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>
35 #include <asm/processor.h>
36 #include <asm/pgtable.h>
38 #include <asm/mmu_context.h>
40 #include <asm/types.h>
41 #include <asm/system.h>
42 #include <asm/uaccess.h>
43 #include <asm/machdep.h>
45 #include <asm/abs_addr.h>
46 #include <asm/tlbflush.h>
50 #include <asm/cacheflush.h>
51 #include <asm/cputable.h>
52 #include <asm/sections.h>
57 #define DBG(fmt...) udbg_printf(fmt)
63 #define DBG_LOW(fmt...) udbg_printf(fmt)
65 #define DBG_LOW(fmt...)
72 * Note: pte --> Linux PTE
73 * HPTE --> PowerPC Hashed Page Table Entry
76 * htab_initialize is called with the MMU off (of course), but
77 * the kernel has been copied down to zero so it can directly
78 * reference global data. At this point it is very difficult
79 * to print debug info.
84 extern unsigned long dart_tablebase;
85 #endif /* CONFIG_U3_DART */
87 static unsigned long _SDR1;
88 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
90 struct hash_pte *htab_address;
91 unsigned long htab_size_bytes;
92 unsigned long htab_hash_mask;
93 int mmu_linear_psize = MMU_PAGE_4K;
94 int mmu_virtual_psize = MMU_PAGE_4K;
95 int mmu_vmalloc_psize = MMU_PAGE_4K;
96 int mmu_io_psize = MMU_PAGE_4K;
97 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
98 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
99 u16 mmu_slb_size = 64;
100 #ifdef CONFIG_HUGETLB_PAGE
101 int mmu_huge_psize = MMU_PAGE_16M;
102 unsigned int HPAGE_SHIFT;
104 #ifdef CONFIG_PPC_64K_PAGES
105 int mmu_ci_restrictions;
107 #ifdef CONFIG_DEBUG_PAGEALLOC
108 static u8 *linear_map_hash_slots;
109 static unsigned long linear_map_hash_count;
110 static DEFINE_SPINLOCK(linear_map_hash_lock);
111 #endif /* CONFIG_DEBUG_PAGEALLOC */
113 /* There are definitions of page sizes arrays to be used when none
114 * is provided by the firmware.
117 /* Pre-POWER4 CPUs (4k pages only)
119 struct mmu_psize_def mmu_psize_defaults_old[] = {
129 /* POWER4, GPUL, POWER5
131 * Support for 16Mb large pages
133 struct mmu_psize_def mmu_psize_defaults_gp[] = {
151 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
152 unsigned long pstart, unsigned long mode,
153 int psize, int ssize)
155 unsigned long vaddr, paddr;
156 unsigned int step, shift;
157 unsigned long tmp_mode;
160 shift = mmu_psize_defs[psize].shift;
163 for (vaddr = vstart, paddr = pstart; vaddr < vend;
164 vaddr += step, paddr += step) {
165 unsigned long hash, hpteg;
166 unsigned long vsid = get_kernel_vsid(vaddr, ssize);
167 unsigned long va = hpt_va(vaddr, vsid, ssize);
171 /* Make non-kernel text non-executable */
172 if (!in_kernel_text(vaddr))
173 tmp_mode = mode | HPTE_R_N;
175 hash = hpt_hash(va, shift, ssize);
176 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
178 DBG("htab_bolt_mapping: calling %p\n", ppc_md.hpte_insert);
180 BUG_ON(!ppc_md.hpte_insert);
181 ret = ppc_md.hpte_insert(hpteg, va, paddr,
182 tmp_mode, HPTE_V_BOLTED, psize, ssize);
186 #ifdef CONFIG_DEBUG_PAGEALLOC
187 if ((paddr >> PAGE_SHIFT) < linear_map_hash_count)
188 linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
189 #endif /* CONFIG_DEBUG_PAGEALLOC */
191 return ret < 0 ? ret : 0;
194 static int __init htab_dt_scan_seg_sizes(unsigned long node,
195 const char *uname, int depth,
198 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
200 unsigned long size = 0;
