2 * PowerPC64 port by Mike Corrigan and Dave Engebretsen
3 * {mikejc|engebret}@us.ibm.com
5 * Copyright (c) 2000 Mike Corrigan <mikejc@us.ibm.com>
7 * SMP scalability work:
8 * Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
13 * PowerPC Hashed Page Table functions
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
24 #include <linux/spinlock.h>
25 #include <linux/errno.h>
26 #include <linux/sched.h>
27 #include <linux/proc_fs.h>
28 #include <linux/stat.h>
29 #include <linux/sysctl.h>
30 #include <linux/ctype.h>
31 #include <linux/cache.h>
32 #include <linux/init.h>
33 #include <linux/signal.h>
34 #include <linux/lmb.h>
36 #include <asm/processor.h>
37 #include <asm/pgtable.h>
39 #include <asm/mmu_context.h>
41 #include <asm/types.h>
42 #include <asm/system.h>
43 #include <asm/uaccess.h>
44 #include <asm/machdep.h>
46 #include <asm/abs_addr.h>
47 #include <asm/tlbflush.h>
51 #include <asm/cacheflush.h>
52 #include <asm/cputable.h>
53 #include <asm/sections.h>
58 #define DBG(fmt...) udbg_printf(fmt)
64 #define DBG_LOW(fmt...) udbg_printf(fmt)
66 #define DBG_LOW(fmt...)
74 * Note: pte --> Linux PTE
75 * HPTE --> PowerPC Hashed Page Table Entry
78 * htab_initialize is called with the MMU off (of course), but
79 * the kernel has been copied down to zero so it can directly
80 * reference global data. At this point it is very difficult
81 * to print debug info.
86 extern unsigned long dart_tablebase;
87 #endif /* CONFIG_U3_DART */
89 static unsigned long _SDR1;
90 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
92 struct hash_pte *htab_address;
93 unsigned long htab_size_bytes;
94 unsigned long htab_hash_mask;
95 int mmu_linear_psize = MMU_PAGE_4K;
96 int mmu_virtual_psize = MMU_PAGE_4K;
97 int mmu_vmalloc_psize = MMU_PAGE_4K;
98 #ifdef CONFIG_SPARSEMEM_VMEMMAP
99 int mmu_vmemmap_psize = MMU_PAGE_4K;
101 int mmu_io_psize = MMU_PAGE_4K;
102 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
103 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
104 u16 mmu_slb_size = 64;
105 #ifdef CONFIG_HUGETLB_PAGE
106 unsigned int HPAGE_SHIFT;
108 #ifdef CONFIG_PPC_64K_PAGES
109 int mmu_ci_restrictions;
111 #ifdef CONFIG_DEBUG_PAGEALLOC
112 static u8 *linear_map_hash_slots;
113 static unsigned long linear_map_hash_count;
114 static DEFINE_SPINLOCK(linear_map_hash_lock);
115 #endif /* CONFIG_DEBUG_PAGEALLOC */
117 /* There are definitions of page sizes arrays to be used when none
118 * is provided by the firmware.
