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)
9 * Copyright 2006-2007 Helge Deller (deller@gmx.de)
14 #include <linux/module.h>
16 #include <linux/bootmem.h>
17 #include <linux/memblock.h>
18 #include <linux/gfp.h>
19 #include <linux/delay.h>
20 #include <linux/init.h>
21 #include <linux/pci.h> /* for hppa_dma_ops and pcxl_dma_ops */
22 #include <linux/initrd.h>
23 #include <linux/swap.h>
24 #include <linux/unistd.h>
25 #include <linux/nodemask.h> /* for node_online_map */
26 #include <linux/pagemap.h> /* for release_pages */
27 #include <linux/compat.h>
29 #include <asm/pgalloc.h>
30 #include <asm/pgtable.h>
32 #include <asm/pdc_chassis.h>
33 #include <asm/mmzone.h>
34 #include <asm/sections.h>
35 #include <asm/msgbuf.h>
37 extern int data_start;
38 extern void parisc_kernel_start(void); /* Kernel entry point in head.S */
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)));
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)));
52 #ifdef CONFIG_DISCONTIGMEM
53 struct node_map_data node_data[MAX_NUMNODES] __read_mostly;
54 signed char pfnnid_map[PFNNID_MAP_MAX] __read_mostly;
57 static struct resource data_resource = {
58 .name = "Kernel data",
59 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
62 static struct resource code_resource = {
63 .name = "Kernel code",
64 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
67 static struct resource pdcdata_resource = {
68 .name = "PDC data (Page Zero)",
71 .flags = IORESOURCE_BUSY | IORESOURCE_MEM,
74 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
76 /* The following array is initialized from the firmware specific
77 * information retrieved in kernel/inventory.c.
80 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
81 int npmem_ranges __read_mostly;
84 * get_memblock() allocates pages via memblock.
85 * We can't use memblock_find_in_range(0, KERNEL_INITIAL_SIZE) here since it
86 * doesn't allocate from bottom to top which is needed because we only created
87 * the initial mapping up to KERNEL_INITIAL_SIZE in the assembly bootup code.
89 static void * __init get_memblock(unsigned long size)
91 static phys_addr_t search_addr __initdata;
95 search_addr = PAGE_ALIGN(__pa((unsigned long) &_end));
96 search_addr = ALIGN(search_addr, size);
97 while (!memblock_is_region_memory(search_addr, size) ||
98 memblock_is_region_reserved(search_addr, size)) {
104 memblock_reserve(phys, size);
106 panic("get_memblock() failed.\n");
108 memset(__va(phys), 0, size);
114 #define MAX_MEM (~0UL)
115 #else /* !CONFIG_64BIT */
116 #define MAX_MEM (3584U*1024U*1024U)
117 #endif /* !CONFIG_64BIT */
119 static unsigned long mem_limit __read_mostly = MAX_MEM;
121 static void __init mem_limit_func(void)
126 /* We need this before __setup() functions are called */
129 for (cp = boot_command_line; *cp; ) {
130 if (memcmp(cp, "mem=", 4) == 0) {
132 limit = memparse(cp, &end);
137 while (*cp != ' ' && *cp)
144 if (limit < mem_limit)
148 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
150 static void __init setup_bootmem(void)
152 unsigned long mem_max;
153 #ifndef CONFIG_DISCONTIGMEM
154 physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
157 int i, sysram_resource_count;
159 disable_sr_hashing(); /* Turn off space register hashing */
162 * Sort the ranges. Since the number of ranges is typically
163 * small, and performance is not an issue here, just do
164 * a simple insertion sort.
