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 */
88 unsigned long htab_hash_mask;
90 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
91 int mmu_linear_psize = MMU_PAGE_4K;
92 int mmu_virtual_psize = MMU_PAGE_4K;
93 #ifdef CONFIG_HUGETLB_PAGE
94 int mmu_huge_psize = MMU_PAGE_16M;
95 unsigned int HPAGE_SHIFT;
98 /* There are definitions of page sizes arrays to be used when none
99 * is provided by the firmware.
102 /* Pre-POWER4 CPUs (4k pages only)
104 struct mmu_psize_def mmu_psize_defaults_old[] = {
114 /* POWER4, GPUL, POWER5
116 * Support for 16Mb large pages
118 struct mmu_psize_def mmu_psize_defaults_gp[] = {
136 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
137 unsigned long pstart, unsigned long mode, int psize)
139 unsigned long vaddr, paddr;
140 unsigned int step, shift;
141 unsigned long tmp_mode;
144 shift = mmu_psize_defs[psize].shift;
147 for (vaddr = vstart, paddr = pstart; vaddr < vend;
148 vaddr += step, paddr += step) {
149 unsigned long vpn, hash, hpteg;
150 unsigned long vsid = get_kernel_vsid(vaddr);
151 unsigned long va = (vsid << 28) | (vaddr & 0x0fffffff);
156 /* Make non-kernel text non-executable */
157 if (!in_kernel_text(vaddr))
158 tmp_mode = mode | HPTE_R_N;
160 hash = hpt_hash(va, shift);
161 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
163 /* The crap below can be cleaned once ppd_md.probe() can
164 * set up the hash callbacks, thus we can just used the
165 * normal insert callback here.
167 #ifdef CONFIG_PPC_ISERIES
168 if (systemcfg->platform == PLATFORM_ISERIES_LPAR)
169 ret = iSeries_hpte_insert(hpteg, va,
176 #ifdef CONFIG_PPC_PSERIES
177 if (systemcfg->platform & PLATFORM_LPAR)
178 ret = pSeries_lpar_hpte_insert(hpteg, va,
185 #ifdef CONFIG_PPC_MULTIPLATFORM
186 ret = native_hpte_insert(hpteg, va,
188 tmp_mode, HPTE_V_BOLTED,
194 return ret < 0 ? ret : 0;
197 static int __init htab_dt_scan_page_sizes(unsigned long node,
198 const char *uname, int depth,
201 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
203 unsigned long size = 0;
205 /* We are scanning "cpu" nodes only */
206 if (type == NULL || strcmp(type, "cpu") != 0)
209 prop = (u32 *)of_get_flat_dt_prop(node,
210 "ibm,segment-page-sizes", &size);
212 DBG("Page sizes from device-tree:\n");
214 cur_cpu_spec->cpu_features &= ~(CPU_FTR_16M_PAGE);
216 unsigned int shift = prop[0];
217 unsigned int slbenc = prop[1];
218 unsigned int lpnum = prop[2];
219 unsigned int lpenc = 0;
220 struct mmu_psize_def *def;
223 size -= 3; prop += 3;
224 while(size > 0 && lpnum) {
225 if (prop[0] == shift)
227 prop += 2; size -= 2;
242 cur_cpu_spec->cpu_features |= CPU_FTR_16M_PAGE;
250 def = &mmu_psize_defs[idx];
255 def->avpnm = (1 << (shift - 23)) - 1;
258 /* We don't know for sure what's up with tlbiel, so
259 * for now we only set it for 4K and 64K pages
261 if (idx == MMU_PAGE_4K || idx == MMU_PAGE_64K)
266 DBG(" %d: shift=%02x, sllp=%04x, avpnm=%08x, "
267 "tlbiel=%d, penc=%d\n",
268 idx, shift, def->sllp, def->avpnm, def->tlbiel,
277 static void __init htab_init_page_sizes(void)
281 /* Default to 4K pages only */
282 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
283 sizeof(mmu_psize_defaults_old));
286 * Try to find the available page sizes in the device-tree
288 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
289 if (rc != 0) /* Found */
293 * Not in the device-tree, let's fallback on known size
294 * list for 16M capable GP & GR
296 if ((systemcfg->platform != PLATFORM_ISERIES_LPAR) &&
297 cpu_has_feature(CPU_FTR_16M_PAGE))
298 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
299 sizeof(mmu_psize_defaults_gp));
302 * Pick a size for the linear mapping. Currently, we only support
303 * 16M, 1M and 4K which is the default
305 if (mmu_psize_defs[MMU_PAGE_16M].shift)
306 mmu_linear_psize = MMU_PAGE_16M;
307 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
308 mmu_linear_psize = MMU_PAGE_1M;
311 * Pick a size for the ordinary pages. Default is 4K, we support