121 /* Pre-POWER4 CPUs (4k pages only)
123 static struct mmu_psize_def mmu_psize_defaults_old[] = {
133 /* POWER4, GPUL, POWER5
135 * Support for 16Mb large pages
137 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
154 static unsigned long htab_convert_pte_flags(unsigned long pteflags)
156 unsigned long rflags = pteflags & 0x1fa;
158 /* _PAGE_EXEC -> NOEXEC */
159 if ((pteflags & _PAGE_EXEC) == 0)
162 /* PP bits. PAGE_USER is already PP bit 0x2, so we only
163 * need to add in 0x1 if it's a read-only user page
165 if ((pteflags & _PAGE_USER) && !((pteflags & _PAGE_RW) &&
166 (pteflags & _PAGE_DIRTY)))
170 return rflags | HPTE_R_C;
173 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
174 unsigned long pstart, unsigned long prot,
175 int psize, int ssize)
177 unsigned long vaddr, paddr;
178 unsigned int step, shift;
181 shift = mmu_psize_defs[psize].shift;
184 prot = htab_convert_pte_flags(prot);
186 DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
187 vstart, vend, pstart, prot, psize, ssize);
189 for (vaddr = vstart, paddr = pstart; vaddr < vend;
190 vaddr += step, paddr += step) {
191 unsigned long hash, hpteg;
192 unsigned long vsid = get_kernel_vsid(vaddr, ssize);
193 unsigned long va = hpt_va(vaddr, vsid, ssize);
195 hash = hpt_hash(va, shift, ssize);
196 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
198 BUG_ON(!ppc_md.hpte_insert);
199 ret = ppc_md.hpte_insert(hpteg, va, paddr, prot,
200 HPTE_V_BOLTED, psize, ssize);
204 #ifdef CONFIG_DEBUG_PAGEALLOC
205 if ((paddr >> PAGE_SHIFT) < linear_map_hash_count)
206 linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
207 #endif /* CONFIG_DEBUG_PAGEALLOC */
209 return ret < 0 ? ret : 0;
212 #ifdef CONFIG_MEMORY_HOTPLUG
213 static int htab_remove_mapping(unsigned long vstart, unsigned long vend,
214 int psize, int ssize)
217 unsigned int step, shift;
219 shift = mmu_psize_defs[psize].shift;
222 if (!ppc_md.hpte_removebolted) {
223 printk(KERN_WARNING "Platform doesn't implement "
224 "hpte_removebolted\n");
228 for (vaddr = vstart; vaddr < vend; vaddr += step)
229 ppc_md.hpte_removebolted(vaddr, psize, ssize);
233 #endif /* CONFIG_MEMORY_HOTPLUG */
235 static int __init htab_dt_scan_seg_sizes(unsigned long node,
236 const char *uname, int depth,
239 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
241 unsigned long size = 0;
243 /* We are scanning "cpu" nodes only */
244 if (type == NULL || strcmp(type, "cpu") != 0)
247 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,processor-segment-sizes",
251 for (; size >= 4; size -= 4, ++prop) {
253 DBG("1T segment support detected\n");
254 cur_cpu_spec->cpu_features |= CPU_FTR_1T_SEGMENT;
258 cur_cpu_spec->cpu_features &= ~CPU_FTR_NO_SLBIE_B;
262 static void __init htab_init_seg_sizes(void)
264 of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
267 static int __init htab_dt_scan_page_sizes(unsigned long node,
268 const char *uname, int depth,
271 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
273 unsigned long size = 0;
275 /* We are scanning "cpu" nodes only */
276 if (type == NULL || strcmp(type, "cpu") != 0)
279 prop = (u32 *)of_get_flat_dt_prop(node,
280 "ibm,segment-page-sizes", &size);
282 DBG("Page sizes from device-tree:\n");
284 cur_cpu_spec->cpu_features &= ~(CPU_FTR_16M_PAGE);
286 unsigned int shift = prop[0];
287 unsigned int slbenc = prop[1];
288 unsigned int lpnum = prop[2];
289 unsigned int lpenc = 0;
290 struct mmu_psize_def *def;
293 size -= 3; prop += 3;
294 while(size > 0 && lpnum) {
295 if (prop[0] == shift)
297 prop += 2; size -= 2;
312 cur_cpu_spec->cpu_features |= CPU_FTR_16M_PAGE;
320 def = &mmu_psize_defs[idx];
325 def->avpnm = (1 << (shift - 23)) - 1;
328 /* We don't know for sure what's up with tlbiel, so
329 * for now we only set it for 4K and 64K pages
331 if (idx == MMU_PAGE_4K || idx == MMU_PAGE_64K)
336 DBG(" %d: shift=%02x, sllp=%04x, avpnm=%08x, "
337 "tlbiel=%d, penc=%d\n",
338 idx, shift, def->sllp, def->avpnm, def->tlbiel,
346 /* Scan for 16G memory blocks that have been set aside for huge pages
347 * and reserve those blocks for 16G huge pages.
