2 * linux/arch/parisc/mm/init.c
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)
12 #include <linux/config.h>
14 #include <linux/module.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>
24 #include <asm/pgalloc.h>
26 #include <asm/pdc_chassis.h>
27 #include <asm/mmzone.h>
29 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
31 extern char _text; /* start of kernel code, defined by linker */
32 extern int data_start;
33 extern char _end; /* end of BSS, defined by linker */
34 extern char __init_begin, __init_end;
36 #ifdef CONFIG_DISCONTIGMEM
37 struct node_map_data node_data[MAX_NUMNODES];
38 bootmem_data_t bmem_data[MAX_NUMNODES];
39 unsigned char pfnnid_map[PFNNID_MAP_MAX];
42 static struct resource data_resource = {
43 .name = "Kernel data",
44 .flags = IORESOURCE_BUSY | IORESOURCE_MEM,
47 static struct resource code_resource = {
48 .name = "Kernel code",
49 .flags = IORESOURCE_BUSY | IORESOURCE_MEM,
52 static struct resource pdcdata_resource = {
53 .name = "PDC data (Page Zero)",
56 .flags = IORESOURCE_BUSY | IORESOURCE_MEM,
59 static struct resource sysram_resources[MAX_PHYSMEM_RANGES];
61 static unsigned long max_pfn;
63 /* The following array is initialized from the firmware specific
64 * information retrieved in kernel/inventory.c.
67 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES];
71 #define MAX_MEM (~0UL)
73 #define MAX_MEM (3584U*1024U*1024U)
74 #endif /* !__LP64__ */
76 static unsigned long mem_limit = MAX_MEM;
78 static void __init mem_limit_func(void)
82 extern char saved_command_line[];
84 /* We need this before __setup() functions are called */
87 for (cp = saved_command_line; *cp; ) {
88 if (memcmp(cp, "mem=", 4) == 0) {
90 limit = memparse(cp, &end);
95 while (*cp != ' ' && *cp)
102 if (limit < mem_limit)
106 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
108 static void __init setup_bootmem(void)
110 unsigned long bootmap_size;
111 unsigned long mem_max;
112 unsigned long bootmap_pages;
113 unsigned long bootmap_start_pfn;
114 unsigned long bootmap_pfn;
115 #ifndef CONFIG_DISCONTIGMEM
116 physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
119 int i, sysram_resource_count;
121 disable_sr_hashing(); /* Turn off space register hashing */
124 * Sort the ranges. Since the number of ranges is typically
125 * small, and performance is not an issue here, just do
126 * a simple insertion sort.
129 for (i = 1; i < npmem_ranges; i++) {
132 for (j = i; j > 0; j--) {
135 if (pmem_ranges[j-1].start_pfn <
136 pmem_ranges[j].start_pfn) {
140 tmp = pmem_ranges[j-1].start_pfn;
141 pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
142 pmem_ranges[j].start_pfn = tmp;
143 tmp = pmem_ranges[j-1].pages;
144 pmem_ranges[j-1].pages = pmem_ranges[j].pages;
145 pmem_ranges[j].pages = tmp;
149 #ifndef CONFIG_DISCONTIGMEM
151 * Throw out ranges that are too far apart (controlled by
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) {
160 printk("Large gap in memory detected (%ld pages). "
161 "Consider turning on CONFIG_DISCONTIGMEM\n",
162 pmem_ranges[i].start_pfn -
163 (pmem_ranges[i-1].start_pfn +
164 pmem_ranges[i-1].pages));
170 if (npmem_ranges > 1) {
172 /* Print the memory ranges */
174 printk(KERN_INFO "Memory Ranges:\n");
176 for (i = 0; i < npmem_ranges; i++) {
180 size = (pmem_ranges[i].pages << PAGE_SHIFT);
181 start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
182 printk(KERN_INFO "%2d) Start 0x%016lx End 0x%016lx Size %6ld Mb\n",
183 i,start, start + (size - 1), size >> 20);
187 sysram_resource_count = npmem_ranges;
188 for (i = 0; i < sysram_resource_count; i++) {
189 struct resource *res = &sysram_resources[i];
190 res->name = "System RAM";
191 res->start = pmem_ranges[i].start_pfn << PAGE_SHIFT;
192 res->end = res->start + (pmem_ranges[i].pages << PAGE_SHIFT)-1;
193 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
194 request_resource(&iomem_resource, res);
198 * For 32 bit kernels we limit the amount of memory we can
199 * support, in order to preserve enough kernel address space
200 * for other purposes. For 64 bit kernels we don't normally
201 * limit the memory, but this mechanism can be used to
202 * artificially limit the amount of memory (and it is written
203 * to work with multiple memory ranges).