202 /* We are scanning "cpu" nodes only */
203 if (type == NULL || strcmp(type, "cpu") != 0)
206 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,processor-segment-sizes",
210 for (; size >= 4; size -= 4, ++prop) {
212 DBG("1T segment support detected\n");
213 cur_cpu_spec->cpu_features |= CPU_FTR_1T_SEGMENT;
217 cur_cpu_spec->cpu_features &= ~CPU_FTR_NO_SLBIE_B;
221 static void __init htab_init_seg_sizes(void)
223 of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
226 static int __init htab_dt_scan_page_sizes(unsigned long node,
227 const char *uname, int depth,
230 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
232 unsigned long size = 0;
234 /* We are scanning "cpu" nodes only */
235 if (type == NULL || strcmp(type, "cpu") != 0)
238 prop = (u32 *)of_get_flat_dt_prop(node,
239 "ibm,segment-page-sizes", &size);
241 DBG("Page sizes from device-tree:\n");
243 cur_cpu_spec->cpu_features &= ~(CPU_FTR_16M_PAGE);
245 unsigned int shift = prop[0];
246 unsigned int slbenc = prop[1];
247 unsigned int lpnum = prop[2];
248 unsigned int lpenc = 0;
249 struct mmu_psize_def *def;
252 size -= 3; prop += 3;
253 while(size > 0 && lpnum) {
254 if (prop[0] == shift)
256 prop += 2; size -= 2;
271 cur_cpu_spec->cpu_features |= CPU_FTR_16M_PAGE;
279 def = &mmu_psize_defs[idx];
284 def->avpnm = (1 << (shift - 23)) - 1;
287 /* We don't know for sure what's up with tlbiel, so
288 * for now we only set it for 4K and 64K pages
290 if (idx == MMU_PAGE_4K || idx == MMU_PAGE_64K)
295 DBG(" %d: shift=%02x, sllp=%04x, avpnm=%08x, "
296 "tlbiel=%d, penc=%d\n",
297 idx, shift, def->sllp, def->avpnm, def->tlbiel,
305 static void __init htab_init_page_sizes(void)
309 /* Default to 4K pages only */
310 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
311 sizeof(mmu_psize_defaults_old));
314 * Try to find the available page sizes in the device-tree
316 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
317 if (rc != 0) /* Found */
321 * Not in the device-tree, let's fallback on known size
322 * list for 16M capable GP & GR
324 if (cpu_has_feature(CPU_FTR_16M_PAGE))
325 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
326 sizeof(mmu_psize_defaults_gp));
328 #ifndef CONFIG_DEBUG_PAGEALLOC
330 * Pick a size for the linear mapping. Currently, we only support
331 * 16M, 1M and 4K which is the default
333 if (mmu_psize_defs[MMU_PAGE_16M].shift)
334 mmu_linear_psize = MMU_PAGE_16M;
335 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
336 mmu_linear_psize = MMU_PAGE_1M;
337 #endif /* CONFIG_DEBUG_PAGEALLOC */
339 #ifdef CONFIG_PPC_64K_PAGES
341 * Pick a size for the ordinary pages. Default is 4K, we support
342 * 64K for user mappings and vmalloc if supported by the processor.
343 * We only use 64k for ioremap if the processor
344 * (and firmware) support cache-inhibited large pages.
345 * If not, we use 4k and set mmu_ci_restrictions so that
346 * hash_page knows to switch processes that use cache-inhibited
347 * mappings to 4k pages.
349 if (mmu_psize_defs[MMU_PAGE_64K].shift) {
350 mmu_virtual_psize = MMU_PAGE_64K;
351 mmu_vmalloc_psize = MMU_PAGE_64K;
352 if (mmu_linear_psize == MMU_PAGE_4K)
353 mmu_linear_psize = MMU_PAGE_64K;
354 if (cpu_has_feature(CPU_FTR_CI_LARGE_PAGE)) {
356 * Don't use 64k pages for ioremap on pSeries, since
357 * that would stop us accessing the HEA ethernet.