349 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
350 const char *uname, int depth,
352 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
353 unsigned long *addr_prop;
354 u32 *page_count_prop;
355 unsigned int expected_pages;
356 long unsigned int phys_addr;
357 long unsigned int block_size;
359 /* We are scanning "memory" nodes only */
360 if (type == NULL || strcmp(type, "memory") != 0)
363 /* This property is the log base 2 of the number of virtual pages that
364 * will represent this memory block. */
365 page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
366 if (page_count_prop == NULL)
368 expected_pages = (1 << page_count_prop[0]);
369 addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
370 if (addr_prop == NULL)
372 phys_addr = addr_prop[0];
373 block_size = addr_prop[1];
374 if (block_size != (16 * GB))
376 printk(KERN_INFO "Huge page(16GB) memory: "
377 "addr = 0x%lX size = 0x%lX pages = %d\n",
378 phys_addr, block_size, expected_pages);
379 lmb_reserve(phys_addr, block_size * expected_pages);
380 add_gpage(phys_addr, block_size, expected_pages);
384 static void __init htab_init_page_sizes(void)
388 /* Default to 4K pages only */
389 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
390 sizeof(mmu_psize_defaults_old));
393 * Try to find the available page sizes in the device-tree
395 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
396 if (rc != 0) /* Found */
400 * Not in the device-tree, let's fallback on known size
401 * list for 16M capable GP & GR
403 if (cpu_has_feature(CPU_FTR_16M_PAGE))
404 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
405 sizeof(mmu_psize_defaults_gp));
407 #ifndef CONFIG_DEBUG_PAGEALLOC
409 * Pick a size for the linear mapping. Currently, we only support
410 * 16M, 1M and 4K which is the default
412 if (mmu_psize_defs[MMU_PAGE_16M].shift)
413 mmu_linear_psize = MMU_PAGE_16M;
414 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
415 mmu_linear_psize = MMU_PAGE_1M;
416 #endif /* CONFIG_DEBUG_PAGEALLOC */
418 #ifdef CONFIG_PPC_64K_PAGES
420 * Pick a size for the ordinary pages. Default is 4K, we support
421 * 64K for user mappings and vmalloc if supported by the processor.
422 * We only use 64k for ioremap if the processor
423 * (and firmware) support cache-inhibited large pages.
424 * If not, we use 4k and set mmu_ci_restrictions so that
425 * hash_page knows to switch processes that use cache-inhibited
426 * mappings to 4k pages.
428 if (mmu_psize_defs[MMU_PAGE_64K].shift) {
429 mmu_virtual_psize = MMU_PAGE_64K;
430 mmu_vmalloc_psize = MMU_PAGE_64K;
431 if (mmu_linear_psize == MMU_PAGE_4K)
432 mmu_linear_psize = MMU_PAGE_64K;
433 if (cpu_has_feature(CPU_FTR_CI_LARGE_PAGE)) {
435 * Don't use 64k pages for ioremap on pSeries, since
436 * that would stop us accessing the HEA ethernet.
438 if (!machine_is(pseries))
439 mmu_io_psize = MMU_PAGE_64K;
441 mmu_ci_restrictions = 1;
443 #endif /* CONFIG_PPC_64K_PAGES */
445 #ifdef CONFIG_SPARSEMEM_VMEMMAP
446 /* We try to use 16M pages for vmemmap if that is supported
447 * and we have at least 1G of RAM at boot
449 if (mmu_psize_defs[MMU_PAGE_16M].shift &&
450 lmb_phys_mem_size() >= 0x40000000)
451 mmu_vmemmap_psize = MMU_PAGE_16M;
452 else if (mmu_psize_defs[MMU_PAGE_64K].shift)
453 mmu_vmemmap_psize = MMU_PAGE_64K;
455 mmu_vmemmap_psize = MMU_PAGE_4K;
456 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
458 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
459 "virtual = %d, io = %d"
460 #ifdef CONFIG_SPARSEMEM_VMEMMAP
464 mmu_psize_defs[mmu_linear_psize].shift,
465 mmu_psize_defs[mmu_virtual_psize].shift,
466 mmu_psize_defs[mmu_io_psize].shift
467 #ifdef CONFIG_SPARSEMEM_VMEMMAP
468 ,mmu_psize_defs[mmu_vmemmap_psize].shift
472 #ifdef CONFIG_HUGETLB_PAGE
473 /* Reserve 16G huge page memory sections for huge pages */
