blob: 2b8e5ed28831a46aa49bedfae5dc5bf8e3228537 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Becky Bruce41151e72011-06-28 09:54:48 +00002 * PPC Huge TLB Page Support for Kernel.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
Becky Bruce41151e72011-06-28 09:54:48 +00005 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
9 */
10
Linus Torvalds1da177e2005-04-16 15:20:36 -070011#include <linux/mm.h>
David Gibson883a3e52009-10-26 19:24:31 +000012#include <linux/io.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090013#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/hugetlb.h>
Paul Mackerras342d3db2011-12-12 12:38:05 +000015#include <linux/export.h>
Becky Bruce41151e72011-06-28 09:54:48 +000016#include <linux/of_fdt.h>
17#include <linux/memblock.h>
18#include <linux/bootmem.h>
Kumar Gala13020be2011-11-24 09:40:07 +000019#include <linux/moduleparam.h>
David Gibson883a3e52009-10-26 19:24:31 +000020#include <asm/pgtable.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <asm/pgalloc.h>
22#include <asm/tlb.h>
Becky Bruce41151e72011-06-28 09:54:48 +000023#include <asm/setup.h>
Aneesh Kumar K.V29409992013-06-20 14:30:16 +053024#include <asm/hugetlb.h>
25
26#ifdef CONFIG_HUGETLB_PAGE
Linus Torvalds1da177e2005-04-16 15:20:36 -070027
Jon Tollefson91224342008-07-23 21:27:55 -070028#define PAGE_SHIFT_64K 16
29#define PAGE_SHIFT_16M 24
30#define PAGE_SHIFT_16G 34
Jon Tollefson4ec161c2008-01-04 09:59:50 +110031
Becky Bruce41151e72011-06-28 09:54:48 +000032unsigned int HPAGE_SHIFT;
33
34/*
35 * Tracks gpages after the device tree is scanned and before the
Becky Brucea6146882011-10-10 10:50:43 +000036 * huge_boot_pages list is ready. On non-Freescale implementations, this is
37 * just used to track 16G pages and so is a single array. FSL-based
38 * implementations may have more than one gpage size, so we need multiple
39 * arrays
Becky Bruce41151e72011-06-28 09:54:48 +000040 */
Becky Bruce881fde12011-10-10 10:50:40 +000041#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +000042#define MAX_NUMBER_GPAGES 128
43struct psize_gpages {
44 u64 gpage_list[MAX_NUMBER_GPAGES];
45 unsigned int nr_gpages;
46};
47static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
Becky Bruce881fde12011-10-10 10:50:40 +000048#else
49#define MAX_NUMBER_GPAGES 1024
50static u64 gpage_freearray[MAX_NUMBER_GPAGES];
51static unsigned nr_gpages;
Becky Bruce41151e72011-06-28 09:54:48 +000052#endif
David Gibsonf10a04c2006-04-28 15:02:51 +100053
David Gibsona4fe3ce2009-10-26 19:24:31 +000054#define hugepd_none(hpd) ((hpd).pd == 0)
55
Aneesh Kumar K.Ve2b3d202013-04-28 09:37:30 +000056#ifdef CONFIG_PPC_BOOK3S_64
57/*
58 * At this point we do the placement change only for BOOK3S 64. This would
59 * possibly work on other subarchs.
60 */
61
62/*
63 * We have PGD_INDEX_SIZ = 12 and PTE_INDEX_SIZE = 8, so that we can have
64 * 16GB hugepage pte in PGD and 16MB hugepage pte at PMD;
Aneesh Kumar K.V06743522014-11-05 21:57:39 +053065 *
66 * Defined in such a way that we can optimize away code block at build time
67 * if CONFIG_HUGETLB_PAGE=n.