167 for (i = 1; i < npmem_ranges; i++) {
170 for (j = i; j > 0; j--) {
173 if (pmem_ranges[j-1].start_pfn <
174 pmem_ranges[j].start_pfn) {
178 tmp = pmem_ranges[j-1].start_pfn;
179 pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
180 pmem_ranges[j].start_pfn = tmp;
181 tmp = pmem_ranges[j-1].pages;
182 pmem_ranges[j-1].pages = pmem_ranges[j].pages;
183 pmem_ranges[j].pages = tmp;
187 #ifndef CONFIG_DISCONTIGMEM
189 * Throw out ranges that are too far apart (controlled by
193 for (i = 1; i < npmem_ranges; i++) {
194 if (pmem_ranges[i].start_pfn -
195 (pmem_ranges[i-1].start_pfn +
196 pmem_ranges[i-1].pages) > MAX_GAP) {
198 printk("Large gap in memory detected (%ld pages). "
199 "Consider turning on CONFIG_DISCONTIGMEM\n",
200 pmem_ranges[i].start_pfn -
201 (pmem_ranges[i-1].start_pfn +
202 pmem_ranges[i-1].pages));
208 /* Print the memory ranges */
209 pr_info("Memory Ranges:\n");
211 for (i = 0; i < npmem_ranges; i++) {
212 struct resource *res = &sysram_resources[i];
216 size = (pmem_ranges[i].pages << PAGE_SHIFT);
217 start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
218 pr_info("%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
219 i, start, start + (size - 1), size >> 20);
221 /* request memory resource */
222 res->name = "System RAM";
224 res->end = start + size - 1;
225 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
226 request_resource(&iomem_resource, res);
229 sysram_resource_count = npmem_ranges;
232 * For 32 bit kernels we limit the amount of memory we can
233 * support, in order to preserve enough kernel address space
234 * for other purposes. For 64 bit kernels we don't normally
235 * limit the memory, but this mechanism can be used to
236 * artificially limit the amount of memory (and it is written
237 * to work with multiple memory ranges).
240 mem_limit_func(); /* check for "mem=" argument */
243 for (i = 0; i < npmem_ranges; i++) {
246 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
247 if ((mem_max + rsize) > mem_limit) {
248 printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
249 if (mem_max == mem_limit)
252 pmem_ranges[i].pages = (mem_limit >> PAGE_SHIFT)
253 - (mem_max >> PAGE_SHIFT);
254 npmem_ranges = i + 1;
262 printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
264 #ifndef CONFIG_DISCONTIGMEM
265 /* Merge the ranges, keeping track of the holes */
268 unsigned long end_pfn;
269 unsigned long hole_pages;
272 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
273 for (i = 1; i < npmem_ranges; i++) {
275 hole_pages = pmem_ranges[i].start_pfn - end_pfn;
277 pmem_holes[npmem_holes].start_pfn = end_pfn;
278 pmem_holes[npmem_holes++].pages = hole_pages;
279 end_pfn += hole_pages;
281 end_pfn += pmem_ranges[i].pages;
284 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
289 #ifdef CONFIG_DISCONTIGMEM
290 for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
291 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
293 memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
295 for (i = 0; i < npmem_ranges; i++) {
296 node_set_state(i, N_NORMAL_MEMORY);
302 * Initialize and free the full range of memory in each range.
306 for (i = 0; i < npmem_ranges; i++) {
307 unsigned long start_pfn;
308 unsigned long npages;
312 start_pfn = pmem_ranges[i].start_pfn;
313 npages = pmem_ranges[i].pages;
315 start = start_pfn << PAGE_SHIFT;
316 size = npages << PAGE_SHIFT;
318 /* add system RAM memblock */
319 memblock_add(start, size);
321 if ((start_pfn + npages) > max_pfn)
322 max_pfn = start_pfn + npages;
325 /* IOMMU is always used to access "high mem" on those boxes
326 * that can support enough mem that a PCI device couldn't
327 * directly DMA to any physical addresses.
328 * ISA DMA support will need to revisit this.
