1 // SPDX-License-Identifier: GPL-2.0
3 * linux/arch/parisc/mm/init.c
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)
15 #include <linux/module.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>
28 #include <asm/pgalloc.h>
29 #include <asm/pgtable.h>
31 #include <asm/pdc_chassis.h>
32 #include <asm/mmzone.h>
33 #include <asm/sections.h>
34 #include <asm/msgbuf.h>
36 extern int data_start;
37 extern void parisc_kernel_start(void); /* Kernel entry point in head.S */
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)));
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)));
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;
56 static struct resource data_resource = {
57 .name = "Kernel data",
58 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
61 static struct resource code_resource = {
62 .name = "Kernel code",
63 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM,
66 static struct resource pdcdata_resource = {
67 .name = "PDC data (Page Zero)",
70 .flags = IORESOURCE_BUSY | IORESOURCE_MEM,
73 static struct resource sysram_resources[MAX_PHYSMEM_RANGES] __read_mostly;
75 /* The following array is initialized from the firmware specific
76 * information retrieved in kernel/inventory.c.
79 physmem_range_t pmem_ranges[MAX_PHYSMEM_RANGES] __read_mostly;
80 int npmem_ranges __read_mostly;
83 * get_memblock() allocates pages via memblock.
84 * We can't use memblock_find_in_range(0, KERNEL_INITIAL_SIZE) here since it
85 * doesn't allocate from bottom to top which is needed because we only created
86 * the initial mapping up to KERNEL_INITIAL_SIZE in the assembly bootup code.
88 static void * __init get_memblock(unsigned long size)
90 static phys_addr_t search_addr __initdata;
94 search_addr = PAGE_ALIGN(__pa((unsigned long) &_end));
95 search_addr = ALIGN(search_addr, size);
96 while (!memblock_is_region_memory(search_addr, size) ||
97 memblock_is_region_reserved(search_addr, size)) {
103 memblock_reserve(phys, size);
105 panic("get_memblock() failed.\n");
107 memset(__va(phys), 0, size);
113 #define MAX_MEM (~0UL)
114 #else /* !CONFIG_64BIT */
115 #define MAX_MEM (3584U*1024U*1024U)
116 #endif /* !CONFIG_64BIT */
118 static unsigned long mem_limit __read_mostly = MAX_MEM;
120 static void __init mem_limit_func(void)
125 /* We need this before __setup() functions are called */
128 for (cp = boot_command_line; *cp; ) {
129 if (memcmp(cp, "mem=", 4) == 0) {
131 limit = memparse(cp, &end);
136 while (*cp != ' ' && *cp)
143 if (limit < mem_limit)
147 #define MAX_GAP (0x40000000UL >> PAGE_SHIFT)
149 static void __init setup_bootmem(void)
151 unsigned long mem_max;
152 #ifndef CONFIG_DISCONTIGMEM
153 physmem_range_t pmem_holes[MAX_PHYSMEM_RANGES - 1];
156 int i, sysram_resource_count;
158 disable_sr_hashing(); /* Turn off space register hashing */
161 * Sort the ranges. Since the number of ranges is typically
162 * small, and performance is not an issue here, just do
163 * a simple insertion sort.
