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/config.h>
25 #include <linux/spinlock.h>
26 #include <linux/errno.h>
27 #include <linux/sched.h>
28 #include <linux/proc_fs.h>
29 #include <linux/stat.h>
30 #include <linux/sysctl.h>
31 #include <linux/ctype.h>
32 #include <linux/cache.h>
33 #include <linux/init.h>
34 #include <linux/signal.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/abs_addr.h>
54 #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];
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 #ifdef CONFIG_HUGETLB_PAGE
98 int mmu_huge_psize = MMU_PAGE_16M;
99 unsigned int HPAGE_SHIFT;
101 #ifdef CONFIG_PPC_64K_PAGES
102 int mmu_ci_restrictions;
105 /* There are definitions of page sizes arrays to be used when none
106 * is provided by the firmware.
109 /* Pre-POWER4 CPUs (4k pages only)
111 struct mmu_psize_def mmu_psize_defaults_old[] = {
121 /* POWER4, GPUL, POWER5
123 * Support for 16Mb large pages
125 struct mmu_psize_def mmu_psize_defaults_gp[] = {
143 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
144 unsigned long pstart, unsigned long mode, int psize)
146 unsigned long vaddr, paddr;
147 unsigned int step, shift;
148 unsigned long tmp_mode;
151 shift = mmu_psize_defs[psize].shift;
154 for (vaddr = vstart, paddr = pstart; vaddr < vend;
155 vaddr += step, paddr += step) {
156 unsigned long vpn, hash, hpteg;
157 unsigned long vsid = get_kernel_vsid(vaddr);
158 unsigned long va = (vsid << 28) | (vaddr & 0x0fffffff);
163 /* Make non-kernel text non-executable */
164 if (!in_kernel_text(vaddr))
165 tmp_mode = mode | HPTE_R_N;
167 hash = hpt_hash(va, shift);
168 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
170 DBG("htab_bolt_mapping: calling %p\n", ppc_md.hpte_insert);
172 BUG_ON(!ppc_md.hpte_insert);
173 ret = ppc_md.hpte_insert(hpteg, va, paddr,
174 tmp_mode, HPTE_V_BOLTED, psize);
179 return ret < 0 ? ret : 0;
182 static int __init htab_dt_scan_page_sizes(unsigned long node,
183 const char *uname, int depth,
186 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
188 unsigned long size = 0;
190 /* We are scanning "cpu" nodes only */
191 if (type == NULL || strcmp(type, "cpu") != 0)
194 prop = (u32 *)of_get_flat_dt_prop(node,
195 "ibm,segment-page-sizes", &size);
197 DBG("Page sizes from device-tree:\n");
199 cur_cpu_spec->cpu_features &= ~(CPU_FTR_16M_PAGE);
201 unsigned int shift = prop[0];
202 unsigned int slbenc = prop[1];
203 unsigned int lpnum = prop[2];
204 unsigned int lpenc = 0;
205 struct mmu_psize_def *def;
208 size -= 3; prop += 3;
209 while(size > 0 && lpnum) {
210 if (prop[0] == shift)
212 prop += 2; size -= 2;
227 cur_cpu_spec->cpu_features |= CPU_FTR_16M_PAGE;
235 def = &mmu_psize_defs[idx];
240 def->avpnm = (1 << (shift - 23)) - 1;
243 /* We don't know for sure what's up with tlbiel, so
244 * for now we only set it for 4K and 64K pages
246 if (idx == MMU_PAGE_4K || idx == MMU_PAGE_64K)
251 DBG(" %d: shift=%02x, sllp=%04x, avpnm=%08x, "
252 "tlbiel=%d, penc=%d\n",
253 idx, shift, def->sllp, def->avpnm, def->tlbiel,
262 static void __init htab_init_page_sizes(void)
266 /* Default to 4K pages only */
267 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
268 sizeof(mmu_psize_defaults_old));
271 * Try to find the available page sizes in the device-tree
273 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
274 if (rc != 0) /* Found */
278 * Not in the device-tree, let's fallback on known size
279 * list for 16M capable GP & GR
281 if (cpu_has_feature(CPU_FTR_16M_PAGE) && !machine_is(iseries))
282 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
283 sizeof(mmu_psize_defaults_gp));
286 * Pick a size for the linear mapping. Currently, we only support
287 * 16M, 1M and 4K which is the default
289 if (mmu_psize_defs[MMU_PAGE_16M].shift)
290 mmu_linear_psize = MMU_PAGE_16M;
291 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
292 mmu_linear_psize = MMU_PAGE_1M;
294 #ifdef CONFIG_PPC_64K_PAGES
296 * Pick a size for the ordinary pages. Default is 4K, we support
297 * 64K for user mappings and vmalloc if supported by the processor.