312 * 64K if cache inhibited large pages are supported by the
315 #ifdef CONFIG_PPC_64K_PAGES
316 if (mmu_psize_defs[MMU_PAGE_64K].shift &&
317 cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
318 mmu_virtual_psize = MMU_PAGE_64K;
321 printk(KERN_INFO "Page orders: linear mapping = %d, others = %d\n",
322 mmu_psize_defs[mmu_linear_psize].shift,
323 mmu_psize_defs[mmu_virtual_psize].shift);
325 #ifdef CONFIG_HUGETLB_PAGE
326 /* Init large page size. Currently, we pick 16M or 1M depending
327 * on what is available
329 if (mmu_psize_defs[MMU_PAGE_16M].shift)
330 mmu_huge_psize = MMU_PAGE_16M;
331 /* With 4k/4level pagetables, we can't (for now) cope with a
332 * huge page size < PMD_SIZE */
333 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
334 mmu_huge_psize = MMU_PAGE_1M;
336 /* Calculate HPAGE_SHIFT and sanity check it */
337 if (mmu_psize_defs[mmu_huge_psize].shift > MIN_HUGEPTE_SHIFT &&
338 mmu_psize_defs[mmu_huge_psize].shift < SID_SHIFT)
339 HPAGE_SHIFT = mmu_psize_defs[mmu_huge_psize].shift;
341 HPAGE_SHIFT = 0; /* No huge pages dude ! */
342 #endif /* CONFIG_HUGETLB_PAGE */
345 static int __init htab_dt_scan_pftsize(unsigned long node,
346 const char *uname, int depth,
349 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
352 /* We are scanning "cpu" nodes only */
353 if (type == NULL || strcmp(type, "cpu") != 0)
356 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
358 /* pft_size[0] is the NUMA CEC cookie */
359 ppc64_pft_size = prop[1];
365 static unsigned long __init htab_get_table_size(void)
367 unsigned long rnd_mem_size, pteg_count;
369 /* If hash size isn't already provided by the platform, we try to
370 * retreive it from the device-tree. If it's not there neither, we
371 * calculate it now based on the total RAM size
373 if (ppc64_pft_size == 0)
374 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
376 return 1UL << ppc64_pft_size;
378 /* round mem_size up to next power of 2 */
379 rnd_mem_size = 1UL << __ilog2(systemcfg->physicalMemorySize);
380 if (rnd_mem_size < systemcfg->physicalMemorySize)
384 pteg_count = max(rnd_mem_size >> (12 + 1), 1UL << 11);
386 return pteg_count << 7;
389 void __init htab_initialize(void)
391 unsigned long table, htab_size_bytes;
392 unsigned long pteg_count;
393 unsigned long mode_rw;
394 unsigned long base = 0, size = 0;
397 extern unsigned long tce_alloc_start, tce_alloc_end;
399 DBG(" -> htab_initialize()\n");
401 /* Initialize page sizes */
402 htab_init_page_sizes();
405 * Calculate the required size of the htab. We want the number of
406 * PTEGs to equal one half the number of real pages.
408 htab_size_bytes = htab_get_table_size();
409 pteg_count = htab_size_bytes >> 7;
411 htab_hash_mask = pteg_count - 1;
413 if (systemcfg->platform & PLATFORM_LPAR) {
414 /* Using a hypervisor which owns the htab */
418 /* Find storage for the HPT. Must be contiguous in
419 * the absolute address space.
421 table = lmb_alloc(htab_size_bytes, htab_size_bytes);
424 DBG("Hash table allocated at %lx, size: %lx\n", table,
427 htab_address = abs_to_virt(table);
429 /* htab absolute addr + encoded htabsize */
430 _SDR1 = table + __ilog2(pteg_count) - 11;
432 /* Initialize the HPT with no entries */
433 memset((void *)table, 0, htab_size_bytes);
436 mode_rw = _PAGE_ACCESSED | _PAGE_DIRTY | _PAGE_COHERENT | PP_RWXX;
438 /* On U3 based machines, we need to reserve the DART area and
439 * _NOT_ map it to avoid cache paradoxes as it's remapped non
443 /* create bolted the linear mapping in the hash table */
444 for (i=0; i < lmb.memory.cnt; i++) {
445 base = lmb.memory.region[i].base + KERNELBASE;
446 size = lmb.memory.region[i].size;
448 DBG("creating mapping for region: %lx : %lx\n", base, size);
450 #ifdef CONFIG_U3_DART
451 /* Do not map the DART space. Fortunately, it will be aligned
452 * in such a way that it will not cross two lmb regions and
453 * will fit within a single 16Mb page.