474 of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
476 /* Set default large page size. Currently, we pick 16M or 1M depending
477 * on what is available
479 if (mmu_psize_defs[MMU_PAGE_16M].shift)
480 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
481 /* With 4k/4level pagetables, we can't (for now) cope with a
482 * huge page size < PMD_SIZE */
483 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
484 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
485 #endif /* CONFIG_HUGETLB_PAGE */
488 static int __init htab_dt_scan_pftsize(unsigned long node,
489 const char *uname, int depth,
492 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
495 /* We are scanning "cpu" nodes only */
496 if (type == NULL || strcmp(type, "cpu") != 0)
499 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
501 /* pft_size[0] is the NUMA CEC cookie */
502 ppc64_pft_size = prop[1];
508 static unsigned long __init htab_get_table_size(void)
510 unsigned long mem_size, rnd_mem_size, pteg_count;
512 /* If hash size isn't already provided by the platform, we try to
513 * retrieve it from the device-tree. If it's not there neither, we
514 * calculate it now based on the total RAM size
516 if (ppc64_pft_size == 0)
517 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
519 return 1UL << ppc64_pft_size;
521 /* round mem_size up to next power of 2 */
522 mem_size = lmb_phys_mem_size();
523 rnd_mem_size = 1UL << __ilog2(mem_size);
524 if (rnd_mem_size < mem_size)
528 pteg_count = max(rnd_mem_size >> (12 + 1), 1UL << 11);
530 return pteg_count << 7;
533 #ifdef CONFIG_MEMORY_HOTPLUG
534 void create_section_mapping(unsigned long start, unsigned long end)
536 BUG_ON(htab_bolt_mapping(start, end, __pa(start),
537 PAGE_KERNEL, mmu_linear_psize,
541 int remove_section_mapping(unsigned long start, unsigned long end)
543 return htab_remove_mapping(start, end, mmu_linear_psize,
546 #endif /* CONFIG_MEMORY_HOTPLUG */
548 static inline void make_bl(unsigned int *insn_addr, void *func)
550 unsigned long funcp = *((unsigned long *)func);
551 int offset = funcp - (unsigned long)insn_addr;
553 *insn_addr = (unsigned int)(0x48000001 | (offset & 0x03fffffc));
554 flush_icache_range((unsigned long)insn_addr, 4+
555 (unsigned long)insn_addr);
558 static void __init htab_finish_init(void)
560 extern unsigned int *htab_call_hpte_insert1;
561 extern unsigned int *htab_call_hpte_insert2;
562 extern unsigned int *htab_call_hpte_remove;
563 extern unsigned int *htab_call_hpte_updatepp;
565 #ifdef CONFIG_PPC_HAS_HASH_64K
566 extern unsigned int *ht64_call_hpte_insert1;
567 extern unsigned int *ht64_call_hpte_insert2;
568 extern unsigned int *ht64_call_hpte_remove;
569 extern unsigned int *ht64_call_hpte_updatepp;
571 make_bl(ht64_call_hpte_insert1, ppc_md.hpte_insert);
572 make_bl(ht64_call_hpte_insert2, ppc_md.hpte_insert);
573 make_bl(ht64_call_hpte_remove, ppc_md.hpte_remove);
574 make_bl(ht64_call_hpte_updatepp, ppc_md.hpte_updatepp);
575 #endif /* CONFIG_PPC_HAS_HASH_64K */
577 make_bl(htab_call_hpte_insert1, ppc_md.hpte_insert);
578 make_bl(htab_call_hpte_insert2, ppc_md.hpte_insert);
579 make_bl(htab_call_hpte_remove, ppc_md.hpte_remove);
580 make_bl(htab_call_hpte_updatepp, ppc_md.hpte_updatepp);
583 void __init htab_initialize(void)
586 unsigned long pteg_count;
587 unsigned long prot, tprot;
588 unsigned long base = 0, size = 0, limit;
591 DBG(" -> htab_initialize()\n");
593 /* Initialize segment sizes */
594 htab_init_seg_sizes();
596 /* Initialize page sizes */
597 htab_init_page_sizes();
599 if (cpu_has_feature(CPU_FTR_1T_SEGMENT)) {
600 mmu_kernel_ssize = MMU_SEGSIZE_1T;
601 mmu_highuser_ssize = MMU_SEGSIZE_1T;
602 printk(KERN_INFO "Using 1TB segments\n");
606 * Calculate the required size of the htab. We want the number of
607 * PTEGs to equal one half the number of real pages.