298 * We only use 64k for ioremap if the processor
299 * (and firmware) support cache-inhibited large pages.
300 * If not, we use 4k and set mmu_ci_restrictions so that
301 * hash_page knows to switch processes that use cache-inhibited
302 * mappings to 4k pages.
304 if (mmu_psize_defs[MMU_PAGE_64K].shift) {
305 mmu_virtual_psize = MMU_PAGE_64K;
306 mmu_vmalloc_psize = MMU_PAGE_64K;
307 if (cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
308 mmu_io_psize = MMU_PAGE_64K;
310 mmu_ci_restrictions = 1;
314 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
315 "virtual = %d, io = %d\n",
316 mmu_psize_defs[mmu_linear_psize].shift,
317 mmu_psize_defs[mmu_virtual_psize].shift,
318 mmu_psize_defs[mmu_io_psize].shift);
320 #ifdef CONFIG_HUGETLB_PAGE
321 /* Init large page size. Currently, we pick 16M or 1M depending
322 * on what is available
324 if (mmu_psize_defs[MMU_PAGE_16M].shift)
325 mmu_huge_psize = MMU_PAGE_16M;
326 /* With 4k/4level pagetables, we can't (for now) cope with a
327 * huge page size < PMD_SIZE */
328 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
329 mmu_huge_psize = MMU_PAGE_1M;
331 /* Calculate HPAGE_SHIFT and sanity check it */
332 if (mmu_psize_defs[mmu_huge_psize].shift > MIN_HUGEPTE_SHIFT &&
333 mmu_psize_defs[mmu_huge_psize].shift < SID_SHIFT)
334 HPAGE_SHIFT = mmu_psize_defs[mmu_huge_psize].shift;
336 HPAGE_SHIFT = 0; /* No huge pages dude ! */
337 #endif /* CONFIG_HUGETLB_PAGE */
340 static int __init htab_dt_scan_pftsize(unsigned long node,
341 const char *uname, int depth,
344 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
347 /* We are scanning "cpu" nodes only */
348 if (type == NULL || strcmp(type, "cpu") != 0)
351 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
353 /* pft_size[0] is the NUMA CEC cookie */
354 ppc64_pft_size = prop[1];
360 static unsigned long __init htab_get_table_size(void)
362 unsigned long mem_size, rnd_mem_size, pteg_count;
364 /* If hash size isn't already provided by the platform, we try to
365 * retrieve it from the device-tree. If it's not there neither, we
366 * calculate it now based on the total RAM size
368 if (ppc64_pft_size == 0)
369 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
371 return 1UL << ppc64_pft_size;
373 /* round mem_size up to next power of 2 */
374 mem_size = lmb_phys_mem_size();
375 rnd_mem_size = 1UL << __ilog2(mem_size);
376 if (rnd_mem_size < mem_size)
380 pteg_count = max(rnd_mem_size >> (12 + 1), 1UL << 11);
382 return pteg_count << 7;
385 #ifdef CONFIG_MEMORY_HOTPLUG
386 void create_section_mapping(unsigned long start, unsigned long end)
388 BUG_ON(htab_bolt_mapping(start, end, __pa(start),
389 _PAGE_ACCESSED | _PAGE_DIRTY | _PAGE_COHERENT | PP_RWXX,
392 #endif /* CONFIG_MEMORY_HOTPLUG */
394 static inline void make_bl(unsigned int *insn_addr, void *func)
396 unsigned long funcp = *((unsigned long *)func);
397 int offset = funcp - (unsigned long)insn_addr;
399 *insn_addr = (unsigned int)(0x48000001 | (offset & 0x03fffffc));