359 if (!machine_is(pseries))
360 mmu_io_psize = MMU_PAGE_64K;
362 mmu_ci_restrictions = 1;
364 #endif /* CONFIG_PPC_64K_PAGES */
366 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
367 "virtual = %d, io = %d\n",
368 mmu_psize_defs[mmu_linear_psize].shift,
369 mmu_psize_defs[mmu_virtual_psize].shift,
370 mmu_psize_defs[mmu_io_psize].shift);
372 #ifdef CONFIG_HUGETLB_PAGE
373 /* Init large page size. Currently, we pick 16M or 1M depending
374 * on what is available
376 if (mmu_psize_defs[MMU_PAGE_16M].shift)
377 set_huge_psize(MMU_PAGE_16M);
378 /* With 4k/4level pagetables, we can't (for now) cope with a
379 * huge page size < PMD_SIZE */
380 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
381 set_huge_psize(MMU_PAGE_1M);
382 #endif /* CONFIG_HUGETLB_PAGE */
385 static int __init htab_dt_scan_pftsize(unsigned long node,
386 const char *uname, int depth,
389 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
392 /* We are scanning "cpu" nodes only */
393 if (type == NULL || strcmp(type, "cpu") != 0)
396 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
398 /* pft_size[0] is the NUMA CEC cookie */
399 ppc64_pft_size = prop[1];
405 static unsigned long __init htab_get_table_size(void)
407 unsigned long mem_size, rnd_mem_size, pteg_count;
409 /* If hash size isn't already provided by the platform, we try to
410 * retrieve it from the device-tree. If it's not there neither, we
411 * calculate it now based on the total RAM size
413 if (ppc64_pft_size == 0)
414 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
416 return 1UL << ppc64_pft_size;
418 /* round mem_size up to next power of 2 */
419 mem_size = lmb_phys_mem_size();
420 rnd_mem_size = 1UL << __ilog2(mem_size);
421 if (rnd_mem_size < mem_size)
425 pteg_count = max(rnd_mem_size >> (12 + 1), 1UL << 11);
427 return pteg_count << 7;
430 #ifdef CONFIG_MEMORY_HOTPLUG
431 void create_section_mapping(unsigned long start, unsigned long end)
433 BUG_ON(htab_bolt_mapping(start, end, __pa(start),
434 _PAGE_ACCESSED | _PAGE_DIRTY | _PAGE_COHERENT | PP_RWXX,
435 mmu_linear_psize, mmu_kernel_ssize));
437 #endif /* CONFIG_MEMORY_HOTPLUG */
439 static inline void make_bl(unsigned int *insn_addr, void *func)
441 unsigned long funcp = *((unsigned long *)func);
442 int offset = funcp - (unsigned long)insn_addr;
444 *insn_addr = (unsigned int)(0x48000001 | (offset & 0x03fffffc));
445 flush_icache_range((unsigned long)insn_addr, 4+
446 (unsigned long)insn_addr);
449 static void __init htab_finish_init(void)
451 extern unsigned int *htab_call_hpte_insert1;
452 extern unsigned int *htab_call_hpte_insert2;
453 extern unsigned int *htab_call_hpte_remove;
454 extern unsigned int *htab_call_hpte_updatepp;
456 #ifdef CONFIG_PPC_HAS_HASH_64K
457 extern unsigned int *ht64_call_hpte_insert1;
458 extern unsigned int *ht64_call_hpte_insert2;
459 extern unsigned int *ht64_call_hpte_remove;
460 extern unsigned int *ht64_call_hpte_updatepp;
462 make_bl(ht64_call_hpte_insert1, ppc_md.hpte_insert);
463 make_bl(ht64_call_hpte_insert2, ppc_md.hpte_insert);
464 make_bl(ht64_call_hpte_remove, ppc_md.hpte_remove);
465 make_bl(ht64_call_hpte_updatepp, ppc_md.hpte_updatepp);
466 #endif /* CONFIG_PPC_HAS_HASH_64K */
468 make_bl(htab_call_hpte_insert1, ppc_md.hpte_insert);
469 make_bl(htab_call_hpte_insert2, ppc_md.hpte_insert);
470 make_bl(htab_call_hpte_remove, ppc_md.hpte_remove);
471 make_bl(htab_call_hpte_updatepp, ppc_md.hpte_updatepp);
474 void __init htab_initialize(void)
477 unsigned long pteg_count;
478 unsigned long mode_rw;
479 unsigned long base = 0, size = 0, limit;
482 extern unsigned long tce_alloc_start, tce_alloc_end;
484 DBG(" -> htab_initialize()\n");
486 /* Initialize segment sizes */
487 htab_init_seg_sizes();
489 /* Initialize page sizes */
490 htab_init_page_sizes();
492 if (cpu_has_feature(CPU_FTR_1T_SEGMENT)) {
493 mmu_kernel_ssize = MMU_SEGSIZE_1T;
494 mmu_highuser_ssize = MMU_SEGSIZE_1T;
495 printk(KERN_INFO "Using 1TB segments\n");
499 * Calculate the required size of the htab. We want the number of
500 * PTEGs to equal one half the number of real pages.