454 * The DART space is assumed to be a full 16Mb region even if
455 * we only use 2Mb of that space. We will use more of it later
456 * for AGP GART. We have to use a full 16Mb large page.
458 DBG("DART base: %lx\n", dart_tablebase);
460 if (dart_tablebase != 0 && dart_tablebase >= base
461 && dart_tablebase < (base + size)) {
462 if (base != dart_tablebase)
463 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
466 if ((base + size) > (dart_tablebase + 16*MB))
467 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
469 dart_tablebase+16*MB,
474 #endif /* CONFIG_U3_DART */
475 BUG_ON(htab_bolt_mapping(base, base + size, base,
476 mode_rw, mmu_linear_psize));
480 * If we have a memory_limit and we've allocated TCEs then we need to
481 * explicitly map the TCE area at the top of RAM. We also cope with the
482 * case that the TCEs start below memory_limit.
483 * tce_alloc_start/end are 16MB aligned so the mapping should work
484 * for either 4K or 16MB pages.
486 if (tce_alloc_start) {
487 tce_alloc_start += KERNELBASE;
488 tce_alloc_end += KERNELBASE;
490 if (base + size >= tce_alloc_start)
491 tce_alloc_start = base + size + 1;
493 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
494 tce_alloc_start, mode_rw,
498 DBG(" <- htab_initialize()\n");
504 * Called by asm hashtable.S for doing lazy icache flush
506 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
510 if (!pfn_valid(pte_pfn(pte)))
513 page = pte_page(pte);
516 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
518 __flush_dcache_icache(page_address(page));
519 set_bit(PG_arch_1, &page->flags);
528 * 1 - normal page fault
529 * -1 - critical hash insertion error
531 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
535 struct mm_struct *mm;
538 int rc, user_region = 0, local = 0;
540 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
543 if ((ea & ~REGION_MASK) >= PGTABLE_RANGE) {
544 DBG_LOW(" out of pgtable range !\n");
548 /* Get region & vsid */
549 switch (REGION_ID(ea)) {
554 DBG_LOW(" user region with no mm !\n");
557 vsid = get_vsid(mm->context.id, ea);
559 case VMALLOC_REGION_ID:
561 vsid = get_kernel_vsid(ea);
565 * Send the problem up to do_page_fault
569 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
576 /* Check CPU locality */
577 tmp = cpumask_of_cpu(smp_processor_id());
578 if (user_region && cpus_equal(mm->cpu_vm_mask, tmp))
581 /* Handle hugepage regions */
582 if (unlikely(in_hugepage_area(mm->context, ea))) {
583 DBG_LOW(" -> huge page !\n");
584 return hash_huge_page(mm, access, ea, vsid, local);
587 /* Get PTE and page size from page tables */
588 ptep = find_linux_pte(pgdir, ea);
589 if (ptep == NULL || !pte_present(*ptep)) {
590 DBG_LOW(" no PTE !\n");
594 #ifndef CONFIG_PPC_64K_PAGES
595 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
597 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
598 pte_val(*(ptep + PTRS_PER_PTE)));
600 /* Pre-check access permissions (will be re-checked atomically
601 * in __hash_page_XX but this pre-check is a fast path
603 if (access & ~pte_val(*ptep)) {
604 DBG_LOW(" no access !\n");
608 /* Do actual hashing */
609 #ifndef CONFIG_PPC_64K_PAGES
610 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local);
612 if (mmu_virtual_psize == MMU_PAGE_64K)
613 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local);
615 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local);
616 #endif /* CONFIG_PPC_64K_PAGES */
618 #ifndef CONFIG_PPC_64K_PAGES
619 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
621 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