206 mem_limit_func(); /* check for "mem=" argument */
210 for (i = 0; i < npmem_ranges; i++) {
213 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
214 if ((mem_max + rsize) > mem_limit) {
215 printk(KERN_WARNING "Memory truncated to %ld Mb\n", mem_limit >> 20);
216 if (mem_max == mem_limit)
219 pmem_ranges[i].pages = (mem_limit >> PAGE_SHIFT)
220 - (mem_max >> PAGE_SHIFT);
221 npmem_ranges = i + 1;
224 num_physpages += pmem_ranges[i].pages;
227 num_physpages += pmem_ranges[i].pages;
231 printk(KERN_INFO "Total Memory: %ld Mb\n",mem_max >> 20);
233 #ifndef CONFIG_DISCONTIGMEM
234 /* Merge the ranges, keeping track of the holes */
237 unsigned long end_pfn;
238 unsigned long hole_pages;
241 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
242 for (i = 1; i < npmem_ranges; i++) {
244 hole_pages = pmem_ranges[i].start_pfn - end_pfn;
246 pmem_holes[npmem_holes].start_pfn = end_pfn;
247 pmem_holes[npmem_holes++].pages = hole_pages;
248 end_pfn += hole_pages;
250 end_pfn += pmem_ranges[i].pages;
253 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
259 for (i = 0; i < npmem_ranges; i++)
260 bootmap_pages += bootmem_bootmap_pages(pmem_ranges[i].pages);
262 bootmap_start_pfn = PAGE_ALIGN(__pa((unsigned long) &_end)) >> PAGE_SHIFT;
264 #ifdef CONFIG_DISCONTIGMEM
265 for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
266 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
267 NODE_DATA(i)->bdata = &bmem_data[i];
269 memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
271 numnodes = npmem_ranges;
273 for (i = 0; i < npmem_ranges; i++)
278 * Initialize and free the full range of memory in each range.
279 * Note that the only writing these routines do are to the bootmap,
280 * and we've made sure to locate the bootmap properly so that they
281 * won't be writing over anything important.
284 bootmap_pfn = bootmap_start_pfn;
286 for (i = 0; i < npmem_ranges; i++) {
287 unsigned long start_pfn;
288 unsigned long npages;
290 start_pfn = pmem_ranges[i].start_pfn;
291 npages = pmem_ranges[i].pages;
293 bootmap_size = init_bootmem_node(NODE_DATA(i),
296 (start_pfn + npages) );
297 free_bootmem_node(NODE_DATA(i),
298 (start_pfn << PAGE_SHIFT),
299 (npages << PAGE_SHIFT) );
300 bootmap_pfn += (bootmap_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
301 if ((start_pfn + npages) > max_pfn)
302 max_pfn = start_pfn + npages;
305 if ((bootmap_pfn - bootmap_start_pfn) != bootmap_pages) {
306 printk(KERN_WARNING "WARNING! bootmap sizing is messed up!\n");
310 /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
312 #define PDC_CONSOLE_IO_IODC_SIZE 32768
314 reserve_bootmem_node(NODE_DATA(0), 0UL,
315 (unsigned long)(PAGE0->mem_free + PDC_CONSOLE_IO_IODC_SIZE));
316 reserve_bootmem_node(NODE_DATA(0),__pa((unsigned long)&_text),
317 (unsigned long)(&_end - &_text));
318 reserve_bootmem_node(NODE_DATA(0), (bootmap_start_pfn << PAGE_SHIFT),
319 ((bootmap_pfn - bootmap_start_pfn) << PAGE_SHIFT));
321 #ifndef CONFIG_DISCONTIGMEM
323 /* reserve the holes */
325 for (i = 0; i < npmem_holes; i++) {
326 reserve_bootmem_node(NODE_DATA(0),
327 (pmem_holes[i].start_pfn << PAGE_SHIFT),
328 (pmem_holes[i].pages << PAGE_SHIFT));
332 #ifdef CONFIG_BLK_DEV_INITRD
334 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
335 if (__pa(initrd_start) < mem_max) {
336 unsigned long initrd_reserve;
338 if (__pa(initrd_end) > mem_max) {
339 initrd_reserve = mem_max - __pa(initrd_start);
341 initrd_reserve = initrd_end - initrd_start;