330 max_low_pfn = max_pfn;
332 /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
334 #define PDC_CONSOLE_IO_IODC_SIZE 32768
336 memblock_reserve(0UL, (unsigned long)(PAGE0->mem_free +
337 PDC_CONSOLE_IO_IODC_SIZE));
338 memblock_reserve(__pa(KERNEL_BINARY_TEXT_START),
339 (unsigned long)(_end - KERNEL_BINARY_TEXT_START));
341 #ifndef CONFIG_DISCONTIGMEM
343 /* reserve the holes */
345 for (i = 0; i < npmem_holes; i++) {
346 memblock_reserve((pmem_holes[i].start_pfn << PAGE_SHIFT),
347 (pmem_holes[i].pages << PAGE_SHIFT));
351 #ifdef CONFIG_BLK_DEV_INITRD
353 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
354 if (__pa(initrd_start) < mem_max) {
355 unsigned long initrd_reserve;
357 if (__pa(initrd_end) > mem_max) {
358 initrd_reserve = mem_max - __pa(initrd_start);
360 initrd_reserve = initrd_end - initrd_start;
362 initrd_below_start_ok = 1;
363 printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
365 memblock_reserve(__pa(initrd_start), initrd_reserve);
370 data_resource.start = virt_to_phys(&data_start);
371 data_resource.end = virt_to_phys(_end) - 1;
372 code_resource.start = virt_to_phys(_text);
373 code_resource.end = virt_to_phys(&data_start)-1;
375 /* We don't know which region the kernel will be in, so try
378 for (i = 0; i < sysram_resource_count; i++) {
379 struct resource *res = &sysram_resources[i];
380 request_resource(res, &code_resource);
381 request_resource(res, &data_resource);
383 request_resource(&sysram_resources[0], &pdcdata_resource);
386 static int __init parisc_text_address(unsigned long vaddr)
388 static unsigned long head_ptr __initdata;
391 head_ptr = PAGE_MASK & (unsigned long)
392 dereference_function_descriptor(&parisc_kernel_start);
394 return core_kernel_text(vaddr) || vaddr == head_ptr;
397 static void __init map_pages(unsigned long start_vaddr,
398 unsigned long start_paddr, unsigned long size,
399 pgprot_t pgprot, int force)
404 unsigned long end_paddr;
405 unsigned long start_pmd;
406 unsigned long start_pte;
409 unsigned long address;
411 unsigned long ro_start;
412 unsigned long ro_end;
413 unsigned long kernel_end;
415 ro_start = __pa((unsigned long)_text);
416 ro_end = __pa((unsigned long)&data_start);
417 kernel_end = __pa((unsigned long)&_end);
419 end_paddr = start_paddr + size;
421 pg_dir = pgd_offset_k(start_vaddr);
423 #if PTRS_PER_PMD == 1
426 start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
428 start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
430 address = start_paddr;
432 while (address < end_paddr) {
433 #if PTRS_PER_PMD == 1
434 pmd = (pmd_t *)__pa(pg_dir);
436 pmd = (pmd_t *)pgd_address(*pg_dir);
439 * pmd is physical at this point
443 pmd = (pmd_t *) get_memblock(PAGE_SIZE << PMD_ORDER);
444 pmd = (pmd_t *) __pa(pmd);
447 pgd_populate(NULL, pg_dir, __va(pmd));
451 /* now change pmd to kernel virtual addresses */
453 pmd = (pmd_t *)__va(pmd) + start_pmd;
454 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
457 * pg_table is physical at this point
460 pg_table = (pte_t *)pmd_address(*pmd);
462 pg_table = (pte_t *) get_memblock(PAGE_SIZE);
463 pg_table = (pte_t *) __pa(pg_table);
466 pmd_populate_kernel(NULL, pmd, __va(pg_table));
468 /* now change pg_table to kernel virtual addresses */
470 pg_table = (pte_t *) __va(pg_table) + start_pte;
471 for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
475 pte = __mk_pte(address, pgprot);
476 else if (parisc_text_address(vaddr)) {
477 pte = __mk_pte(address, PAGE_KERNEL_EXEC);
478 if (address >= ro_start && address < kernel_end)
479 pte = pte_mkhuge(pte);
482 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
483 if (address >= ro_start && address < ro_end) {
484 pte = __mk_pte(address, PAGE_KERNEL_EXEC);
485 pte = pte_mkhuge(pte);
489 pte = __mk_pte(address, pgprot);
490 if (address >= ro_start && address < kernel_end)
491 pte = pte_mkhuge(pte);
494 if (address >= end_paddr)
497 set_pte(pg_table, pte);
499 address += PAGE_SIZE;
504 if (address >= end_paddr)
511 void free_initmem(void)
513 unsigned long init_begin = (unsigned long)__init_begin;
514 unsigned long init_end = (unsigned long)__init_end;
516 /* The init text pages are marked R-X. We have to
517 * flush the icache and mark them RW-
519 * This is tricky, because map_pages is in the init section.