166 for (i = 1; i < npmem_ranges; i++) {
169 for (j = i; j > 0; j--) {
172 if (pmem_ranges[j-1].start_pfn <
173 pmem_ranges[j].start_pfn) {
177 tmp = pmem_ranges[j-1].start_pfn;
178 pmem_ranges[j-1].start_pfn = pmem_ranges[j].start_pfn;
179 pmem_ranges[j].start_pfn = tmp;
180 tmp = pmem_ranges[j-1].pages;
181 pmem_ranges[j-1].pages = pmem_ranges[j].pages;
182 pmem_ranges[j].pages = tmp;
186 #ifndef CONFIG_DISCONTIGMEM
188 * Throw out ranges that are too far apart (controlled by
192 for (i = 1; i < npmem_ranges; i++) {
193 if (pmem_ranges[i].start_pfn -
194 (pmem_ranges[i-1].start_pfn +
195 pmem_ranges[i-1].pages) > MAX_GAP) {
197 printk("Large gap in memory detected (%ld pages). "
198 "Consider turning on CONFIG_DISCONTIGMEM\n",
199 pmem_ranges[i].start_pfn -
200 (pmem_ranges[i-1].start_pfn +
201 pmem_ranges[i-1].pages));
207 /* Print the memory ranges */
208 pr_info("Memory Ranges:\n");
210 for (i = 0; i < npmem_ranges; i++) {
211 struct resource *res = &sysram_resources[i];
215 size = (pmem_ranges[i].pages << PAGE_SHIFT);
216 start = (pmem_ranges[i].start_pfn << PAGE_SHIFT);
217 pr_info("%2d) Start 0x%016lx End 0x%016lx Size %6ld MB\n",
218 i, start, start + (size - 1), size >> 20);
220 /* request memory resource */
221 res->name = "System RAM";
223 res->end = start + size - 1;
224 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
225 request_resource(&iomem_resource, res);
228 sysram_resource_count = npmem_ranges;
231 * For 32 bit kernels we limit the amount of memory we can
232 * support, in order to preserve enough kernel address space
233 * for other purposes. For 64 bit kernels we don't normally
234 * limit the memory, but this mechanism can be used to
235 * artificially limit the amount of memory (and it is written
236 * to work with multiple memory ranges).
239 mem_limit_func(); /* check for "mem=" argument */
242 for (i = 0; i < npmem_ranges; i++) {
245 rsize = pmem_ranges[i].pages << PAGE_SHIFT;
246 if ((mem_max + rsize) > mem_limit) {
247 printk(KERN_WARNING "Memory truncated to %ld MB\n", mem_limit >> 20);
248 if (mem_max == mem_limit)
251 pmem_ranges[i].pages = (mem_limit >> PAGE_SHIFT)
252 - (mem_max >> PAGE_SHIFT);
253 npmem_ranges = i + 1;
261 printk(KERN_INFO "Total Memory: %ld MB\n",mem_max >> 20);
263 #ifndef CONFIG_DISCONTIGMEM
264 /* Merge the ranges, keeping track of the holes */
267 unsigned long end_pfn;
268 unsigned long hole_pages;
271 end_pfn = pmem_ranges[0].start_pfn + pmem_ranges[0].pages;
272 for (i = 1; i < npmem_ranges; i++) {
274 hole_pages = pmem_ranges[i].start_pfn - end_pfn;
276 pmem_holes[npmem_holes].start_pfn = end_pfn;
277 pmem_holes[npmem_holes++].pages = hole_pages;
278 end_pfn += hole_pages;
280 end_pfn += pmem_ranges[i].pages;
283 pmem_ranges[0].pages = end_pfn - pmem_ranges[0].start_pfn;
288 #ifdef CONFIG_DISCONTIGMEM
289 for (i = 0; i < MAX_PHYSMEM_RANGES; i++) {
290 memset(NODE_DATA(i), 0, sizeof(pg_data_t));
292 memset(pfnnid_map, 0xff, sizeof(pfnnid_map));
294 for (i = 0; i < npmem_ranges; i++) {
295 node_set_state(i, N_NORMAL_MEMORY);
301 * Initialize and free the full range of memory in each range.
305 for (i = 0; i < npmem_ranges; i++) {
306 unsigned long start_pfn;
307 unsigned long npages;
311 start_pfn = pmem_ranges[i].start_pfn;
312 npages = pmem_ranges[i].pages;
314 start = start_pfn << PAGE_SHIFT;
315 size = npages << PAGE_SHIFT;
317 /* add system RAM memblock */
318 memblock_add(start, size);
320 if ((start_pfn + npages) > max_pfn)
321 max_pfn = start_pfn + npages;
324 /* IOMMU is always used to access "high mem" on those boxes
325 * that can support enough mem that a PCI device couldn't
326 * directly DMA to any physical addresses.