622 pte_val(*(ptep + PTRS_PER_PTE)));
624 DBG_LOW(" -> rc=%d\n", rc);
628 void hash_preload(struct mm_struct *mm, unsigned long ea,
629 unsigned long access, unsigned long trap)
638 /* We don't want huge pages prefaulted for now
640 if (unlikely(in_hugepage_area(mm->context, ea)))
643 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
644 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
646 /* Get PTE, VSID, access mask */
650 ptep = find_linux_pte(pgdir, ea);
653 vsid = get_vsid(mm->context.id, ea);
656 local_irq_save(flags);
657 mask = cpumask_of_cpu(smp_processor_id());
658 if (cpus_equal(mm->cpu_vm_mask, mask))
660 #ifndef CONFIG_PPC_64K_PAGES
661 __hash_page_4K(ea, access, vsid, ptep, trap, local);
663 if (mmu_virtual_psize == MMU_PAGE_64K)
664 __hash_page_64K(ea, access, vsid, ptep, trap, local);
666 __hash_page_4K(ea, access, vsid, ptep, trap, local);
667 #endif /* CONFIG_PPC_64K_PAGES */
668 local_irq_restore(flags);
671 void flush_hash_page(unsigned long va, real_pte_t pte, int psize, int local)
673 unsigned long hash, index, shift, hidx, slot;
675 DBG_LOW("flush_hash_page(va=%016x)\n", va);
676 pte_iterate_hashed_subpages(pte, psize, va, index, shift) {
677 hash = hpt_hash(va, shift);
678 hidx = __rpte_to_hidx(pte, index);
679 if (hidx & _PTEIDX_SECONDARY)
681 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
682 slot += hidx & _PTEIDX_GROUP_IX;
683 DBG_LOW(" sub %d: hash=%x, hidx=%x\n", index, slot, hidx);
684 ppc_md.hpte_invalidate(slot, va, psize, local);
685 } pte_iterate_hashed_end();
688 void flush_hash_range(unsigned long number, int local)
690 if (ppc_md.flush_hash_range)
691 ppc_md.flush_hash_range(number, local);
694 struct ppc64_tlb_batch *batch =
695 &__get_cpu_var(ppc64_tlb_batch);
697 for (i = 0; i < number; i++)
698 flush_hash_page(batch->vaddr[i], batch->pte[i],
699 batch->psize, local);
703 static inline void make_bl(unsigned int *insn_addr, void *func)
705 unsigned long funcp = *((unsigned long *)func);
706 int offset = funcp - (unsigned long)insn_addr;
708 *insn_addr = (unsigned int)(0x48000001 | (offset & 0x03fffffc));
709 flush_icache_range((unsigned long)insn_addr, 4+
710 (unsigned long)insn_addr);
714 * low_hash_fault is called when we the low level hash code failed
715 * to instert a PTE due to an hypervisor error
717 void low_hash_fault(struct pt_regs *regs, unsigned long address)
719 if (user_mode(regs)) {
722 info.si_signo = SIGBUS;
724 info.si_code = BUS_ADRERR;
725 info.si_addr = (void __user *)address;
726 force_sig_info(SIGBUS, &info, current);
729 bad_page_fault(regs, address, SIGBUS);
732 void __init htab_finish_init(void)
734 extern unsigned int *htab_call_hpte_insert1;
735 extern unsigned int *htab_call_hpte_insert2;
736 extern unsigned int *htab_call_hpte_remove;
737 extern unsigned int *htab_call_hpte_updatepp;
739 #ifdef CONFIG_PPC_64K_PAGES
740 extern unsigned int *ht64_call_hpte_insert1;
741 extern unsigned int *ht64_call_hpte_insert2;
742 extern unsigned int *ht64_call_hpte_remove;
743 extern unsigned int *ht64_call_hpte_updatepp;
745 make_bl(ht64_call_hpte_insert1, ppc_md.hpte_insert);
746 make_bl(ht64_call_hpte_insert2, ppc_md.hpte_insert);
747 make_bl(ht64_call_hpte_remove, ppc_md.hpte_remove);
748 make_bl(ht64_call_hpte_updatepp, ppc_md.hpte_updatepp);
749 #endif /* CONFIG_PPC_64K_PAGES */
751 make_bl(htab_call_hpte_insert1, ppc_md.hpte_insert);
752 make_bl(htab_call_hpte_insert2, ppc_md.hpte_insert);
753 make_bl(htab_call_hpte_remove, ppc_md.hpte_remove);
754 make_bl(htab_call_hpte_updatepp, ppc_md.hpte_updatepp);