520 * Do a dummy remap of the data section first (the data
521 * section is already PAGE_KERNEL) to pull in the TLB entries
523 map_pages(init_begin, __pa(init_begin), init_end - init_begin,
525 /* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
527 map_pages(init_begin, __pa(init_begin), init_end - init_begin,
530 /* force the kernel to see the new TLB entries */
531 __flush_tlb_range(0, init_begin, init_end);
533 /* finally dump all the instructions which were cached, since the
534 * pages are no-longer executable */
535 flush_icache_range(init_begin, init_end);
537 free_initmem_default(POISON_FREE_INITMEM);
539 /* set up a new led state on systems shipped LED State panel */
540 pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
544 #ifdef CONFIG_DEBUG_RODATA
545 void mark_rodata_ro(void)
547 /* rodata memory was already mapped with KERNEL_RO access rights by
548 pagetable_init() and map_pages(). No need to do additional stuff here */
549 printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
550 (unsigned long)(__end_rodata - __start_rodata) >> 10);
556 * Just an arbitrary offset to serve as a "hole" between mapping areas
557 * (between top of physical memory and a potential pcxl dma mapping
558 * area, and below the vmalloc mapping area).
560 * The current 32K value just means that there will be a 32K "hole"
561 * between mapping areas. That means that any out-of-bounds memory
562 * accesses will hopefully be caught. The vmalloc() routines leaves
563 * a hole of 4kB between each vmalloced area for the same reason.
566 /* Leave room for gateway page expansion */
567 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
568 #error KERNEL_MAP_START is in gateway reserved region
570 #define MAP_START (KERNEL_MAP_START)
572 #define VM_MAP_OFFSET (32*1024)
573 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
574 & ~(VM_MAP_OFFSET-1)))
576 void *parisc_vmalloc_start __read_mostly;
577 EXPORT_SYMBOL(parisc_vmalloc_start);
580 unsigned long pcxl_dma_start __read_mostly;
583 void __init mem_init(void)
585 /* Do sanity checks on IPC (compat) structures */
586 BUILD_BUG_ON(sizeof(struct ipc64_perm) != 48);
588 BUILD_BUG_ON(sizeof(struct semid64_ds) != 80);
589 BUILD_BUG_ON(sizeof(struct msqid64_ds) != 104);
590 BUILD_BUG_ON(sizeof(struct shmid64_ds) != 104);
593 BUILD_BUG_ON(sizeof(struct compat_ipc64_perm) != sizeof(struct ipc64_perm));
594 BUILD_BUG_ON(sizeof(struct compat_semid64_ds) != 80);
595 BUILD_BUG_ON(sizeof(struct compat_msqid64_ds) != 104);
596 BUILD_BUG_ON(sizeof(struct compat_shmid64_ds) != 104);
599 /* Do sanity checks on page table constants */
600 BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
601 BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
602 BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
603 BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
606 high_memory = __va((max_pfn << PAGE_SHIFT));
607 set_max_mapnr(max_low_pfn);
611 if (hppa_dma_ops == &pcxl_dma_ops) {
612 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
613 parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
614 + PCXL_DMA_MAP_SIZE);
617 parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
620 parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
623 mem_init_print_info(NULL);
624 #ifdef CONFIG_DEBUG_KERNEL /* double-sanity-check paranoia */
625 printk("virtual kernel memory layout:\n"
626 " vmalloc : 0x%p - 0x%p (%4ld MB)\n"
627 " memory : 0x%p - 0x%p (%4ld MB)\n"
628 " .init : 0x%p - 0x%p (%4ld kB)\n"
629 " .data : 0x%p - 0x%p (%4ld kB)\n"
630 " .text : 0x%p - 0x%p (%4ld kB)\n",