327 * ISA DMA support will need to revisit this.
329 max_low_pfn = max_pfn;
331 /* reserve PAGE0 pdc memory, kernel text/data/bss & bootmap */
333 #define PDC_CONSOLE_IO_IODC_SIZE 32768
335 memblock_reserve(0UL, (unsigned long)(PAGE0->mem_free +
336 PDC_CONSOLE_IO_IODC_SIZE));
337 memblock_reserve(__pa(KERNEL_BINARY_TEXT_START),
338 (unsigned long)(_end - KERNEL_BINARY_TEXT_START));
340 #ifndef CONFIG_DISCONTIGMEM
342 /* reserve the holes */
344 for (i = 0; i < npmem_holes; i++) {
345 memblock_reserve((pmem_holes[i].start_pfn << PAGE_SHIFT),
346 (pmem_holes[i].pages << PAGE_SHIFT));
350 #ifdef CONFIG_BLK_DEV_INITRD
352 printk(KERN_INFO "initrd: %08lx-%08lx\n", initrd_start, initrd_end);
353 if (__pa(initrd_start) < mem_max) {
354 unsigned long initrd_reserve;
356 if (__pa(initrd_end) > mem_max) {
357 initrd_reserve = mem_max - __pa(initrd_start);
359 initrd_reserve = initrd_end - initrd_start;
361 initrd_below_start_ok = 1;
362 printk(KERN_INFO "initrd: reserving %08lx-%08lx (mem_max %08lx)\n", __pa(initrd_start), __pa(initrd_start) + initrd_reserve, mem_max);
364 memblock_reserve(__pa(initrd_start), initrd_reserve);
369 data_resource.start = virt_to_phys(&data_start);
370 data_resource.end = virt_to_phys(_end) - 1;
371 code_resource.start = virt_to_phys(_text);
372 code_resource.end = virt_to_phys(&data_start)-1;
374 /* We don't know which region the kernel will be in, so try
377 for (i = 0; i < sysram_resource_count; i++) {
378 struct resource *res = &sysram_resources[i];
379 request_resource(res, &code_resource);
380 request_resource(res, &data_resource);
382 request_resource(&sysram_resources[0], &pdcdata_resource);
384 /* Initialize Page Deallocation Table (PDT) and check for bad memory. */
388 static int __init parisc_text_address(unsigned long vaddr)
390 static unsigned long head_ptr __initdata;
393 head_ptr = PAGE_MASK & (unsigned long)
394 dereference_function_descriptor(&parisc_kernel_start);
396 return core_kernel_text(vaddr) || vaddr == head_ptr;
399 static void __init map_pages(unsigned long start_vaddr,
400 unsigned long start_paddr, unsigned long size,
401 pgprot_t pgprot, int force)
406 unsigned long end_paddr;
407 unsigned long start_pmd;
408 unsigned long start_pte;
411 unsigned long address;
413 unsigned long ro_start;
414 unsigned long ro_end;
415 unsigned long kernel_end;
417 ro_start = __pa((unsigned long)_text);
418 ro_end = __pa((unsigned long)&data_start);
419 kernel_end = __pa((unsigned long)&_end);
421 end_paddr = start_paddr + size;
423 pg_dir = pgd_offset_k(start_vaddr);
425 #if PTRS_PER_PMD == 1
428 start_pmd = ((start_vaddr >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
430 start_pte = ((start_vaddr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
432 address = start_paddr;
434 while (address < end_paddr) {
435 #if PTRS_PER_PMD == 1
436 pmd = (pmd_t *)__pa(pg_dir);
438 pmd = (pmd_t *)pgd_address(*pg_dir);
441 * pmd is physical at this point
445 pmd = (pmd_t *) get_memblock(PAGE_SIZE << PMD_ORDER);
446 pmd = (pmd_t *) __pa(pmd);
449 pgd_populate(NULL, pg_dir, __va(pmd));