343 initrd_below_start_ok = 1;
344 printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
346 reserve_bootmem_node(NODE_DATA(0),__pa(initrd_start), initrd_reserve);
351 data_resource.start = virt_to_phys(&data_start);
352 data_resource.end = virt_to_phys(&_end)-1;
353 code_resource.start = virt_to_phys(&_text);
354 code_resource.end = virt_to_phys(&data_start)-1;
356 /* We don't know which region the kernel will be in, so try
359 for (i = 0; i < sysram_resource_count; i++) {
360 struct resource *res = &sysram_resources[i];
361 request_resource(res, &code_resource);
362 request_resource(res, &data_resource);
364 request_resource(&sysram_resources[0], &pdcdata_resource);
367 void free_initmem(void)
371 printk(KERN_INFO "NOT FREEING INITMEM (%dk)\n",
372 (&__init_end - &__init_begin) >> 10);
377 printk(KERN_INFO "Freeing unused kernel memory: ");
380 /* Attempt to catch anyone trying to execute code here
381 * by filling the page with BRK insns.
383 * If we disable interrupts for all CPUs, then IPI stops working.
384 * Kinda breaks the global cache flushing.
388 memset(&__init_begin, 0x00,
389 (unsigned long)&__init_end - (unsigned long)&__init_begin);
392 asm volatile("sync" : : );
393 flush_icache_range((unsigned long)&__init_begin, (unsigned long)&__init_end);
394 asm volatile("sync" : : );
399 addr = (unsigned long)(&__init_begin);
400 for (; addr < (unsigned long)(&__init_end); addr += PAGE_SIZE) {
401 ClearPageReserved(virt_to_page(addr));
402 set_page_count(virt_to_page(addr), 1);
408 /* set up a new led state on systems shipped LED State panel */
409 pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
411 printk("%luk freed\n", (unsigned long)(&__init_end - &__init_begin) >> 10);
416 * Just an arbitrary offset to serve as a "hole" between mapping areas
417 * (between top of physical memory and a potential pcxl dma mapping
418 * area, and below the vmalloc mapping area).
420 * The current 32K value just means that there will be a 32K "hole"
421 * between mapping areas. That means that any out-of-bounds memory
422 * accesses will hopefully be caught. The vmalloc() routines leaves
423 * a hole of 4kB between each vmalloced area for the same reason.
426 /* Leave room for gateway page expansion */
427 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
428 #error KERNEL_MAP_START is in gateway reserved region
430 #define MAP_START (KERNEL_MAP_START)
432 #define VM_MAP_OFFSET (32*1024)
433 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
434 & ~(VM_MAP_OFFSET-1)))
437 EXPORT_SYMBOL(vmalloc_start);
440 unsigned long pcxl_dma_start;
443 void __init mem_init(void)
445 high_memory = __va((max_pfn << PAGE_SHIFT));
447 #ifndef CONFIG_DISCONTIGMEM
448 max_mapnr = page_to_pfn(virt_to_page(high_memory - 1)) + 1;
449 mem_map = zone_table[ZONE_DMA]->zone_mem_map;
450 totalram_pages += free_all_bootmem();
455 for (i = 0; i < npmem_ranges; i++)
456 totalram_pages += free_all_bootmem_node(NODE_DATA(i));
460 printk(KERN_INFO "Memory: %luk available\n", num_physpages << (PAGE_SHIFT-10));
463 if (hppa_dma_ops == &pcxl_dma_ops) {
464 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
465 vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start + PCXL_DMA_MAP_SIZE);
468 vmalloc_start = SET_MAP_OFFSET(MAP_START);
471 vmalloc_start = SET_MAP_OFFSET(MAP_START);
476 int do_check_pgt_cache(int low, int high)
481 unsigned long *empty_zero_page;
485 int i,free = 0,total = 0,reserved = 0;
