blob: d692ae31cfc76fa1d8c53a94a97ac0bff8daa97c [file] [log] [blame]
Paul Mackerras14cf11a2005-09-26 16:04:21 +10001/*
2 * This file contains ioremap and related functions for 64-bit machines.
3 *
4 * Derived from arch/ppc64/mm/init.c
5 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
6 *
7 * Modifications by Paul Mackerras (PowerMac) (paulus@samba.org)
8 * and Cort Dougan (PReP) (cort@cs.nmt.edu)
9 * Copyright (C) 1996 Paul Mackerras
Paul Mackerras14cf11a2005-09-26 16:04:21 +100010 *
11 * Derived from "arch/i386/mm/init.c"
12 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
13 *
14 * Dave Engebretsen <engebret@us.ibm.com>
15 * Rework for PPC64 port.
16 *
17 * This program is free software; you can redistribute it and/or
18 * modify it under the terms of the GNU General Public License
19 * as published by the Free Software Foundation; either version
20 * 2 of the License, or (at your option) any later version.
21 *
22 */
23
Paul Mackerras14cf11a2005-09-26 16:04:21 +100024#include <linux/signal.h>
25#include <linux/sched.h>
26#include <linux/kernel.h>
27#include <linux/errno.h>
28#include <linux/string.h>
Paul Gortmaker66b15db2011-05-27 10:46:24 -040029#include <linux/export.h>
Paul Mackerras14cf11a2005-09-26 16:04:21 +100030#include <linux/types.h>
31#include <linux/mman.h>
32#include <linux/mm.h>
33#include <linux/swap.h>
34#include <linux/stddef.h>
35#include <linux/vmalloc.h>
Yinghai Lu95f72d12010-07-12 14:36:09 +100036#include <linux/memblock.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090037#include <linux/slab.h>
Aneesh Kumar K.V06743522014-11-05 21:57:39 +053038#include <linux/hugetlb.h>
Paul Mackerras14cf11a2005-09-26 16:04:21 +100039
40#include <asm/pgalloc.h>
41#include <asm/page.h>
42#include <asm/prom.h>
Paul Mackerras14cf11a2005-09-26 16:04:21 +100043#include <asm/io.h>
44#include <asm/mmu_context.h>
45#include <asm/pgtable.h>
46#include <asm/mmu.h>
Paul Mackerras14cf11a2005-09-26 16:04:21 +100047#include <asm/smp.h>
48#include <asm/machdep.h>
49#include <asm/tlb.h>
Paul Mackerras14cf11a2005-09-26 16:04:21 +100050#include <asm/processor.h>
Paul Mackerras14cf11a2005-09-26 16:04:21 +100051#include <asm/cputable.h>
Paul Mackerras14cf11a2005-09-26 16:04:21 +100052#include <asm/sections.h>
Stephen Rothwell5e203d62006-09-25 13:36:31 +100053#include <asm/firmware.h>
Anton Blanchard68cf0d62014-09-17 22:15:35 +100054#include <asm/dma.h>
David Gibson800fc3e2005-11-16 15:43:48 +110055
56#include "mmu_decl.h"
Paul Mackerras14cf11a2005-09-26 16:04:21 +100057
Aneesh Kumar K.V9e813302014-08-13 12:32:04 +053058#define CREATE_TRACE_POINTS
59#include <trace/events/thp.h>
60
Aneesh Kumar K.V78f1dbd2012-09-10 02:52:57 +000061/* Some sanity checking */
62#if TASK_SIZE_USER64 > PGTABLE_RANGE
63#error TASK_SIZE_USER64 exceeds pagetable range
64#endif
Paul Mackerras14cf11a2005-09-26 16:04:21 +100065
Aneesh Kumar K.V78f1dbd2012-09-10 02:52:57 +000066#ifdef CONFIG_PPC_STD_MMU_64
Aneesh Kumar K.Vaf81d782013-03-13 03:34:55 +000067#if TASK_SIZE_USER64 > (1UL << (ESID_BITS + SID_SHIFT))
Aneesh Kumar K.V78f1dbd2012-09-10 02:52:57 +000068#error TASK_SIZE_USER64 exceeds user VSID range
69#endif
70#endif
71
72unsigned long ioremap_bot = IOREMAP_BASE;
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +000073
74#ifdef CONFIG_PPC_MMU_NOHASH
Scott Wood7d176222014-08-01 22:07:40 -050075static __ref void *early_alloc_pgtable(unsigned long size)
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +000076{
77 void *pt;
78
Anton Blanchard10239732014-09-17 22:15:33 +100079 pt = __va(memblock_alloc_base(size, size, __pa(MAX_DMA_ADDRESS)));
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +000080 memset(pt, 0, size);
81
82 return pt;
83}
84#endif /* CONFIG_PPC_MMU_NOHASH */
85
Paul Mackerras14cf11a2005-09-26 16:04:21 +100086/*
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +000087 * map_kernel_page currently only called by __ioremap
88 * map_kernel_page adds an entry to the ioremap page table
Paul Mackerras14cf11a2005-09-26 16:04:21 +100089 * and adds an entry to the HPT, possibly bolting it
90 */
Benjamin Herrenschmidt32a74942009-07-23 23:15:58 +000091int map_kernel_page(unsigned long ea, unsigned long pa, int flags)
Paul Mackerras14cf11a2005-09-26 16:04:21 +100092{
93 pgd_t *pgdp;
94 pud_t *pudp;