453 /* now change pmd to kernel virtual addresses */
455 pmd = (pmd_t *)__va(pmd) + start_pmd;
456 for (tmp1 = start_pmd; tmp1 < PTRS_PER_PMD; tmp1++, pmd++) {
459 * pg_table is physical at this point
462 pg_table = (pte_t *)pmd_address(*pmd);
464 pg_table = (pte_t *) get_memblock(PAGE_SIZE);
465 pg_table = (pte_t *) __pa(pg_table);
468 pmd_populate_kernel(NULL, pmd, __va(pg_table));
470 /* now change pg_table to kernel virtual addresses */
472 pg_table = (pte_t *) __va(pg_table) + start_pte;
473 for (tmp2 = start_pte; tmp2 < PTRS_PER_PTE; tmp2++, pg_table++) {
477 pte = __mk_pte(address, pgprot);
478 else if (parisc_text_address(vaddr)) {
479 pte = __mk_pte(address, PAGE_KERNEL_EXEC);
480 if (address >= ro_start && address < kernel_end)
481 pte = pte_mkhuge(pte);
484 #if defined(CONFIG_PARISC_PAGE_SIZE_4KB)
485 if (address >= ro_start && address < ro_end) {
486 pte = __mk_pte(address, PAGE_KERNEL_EXEC);
487 pte = pte_mkhuge(pte);
491 pte = __mk_pte(address, pgprot);
492 if (address >= ro_start && address < kernel_end)
493 pte = pte_mkhuge(pte);
496 if (address >= end_paddr)
499 set_pte(pg_table, pte);
501 address += PAGE_SIZE;
506 if (address >= end_paddr)
513 void __init set_kernel_text_rw(int enable_read_write)
515 unsigned long start = (unsigned long)__init_begin;
516 unsigned long end = (unsigned long)_etext;
518 map_pages(start, __pa(start), end-start,
519 PAGE_KERNEL_RWX, enable_read_write ? 1:0);
521 /* force the kernel to see the new page table entries */
526 void __ref free_initmem(void)
528 unsigned long init_begin = (unsigned long)__init_begin;
529 unsigned long init_end = (unsigned long)__init_end;
531 /* The init text pages are marked R-X. We have to
532 * flush the icache and mark them RW-
534 * This is tricky, because map_pages is in the init section.
535 * Do a dummy remap of the data section first (the data
536 * section is already PAGE_KERNEL) to pull in the TLB entries
538 map_pages(init_begin, __pa(init_begin), init_end - init_begin,
540 /* now remap at PAGE_KERNEL since the TLB is pre-primed to execute
542 map_pages(init_begin, __pa(init_begin), init_end - init_begin,
545 /* force the kernel to see the new TLB entries */
546 __flush_tlb_range(0, init_begin, init_end);
548 /* finally dump all the instructions which were cached, since the
549 * pages are no-longer executable */
550 flush_icache_range(init_begin, init_end);
552 free_initmem_default(POISON_FREE_INITMEM);
554 /* set up a new led state on systems shipped LED State panel */
555 pdc_chassis_send_status(PDC_CHASSIS_DIRECT_BCOMPLETE);
559 #ifdef CONFIG_STRICT_KERNEL_RWX
560 void mark_rodata_ro(void)
562 /* rodata memory was already mapped with KERNEL_RO access rights by
563 pagetable_init() and map_pages(). No need to do additional stuff here */
564 printk (KERN_INFO "Write protecting the kernel read-only data: %luk\n",
565 (unsigned long)(__end_rodata - __start_rodata) >> 10);
571 * Just an arbitrary offset to serve as a "hole" between mapping areas
572 * (between top of physical memory and a potential pcxl dma mapping
573 * area, and below the vmalloc mapping area).