486 int shared = 0, cached = 0;
488 printk(KERN_INFO "Mem-info:\n");
490 printk(KERN_INFO "Free swap: %6ldkB\n",
491 nr_swap_pages<<(PAGE_SHIFT-10));
492 #ifndef CONFIG_DISCONTIGMEM
496 if (PageReserved(mem_map+i))
498 else if (PageSwapCache(mem_map+i))
500 else if (!page_count(&mem_map[i]))
503 shared += page_count(&mem_map[i]) - 1;
506 for (i = 0; i < npmem_ranges; i++) {
509 for (j = node_start_pfn(i); j < node_end_pfn(i); j++) {
512 p = node_mem_map(i) + j - node_start_pfn(i);
517 else if (PageSwapCache(p))
519 else if (!page_count(p))
522 shared += page_count(p) - 1;
526 printk(KERN_INFO "%d pages of RAM\n", total);
527 printk(KERN_INFO "%d reserved pages\n", reserved);
528 printk(KERN_INFO "%d pages shared\n", shared);
529 printk(KERN_INFO "%d pages swap cached\n", cached);
532 #ifdef CONFIG_DISCONTIGMEM
537 for (i = 0; i < npmem_ranges; i++) {
538 for (j = 0; j < MAX_NR_ZONES; j++) {
539 zl = NODE_DATA(i)->node_zonelists + j;
541 printk("Zone list for zone %d on node %d: ", j, i);
542 for (k = 0; zl->zones[k] != NULL; k++)
543 printk("[%d/%s] ", zl->zones[k]->zone_pgdat->node_id, zl->zones[k]->name);
552 static void __init map_pages(unsigned long start_vaddr, unsigned long start_paddr, unsigned long size, pgprot_t pgprot)
557 unsigned long end_paddr;
558 unsigned long start_pmd;
559 unsigned long start_pte;
562 unsigned long address;
563 unsigned long ro_start;
564 unsigned long ro_end;
565 unsigned long fv_addr;
566 unsigned long gw_addr;
567 extern const unsigned long fault_vector_20;
568 extern void * const linux_gateway_page;
570 ro_start = __pa((unsigned long)&_text);
571 ro_end = __pa((unsigned long)&data_start);
572 fv_addr = __pa((unsigned long)&fault_vector_20) & PAGE_MASK;
573 gw_addr = __pa((unsigned long)&linux_gateway_page) & PAGE_MASK;
575 end_paddr = start_paddr + size;
577 pg_dir = pgd_offset_k(start_vaddr);
579 #if PTRS_PER_PMD == 1
582 start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
584 start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
586 address = start_paddr;
587 while (address < end_paddr) {
588 #if PTRS_PER_PMD == 1
589 pmd = (pmd_t *)__pa(pg_dir);
591 pmd = (pmd_t *)pgd_address(*pg_dir);
594 * pmd is physical at this point
598 pmd = (pmd_t *) alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE << PMD_ORDER);
599 pmd = (pmd_t *) __pa(pmd);
602 pgd_populate(NULL, pg_dir, __va(pmd));
606 /* now change pmd to kernel virtual addresses */
608 pmd = (pmd_t *)__va(pmd) + start_pmd;
609 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++,pmd++) {
612 * pg_table is physical at this point
615 pg_table = (pte_t *)pmd_address(*pmd);
618 alloc_bootmem_low_pages_node(NODE_DATA(0),PAGE_SIZE);
619 pg_table = (pte_t *) __pa(pg_table);
622 pmd_populate_kernel(NULL, pmd, __va(pg_table));
624 /* now change pg_table to kernel virtual addresses */
626 pg_table = (pte_t *) __va(pg_table) + start_pte;
627 for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++,pg_table++) {
631 * Map the fault vector writable so we can
632 * write the HPMC checksum.
634 if (address >= ro_start && address < ro_end
635 && address != fv_addr
636 && address != gw_addr)
637 pte = __mk_pte(address, PAGE_KERNEL_RO);
639 pte = __mk_pte(address, pgprot);
641 if (address >= end_paddr)
644 set_pte(pg_table, pte);
646 address += PAGE_SIZE;
650 if (address >= end_paddr)
658 * pagetable_init() sets up the page tables
660 * Note that gateway_init() places the Linux gateway page at page 0.