95 pmd_t *pmdp;
96 pte_t *ptep;
Paul Mackerras14cf11a2005-09-26 16:04:21 +100097
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +000098 if (slab_is_available()) {
Paul Mackerras14cf11a2005-09-26 16:04:21 +100099 pgdp = pgd_offset_k(ea);
100 pudp = pud_alloc(&init_mm, pgdp, ea);
101 if (!pudp)
102 return -ENOMEM;
103 pmdp = pmd_alloc(&init_mm, pudp, ea);
104 if (!pmdp)
105 return -ENOMEM;
Paul Mackerras23fd0772005-10-31 13:37:12 +1100106 ptep = pte_alloc_kernel(pmdp, ea);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000107 if (!ptep)
108 return -ENOMEM;
109 set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT,
110 __pgprot(flags)));
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000111 } else {
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +0000112#ifdef CONFIG_PPC_MMU_NOHASH
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +0000113 pgdp = pgd_offset_k(ea);
114#ifdef PUD_TABLE_SIZE
115 if (pgd_none(*pgdp)) {
116 pudp = early_alloc_pgtable(PUD_TABLE_SIZE);
117 BUG_ON(pudp == NULL);
118 pgd_populate(&init_mm, pgdp, pudp);
119 }
120#endif /* PUD_TABLE_SIZE */
121 pudp = pud_offset(pgdp, ea);
122 if (pud_none(*pudp)) {
123 pmdp = early_alloc_pgtable(PMD_TABLE_SIZE);
124 BUG_ON(pmdp == NULL);
125 pud_populate(&init_mm, pudp, pmdp);
126 }
127 pmdp = pmd_offset(pudp, ea);
128 if (!pmd_present(*pmdp)) {
129 ptep = early_alloc_pgtable(PAGE_SIZE);
130 BUG_ON(ptep == NULL);
131 pmd_populate_kernel(&init_mm, pmdp, ptep);
132 }
133 ptep = pte_offset_kernel(pmdp, ea);
134 set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT,
135 __pgprot(flags)));
136#else /* CONFIG_PPC_MMU_NOHASH */
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000137 /*
138 * If the mm subsystem is not fully up, we cannot create a
139 * linux page table entry for this mapping. Simply bolt an
140 * entry in the hardware page table.
Benjamin Herrenschmidt3c726f82005-11-07 11:06:55 +1100141 *
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000142 */
Paul Mackerras1189be62007-10-11 20:37:10 +1000143 if (htab_bolt_mapping(ea, ea + PAGE_SIZE, pa, flags,
144 mmu_io_psize, mmu_kernel_ssize)) {
Benjamin Herrenschmidt77ac1662005-11-10 11:12:11 +1100145 printk(KERN_ERR "Failed to do bolted mapping IO "
146 "memory at %016lx !\n", pa);
147 return -ENOMEM;
148 }
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +0000149#endif /* !CONFIG_PPC_MMU_NOHASH */
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000150 }
Scott Wood47ce8af2013-10-11 19:22:37 -0500151
Scott Wood47ce8af2013-10-11 19:22:37 -0500152 smp_wmb();
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000153 return 0;
154}
155
156
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000157/**
158 * __ioremap_at - Low level function to establish the page tables
159 * for an IO mapping
160 */
161void __iomem * __ioremap_at(phys_addr_t pa, void *ea, unsigned long size,
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000162 unsigned long flags)
163{
164 unsigned long i;
165
Benjamin Herrenschmidta1f242f2008-07-23 21:27:08 -0700166 /* Make sure we have the base flags */
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000167 if ((flags & _PAGE_PRESENT) == 0)
168 flags |= pgprot_val(PAGE_KERNEL);
169
Benjamin Herrenschmidta1f242f2008-07-23 21:27:08 -0700170 /* Non-cacheable page cannot be coherent */
171 if (flags & _PAGE_NO_CACHE)
172 flags &= ~_PAGE_COHERENT;
173
174 /* We don't support the 4K PFN hack with ioremap */
175 if (flags & _PAGE_4K_PFN)
176 return NULL;
177
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000178 WARN_ON(pa & ~PAGE_MASK);
179 WARN_ON(((unsigned long)ea) & ~PAGE_MASK);
180 WARN_ON(size & ~PAGE_MASK);
181
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000182 for (i = 0; i < size; i += PAGE_SIZE)
Benjamin Herrenschmidta2450672009-07-23 23:15:16 +0000183 if (map_kernel_page((unsigned long)ea+i, pa+i, flags))
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000184 return NULL;
185
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000186 return (void __iomem *)ea;
187}
188
189/**
190 * __iounmap_from - Low level function to tear down the page tables
191 * for an IO mapping. This is used for mappings that
192 * are manipulated manually, like partial unmapping of
193 * PCI IOs or ISA space.