575 * The current 32K value just means that there will be a 32K "hole"
576 * between mapping areas. That means that any out-of-bounds memory
577 * accesses will hopefully be caught. The vmalloc() routines leaves
578 * a hole of 4kB between each vmalloced area for the same reason.
581 /* Leave room for gateway page expansion */
582 #if KERNEL_MAP_START < GATEWAY_PAGE_SIZE
583 #error KERNEL_MAP_START is in gateway reserved region
585 #define MAP_START (KERNEL_MAP_START)
587 #define VM_MAP_OFFSET (32*1024)
588 #define SET_MAP_OFFSET(x) ((void *)(((unsigned long)(x) + VM_MAP_OFFSET) \
589 & ~(VM_MAP_OFFSET-1)))
591 void *parisc_vmalloc_start __read_mostly;
592 EXPORT_SYMBOL(parisc_vmalloc_start);
595 unsigned long pcxl_dma_start __read_mostly;
598 void __init mem_init(void)
600 /* Do sanity checks on IPC (compat) structures */
601 BUILD_BUG_ON(sizeof(struct ipc64_perm) != 48);
603 BUILD_BUG_ON(sizeof(struct semid64_ds) != 80);
604 BUILD_BUG_ON(sizeof(struct msqid64_ds) != 104);
605 BUILD_BUG_ON(sizeof(struct shmid64_ds) != 104);
608 BUILD_BUG_ON(sizeof(struct compat_ipc64_perm) != sizeof(struct ipc64_perm));
609 BUILD_BUG_ON(sizeof(struct compat_semid64_ds) != 80);
610 BUILD_BUG_ON(sizeof(struct compat_msqid64_ds) != 104);
611 BUILD_BUG_ON(sizeof(struct compat_shmid64_ds) != 104);
614 /* Do sanity checks on page table constants */
615 BUILD_BUG_ON(PTE_ENTRY_SIZE != sizeof(pte_t));
616 BUILD_BUG_ON(PMD_ENTRY_SIZE != sizeof(pmd_t));
617 BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
618 BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
621 high_memory = __va((max_pfn << PAGE_SHIFT));
622 set_max_mapnr(page_to_pfn(virt_to_page(high_memory - 1)) + 1);
626 if (boot_cpu_data.cpu_type == pcxl2 || boot_cpu_data.cpu_type == pcxl) {
627 pcxl_dma_start = (unsigned long)SET_MAP_OFFSET(MAP_START);
628 parisc_vmalloc_start = SET_MAP_OFFSET(pcxl_dma_start
629 + PCXL_DMA_MAP_SIZE);
632 parisc_vmalloc_start = SET_MAP_OFFSET(MAP_START);
634 mem_init_print_info(NULL);
638 * Do not expose the virtual kernel memory layout to userspace.
639 * But keep code for debugging purposes.
641 printk("virtual kernel memory layout:\n"
642 " vmalloc : 0x%px - 0x%px (%4ld MB)\n"
643 " memory : 0x%px - 0x%px (%4ld MB)\n"
644 " .init : 0x%px - 0x%px (%4ld kB)\n"
645 " .data : 0x%px - 0x%px (%4ld kB)\n"
646 " .text : 0x%px - 0x%px (%4ld kB)\n",
648 (void*)VMALLOC_START, (void*)VMALLOC_END,
649 (VMALLOC_END - VMALLOC_START) >> 20,
651 __va(0), high_memory,
652 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
654 __init_begin, __init_end,
655 ((unsigned long)__init_end - (unsigned long)__init_begin) >> 10,
658 ((unsigned long)_edata - (unsigned long)_etext) >> 10,
661 ((unsigned long)_etext - (unsigned long)_text) >> 10);
665 unsigned long *empty_zero_page __read_mostly;
666 EXPORT_SYMBOL(empty_zero_page);
669 * pagetable_init() sets up the page tables
671 * Note that gateway_init() places the Linux gateway page at page 0.