632 (void*)VMALLOC_START, (void*)VMALLOC_END,
633 (VMALLOC_END - VMALLOC_START) >> 20,
635 __va(0), high_memory,
636 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
638 __init_begin, __init_end,
639 ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
642 ((unsigned long)_edata - (unsigned long)_etext) >> 10,
645 ((unsigned long)_etext - (unsigned long)_text) >> 10);
649 unsigned long *empty_zero_page __read_mostly;
650 EXPORT_SYMBOL(empty_zero_page);
652 void show_mem(unsigned int filter)
654 int total = 0,reserved = 0;
657 printk(KERN_INFO "Mem-info:\n");
658 show_free_areas(filter);
660 for_each_online_pgdat(pgdat) {
664 pgdat_resize_lock(pgdat, &flags);
665 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
666 struct zone *zone = &pgdat->node_zones[zoneid];
667 if (!populated_zone(zone))
670 total += zone->present_pages;
671 reserved = zone->present_pages - zone->managed_pages;
673 pgdat_resize_unlock(pgdat, &flags);
676 printk(KERN_INFO "%d pages of RAM\n", total);
677 printk(KERN_INFO "%d reserved pages\n", reserved);
679 #ifdef CONFIG_DISCONTIGMEM
684 for (i = 0; i < npmem_ranges; i++) {
685 zl = node_zonelist(i, 0);
686 for (j = 0; j < MAX_NR_ZONES; j++) {
690 printk("Zone list for zone %d on node %d: ", j, i);
691 for_each_zone_zonelist(zone, z, zl, j)
692 printk("[%d/%s] ", zone_to_nid(zone),
702 * pagetable_init() sets up the page tables
704 * Note that gateway_init() places the Linux gateway page at page 0.
705 * Since gateway pages cannot be dereferenced this has the desirable
706 * side effect of trapping those pesky NULL-reference errors in the
709 static void __init pagetable_init(void)
713 /* Map each physical memory range to its kernel vaddr */
715 for (range = 0; range < npmem_ranges; range++) {
716 unsigned long start_paddr;
717 unsigned long end_paddr;
720 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
721 size = pmem_ranges[range].pages << PAGE_SHIFT;
722 end_paddr = start_paddr + size;
724 map_pages((unsigned long)__va(start_paddr), start_paddr,
725 size, PAGE_KERNEL, 0);
728 #ifdef CONFIG_BLK_DEV_INITRD
729 if (initrd_end && initrd_end > mem_limit) {
730 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
731 map_pages(initrd_start, __pa(initrd_start),
732 initrd_end - initrd_start, PAGE_KERNEL, 0);
736 empty_zero_page = get_memblock(PAGE_SIZE);
739 static void __init gateway_init(void)
741 unsigned long linux_gateway_page_addr;
742 /* FIXME: This is 'const' in order to trick the compiler
743 into not treating it as DP-relative data. */
744 extern void * const linux_gateway_page;
746 linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
749 * Setup Linux Gateway page.
751 * The Linux gateway page will reside in kernel space (on virtual
752 * page 0), so it doesn't need to be aliased into user space.
755 map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
756 PAGE_SIZE, PAGE_GATEWAY, 1);
759 void __init paging_init(void)
766 flush_cache_all_local(); /* start with known state */
767 flush_tlb_all_local(NULL);
769 for (i = 0; i < npmem_ranges; i++) {
770 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
772 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
774 #ifdef CONFIG_DISCONTIGMEM
775 /* Need to initialize the pfnnid_map before we can initialize
779 for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
780 j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
787 free_area_init_node(i, zones_size,
788 pmem_ranges[i].start_pfn, NULL);
795 * Currently, all PA20 chips have 18 bit protection IDs, which is the
796 * limiting factor (space ids are 32 bits).