194 */
195void __iounmap_at(void *ea, unsigned long size)
196{
197 WARN_ON(((unsigned long)ea) & ~PAGE_MASK);
198 WARN_ON(size & ~PAGE_MASK);
199
200 unmap_kernel_range((unsigned long)ea, size);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000201}
202
Benjamin Herrenschmidt1cdab552009-02-22 16:19:14 +0000203void __iomem * __ioremap_caller(phys_addr_t addr, unsigned long size,
204 unsigned long flags, void *caller)
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000205{
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000206 phys_addr_t paligned;
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000207 void __iomem *ret;
208
209 /*
210 * Choose an address to map it to.
211 * Once the imalloc system is running, we use it.
212 * Before that, we map using addresses going
213 * up from ioremap_bot. imalloc will use
214 * the addresses from ioremap_bot through
215 * IMALLOC_END
216 *
217 */
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000218 paligned = addr & PAGE_MASK;
219 size = PAGE_ALIGN(addr + size) - paligned;
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000220
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000221 if ((size == 0) || (paligned == 0))
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000222 return NULL;
223
Michael Ellermanf691fa12015-03-30 14:10:37 +1100224 if (slab_is_available()) {
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000225 struct vm_struct *area;
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000226
Benjamin Herrenschmidt1cdab552009-02-22 16:19:14 +0000227 area = __get_vm_area_caller(size, VM_IOREMAP,
228 ioremap_bot, IOREMAP_END,
229 caller);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000230 if (area == NULL)
231 return NULL;
Michael Ellerman7a9d1252010-11-28 18:26:36 +0000232
233 area->phys_addr = paligned;
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000234 ret = __ioremap_at(paligned, area->addr, size, flags);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000235 if (!ret)
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000236 vunmap(area->addr);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000237 } else {
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000238 ret = __ioremap_at(paligned, (void *)ioremap_bot, size, flags);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000239 if (ret)
240 ioremap_bot += size;
241 }
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000242
243 if (ret)
244 ret += addr & ~PAGE_MASK;
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000245 return ret;
246}
247
Benjamin Herrenschmidt1cdab552009-02-22 16:19:14 +0000248void __iomem * __ioremap(phys_addr_t addr, unsigned long size,
249 unsigned long flags)
250{
251 return __ioremap_caller(addr, size, flags, __builtin_return_address(0));
252}
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100253
Benjamin Herrenschmidt68a64352006-11-13 09:27:39 +1100254void __iomem * ioremap(phys_addr_t addr, unsigned long size)
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100255{
256 unsigned long flags = _PAGE_NO_CACHE | _PAGE_GUARDED;
Benjamin Herrenschmidt1cdab552009-02-22 16:19:14 +0000257 void *caller = __builtin_return_address(0);
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100258
259 if (ppc_md.ioremap)
Benjamin Herrenschmidt1cdab552009-02-22 16:19:14 +0000260 return ppc_md.ioremap(addr, size, flags, caller);
261 return __ioremap_caller(addr, size, flags, caller);
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100262}
263
Anton Blanchardbe135f42011-05-08 21:41:59 +0000264void __iomem * ioremap_wc(phys_addr_t addr, unsigned long size)
265{
266 unsigned long flags = _PAGE_NO_CACHE;
267 void *caller = __builtin_return_address(0);
268
269 if (ppc_md.ioremap)
270 return ppc_md.ioremap(addr, size, flags, caller);
271 return __ioremap_caller(addr, size, flags, caller);
272}
273
Anton Blanchard40f1ce72011-05-08 21:43:47 +0000274void __iomem * ioremap_prot(phys_addr_t addr, unsigned long size,
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100275 unsigned long flags)
276{
Benjamin Herrenschmidt1cdab552009-02-22 16:19:14 +0000277 void *caller = __builtin_return_address(0);
278
Benjamin Herrenschmidta1f242f2008-07-23 21:27:08 -0700279 /* writeable implies dirty for kernel addresses */
280 if (flags & _PAGE_RW)
281 flags |= _PAGE_DIRTY;
282
283 /* we don't want to let _PAGE_USER and _PAGE_EXEC leak out */
284 flags &= ~(_PAGE_USER | _PAGE_EXEC);
285
Benjamin Herrenschmidt55052ee2010-04-07 14:39:36 +1000286#ifdef _PAGE_BAP_SR
287 /* _PAGE_USER contains _PAGE_BAP_SR on BookE using the new PTE format
288 * which means that we just cleared supervisor access... oops ;-) This
289 * restores it
290 */
291 flags |= _PAGE_BAP_SR;
292#endif
293
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100294 if (ppc_md.ioremap)
Benjamin Herrenschmidt1cdab552009-02-22 16:19:14 +0000295 return ppc_md.ioremap(addr, size, flags, caller);
296 return __ioremap_caller(addr, size, flags, caller);
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100297}
298
299
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000300/*
301 * Unmap an IO region and remove it from imalloc'd list.