400 flush_icache_range((unsigned long)insn_addr, 4+
401 (unsigned long)insn_addr);
404 static void __init htab_finish_init(void)
406 extern unsigned int *htab_call_hpte_insert1;
407 extern unsigned int *htab_call_hpte_insert2;
408 extern unsigned int *htab_call_hpte_remove;
409 extern unsigned int *htab_call_hpte_updatepp;
411 #ifdef CONFIG_PPC_64K_PAGES
412 extern unsigned int *ht64_call_hpte_insert1;
413 extern unsigned int *ht64_call_hpte_insert2;
414 extern unsigned int *ht64_call_hpte_remove;
415 extern unsigned int *ht64_call_hpte_updatepp;
417 make_bl(ht64_call_hpte_insert1, ppc_md.hpte_insert);
418 make_bl(ht64_call_hpte_insert2, ppc_md.hpte_insert);
419 make_bl(ht64_call_hpte_remove, ppc_md.hpte_remove);
420 make_bl(ht64_call_hpte_updatepp, ppc_md.hpte_updatepp);
421 #endif /* CONFIG_PPC_64K_PAGES */
423 make_bl(htab_call_hpte_insert1, ppc_md.hpte_insert);
424 make_bl(htab_call_hpte_insert2, ppc_md.hpte_insert);
425 make_bl(htab_call_hpte_remove, ppc_md.hpte_remove);
426 make_bl(htab_call_hpte_updatepp, ppc_md.hpte_updatepp);
429 void __init htab_initialize(void)
432 unsigned long pteg_count;
433 unsigned long mode_rw;
434 unsigned long base = 0, size = 0;
437 extern unsigned long tce_alloc_start, tce_alloc_end;
439 DBG(" -> htab_initialize()\n");
441 /* Initialize page sizes */
442 htab_init_page_sizes();
445 * Calculate the required size of the htab. We want the number of
446 * PTEGs to equal one half the number of real pages.
448 htab_size_bytes = htab_get_table_size();
449 pteg_count = htab_size_bytes >> 7;
451 htab_hash_mask = pteg_count - 1;
453 if (firmware_has_feature(FW_FEATURE_LPAR)) {
454 /* Using a hypervisor which owns the htab */
458 /* Find storage for the HPT. Must be contiguous in
459 * the absolute address space.
461 table = lmb_alloc(htab_size_bytes, htab_size_bytes);
463 DBG("Hash table allocated at %lx, size: %lx\n", table,
466 htab_address = abs_to_virt(table);
468 /* htab absolute addr + encoded htabsize */
469 _SDR1 = table + __ilog2(pteg_count) - 11;
471 /* Initialize the HPT with no entries */
472 memset((void *)table, 0, htab_size_bytes);
475 mtspr(SPRN_SDR1, _SDR1);
478 mode_rw = _PAGE_ACCESSED | _PAGE_DIRTY | _PAGE_COHERENT | PP_RWXX;
480 /* On U3 based machines, we need to reserve the DART area and
481 * _NOT_ map it to avoid cache paradoxes as it's remapped non
485 /* create bolted the linear mapping in the hash table */
486 for (i=0; i < lmb.memory.cnt; i++) {
487 base = (unsigned long)__va(lmb.memory.region[i].base);
488 size = lmb.memory.region[i].size;
490 DBG("creating mapping for region: %lx : %lx\n", base, size);
492 #ifdef CONFIG_U3_DART
493 /* Do not map the DART space. Fortunately, it will be aligned
494 * in such a way that it will not cross two lmb regions and
495 * will fit within a single 16Mb page.
496 * The DART space is assumed to be a full 16Mb region even if
497 * we only use 2Mb of that space. We will use more of it later
498 * for AGP GART. We have to use a full 16Mb large page.