609 htab_size_bytes = htab_get_table_size();
610 pteg_count = htab_size_bytes >> 7;
612 htab_hash_mask = pteg_count - 1;
614 if (firmware_has_feature(FW_FEATURE_LPAR)) {
615 /* Using a hypervisor which owns the htab */
619 /* Find storage for the HPT. Must be contiguous in
620 * the absolute address space. On cell we want it to be
621 * in the first 2 Gig so we can use it for IOMMU hacks.
623 if (machine_is(cell))
628 table = lmb_alloc_base(htab_size_bytes, htab_size_bytes, limit);
630 DBG("Hash table allocated at %lx, size: %lx\n", table,
633 htab_address = abs_to_virt(table);
635 /* htab absolute addr + encoded htabsize */
636 _SDR1 = table + __ilog2(pteg_count) - 11;
638 /* Initialize the HPT with no entries */
639 memset((void *)table, 0, htab_size_bytes);
642 mtspr(SPRN_SDR1, _SDR1);
647 #ifdef CONFIG_DEBUG_PAGEALLOC
648 linear_map_hash_count = lmb_end_of_DRAM() >> PAGE_SHIFT;
649 linear_map_hash_slots = __va(lmb_alloc_base(linear_map_hash_count,
651 memset(linear_map_hash_slots, 0, linear_map_hash_count);
652 #endif /* CONFIG_DEBUG_PAGEALLOC */
654 /* On U3 based machines, we need to reserve the DART area and
655 * _NOT_ map it to avoid cache paradoxes as it's remapped non
659 /* create bolted the linear mapping in the hash table */
660 for (i=0; i < lmb.memory.cnt; i++) {
661 base = (unsigned long)__va(lmb.memory.region[i].base);
662 size = lmb.memory.region[i].size;
663 tprot = prot | (in_kernel_text(base) ? _PAGE_EXEC : 0);
665 DBG("creating mapping for region: %lx..%lx (prot: %x)\n",
668 #ifdef CONFIG_U3_DART
669 /* Do not map the DART space. Fortunately, it will be aligned
670 * in such a way that it will not cross two lmb regions and
671 * will fit within a single 16Mb page.
672 * The DART space is assumed to be a full 16Mb region even if
673 * we only use 2Mb of that space. We will use more of it later
674 * for AGP GART. We have to use a full 16Mb large page.
676 DBG("DART base: %lx\n", dart_tablebase);
678 if (dart_tablebase != 0 && dart_tablebase >= base
679 && dart_tablebase < (base + size)) {
680 unsigned long dart_table_end = dart_tablebase + 16 * MB;
681 if (base != dart_tablebase)
682 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
686 if ((base + size) > dart_table_end)
687 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
689 __pa(dart_table_end),
695 #endif /* CONFIG_U3_DART */
696 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
697 tprot, mmu_linear_psize, mmu_kernel_ssize));
701 * If we have a memory_limit and we've allocated TCEs then we need to
702 * explicitly map the TCE area at the top of RAM. We also cope with the
703 * case that the TCEs start below memory_limit.
704 * tce_alloc_start/end are 16MB aligned so the mapping should work
705 * for either 4K or 16MB pages.