302 * Access to IO memory should be serialized by driver.
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000303 */
Benjamin Herrenschmidt68a64352006-11-13 09:27:39 +1100304void __iounmap(volatile void __iomem *token)
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000305{
306 void *addr;
307
Michael Ellermanf691fa12015-03-30 14:10:37 +1100308 if (!slab_is_available())
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000309 return;
310
Benjamin Herrenschmidt3d5134e2007-06-04 15:15:36 +1000311 addr = (void *) ((unsigned long __force)
312 PCI_FIX_ADDR(token) & PAGE_MASK);
313 if ((unsigned long)addr < ioremap_bot) {
314 printk(KERN_WARNING "Attempt to iounmap early bolted mapping"
315 " at 0x%p\n", addr);
316 return;
317 }
318 vunmap(addr);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000319}
320
Benjamin Herrenschmidt68a64352006-11-13 09:27:39 +1100321void iounmap(volatile void __iomem *token)
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100322{
323 if (ppc_md.iounmap)
324 ppc_md.iounmap(token);
325 else
326 __iounmap(token);
327}
328
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000329EXPORT_SYMBOL(ioremap);
Anton Blanchardbe135f42011-05-08 21:41:59 +0000330EXPORT_SYMBOL(ioremap_wc);
Anton Blanchard40f1ce72011-05-08 21:43:47 +0000331EXPORT_SYMBOL(ioremap_prot);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000332EXPORT_SYMBOL(__ioremap);
Olof Johanssona302cb92007-08-31 13:58:51 +1000333EXPORT_SYMBOL(__ioremap_at);
Paul Mackerras14cf11a2005-09-26 16:04:21 +1000334EXPORT_SYMBOL(iounmap);
Benjamin Herrenschmidt4cb3cee2006-11-11 17:25:10 +1100335EXPORT_SYMBOL(__iounmap);
Olof Johanssona302cb92007-08-31 13:58:51 +1000336EXPORT_SYMBOL(__iounmap_at);
Aneesh Kumar K.V5c1f6ee2013-04-28 09:37:33 +0000337
Aneesh Kumar K.V06743522014-11-05 21:57:39 +0530338#ifndef __PAGETABLE_PUD_FOLDED
339/* 4 level page table */
340struct page *pgd_page(pgd_t pgd)
341{
342 if (pgd_huge(pgd))
343 return pte_page(pgd_pte(pgd));
344 return virt_to_page(pgd_page_vaddr(pgd));
345}
346#endif
347
348struct page *pud_page(pud_t pud)
349{
350 if (pud_huge(pud))
351 return pte_page(pud_pte(pud));
352 return virt_to_page(pud_page_vaddr(pud));
353}
354
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530355/*
356 * For hugepage we have pfn in the pmd, we use PTE_RPN_SHIFT bits for flags
357 * For PTE page, we have a PTE_FRAG_SIZE (4K) aligned virtual address.
358 */
359struct page *pmd_page(pmd_t pmd)
360{
Aneesh Kumar K.V06743522014-11-05 21:57:39 +0530361 if (pmd_trans_huge(pmd) || pmd_huge(pmd))
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530362 return pfn_to_page(pmd_pfn(pmd));
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530363 return virt_to_page(pmd_page_vaddr(pmd));
364}
365
Aneesh Kumar K.V5c1f6ee2013-04-28 09:37:33 +0000366#ifdef CONFIG_PPC_64K_PAGES
367static pte_t *get_from_cache(struct mm_struct *mm)
368{
369 void *pte_frag, *ret;
370
371 spin_lock(&mm->page_table_lock);
372 ret = mm->context.pte_frag;
373 if (ret) {
374 pte_frag = ret + PTE_FRAG_SIZE;
375 /*
376 * If we have taken up all the fragments mark PTE page NULL
377 */
378 if (((unsigned long)pte_frag & ~PAGE_MASK) == 0)
379 pte_frag = NULL;
380 mm->context.pte_frag = pte_frag;
381 }
382 spin_unlock(&mm->page_table_lock);
383 return (pte_t *)ret;
384}
385
386static pte_t *__alloc_for_cache(struct mm_struct *mm, int kernel)
387{
388 void *ret = NULL;
389 struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
390 __GFP_REPEAT | __GFP_ZERO);
391 if (!page)
392 return NULL;
Kirill A. Shutemov4f8049432013-11-14 14:31:38 -0800393 if (!kernel && !pgtable_page_ctor(page)) {
394 __free_page(page);
395 return NULL;
396 }
Aneesh Kumar K.V5c1f6ee2013-04-28 09:37:33 +0000397
398 ret = page_address(page);
399 spin_lock(&mm->page_table_lock);
400 /*
401 * If we find pgtable_page set, we return
402 * the allocated page with single fragement
403 * count.