502 htab_size_bytes = htab_get_table_size();
503 pteg_count = htab_size_bytes >> 7;
505 htab_hash_mask = pteg_count - 1;
507 if (firmware_has_feature(FW_FEATURE_LPAR)) {
508 /* Using a hypervisor which owns the htab */
512 /* Find storage for the HPT. Must be contiguous in
513 * the absolute address space. On cell we want it to be
514 * in the first 2 Gig so we can use it for IOMMU hacks.
516 if (machine_is(cell))
521 table = lmb_alloc_base(htab_size_bytes, htab_size_bytes, limit);
523 DBG("Hash table allocated at %lx, size: %lx\n", table,
526 htab_address = abs_to_virt(table);
528 /* htab absolute addr + encoded htabsize */
529 _SDR1 = table + __ilog2(pteg_count) - 11;
531 /* Initialize the HPT with no entries */
532 memset((void *)table, 0, htab_size_bytes);
535 mtspr(SPRN_SDR1, _SDR1);
538 mode_rw = _PAGE_ACCESSED | _PAGE_DIRTY | _PAGE_COHERENT | PP_RWXX;
540 #ifdef CONFIG_DEBUG_PAGEALLOC
541 linear_map_hash_count = lmb_end_of_DRAM() >> PAGE_SHIFT;
542 linear_map_hash_slots = __va(lmb_alloc_base(linear_map_hash_count,
544 memset(linear_map_hash_slots, 0, linear_map_hash_count);
545 #endif /* CONFIG_DEBUG_PAGEALLOC */
547 /* On U3 based machines, we need to reserve the DART area and
548 * _NOT_ map it to avoid cache paradoxes as it's remapped non
552 /* create bolted the linear mapping in the hash table */
553 for (i=0; i < lmb.memory.cnt; i++) {
554 base = (unsigned long)__va(lmb.memory.region[i].base);
555 size = lmb.memory.region[i].size;
557 DBG("creating mapping for region: %lx : %lx\n", base, size);
559 #ifdef CONFIG_U3_DART
560 /* Do not map the DART space. Fortunately, it will be aligned
561 * in such a way that it will not cross two lmb regions and
562 * will fit within a single 16Mb page.
563 * The DART space is assumed to be a full 16Mb region even if
564 * we only use 2Mb of that space. We will use more of it later
565 * for AGP GART. We have to use a full 16Mb large page.
567 DBG("DART base: %lx\n", dart_tablebase);
569 if (dart_tablebase != 0 && dart_tablebase >= base
570 && dart_tablebase < (base + size)) {
571 unsigned long dart_table_end = dart_tablebase + 16 * MB;
572 if (base != dart_tablebase)
573 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
577 if ((base + size) > dart_table_end)
578 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
580 __pa(dart_table_end),
586 #endif /* CONFIG_U3_DART */
587 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
588 mode_rw, mmu_linear_psize, mmu_kernel_ssize));
592 * If we have a memory_limit and we've allocated TCEs then we need to
593 * explicitly map the TCE area at the top of RAM. We also cope with the
594 * case that the TCEs start below memory_limit.
595 * tce_alloc_start/end are 16MB aligned so the mapping should work
596 * for either 4K or 16MB pages.