707 if (tce_alloc_start) {
708 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
709 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
711 if (base + size >= tce_alloc_start)
712 tce_alloc_start = base + size + 1;
714 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
715 __pa(tce_alloc_start), prot,
716 mmu_linear_psize, mmu_kernel_ssize));
721 DBG(" <- htab_initialize()\n");
726 void htab_initialize_secondary(void)
728 if (!firmware_has_feature(FW_FEATURE_LPAR))
729 mtspr(SPRN_SDR1, _SDR1);
733 * Called by asm hashtable.S for doing lazy icache flush
735 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
739 if (!pfn_valid(pte_pfn(pte)))
742 page = pte_page(pte);
745 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
747 __flush_dcache_icache(page_address(page));
748 set_bit(PG_arch_1, &page->flags);
755 #ifdef CONFIG_PPC_MM_SLICES
756 unsigned int get_paca_psize(unsigned long addr)
758 unsigned long index, slices;
760 if (addr < SLICE_LOW_TOP) {
761 slices = get_paca()->context.low_slices_psize;
762 index = GET_LOW_SLICE_INDEX(addr);
764 slices = get_paca()->context.high_slices_psize;
765 index = GET_HIGH_SLICE_INDEX(addr);
767 return (slices >> (index * 4)) & 0xF;
771 unsigned int get_paca_psize(unsigned long addr)
773 return get_paca()->context.user_psize;
778 * Demote a segment to using 4k pages.
779 * For now this makes the whole process use 4k pages.
781 #ifdef CONFIG_PPC_64K_PAGES
782 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
784 if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
786 slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
787 #ifdef CONFIG_SPU_BASE
788 spu_flush_all_slbs(mm);
790 if (get_paca_psize(addr) != MMU_PAGE_4K) {
791 get_paca()->context = mm->context;
792 slb_flush_and_rebolt();
795 #endif /* CONFIG_PPC_64K_PAGES */
797 #ifdef CONFIG_PPC_SUBPAGE_PROT
799 * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
800 * Userspace sets the subpage permissions using the subpage_prot system call.
802 * Result is 0: full permissions, _PAGE_RW: read-only,
803 * _PAGE_USER or _PAGE_USER|_PAGE_RW: no access.
805 static int subpage_protection(pgd_t *pgdir, unsigned long ea)
807 struct subpage_prot_table *spt = pgd_subpage_prot(pgdir);
811 if (ea >= spt->maxaddr)
813 if (ea < 0x100000000) {
814 /* addresses below 4GB use spt->low_prot */
815 sbpm = spt->low_prot;
817 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
821 sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
824 spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
826 /* extract 2-bit bitfield for this 4k subpage */
827 spp >>= 30 - 2 * ((ea >> 12) & 0xf);
829 /* turn 0,1,2,3 into combination of _PAGE_USER and _PAGE_RW */
830 spp = ((spp & 2) ? _PAGE_USER : 0) | ((spp & 1) ? _PAGE_RW : 0);
834 #else /* CONFIG_PPC_SUBPAGE_PROT */
835 static inline int subpage_protection(pgd_t *pgdir, unsigned long ea)
843 * 1 - normal page fault
844 * -1 - critical hash insertion error
845 * -2 - access not permitted by subpage protection mechanism
847 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
851 struct mm_struct *mm;
854 int rc, user_region = 0, local = 0;
857 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
860 if ((ea & ~REGION_MASK) >= PGTABLE_RANGE) {
861 DBG_LOW(" out of pgtable range !\n");
865 /* Get region & vsid */
866 switch (REGION_ID(ea)) {
871 DBG_LOW(" user region with no mm !\n");
874 psize = get_slice_psize(mm, ea);
875 ssize = user_segment_size(ea);
876 vsid = get_vsid(mm->context.id, ea, ssize);
878 case VMALLOC_REGION_ID:
880 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
881 if (ea < VMALLOC_END)
882 psize = mmu_vmalloc_psize;
884 psize = mmu_io_psize;
885 ssize = mmu_kernel_ssize;
889 * Send the problem up to do_page_fault
893 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
900 /* Check CPU locality */
901 tmp = cpumask_of_cpu(smp_processor_id());
902 if (user_region && cpus_equal(mm->cpu_vm_mask, tmp))
905 #ifdef CONFIG_HUGETLB_PAGE
906 /* Handle hugepage regions */
907 if (HPAGE_SHIFT && mmu_huge_psizes[psize]) {
908 DBG_LOW(" -> huge page !\n");
909 return hash_huge_page(mm, access, ea, vsid, local, trap);
911 #endif /* CONFIG_HUGETLB_PAGE */
913 #ifndef CONFIG_PPC_64K_PAGES