404 */
405 if (likely(!mm->context.pte_frag)) {
406 atomic_set(&page->_count, PTE_FRAG_NR);
407 mm->context.pte_frag = ret + PTE_FRAG_SIZE;
408 }
409 spin_unlock(&mm->page_table_lock);
410
Aneesh Kumar K.V5c1f6ee2013-04-28 09:37:33 +0000411 return (pte_t *)ret;
412}
413
414pte_t *page_table_alloc(struct mm_struct *mm, unsigned long vmaddr, int kernel)
415{
416 pte_t *pte;
417
418 pte = get_from_cache(mm);
419 if (pte)
420 return pte;
421
422 return __alloc_for_cache(mm, kernel);
423}
424
425void page_table_free(struct mm_struct *mm, unsigned long *table, int kernel)
426{
427 struct page *page = virt_to_page(table);
428 if (put_page_testzero(page)) {
429 if (!kernel)
430 pgtable_page_dtor(page);
431 free_hot_cold_page(page, 0);
432 }
433}
434
435#ifdef CONFIG_SMP
436static void page_table_free_rcu(void *table)
437{
438 struct page *page = virt_to_page(table);
439 if (put_page_testzero(page)) {
440 pgtable_page_dtor(page);
441 free_hot_cold_page(page, 0);
442 }
443}
444
445void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int shift)
446{
447 unsigned long pgf = (unsigned long)table;
448
449 BUG_ON(shift > MAX_PGTABLE_INDEX_SIZE);
450 pgf |= shift;
451 tlb_remove_table(tlb, (void *)pgf);
452}
453
454void __tlb_remove_table(void *_table)
455{
456 void *table = (void *)((unsigned long)_table & ~MAX_PGTABLE_INDEX_SIZE);
457 unsigned shift = (unsigned long)_table & MAX_PGTABLE_INDEX_SIZE;
458
459 if (!shift)
460 /* PTE page needs special handling */
461 page_table_free_rcu(table);
462 else {
463 BUG_ON(shift > MAX_PGTABLE_INDEX_SIZE);
464 kmem_cache_free(PGT_CACHE(shift), table);
465 }
466}
467#else
468void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int shift)
469{
470 if (!shift) {
471 /* PTE page needs special handling */
472 struct page *page = virt_to_page(table);
473 if (put_page_testzero(page)) {
474 pgtable_page_dtor(page);
475 free_hot_cold_page(page, 0);
476 }
477 } else {
478 BUG_ON(shift > MAX_PGTABLE_INDEX_SIZE);
479 kmem_cache_free(PGT_CACHE(shift), table);
480 }
481}
482#endif
483#endif /* CONFIG_PPC_64K_PAGES */
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530484
485#ifdef CONFIG_TRANSPARENT_HUGEPAGE
486
487/*
488 * This is called when relaxing access to a hugepage. It's also called in the page
489 * fault path when we don't hit any of the major fault cases, ie, a minor
490 * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
491 * handled those two for us, we additionally deal with missing execute
492 * permission here on some processors
493 */
494int pmdp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
495 pmd_t *pmdp, pmd_t entry, int dirty)
496{
497 int changed;
498#ifdef CONFIG_DEBUG_VM
499 WARN_ON(!pmd_trans_huge(*pmdp));
500 assert_spin_locked(&vma->vm_mm->page_table_lock);
501#endif
502 changed = !pmd_same(*(pmdp), entry);
503 if (changed) {
504 __ptep_set_access_flags(pmdp_ptep(pmdp), pmd_pte(entry));
505 /*
506 * Since we are not supporting SW TLB systems, we don't
507 * have any thing similar to flush_tlb_page_nohash()
508 */
509 }
510 return changed;
511}
512
513unsigned long pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
Aneesh Kumar K.V88247e82014-02-12 09:13:36 +0530514 pmd_t *pmdp, unsigned long clr,
515 unsigned long set)
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530516{
517
518 unsigned long old, tmp;
519
520#ifdef CONFIG_DEBUG_VM
521 WARN_ON(!pmd_trans_huge(*pmdp));
522 assert_spin_locked(&mm->page_table_lock);
523#endif
524
525#ifdef PTE_ATOMIC_UPDATES
526 __asm__ __volatile__(
527 "1: ldarx %0,0,%3\n\
528 andi. %1,%0,%6\n\
529 bne- 1b \n\
530 andc %1,%0,%4 \n\
Aneesh Kumar K.V88247e82014-02-12 09:13:36 +0530531 or %1,%1,%7\n\
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530532 stdcx. %1,0,%3 \n\
533 bne- 1b"
534 : "=&r" (old), "=&r" (tmp), "=m" (*pmdp)
Aneesh Kumar K.V88247e82014-02-12 09:13:36 +0530535 : "r" (pmdp), "r" (clr), "m" (*pmdp), "i" (_PAGE_BUSY), "r" (set)
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530536 : "cc" );
537#else
538 old = pmd_val(*pmdp);
Aneesh Kumar K.V88247e82014-02-12 09:13:36 +0530539 *pmdp = __pmd((old & ~clr) | set);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530540#endif
Aneesh Kumar K.V9e813302014-08-13 12:32:04 +0530541 trace_hugepage_update(addr, old, clr, set);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530542 if (old & _PAGE_HASHPTE)
Aneesh Kumar K.Vfc047952014-08-13 12:32:00 +0530543 hpte_do_hugepage_flush(mm, addr, pmdp, old);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530544 return old;
545}
546
Aneesh Kumar K.V15a25b22015-06-24 16:57:39 -0700547pmd_t pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
548 pmd_t *pmdp)
549{
550 pmd_t pmd;
551
552 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
553 VM_BUG_ON(pmd_trans_huge(*pmdp));
554
555 pmd = *pmdp;
556 pmd_clear(pmdp);
557 /*
558 * Wait for all pending hash_page to finish. This is needed
559 * in case of subpage collapse. When we collapse normal pages
560 * to hugepage, we first clear the pmd, then invalidate all
561 * the PTE entries. The assumption here is that any low level
562 * page fault will see a none pmd and take the slow path that
563 * will wait on mmap_sem. But we could very well be in a
564 * hash_page with local ptep pointer value. Such a hash page
565 * can result in adding new HPTE entries for normal subpages.