598 if (tce_alloc_start) {
599 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
600 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
602 if (base + size >= tce_alloc_start)
603 tce_alloc_start = base + size + 1;
605 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
606 __pa(tce_alloc_start), mode_rw,
607 mmu_linear_psize, mmu_kernel_ssize));
612 DBG(" <- htab_initialize()\n");
617 void htab_initialize_secondary(void)
619 if (!firmware_has_feature(FW_FEATURE_LPAR))
620 mtspr(SPRN_SDR1, _SDR1);
624 * Called by asm hashtable.S for doing lazy icache flush
626 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
630 if (!pfn_valid(pte_pfn(pte)))
633 page = pte_page(pte);
636 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
638 __flush_dcache_icache(page_address(page));
639 set_bit(PG_arch_1, &page->flags);
647 * Demote a segment to using 4k pages.
648 * For now this makes the whole process use 4k pages.
650 #ifdef CONFIG_PPC_64K_PAGES
651 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
653 if (mm->context.user_psize == MMU_PAGE_4K)
655 slice_set_user_psize(mm, MMU_PAGE_4K);
656 #ifdef CONFIG_SPU_BASE
657 spu_flush_all_slbs(mm);
659 if (get_paca()->context.user_psize != MMU_PAGE_4K) {
660 get_paca()->context = mm->context;
661 slb_flush_and_rebolt();
664 #endif /* CONFIG_PPC_64K_PAGES */
666 #ifdef CONFIG_PPC_SUBPAGE_PROT
668 * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
669 * Userspace sets the subpage permissions using the subpage_prot system call.
671 * Result is 0: full permissions, _PAGE_RW: read-only,
672 * _PAGE_USER or _PAGE_USER|_PAGE_RW: no access.
674 static int subpage_protection(pgd_t *pgdir, unsigned long ea)
676 struct subpage_prot_table *spt = pgd_subpage_prot(pgdir);
680 if (ea >= spt->maxaddr)
682 if (ea < 0x100000000) {
683 /* addresses below 4GB use spt->low_prot */
684 sbpm = spt->low_prot;
686 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
690 sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
693 spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
695 /* extract 2-bit bitfield for this 4k subpage */
696 spp >>= 30 - 2 * ((ea >> 12) & 0xf);
698 /* turn 0,1,2,3 into combination of _PAGE_USER and _PAGE_RW */
699 spp = ((spp & 2) ? _PAGE_USER : 0) | ((spp & 1) ? _PAGE_RW : 0);
703 #else /* CONFIG_PPC_SUBPAGE_PROT */
704 static inline int subpage_protection(pgd_t *pgdir, unsigned long ea)
712 * 1 - normal page fault
713 * -1 - critical hash insertion error
714 * -2 - access not permitted by subpage protection mechanism
716 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
720 struct mm_struct *mm;
723 int rc, user_region = 0, local = 0;
726 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
729 if ((ea & ~REGION_MASK) >= PGTABLE_RANGE) {
730 DBG_LOW(" out of pgtable range !\n");
734 /* Get region & vsid */
735 switch (REGION_ID(ea)) {
740 DBG_LOW(" user region with no mm !\n");
743 #ifdef CONFIG_PPC_MM_SLICES
744 psize = get_slice_psize(mm, ea);
746 psize = mm->context.user_psize;
748 ssize = user_segment_size(ea);
749 vsid = get_vsid(mm->context.id, ea, ssize);
751 case VMALLOC_REGION_ID:
753 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
754 if (ea < VMALLOC_END)
755 psize = mmu_vmalloc_psize;
757 psize = mmu_io_psize;
758 ssize = mmu_kernel_ssize;
762 * Send the problem up to do_page_fault
766 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
773 /* Check CPU locality */
774 tmp = cpumask_of_cpu(smp_processor_id());
775 if (user_region && cpus_equal(mm->cpu_vm_mask, tmp))
778 #ifdef CONFIG_HUGETLB_PAGE
779 /* Handle hugepage regions */
780 if (HPAGE_SHIFT && psize == mmu_huge_psize) {
781 DBG_LOW(" -> huge page !\n");
782 return hash_huge_page(mm, access, ea, vsid, local, trap);
784 #endif /* CONFIG_HUGETLB_PAGE */
786 #ifndef CONFIG_PPC_64K_PAGES
787 /* If we use 4K pages and our psize is not 4K, then we are hitting
788 * a special driver mapping, we need to align the address before
791 if (psize != MMU_PAGE_4K)