661 * Since gateway pages cannot be dereferenced this has the desirable
662 * side effect of trapping those pesky NULL-reference errors in the
665 static void __init pagetable_init(void)
669 /* Map each physical memory range to its kernel vaddr */
671 for (range = 0; range < npmem_ranges; range++) {
672 unsigned long start_paddr;
673 unsigned long end_paddr;
676 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
677 end_paddr = start_paddr + (pmem_ranges[range].pages << PAGE_SHIFT);
678 size = pmem_ranges[range].pages << PAGE_SHIFT;
680 map_pages((unsigned long)__va(start_paddr), start_paddr,
684 #ifdef CONFIG_BLK_DEV_INITRD
685 if (initrd_end && initrd_end > mem_limit) {
686 printk("initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
687 map_pages(initrd_start, __pa(initrd_start),
688 initrd_end - initrd_start, PAGE_KERNEL);
692 empty_zero_page = alloc_bootmem_pages(PAGE_SIZE);
693 memset(empty_zero_page, 0, PAGE_SIZE);
696 static void __init gateway_init(void)
698 unsigned long linux_gateway_page_addr;
699 /* FIXME: This is 'const' in order to trick the compiler
700 into not treating it as DP-relative data. */
701 extern void * const linux_gateway_page;
703 linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
706 * Setup Linux Gateway page.
708 * The Linux gateway page will reside in kernel space (on virtual
709 * page 0), so it doesn't need to be aliased into user space.
712 map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
713 PAGE_SIZE, PAGE_GATEWAY);
718 map_hpux_gateway_page(struct task_struct *tsk, struct mm_struct *mm)
723 unsigned long start_pmd;
724 unsigned long start_pte;
725 unsigned long address;
726 unsigned long hpux_gw_page_addr;
727 /* FIXME: This is 'const' in order to trick the compiler
728 into not treating it as DP-relative data. */
729 extern void * const hpux_gateway_page;
731 hpux_gw_page_addr = HPUX_GATEWAY_ADDR & PAGE_MASK;
734 * Setup HP-UX Gateway page.
736 * The HP-UX gateway page resides in the user address space,
737 * so it needs to be aliased into each process.
740 pg_dir = pgd_offset(mm,hpux_gw_page_addr);
742 #if PTRS_PER_PMD == 1
745 start_pmd = ((hpux_gw_page_addr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
747 start_pte = ((hpux_gw_page_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
749 address = __pa(&hpux_gateway_page);
750 #if PTRS_PER_PMD == 1
751 pmd = (pmd_t *)__pa(pg_dir);
753 pmd = (pmd_t *) (PAGE_MASK & pgd_val(*pg_dir));
756 * pmd is physical at this point
760 pmd = (pmd_t *) get_zeroed_page(GFP_KERNEL);
761 pmd = (pmd_t *) __pa(pmd);
764 pgd_val(*pg_dir) = _PAGE_TABLE | (unsigned long) pmd;
766 /* now change pmd to kernel virtual addresses */
768 pmd = (pmd_t *)__va(pmd) + start_pmd;
771 * pg_table is physical at this point
774 pg_table = (pte_t *) (PAGE_MASK & pmd_val(*pmd));
776 pg_table = (pte_t *) __pa(get_zeroed_page(GFP_KERNEL));
778 pmd_val(*pmd) = _PAGE_TABLE | (unsigned long) pg_table;
780 /* now change pg_table to kernel virtual addresses */
782 pg_table = (pte_t *) __va(pg_table) + start_pte;
783 set_pte(pg_table, __mk_pte(address, PAGE_GATEWAY));
785 EXPORT_SYMBOL(map_hpux_gateway_page);
788 extern void flush_tlb_all_local(void);
790 void __init paging_init(void)
797 flush_cache_all_local(); /* start with known state */
798 flush_tlb_all_local();
800 for (i = 0; i < npmem_ranges; i++) {
801 unsigned long zones_size[MAX_NR_ZONES] = { 0, 0, 0 };
803 /* We have an IOMMU, so all memory can go into a single
805 zones_size[ZONE_DMA] = pmem_ranges[i].pages;
807 free_area_init_node(i, NODE_DATA(i), zones_size,
808 pmem_ranges[i].start_pfn, 0);
810 #ifdef CONFIG_DISCONTIGMEM
813 for (j = (node_start_pfn(i) >> PFNNID_SHIFT);
814 j <= (node_end_pfn(i) >> PFNNID_SHIFT);
826 * Currently, all PA20 chips have 18 bit protection id's, which is the
827 * limiting factor (space ids are 32 bits).