566 * That means we could be modifying the page content as we
567 * copy them to a huge page. So wait for parallel hash_page
568 * to finish before invalidating HPTE entries. We can do this
569 * by sending an IPI to all the cpus and executing a dummy
570 * function there.
571 */
572 kick_all_cpus_sync();
573 /*
574 * Now invalidate the hpte entries in the range
575 * covered by pmd. This make sure we take a
576 * fault and will find the pmd as none, which will
577 * result in a major fault which takes mmap_sem and
578 * hence wait for collapse to complete. Without this
579 * the __collapse_huge_page_copy can result in copying
580 * the old content.
581 */
582 flush_tlb_pmd_range(vma->vm_mm, &pmd, address);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530583 return pmd;
584}
585
586int pmdp_test_and_clear_young(struct vm_area_struct *vma,
587 unsigned long address, pmd_t *pmdp)
588{
589 return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp);
590}
591
592/*
593 * We currently remove entries from the hashtable regardless of whether
594 * the entry was young or dirty. The generic routines only flush if the
595 * entry was young or dirty which is not good enough.
596 *
597 * We should be more intelligent about this but for the moment we override
598 * these functions and force a tlb flush unconditionally
599 */
600int pmdp_clear_flush_young(struct vm_area_struct *vma,
601 unsigned long address, pmd_t *pmdp)
602{
603 return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp);
604}
605
606/*
607 * We mark the pmd splitting and invalidate all the hpte
608 * entries for this hugepage.
609 */
610void pmdp_splitting_flush(struct vm_area_struct *vma,
611 unsigned long address, pmd_t *pmdp)
612{
613 unsigned long old, tmp;
614
615 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
616
617#ifdef CONFIG_DEBUG_VM
618 WARN_ON(!pmd_trans_huge(*pmdp));
619 assert_spin_locked(&vma->vm_mm->page_table_lock);
620#endif
621
622#ifdef PTE_ATOMIC_UPDATES
623
624 __asm__ __volatile__(
625 "1: ldarx %0,0,%3\n\
626 andi. %1,%0,%6\n\
627 bne- 1b \n\
628 ori %1,%0,%4 \n\
629 stdcx. %1,0,%3 \n\
630 bne- 1b"
631 : "=&r" (old), "=&r" (tmp), "=m" (*pmdp)
632 : "r" (pmdp), "i" (_PAGE_SPLITTING), "m" (*pmdp), "i" (_PAGE_BUSY)
633 : "cc" );
634#else
635 old = pmd_val(*pmdp);
636 *pmdp = __pmd(old | _PAGE_SPLITTING);
637#endif
638 /*
639 * If we didn't had the splitting flag set, go and flush the
640 * HPTE entries.
641 */
Aneesh Kumar K.V9e813302014-08-13 12:32:04 +0530642 trace_hugepage_splitting(address, old);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530643 if (!(old & _PAGE_SPLITTING)) {
644 /* We need to flush the hpte */
645 if (old & _PAGE_HASHPTE)
Aneesh Kumar K.Vfc047952014-08-13 12:32:00 +0530646 hpte_do_hugepage_flush(vma->vm_mm, address, pmdp, old);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530647 }
Aneesh Kumar K.V346519a2014-03-15 16:17:58 +0530648 /*
649 * This ensures that generic code that rely on IRQ disabling
650 * to prevent a parallel THP split work as expected.