799 #define NR_SPACE_IDS 262144
804 * Currently we have a one-to-one relationship between space IDs and
805 * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
806 * support 15 bit protection IDs, so that is the limiting factor.
807 * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
808 * probably not worth the effort for a special case here.
811 #define NR_SPACE_IDS 32768
813 #endif /* !CONFIG_PA20 */
815 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
816 #define SID_ARRAY_SIZE (NR_SPACE_IDS / (8 * sizeof(long)))
818 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
819 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
820 static unsigned long space_id_index;
821 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
822 static unsigned long dirty_space_ids = 0;
824 static DEFINE_SPINLOCK(sid_lock);
826 unsigned long alloc_sid(void)
830 spin_lock(&sid_lock);
832 if (free_space_ids == 0) {
833 if (dirty_space_ids != 0) {
834 spin_unlock(&sid_lock);
835 flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
836 spin_lock(&sid_lock);
838 BUG_ON(free_space_ids == 0);
843 index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
844 space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
845 space_id_index = index;
847 spin_unlock(&sid_lock);
849 return index << SPACEID_SHIFT;
852 void free_sid(unsigned long spaceid)
854 unsigned long index = spaceid >> SPACEID_SHIFT;
855 unsigned long *dirty_space_offset;
857 dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
858 index &= (BITS_PER_LONG - 1);
860 spin_lock(&sid_lock);
862 BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
864 *dirty_space_offset |= (1L << index);
867 spin_unlock(&sid_lock);
872 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
876 /* NOTE: sid_lock must be held upon entry */
878 *ndirtyptr = dirty_space_ids;
879 if (dirty_space_ids != 0) {
880 for (i = 0; i < SID_ARRAY_SIZE; i++) {
881 dirty_array[i] = dirty_space_id[i];
882 dirty_space_id[i] = 0;
890 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
894 /* NOTE: sid_lock must be held upon entry */
897 for (i = 0; i < SID_ARRAY_SIZE; i++) {
898 space_id[i] ^= dirty_array[i];
901 free_space_ids += ndirty;
906 #else /* CONFIG_SMP */
908 static void recycle_sids(void)
912 /* NOTE: sid_lock must be held upon entry */
914 if (dirty_space_ids != 0) {
915 for (i = 0; i < SID_ARRAY_SIZE; i++) {
916 space_id[i] ^= dirty_space_id[i];
917 dirty_space_id[i] = 0;
920 free_space_ids += dirty_space_ids;
928 * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
929 * purged, we can safely reuse the space ids that were released but
930 * not flushed from the tlb.
935 static unsigned long recycle_ndirty;
936 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
937 static unsigned int recycle_inuse;
939 void flush_tlb_all(void)
944 spin_lock(&sid_lock);
945 __inc_irq_stat(irq_tlb_count);
946 if (dirty_space_ids > RECYCLE_THRESHOLD) {
947 BUG_ON(recycle_inuse); /* FIXME: Use a semaphore/wait queue here */
948 get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
952 spin_unlock(&sid_lock);
953 on_each_cpu(flush_tlb_all_local, NULL, 1);
955 spin_lock(&sid_lock);
956 recycle_sids(recycle_ndirty,recycle_dirty_array);
958 spin_unlock(&sid_lock);
962 void flush_tlb_all(void)
964 spin_lock(&sid_lock);
965 __inc_irq_stat(irq_tlb_count);
966 flush_tlb_all_local(NULL);
968 spin_unlock(&sid_lock);
972 #ifdef CONFIG_BLK_DEV_INITRD
973 void free_initrd_mem(unsigned long start, unsigned long end)
975 free_reserved_area((void *)start, (void *)end, -1, "initrd");