830 #define NR_SPACE_IDS 262144
835 * Currently we have a one-to-one relationship between space id's and
836 * protection id's. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
837 * support 15 bit protection id's, so that is the limiting factor.
838 * PCXT' has 18 bit protection id's, but only 16 bit spaceids, so it's
839 * probably not worth the effort for a special case here.
842 #define NR_SPACE_IDS 32768
844 #endif /* !CONFIG_PA20 */
846 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
847 #define SID_ARRAY_SIZE (NR_SPACE_IDS / (8 * sizeof(long)))
849 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
850 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
851 static unsigned long space_id_index;
852 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
853 static unsigned long dirty_space_ids = 0;
855 static spinlock_t sid_lock = SPIN_LOCK_UNLOCKED;
857 unsigned long alloc_sid(void)
861 spin_lock(&sid_lock);
863 if (free_space_ids == 0) {
864 if (dirty_space_ids != 0) {
865 spin_unlock(&sid_lock);
866 flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
867 spin_lock(&sid_lock);
869 if (free_space_ids == 0)
875 index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
876 space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
877 space_id_index = index;
879 spin_unlock(&sid_lock);
881 return index << SPACEID_SHIFT;
884 void free_sid(unsigned long spaceid)
886 unsigned long index = spaceid >> SPACEID_SHIFT;
887 unsigned long *dirty_space_offset;
889 dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
890 index &= (BITS_PER_LONG - 1);
892 spin_lock(&sid_lock);
894 if (*dirty_space_offset & (1L << index))
895 BUG(); /* attempt to free space id twice */
897 *dirty_space_offset |= (1L << index);
900 spin_unlock(&sid_lock);
905 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
909 /* NOTE: sid_lock must be held upon entry */
911 *ndirtyptr = dirty_space_ids;
912 if (dirty_space_ids != 0) {
913 for (i = 0; i < SID_ARRAY_SIZE; i++) {
914 dirty_array[i] = dirty_space_id[i];
915 dirty_space_id[i] = 0;
923 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
927 /* NOTE: sid_lock must be held upon entry */
930 for (i = 0; i < SID_ARRAY_SIZE; i++) {
931 space_id[i] ^= dirty_array[i];
934 free_space_ids += ndirty;
939 #else /* CONFIG_SMP */
941 static void recycle_sids(void)
945 /* NOTE: sid_lock must be held upon entry */
947 if (dirty_space_ids != 0) {
948 for (i = 0; i < SID_ARRAY_SIZE; i++) {
949 space_id[i] ^= dirty_space_id[i];
950 dirty_space_id[i] = 0;
953 free_space_ids += dirty_space_ids;
961 * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
962 * purged, we can safely reuse the space ids that were released but
963 * not flushed from the tlb.
968 static unsigned long recycle_ndirty;
969 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
970 static unsigned int recycle_inuse = 0;
972 void flush_tlb_all(void)
977 spin_lock(&sid_lock);
978 if (dirty_space_ids > RECYCLE_THRESHOLD) {
980 BUG(); /* FIXME: Use a semaphore/wait queue here */
982 get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
986 spin_unlock(&sid_lock);
987 on_each_cpu((void (*)(void *))flush_tlb_all_local, NULL, 1, 1);
989 spin_lock(&sid_lock);
990 recycle_sids(recycle_ndirty,recycle_dirty_array);
992 spin_unlock(&sid_lock);
996 void flush_tlb_all(void)
998 spin_lock(&sid_lock);
999 flush_tlb_all_local();
1001 spin_unlock(&sid_lock);
1005 #ifdef CONFIG_BLK_DEV_INITRD
1006 void free_initrd_mem(unsigned long start, unsigned long end)
1010 printk(KERN_INFO "Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1011 for (; start < end; start += PAGE_SIZE) {
1012 ClearPageReserved(virt_to_page(start));
1013 set_page_count(virt_to_page(start), 1);