651 */
652 kick_all_cpus_sync();
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530653}
654
655/*
656 * We want to put the pgtable in pmd and use pgtable for tracking
657 * the base page size hptes
658 */
659void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
660 pgtable_t pgtable)
661{
662 pgtable_t *pgtable_slot;
663 assert_spin_locked(&mm->page_table_lock);
664 /*
665 * we store the pgtable in the second half of PMD
666 */
667 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
668 *pgtable_slot = pgtable;
669 /*
670 * expose the deposited pgtable to other cpus.
671 * before we set the hugepage PTE at pmd level
672 * hash fault code looks at the deposted pgtable
673 * to store hash index values.
674 */
675 smp_wmb();
676}
677
678pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
679{
680 pgtable_t pgtable;
681 pgtable_t *pgtable_slot;
682
683 assert_spin_locked(&mm->page_table_lock);
684 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
685 pgtable = *pgtable_slot;
686 /*
687 * Once we withdraw, mark the entry NULL.
688 */
689 *pgtable_slot = NULL;
690 /*
691 * We store HPTE information in the deposited PTE fragment.
692 * zero out the content on withdraw.
693 */
694 memset(pgtable, 0, PTE_FRAG_SIZE);
695 return pgtable;
696}
697
698/*
699 * set a new huge pmd. We should not be called for updating
700 * an existing pmd entry. That should go via pmd_hugepage_update.
701 */
702void set_pmd_at(struct mm_struct *mm, unsigned long addr,
703 pmd_t *pmdp, pmd_t pmd)
704{
705#ifdef CONFIG_DEBUG_VM
Mel Gorman8a0516e2015-02-12 14:58:22 -0800706 WARN_ON((pmd_val(*pmdp) & (_PAGE_PRESENT | _PAGE_USER)) ==
707 (_PAGE_PRESENT | _PAGE_USER));
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530708 assert_spin_locked(&mm->page_table_lock);
709 WARN_ON(!pmd_trans_huge(pmd));
710#endif
Michael Ellerman4f9c53c2015-03-25 20:11:57 +1100711 trace_hugepage_set_pmd(addr, pmd_val(pmd));
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530712 return set_pte_at(mm, addr, pmdp_ptep(pmdp), pmd_pte(pmd));
713}
714
715void pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
716 pmd_t *pmdp)
717{
Aneesh Kumar K.V88247e82014-02-12 09:13:36 +0530718 pmd_hugepage_update(vma->vm_mm, address, pmdp, _PAGE_PRESENT, 0);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530719}
720
721/*
722 * A linux hugepage PMD was changed and the corresponding hash table entries
723 * neesd to be flushed.
724 */
725void hpte_do_hugepage_flush(struct mm_struct *mm, unsigned long addr,
Aneesh Kumar K.Vfc047952014-08-13 12:32:00 +0530726 pmd_t *pmdp, unsigned long old_pmd)
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530727{
Aneesh Kumar K.Vaefa5682014-12-04 11:00:14 +0530728 int ssize;
Aneesh Kumar K.Vf1581bf2014-11-02 21:15:27 +0530729 unsigned int psize;
730 unsigned long vsid;
Aneesh Kumar K.Vaefa5682014-12-04 11:00:14 +0530731 unsigned long flags = 0;
Aneesh Kumar K.Vd557b092014-11-02 21:15:28 +0530732 const struct cpumask *tmp;
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530733
Aneesh Kumar K.Vfa1f8ae2014-08-13 12:31:58 +0530734 /* get the base page size,vsid and segment size */
Aneesh Kumar K.Vfc047952014-08-13 12:32:00 +0530735#ifdef CONFIG_DEBUG_VM
Aneesh Kumar K.Vf1581bf2014-11-02 21:15:27 +0530736 psize = get_slice_psize(mm, addr);
Aneesh Kumar K.Vfc047952014-08-13 12:32:00 +0530737 BUG_ON(psize == MMU_PAGE_16M);
738#endif
739 if (old_pmd & _PAGE_COMBO)
740 psize = MMU_PAGE_4K;
741 else
742 psize = MMU_PAGE_64K;
743
Aneesh Kumar K.Vf1581bf2014-11-02 21:15:27 +0530744 if (!is_kernel_addr(addr)) {
745 ssize = user_segment_size(addr);
746 vsid = get_vsid(mm->context.id, addr, ssize);
Aneesh Kumar K.Vfa1f8ae2014-08-13 12:31:58 +0530747 WARN_ON(vsid == 0);
748 } else {
Aneesh Kumar K.Vf1581bf2014-11-02 21:15:27 +0530749 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
Aneesh Kumar K.Vfa1f8ae2014-08-13 12:31:58 +0530750 ssize = mmu_kernel_ssize;
751 }
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530752
Aneesh Kumar K.Vd557b092014-11-02 21:15:28 +0530753 tmp = cpumask_of(smp_processor_id());
754 if (cpumask_equal(mm_cpumask(mm), tmp))
Aneesh Kumar K.Vaefa5682014-12-04 11:00:14 +0530755 flags |= HPTE_LOCAL_UPDATE;
Aneesh Kumar K.Vd557b092014-11-02 21:15:28 +0530756
Aneesh Kumar K.Vaefa5682014-12-04 11:00:14 +0530757 return flush_hash_hugepage(vsid, addr, pmdp, psize, ssize, flags);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530758}
759
760static pmd_t pmd_set_protbits(pmd_t pmd, pgprot_t pgprot)
761{
Aneesh Kumar K.Vf281b5d2015-12-01 09:06:35 +0530762 return __pmd(pmd_val(pmd) | pgprot_val(pgprot));
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530763}
764
765pmd_t pfn_pmd(unsigned long pfn, pgprot_t pgprot)
766{
Aneesh Kumar K.Vf281b5d2015-12-01 09:06:35 +0530767 unsigned long pmdv;
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530768 /*
Kirill A. Shutemov780fc562015-02-16 16:00:18 -0800769 * For a valid pte, we would have _PAGE_PRESENT always
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530770 * set. We use this to check THP page at pmd level.