792 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
793 #endif /* CONFIG_PPC_64K_PAGES */
795 /* Get PTE and page size from page tables */
796 ptep = find_linux_pte(pgdir, ea);
797 if (ptep == NULL || !pte_present(*ptep)) {
798 DBG_LOW(" no PTE !\n");
802 #ifndef CONFIG_PPC_64K_PAGES
803 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
805 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
806 pte_val(*(ptep + PTRS_PER_PTE)));
808 /* Pre-check access permissions (will be re-checked atomically
809 * in __hash_page_XX but this pre-check is a fast path
811 if (access & ~pte_val(*ptep)) {
812 DBG_LOW(" no access !\n");
816 /* Do actual hashing */
817 #ifdef CONFIG_PPC_64K_PAGES
818 /* If _PAGE_4K_PFN is set, make sure this is a 4k segment */
819 if (pte_val(*ptep) & _PAGE_4K_PFN) {
820 demote_segment_4k(mm, ea);
824 /* If this PTE is non-cacheable and we have restrictions on
825 * using non cacheable large pages, then we switch to 4k
827 if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
828 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
830 demote_segment_4k(mm, ea);
832 } else if (ea < VMALLOC_END) {
834 * some driver did a non-cacheable mapping
835 * in vmalloc space, so switch vmalloc
838 printk(KERN_ALERT "Reducing vmalloc segment "
839 "to 4kB pages because of "
840 "non-cacheable mapping\n");
841 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
842 #ifdef CONFIG_SPU_BASE
843 spu_flush_all_slbs(mm);
848 if (psize != get_paca()->context.user_psize) {
849 get_paca()->context = mm->context;
850 slb_flush_and_rebolt();
852 } else if (get_paca()->vmalloc_sllp !=
853 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
854 get_paca()->vmalloc_sllp =
855 mmu_psize_defs[mmu_vmalloc_psize].sllp;
856 slb_vmalloc_update();
858 #endif /* CONFIG_PPC_64K_PAGES */
860 #ifdef CONFIG_PPC_HAS_HASH_64K
861 if (psize == MMU_PAGE_64K)
862 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
864 #endif /* CONFIG_PPC_HAS_HASH_64K */
866 int spp = subpage_protection(pgdir, ea);
870 rc = __hash_page_4K(ea, access, vsid, ptep, trap,
874 #ifndef CONFIG_PPC_64K_PAGES
875 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
877 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
878 pte_val(*(ptep + PTRS_PER_PTE)));
880 DBG_LOW(" -> rc=%d\n", rc);
883 EXPORT_SYMBOL_GPL(hash_page);
885 void hash_preload(struct mm_struct *mm, unsigned long ea,
886 unsigned long access, unsigned long trap)
896 BUG_ON(REGION_ID(ea) != USER_REGION_ID);
898 #ifdef CONFIG_PPC_MM_SLICES
899 /* We only prefault standard pages for now */
900 if (unlikely(get_slice_psize(mm, ea) != mm->context.user_psize))
904 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
905 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
907 /* Get Linux PTE if available */
911 ptep = find_linux_pte(pgdir, ea);
915 #ifdef CONFIG_PPC_64K_PAGES
916 /* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
917 * a 64K kernel), then we don't preload, hash_page() will take
918 * care of it once we actually try to access the page.
919 * That way we don't have to duplicate all of the logic for segment
920 * page size demotion here
922 if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
924 #endif /* CONFIG_PPC_64K_PAGES */
927 ssize = user_segment_size(ea);
928 vsid = get_vsid(mm->context.id, ea, ssize);
930 /* Hash doesn't like irqs */
931 local_irq_save(flags);
933 /* Is that local to this CPU ? */
934 mask = cpumask_of_cpu(smp_processor_id());
935 if (cpus_equal(mm->cpu_vm_mask, mask))
939 #ifdef CONFIG_PPC_HAS_HASH_64K
940 if (mm->context.user_psize == MMU_PAGE_64K)
941 __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
943 #endif /* CONFIG_PPC_HAS_HASH_64K */
944 __hash_page_4K(ea, access, vsid, ptep, trap, local, ssize,
945 subpage_protection(pgdir, ea));
947 local_irq_restore(flags);
950 /* WARNING: This is called from hash_low_64.S, if you change this prototype,
951 * do not forget to update the assembly call site !