914 /* If we use 4K pages and our psize is not 4K, then we are hitting
915 * a special driver mapping, we need to align the address before
918 if (psize != MMU_PAGE_4K)
919 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
920 #endif /* CONFIG_PPC_64K_PAGES */
922 /* Get PTE and page size from page tables */
923 ptep = find_linux_pte(pgdir, ea);
924 if (ptep == NULL || !pte_present(*ptep)) {
925 DBG_LOW(" no PTE !\n");
929 #ifndef CONFIG_PPC_64K_PAGES
930 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
932 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
933 pte_val(*(ptep + PTRS_PER_PTE)));
935 /* Pre-check access permissions (will be re-checked atomically
936 * in __hash_page_XX but this pre-check is a fast path
938 if (access & ~pte_val(*ptep)) {
939 DBG_LOW(" no access !\n");
943 /* Do actual hashing */
944 #ifdef CONFIG_PPC_64K_PAGES
945 /* If _PAGE_4K_PFN is set, make sure this is a 4k segment */
946 if ((pte_val(*ptep) & _PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
947 demote_segment_4k(mm, ea);
951 /* If this PTE is non-cacheable and we have restrictions on
952 * using non cacheable large pages, then we switch to 4k
954 if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
955 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
957 demote_segment_4k(mm, ea);
959 } else if (ea < VMALLOC_END) {
961 * some driver did a non-cacheable mapping
962 * in vmalloc space, so switch vmalloc
965 printk(KERN_ALERT "Reducing vmalloc segment "
966 "to 4kB pages because of "
967 "non-cacheable mapping\n");
968 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
969 #ifdef CONFIG_SPU_BASE
970 spu_flush_all_slbs(mm);
975 if (psize != get_paca_psize(ea)) {
976 get_paca()->context = mm->context;
977 slb_flush_and_rebolt();
979 } else if (get_paca()->vmalloc_sllp !=
980 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
981 get_paca()->vmalloc_sllp =
982 mmu_psize_defs[mmu_vmalloc_psize].sllp;
983 slb_vmalloc_update();
985 #endif /* CONFIG_PPC_64K_PAGES */
987 #ifdef CONFIG_PPC_HAS_HASH_64K
988 if (psize == MMU_PAGE_64K)
989 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
991 #endif /* CONFIG_PPC_HAS_HASH_64K */
993 int spp = subpage_protection(pgdir, ea);
997 rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1001 #ifndef CONFIG_PPC_64K_PAGES
1002 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1004 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1005 pte_val(*(ptep + PTRS_PER_PTE)));
1007 DBG_LOW(" -> rc=%d\n", rc);
1010 EXPORT_SYMBOL_GPL(hash_page);
1012 void hash_preload(struct mm_struct *mm, unsigned long ea,
1013 unsigned long access, unsigned long trap)
1019 unsigned long flags;
1023 BUG_ON(REGION_ID(ea) != USER_REGION_ID);
1025 #ifdef CONFIG_PPC_MM_SLICES
1026 /* We only prefault standard pages for now */
1027 if (unlikely(get_slice_psize(mm, ea) != mm->context.user_psize))
1031 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1032 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
1034 /* Get Linux PTE if available */
1038 ptep = find_linux_pte(pgdir, ea);
1042 #ifdef CONFIG_PPC_64K_PAGES
1043 /* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
1044 * a 64K kernel), then we don't preload, hash_page() will take
1045 * care of it once we actually try to access the page.
1046 * That way we don't have to duplicate all of the logic for segment
1047 * page size demotion here
1049 if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
1051 #endif /* CONFIG_PPC_64K_PAGES */
1054 ssize = user_segment_size(ea);
1055 vsid = get_vsid(mm->context.id, ea, ssize);
1057 /* Hash doesn't like irqs */
1058 local_irq_save(flags);
1060 /* Is that local to this CPU ? */
1061 mask = cpumask_of_cpu(smp_processor_id());
1062 if (cpus_equal(mm->cpu_vm_mask, mask))
1066 #ifdef CONFIG_PPC_HAS_HASH_64K
1067 if (mm->context.user_psize == MMU_PAGE_64K)
1068 __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1070 #endif /* CONFIG_PPC_HAS_HASH_64K */
1071 __hash_page_4K(ea, access, vsid, ptep, trap, local, ssize,
1072 subpage_protection(pgdir, ea));
1074 local_irq_restore(flags);
1077 /* WARNING: This is called from hash_low_64.S, if you change this prototype,
1078 * do not forget to update the assembly call site !