672 * Since gateway pages cannot be dereferenced this has the desirable
673 * side effect of trapping those pesky NULL-reference errors in the
676 static void __init pagetable_init(void)
680 /* Map each physical memory range to its kernel vaddr */
682 for (range = 0; range < npmem_ranges; range++) {
683 unsigned long start_paddr;
684 unsigned long end_paddr;
687 start_paddr = pmem_ranges[range].start_pfn << PAGE_SHIFT;
688 size = pmem_ranges[range].pages << PAGE_SHIFT;
689 end_paddr = start_paddr + size;
691 map_pages((unsigned long)__va(start_paddr), start_paddr,
692 size, PAGE_KERNEL, 0);
695 #ifdef CONFIG_BLK_DEV_INITRD
696 if (initrd_end && initrd_end > mem_limit) {
697 printk(KERN_INFO "initrd: mapping %08lx-%08lx\n", initrd_start, initrd_end);
698 map_pages(initrd_start, __pa(initrd_start),
699 initrd_end - initrd_start, PAGE_KERNEL, 0);
703 empty_zero_page = get_memblock(PAGE_SIZE);
706 static void __init gateway_init(void)
708 unsigned long linux_gateway_page_addr;
709 /* FIXME: This is 'const' in order to trick the compiler
710 into not treating it as DP-relative data. */
711 extern void * const linux_gateway_page;
713 linux_gateway_page_addr = LINUX_GATEWAY_ADDR & PAGE_MASK;
716 * Setup Linux Gateway page.
718 * The Linux gateway page will reside in kernel space (on virtual
719 * page 0), so it doesn't need to be aliased into user space.
722 map_pages(linux_gateway_page_addr, __pa(&linux_gateway_page),
723 PAGE_SIZE, PAGE_GATEWAY, 1);
726 void __init paging_init(void)
733 flush_cache_all_local(); /* start with known state */
734 flush_tlb_all_local(NULL);
736 for (i = 0; i < npmem_ranges; i++) {
737 unsigned long zones_size[MAX_NR_ZONES] = { 0, };
739 zones_size[ZONE_NORMAL] = pmem_ranges[i].pages;
741 #ifdef CONFIG_DISCONTIGMEM
742 /* Need to initialize the pfnnid_map before we can initialize
746 for (j = (pmem_ranges[i].start_pfn >> PFNNID_SHIFT);
747 j <= ((pmem_ranges[i].start_pfn + pmem_ranges[i].pages) >> PFNNID_SHIFT);
754 free_area_init_node(i, zones_size,
755 pmem_ranges[i].start_pfn, NULL);
762 * Currently, all PA20 chips have 18 bit protection IDs, which is the
763 * limiting factor (space ids are 32 bits).
766 #define NR_SPACE_IDS 262144
771 * Currently we have a one-to-one relationship between space IDs and
772 * protection IDs. Older parisc chips (PCXS, PCXT, PCXL, PCXL2) only
773 * support 15 bit protection IDs, so that is the limiting factor.
774 * PCXT' has 18 bit protection IDs, but only 16 bit spaceids, so it's
775 * probably not worth the effort for a special case here.