771 * leaf pte for huge page, bottom two bits != 00
772 */
Aneesh Kumar K.Vf281b5d2015-12-01 09:06:35 +0530773 pmdv = pfn << PTE_RPN_SHIFT;
774 pmdv |= _PAGE_THP_HUGE;
775 return pmd_set_protbits(__pmd(pmdv), pgprot);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530776}
777
778pmd_t mk_pmd(struct page *page, pgprot_t pgprot)
779{
780 return pfn_pmd(page_to_pfn(page), pgprot);
781}
782
783pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot)
784{
Aneesh Kumar K.Vf281b5d2015-12-01 09:06:35 +0530785 unsigned long pmdv;
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530786
Aneesh Kumar K.Vf281b5d2015-12-01 09:06:35 +0530787 pmdv = pmd_val(pmd);
788 pmdv &= _HPAGE_CHG_MASK;
789 return pmd_set_protbits(__pmd(pmdv), newprot);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530790}
791
792/*
793 * This is called at the end of handling a user page fault, when the
794 * fault has been handled by updating a HUGE PMD entry in the linux page tables.
795 * We use it to preload an HPTE into the hash table corresponding to
796 * the updated linux HUGE PMD entry.
797 */
798void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
799 pmd_t *pmd)
800{
801 return;
802}
803
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700804pmd_t pmdp_huge_get_and_clear(struct mm_struct *mm,
805 unsigned long addr, pmd_t *pmdp)
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530806{
807 pmd_t old_pmd;
808 pgtable_t pgtable;
809 unsigned long old;
810 pgtable_t *pgtable_slot;
811
Aneesh Kumar K.V88247e82014-02-12 09:13:36 +0530812 old = pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530813 old_pmd = __pmd(old);
814 /*
815 * We have pmd == none and we are holding page_table_lock.
816 * So we can safely go and clear the pgtable hash
817 * index info.
818 */
819 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
820 pgtable = *pgtable_slot;
821 /*
822 * Let's zero out old valid and hash index details
823 * hash fault look at them.
824 */
825 memset(pgtable, 0, PTE_FRAG_SIZE);
Aneesh Kumar K.V13bd8172015-05-11 11:56:01 +0530826 /*
827 * Serialize against find_linux_pte_or_hugepte which does lock-less
828 * lookup in page tables with local interrupts disabled. For huge pages
829 * it casts pmd_t to pte_t. Since format of pte_t is different from
830 * pmd_t we want to prevent transit from pmd pointing to page table
831 * to pmd pointing to huge page (and back) while interrupts are disabled.
832 * We clear pmd to possibly replace it with page table pointer in
833 * different code paths. So make sure we wait for the parallel
834 * find_linux_pte_or_hugepage to finish.
835 */
836 kick_all_cpus_sync();
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530837 return old_pmd;
838}
Aneesh Kumar K.V437d4962013-06-20 14:30:26 +0530839
840int has_transparent_hugepage(void)
841{
842 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
843 return 0;
844 /*
845 * We support THP only if PMD_SIZE is 16MB.
846 */
847 if (mmu_psize_defs[MMU_PAGE_16M].shift != PMD_SHIFT)
848 return 0;
849 /*
850 * We need to make sure that we support 16MB hugepage in a segement
851 * with base page size 64K or 4K. We only enable THP with a PAGE_SIZE
852 * of 64K.
853 */
854 /*
855 * If we have 64K HPTE, we will be using that by default
856 */
857 if (mmu_psize_defs[MMU_PAGE_64K].shift &&
858 (mmu_psize_defs[MMU_PAGE_64K].penc[MMU_PAGE_16M] == -1))
859 return 0;
860 /*
861 * Ok we only have 4K HPTE
862 */
863 if (mmu_psize_defs[MMU_PAGE_4K].penc[MMU_PAGE_16M] == -1)
864 return 0;
865
866 return 1;
867}
Aneesh Kumar K.V074c2ea2013-06-20 14:30:15 +0530868#endif /* CONFIG_TRANSPARENT_HUGEPAGE */