500 DBG("DART base: %lx\n", dart_tablebase);
502 if (dart_tablebase != 0 && dart_tablebase >= base
503 && dart_tablebase < (base + size)) {
504 unsigned long dart_table_end = dart_tablebase + 16 * MB;
505 if (base != dart_tablebase)
506 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
509 if ((base + size) > dart_table_end)
510 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
512 __pa(dart_table_end),
517 #endif /* CONFIG_U3_DART */
518 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
519 mode_rw, mmu_linear_psize));
523 * If we have a memory_limit and we've allocated TCEs then we need to
524 * explicitly map the TCE area at the top of RAM. We also cope with the
525 * case that the TCEs start below memory_limit.
526 * tce_alloc_start/end are 16MB aligned so the mapping should work
527 * for either 4K or 16MB pages.
529 if (tce_alloc_start) {
530 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
531 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
533 if (base + size >= tce_alloc_start)
534 tce_alloc_start = base + size + 1;
536 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
537 __pa(tce_alloc_start), mode_rw,
543 DBG(" <- htab_initialize()\n");
548 void htab_initialize_secondary(void)
550 if (!firmware_has_feature(FW_FEATURE_LPAR))
551 mtspr(SPRN_SDR1, _SDR1);
555 * Called by asm hashtable.S for doing lazy icache flush
557 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
561 if (!pfn_valid(pte_pfn(pte)))
564 page = pte_page(pte);
567 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
569 __flush_dcache_icache(page_address(page));
570 set_bit(PG_arch_1, &page->flags);
579 * 1 - normal page fault
580 * -1 - critical hash insertion error
582 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
586 struct mm_struct *mm;
589 int rc, user_region = 0, local = 0;
592 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
595 if ((ea & ~REGION_MASK) >= PGTABLE_RANGE) {
596 DBG_LOW(" out of pgtable range !\n");
600 /* Get region & vsid */
601 switch (REGION_ID(ea)) {
606 DBG_LOW(" user region with no mm !\n");
609 vsid = get_vsid(mm->context.id, ea);
610 psize = mm->context.user_psize;
612 case VMALLOC_REGION_ID:
614 vsid = get_kernel_vsid(ea);
615 if (ea < VMALLOC_END)
616 psize = mmu_vmalloc_psize;
618 psize = mmu_io_psize;
622 * Send the problem up to do_page_fault
626 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
633 /* Check CPU locality */
634 tmp = cpumask_of_cpu(smp_processor_id());
635 if (user_region && cpus_equal(mm->cpu_vm_mask, tmp))
638 /* Handle hugepage regions */
639 if (unlikely(in_hugepage_area(mm->context, ea))) {
640 DBG_LOW(" -> huge page !\n");
641 return hash_huge_page(mm, access, ea, vsid, local, trap);
644 /* Get PTE and page size from page tables */
645 ptep = find_linux_pte(pgdir, ea);
646 if (ptep == NULL || !pte_present(*ptep)) {
647 DBG_LOW(" no PTE !\n");
651 #ifndef CONFIG_PPC_64K_PAGES
652 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
654 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
655 pte_val(*(ptep + PTRS_PER_PTE)));
657 /* Pre-check access permissions (will be re-checked atomically
658 * in __hash_page_XX but this pre-check is a fast path
660 if (access & ~pte_val(*ptep)) {
661 DBG_LOW(" no access !\n");
665 /* Do actual hashing */
666 #ifndef CONFIG_PPC_64K_PAGES
667 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local);
669 if (mmu_ci_restrictions) {
670 /* If this PTE is non-cacheable, switch to 4k */
671 if (psize == MMU_PAGE_64K &&
672 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
675 mm->context.user_psize = MMU_PAGE_4K;
676 mm->context.sllp = SLB_VSID_USER |
677 mmu_psize_defs[MMU_PAGE_4K].sllp;
