)))
743 
744 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
745 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
746 static unsigned long space_id_index;
747 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
748 static unsigned long dirty_space_ids = 0;
749 
750 static DEFINE_SPINLOCK(sid_lock);
751 
752 unsigned long alloc_sid(void)
753 {
754 	unsigned long index;
755 
756 	spin_lock(&sid_lock);
757 
758 	if (free_space_ids == 0) {
759 		if (dirty_space_ids != 0) {
760 			spin_unlock(&sid_lock);
761 			flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
762 			spin_lock(&sid_lock);
763 		}
764 		BUG_ON(free_space_ids == 0);
765 	}
766 
767 	free_space_ids--;
768 
769 	index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
770 	space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
771 	space_id_index = index;
772 
773 	spin_unlock(&sid_lock);
774 
775 	return index << SPACEID_SHIFT;
776 }
777 
778 void free_sid(unsigned long spaceid)
779 {
780 	unsigned long index = spaceid >> SPACEID_SHIFT;
781 	unsigned long *dirty_space_offset;
782 
783 	dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
784 	index &= (BITS_PER_LONG - 1);
785 
786 	spin_lock(&sid_lock);
787 
788 	BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
789 
790 	*dirty_space_offset |= (1L << index);
791 	dirty_space_ids++;
792 
793 	spin_unlock(&sid_lock);
794 }
795 
796 
797 #ifdef CONFIG_SMP
798 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
799 {
800 	int i;
801 
802 	/* NOTE: sid_lock must be held upon entry */
803 
804 	*ndirtyptr = dirty_space_ids;
805 	if (dirty_space_ids != 0) {
806 	    for (i = 0; i < SID_ARRAY_SIZE; i++) {
807 		dirty_array[i] = dirty_space_id[i];
808 		dirty_space_id[i] = 0;
809 	    }
810 	    dirty_space_ids = 0;
811 	}
812 
813 	return;
814 }
815 
816 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
817 {
818 	int i;
819 
820 	/* NOTE: sid_lock must be held upon entry */
821 
822 	if (ndirty != 0) {
823 		for (i = 0; i < SID_ARRAY_SIZE; i++) {
824 			space_id[i] ^= dirty_array[i];
825 		}
826 
827 		free_space_ids += ndirty;
828 		space_id_index = 0;
829 	}
830 }
831 
832 #else /* CONFIG_SMP */
833 
834 static void recycle_sids(void)
835 {
836 	int i;
837 
838 	/* NOTE: sid_lock must be held upon entry */
839 
840 	if (dirty_space_ids != 0) {
841 		for (i = 0; i < SID_ARRAY_SIZE; i++) {
842 			space_id[i] ^= dirty_space_id[i];
843 			dirty_space_id[i] = 0;
844 		}
845 
846 		free_space_ids += dirty_space_ids;
847 		dirty_space_ids = 0;
848 		space_id_index = 0;
849 	}
850 }
851 #endif
852 
853 /*
854  * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
855  * purged, we can safely reuse the space ids that were released but
856  * not flushed from the tlb.
857  */
858 
859 #ifdef CONFIG_SMP
860 
861 static unsigned long recycle_ndirty;
862 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
863 static unsigned int recycle_inuse;
864 
865 void flush_tlb_all(void)
866 {
867 	int do_recycle;
868 
869 	__inc_irq_stat(irq_tlb_count);
870 	do_recycle = 0;
871 	spin_lock(&sid_lock);
872 	if (dirty_space_ids > RECYCLE_THRESHOLD) {
873 	    BUG_ON(recycle_inuse);  /* FIXME: Use a semaphore/wait queue here */
874 	    get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
875 	    recycle_inuse++;
876 	    do_recycle++;
877 	}
878 	spin_unlock(&sid_lock);
879 	on_each_cpu(flush_tlb_all_local, NULL, 1);
880 	if (do_recycle) {
881 	    spin_lock(&sid_lock);
882 	    recycle_sids(recycle_ndirty,recycle_dirty_array);
883 	    recycle_inuse = 0;
884 	    spin_unlock(&sid_lock);
885 	}
886 }
887 #else
888 void flush_tlb_all(void)
889 {
890 	__inc_irq_stat(irq_tlb_count);
891 	spin_lock(&sid_lock);
892 	flush_tlb_all_local(NULL);
893 	recycle_sids();
894 	spin_unlock(&sid_lock);
895 }
896 #endif
897