1080 void flush_hash_page(unsigned long va, real_pte_t pte, int psize, int ssize,
1083 unsigned long hash, index, shift, hidx, slot;
1085 DBG_LOW("flush_hash_page(va=%016x)\n", va);
1086 pte_iterate_hashed_subpages(pte, psize, va, index, shift) {
1087 hash = hpt_hash(va, shift, ssize);
1088 hidx = __rpte_to_hidx(pte, index);
1089 if (hidx & _PTEIDX_SECONDARY)
1091 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1092 slot += hidx & _PTEIDX_GROUP_IX;
1093 DBG_LOW(" sub %d: hash=%x, hidx=%x\n", index, slot, hidx);
1094 ppc_md.hpte_invalidate(slot, va, psize, ssize, local);
1095 } pte_iterate_hashed_end();
1098 void flush_hash_range(unsigned long number, int local)
1100 if (ppc_md.flush_hash_range)
1101 ppc_md.flush_hash_range(number, local);
1104 struct ppc64_tlb_batch *batch =
1105 &__get_cpu_var(ppc64_tlb_batch);
1107 for (i = 0; i < number; i++)
1108 flush_hash_page(batch->vaddr[i], batch->pte[i],
1109 batch->psize, batch->ssize, local);
1114 * low_hash_fault is called when we the low level hash code failed
1115 * to instert a PTE due to an hypervisor error
1117 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1119 if (user_mode(regs)) {
1120 #ifdef CONFIG_PPC_SUBPAGE_PROT
1122 _exception(SIGSEGV, regs, SEGV_ACCERR, address);
1125 _exception(SIGBUS, regs, BUS_ADRERR, address);
1127 bad_page_fault(regs, address, SIGBUS);
1130 #ifdef CONFIG_DEBUG_PAGEALLOC
1131 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1133 unsigned long hash, hpteg;
1134 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1135 unsigned long va = hpt_va(vaddr, vsid, mmu_kernel_ssize);
1136 unsigned long mode = htab_convert_pte_flags(PAGE_KERNEL);
1139 hash = hpt_hash(va, PAGE_SHIFT, mmu_kernel_ssize);
1140 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
1142 ret = ppc_md.hpte_insert(hpteg, va, __pa(vaddr),
1143 mode, HPTE_V_BOLTED,
1144 mmu_linear_psize, mmu_kernel_ssize);
1146 spin_lock(&linear_map_hash_lock);
1147 BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1148 linear_map_hash_slots[lmi] = ret | 0x80;
1149 spin_unlock(&linear_map_hash_lock);
1152 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1154 unsigned long hash, hidx, slot;
1155 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1156 unsigned long va = hpt_va(vaddr, vsid, mmu_kernel_ssize);
1158 hash = hpt_hash(va, PAGE_SHIFT, mmu_kernel_ssize);
1159 spin_lock(&linear_map_hash_lock);
1160 BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1161 hidx = linear_map_hash_slots[lmi] & 0x7f;
1162 linear_map_hash_slots[lmi] = 0;
1163 spin_unlock(&linear_map_hash_lock);
1164 if (hidx & _PTEIDX_SECONDARY)
1166 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1167 slot += hidx & _PTEIDX_GROUP_IX;
1168 ppc_md.hpte_invalidate(slot, va, mmu_linear_psize, mmu_kernel_ssize, 0);
1171 void kernel_map_pages(struct page *page, int numpages, int enable)
1173 unsigned long flags, vaddr, lmi;
1176 local_irq_save(flags);
1177 for (i = 0; i < numpages; i++, page++) {
1178 vaddr = (unsigned long)page_address(page);
1179 lmi = __pa(vaddr) >> PAGE_SHIFT;
1180 if (lmi >= linear_map_hash_count)
1183 kernel_map_linear_page(vaddr, lmi);
1185 kernel_unmap_linear_page(vaddr, lmi);
1187 local_irq_restore(flags);
1189 #endif /* CONFIG_DEBUG_PAGEALLOC */