778 #define NR_SPACE_IDS 32768
780 #endif /* !CONFIG_PA20 */
782 #define RECYCLE_THRESHOLD (NR_SPACE_IDS / 2)
783 #define SID_ARRAY_SIZE (NR_SPACE_IDS / (8 * sizeof(long)))
785 static unsigned long space_id[SID_ARRAY_SIZE] = { 1 }; /* disallow space 0 */
786 static unsigned long dirty_space_id[SID_ARRAY_SIZE];
787 static unsigned long space_id_index;
788 static unsigned long free_space_ids = NR_SPACE_IDS - 1;
789 static unsigned long dirty_space_ids = 0;
791 static DEFINE_SPINLOCK(sid_lock);
793 unsigned long alloc_sid(void)
797 spin_lock(&sid_lock);
799 if (free_space_ids == 0) {
800 if (dirty_space_ids != 0) {
801 spin_unlock(&sid_lock);
802 flush_tlb_all(); /* flush_tlb_all() calls recycle_sids() */
803 spin_lock(&sid_lock);
805 BUG_ON(free_space_ids == 0);
810 index = find_next_zero_bit(space_id, NR_SPACE_IDS, space_id_index);
811 space_id[index >> SHIFT_PER_LONG] |= (1L << (index & (BITS_PER_LONG - 1)));
812 space_id_index = index;
814 spin_unlock(&sid_lock);
816 return index << SPACEID_SHIFT;
819 void free_sid(unsigned long spaceid)
821 unsigned long index = spaceid >> SPACEID_SHIFT;
822 unsigned long *dirty_space_offset;
824 dirty_space_offset = dirty_space_id + (index >> SHIFT_PER_LONG);
825 index &= (BITS_PER_LONG - 1);
827 spin_lock(&sid_lock);
829 BUG_ON(*dirty_space_offset & (1L << index)); /* attempt to free space id twice */
831 *dirty_space_offset |= (1L << index);
834 spin_unlock(&sid_lock);
839 static void get_dirty_sids(unsigned long *ndirtyptr,unsigned long *dirty_array)
843 /* NOTE: sid_lock must be held upon entry */
845 *ndirtyptr = dirty_space_ids;
846 if (dirty_space_ids != 0) {
847 for (i = 0; i < SID_ARRAY_SIZE; i++) {
848 dirty_array[i] = dirty_space_id[i];
849 dirty_space_id[i] = 0;
857 static void recycle_sids(unsigned long ndirty,unsigned long *dirty_array)
861 /* NOTE: sid_lock must be held upon entry */
864 for (i = 0; i < SID_ARRAY_SIZE; i++) {
865 space_id[i] ^= dirty_array[i];
868 free_space_ids += ndirty;
873 #else /* CONFIG_SMP */
875 static void recycle_sids(void)
879 /* NOTE: sid_lock must be held upon entry */
881 if (dirty_space_ids != 0) {
882 for (i = 0; i < SID_ARRAY_SIZE; i++) {
883 space_id[i] ^= dirty_space_id[i];
884 dirty_space_id[i] = 0;
887 free_space_ids += dirty_space_ids;
895 * flush_tlb_all() calls recycle_sids(), since whenever the entire tlb is
896 * purged, we can safely reuse the space ids that were released but
897 * not flushed from the tlb.
902 static unsigned long recycle_ndirty;
903 static unsigned long recycle_dirty_array[SID_ARRAY_SIZE];
904 static unsigned int recycle_inuse;
906 void flush_tlb_all(void)
910 __inc_irq_stat(irq_tlb_count);
912 spin_lock(&sid_lock);
913 if (dirty_space_ids > RECYCLE_THRESHOLD) {
914 BUG_ON(recycle_inuse); /* FIXME: Use a semaphore/wait queue here */
915 get_dirty_sids(&recycle_ndirty,recycle_dirty_array);
919 spin_unlock(&sid_lock);
920 on_each_cpu(flush_tlb_all_local, NULL, 1);
922 spin_lock(&sid_lock);
923 recycle_sids(recycle_ndirty,recycle_dirty_array);
925 spin_unlock(&sid_lock);
929 void flush_tlb_all(void)
931 __inc_irq_stat(irq_tlb_count);
932 spin_lock(&sid_lock);
933 flush_tlb_all_local(NULL);
935 spin_unlock(&sid_lock);
939 #ifdef CONFIG_BLK_DEV_INITRD
940 void free_initrd_mem(unsigned long start, unsigned long end)
942 free_reserved_area((void *)start, (void *)end, -1, "initrd");