953 void flush_hash_page(unsigned long va, real_pte_t pte, int psize, int ssize,
956 unsigned long hash, index, shift, hidx, slot;
958 DBG_LOW("flush_hash_page(va=%016x)\n", va);
959 pte_iterate_hashed_subpages(pte, psize, va, index, shift) {
960 hash = hpt_hash(va, shift, ssize);
961 hidx = __rpte_to_hidx(pte, index);
962 if (hidx & _PTEIDX_SECONDARY)
964 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
965 slot += hidx & _PTEIDX_GROUP_IX;
966 DBG_LOW(" sub %d: hash=%x, hidx=%x\n", index, slot, hidx);
967 ppc_md.hpte_invalidate(slot, va, psize, ssize, local);
968 } pte_iterate_hashed_end();
971 void flush_hash_range(unsigned long number, int local)
973 if (ppc_md.flush_hash_range)
974 ppc_md.flush_hash_range(number, local);
977 struct ppc64_tlb_batch *batch =
978 &__get_cpu_var(ppc64_tlb_batch);
980 for (i = 0; i < number; i++)
981 flush_hash_page(batch->vaddr[i], batch->pte[i],
982 batch->psize, batch->ssize, local);
987 * low_hash_fault is called when we the low level hash code failed
988 * to instert a PTE due to an hypervisor error
990 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
992 if (user_mode(regs)) {
993 #ifdef CONFIG_PPC_SUBPAGE_PROT
995 _exception(SIGSEGV, regs, SEGV_ACCERR, address);
998 _exception(SIGBUS, regs, BUS_ADRERR, address);
1000 bad_page_fault(regs, address, SIGBUS);
1003 #ifdef CONFIG_DEBUG_PAGEALLOC
1004 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1006 unsigned long hash, hpteg;
1007 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1008 unsigned long va = hpt_va(vaddr, vsid, mmu_kernel_ssize);
1009 unsigned long mode = _PAGE_ACCESSED | _PAGE_DIRTY |
1010 _PAGE_COHERENT | PP_RWXX | HPTE_R_N;
1013 hash = hpt_hash(va, PAGE_SHIFT, mmu_kernel_ssize);
1014 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
1016 ret = ppc_md.hpte_insert(hpteg, va, __pa(vaddr),
1017 mode, HPTE_V_BOLTED,
1018 mmu_linear_psize, mmu_kernel_ssize);
1020 spin_lock(&linear_map_hash_lock);
1021 BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1022 linear_map_hash_slots[lmi] = ret | 0x80;
1023 spin_unlock(&linear_map_hash_lock);
1026 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1028 unsigned long hash, hidx, slot;
1029 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1030 unsigned long va = hpt_va(vaddr, vsid, mmu_kernel_ssize);
1032 hash = hpt_hash(va, PAGE_SHIFT, mmu_kernel_ssize);
1033 spin_lock(&linear_map_hash_lock);
1034 BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1035 hidx = linear_map_hash_slots[lmi] & 0x7f;
1036 linear_map_hash_slots[lmi] = 0;
1037 spin_unlock(&linear_map_hash_lock);
1038 if (hidx & _PTEIDX_SECONDARY)
1040 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1041 slot += hidx & _PTEIDX_GROUP_IX;
1042 ppc_md.hpte_invalidate(slot, va, mmu_linear_psize, mmu_kernel_ssize, 0);
1045 void kernel_map_pages(struct page *page, int numpages, int enable)
1047 unsigned long flags, vaddr, lmi;
1050 local_irq_save(flags);
1051 for (i = 0; i < numpages; i++, page++) {
1052 vaddr = (unsigned long)page_address(page);
1053 lmi = __pa(vaddr) >> PAGE_SHIFT;
1054 if (lmi >= linear_map_hash_count)
1057 kernel_map_linear_page(vaddr, lmi);
1059 kernel_unmap_linear_page(vaddr, lmi);
1061 local_irq_restore(flags);
1063 #endif /* CONFIG_DEBUG_PAGEALLOC */