678 } else if (ea < VMALLOC_END) {
680 * some driver did a non-cacheable mapping
681 * in vmalloc space, so switch vmalloc
684 printk(KERN_ALERT "Reducing vmalloc segment "
685 "to 4kB pages because of "
686 "non-cacheable mapping\n");
687 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
691 if (psize != get_paca()->context.user_psize) {
692 get_paca()->context = mm->context;
693 slb_flush_and_rebolt();
695 } else if (get_paca()->vmalloc_sllp !=
696 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
697 get_paca()->vmalloc_sllp =
698 mmu_psize_defs[mmu_vmalloc_psize].sllp;
699 slb_flush_and_rebolt();
702 if (psize == MMU_PAGE_64K)
703 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local);
705 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local);
706 #endif /* CONFIG_PPC_64K_PAGES */
708 #ifndef CONFIG_PPC_64K_PAGES
709 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
711 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
712 pte_val(*(ptep + PTRS_PER_PTE)));
714 DBG_LOW(" -> rc=%d\n", rc);
717 EXPORT_SYMBOL_GPL(hash_page);
719 void hash_preload(struct mm_struct *mm, unsigned long ea,
720 unsigned long access, unsigned long trap)
729 /* We don't want huge pages prefaulted for now
731 if (unlikely(in_hugepage_area(mm->context, ea)))
734 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
735 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
737 /* Get PTE, VSID, access mask */
741 ptep = find_linux_pte(pgdir, ea);
744 vsid = get_vsid(mm->context.id, ea);
747 local_irq_save(flags);
748 mask = cpumask_of_cpu(smp_processor_id());
749 if (cpus_equal(mm->cpu_vm_mask, mask))
751 #ifndef CONFIG_PPC_64K_PAGES
752 __hash_page_4K(ea, access, vsid, ptep, trap, local);
754 if (mmu_ci_restrictions) {
755 /* If this PTE is non-cacheable, switch to 4k */
756 if (mm->context.user_psize == MMU_PAGE_64K &&
757 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
758 mm->context.user_psize = MMU_PAGE_4K;
759 mm->context.sllp = SLB_VSID_USER |
760 mmu_psize_defs[MMU_PAGE_4K].sllp;
761 get_paca()->context = mm->context;
762 slb_flush_and_rebolt();
765 if (mm->context.user_psize == MMU_PAGE_64K)
766 __hash_page_64K(ea, access, vsid, ptep, trap, local);
768 __hash_page_4K(ea, access, vsid, ptep, trap, local);
769 #endif /* CONFIG_PPC_64K_PAGES */
770 local_irq_restore(flags);
773 void flush_hash_page(unsigned long va, real_pte_t pte, int psize, int local)
775 unsigned long hash, index, shift, hidx, slot;
777 DBG_LOW("flush_hash_page(va=%016x)\n", va);
778 pte_iterate_hashed_subpages(pte, psize, va, index, shift) {
779 hash = hpt_hash(va, shift);
780 hidx = __rpte_to_hidx(pte, index);
781 if (hidx & _PTEIDX_SECONDARY)
783 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
784 slot += hidx & _PTEIDX_GROUP_IX;
785 DBG_LOW(" sub %d: hash=%x, hidx=%x\n", index, slot, hidx);
786 ppc_md.hpte_invalidate(slot, va, psize, local);
787 } pte_iterate_hashed_end();
790 void flush_hash_range(unsigned long number, int local)
792 if (ppc_md.flush_hash_range)
793 ppc_md.flush_hash_range(number, local);
796 struct ppc64_tlb_batch *batch =
797 &__get_cpu_var(ppc64_tlb_batch);
799 for (i = 0; i < number; i++)
800 flush_hash_page(batch->vaddr[i], batch->pte[i],
801 batch->psize, local);
806 * low_hash_fault is called when we the low level hash code failed
807 * to instert a PTE due to an hypervisor error
809 void low_hash_fault(struct pt_regs *regs, unsigned long address)
811 if (user_mode(regs)) {
814 info.si_signo = SIGBUS;
816 info.si_code = BUS_ADRERR;
817 info.si_addr = (void __user *)address;
818 force_sig_info(SIGBUS, &info, current);
821 bad_page_fault(regs, address, SIGBUS);