Aneesh Kumar K.Ve2b3d202013-04-28 09:37:30 +000068 */
69int pmd_huge(pmd_t pmd)
70{
71 /*
72 * leaf pte for huge page, bottom two bits != 00
73 */
74 return ((pmd_val(pmd) & 0x3) != 0x0);
75}
76
77int pud_huge(pud_t pud)
78{
79 /*
80 * leaf pte for huge page, bottom two bits != 00
81 */
82 return ((pud_val(pud) & 0x3) != 0x0);
83}
84
85int pgd_huge(pgd_t pgd)
86{
87 /*
88 * leaf pte for huge page, bottom two bits != 00
89 */
90 return ((pgd_val(pgd) & 0x3) != 0x0);
91}
92#else
93int pmd_huge(pmd_t pmd)
94{
95 return 0;
96}
97
98int pud_huge(pud_t pud)
99{
100 return 0;
101}
102
103int pgd_huge(pgd_t pgd)
104{
105 return 0;
106}
107#endif
108
David Gibsona4fe3ce2009-10-26 19:24:31 +0000109pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
110{
Aneesh Kumar K.V12bc9f62013-06-20 14:30:18 +0530111 /* Only called for hugetlbfs pages, hence can ignore THP */
David Gibsona4fe3ce2009-10-26 19:24:31 +0000112 return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
113}
114
115static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
116 unsigned long address, unsigned pdshift, unsigned pshift)
117{
Becky Bruce41151e72011-06-28 09:54:48 +0000118 struct kmem_cache *cachep;
119 pte_t *new;
120
Becky Bruce881fde12011-10-10 10:50:40 +0000121#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000122 int i;
123 int num_hugepd = 1 << (pshift - pdshift);
124 cachep = hugepte_cache;
Becky Bruce881fde12011-10-10 10:50:40 +0000125#else
126 cachep = PGT_CACHE(pdshift - pshift);
Becky Bruce41151e72011-06-28 09:54:48 +0000127#endif
128
129 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
David Gibsonf10a04c2006-04-28 15:02:51 +1000130
David Gibsona4fe3ce2009-10-26 19:24:31 +0000131 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
132 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
133
David Gibsonf10a04c2006-04-28 15:02:51 +1000134 if (! new)
135 return -ENOMEM;
136
137 spin_lock(&mm->page_table_lock);
Becky Bruce881fde12011-10-10 10:50:40 +0000138#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000139 /*
140 * We have multiple higher-level entries that point to the same
141 * actual pte location. Fill in each as we go and backtrack on error.
142 * We need all of these so the DTLB pgtable walk code can find the
143 * right higher-level entry without knowing if it's a hugepage or not.
144 */
145 for (i = 0; i < num_hugepd; i++, hpdp++) {
146 if (unlikely(!hugepd_none(*hpdp)))
147 break;
148 else
Aneesh Kumar K.Vcf9427b2013-04-28 09:37:29 +0000149 /* We use the old format for PPC_FSL_BOOK3E */
Becky Bruce41151e72011-06-28 09:54:48 +0000150 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
151 }
152 /* If we bailed from the for loop early, an error occurred, clean up */
153 if (i < num_hugepd) {
154 for (i = i - 1 ; i >= 0; i--, hpdp--)
155 hpdp->pd = 0;
156 kmem_cache_free(cachep, new);
157 }
Becky Brucea1cd5412011-10-10 10:50:39 +0000158#else
159 if (!hugepd_none(*hpdp))
160 kmem_cache_free(cachep, new);
Aneesh Kumar K.Vcf9427b2013-04-28 09:37:29 +0000161 else {
162#ifdef CONFIG_PPC_BOOK3S_64
163 hpdp->pd = (unsigned long)new |
164 (shift_to_mmu_psize(pshift) << 2);
165#else
Becky Brucea1cd5412011-10-10 10:50:39 +0000166 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
Becky Bruce41151e72011-06-28 09:54:48 +0000167#endif
Aneesh Kumar K.Vcf9427b2013-04-28 09:37:29 +0000168 }
169#endif
David Gibsonf10a04c2006-04-28 15:02:51 +1000170 spin_unlock(&mm->page_table_lock);
171 return 0;
172}
173
Becky Brucea1cd5412011-10-10 10:50:39 +0000174/*
175 * These macros define how to determine which level of the page table holds
176 * the hpdp.
177 */
178#ifdef CONFIG_PPC_FSL_BOOK3E
179#define HUGEPD_PGD_SHIFT PGDIR_SHIFT
180#define HUGEPD_PUD_SHIFT PUD_SHIFT
181#else
182#define HUGEPD_PGD_SHIFT PUD_SHIFT
183#define HUGEPD_PUD_SHIFT PMD_SHIFT
184#endif
185
Aneesh Kumar K.Ve2b3d202013-04-28 09:37:30 +0000186#ifdef CONFIG_PPC_BOOK3S_64
187/*
188 * At this point we do the placement change only for BOOK3S 64. This would
189 * possibly work on other subarchs.
190 */
191pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
192{
193 pgd_t *pg;
194 pud_t *pu;
195 pmd_t *pm;
196 hugepd_t *hpdp = NULL;
197 unsigned pshift = __ffs(sz);
198 unsigned pdshift = PGDIR_SHIFT;
199
200 addr &= ~(sz-1);
201 pg = pgd_offset(mm, addr);
202
203 if (pshift == PGDIR_SHIFT)
204 /* 16GB huge page */
205 return (pte_t *) pg;
206 else if (pshift > PUD_SHIFT)
207 /*
208 * We need to use hugepd table
209 */
210 hpdp = (hugepd_t *)pg;
211 else {
212 pdshift = PUD_SHIFT;
213 pu = pud_alloc(mm, pg, addr);
214 if (pshift == PUD_SHIFT)
215 return (pte_t *)pu;
216 else if (pshift > PMD_SHIFT)
217 hpdp = (hugepd_t *)pu;
218 else {
219 pdshift = PMD_SHIFT;
220 pm = pmd_alloc(mm, pu, addr);
221 if (pshift == PMD_SHIFT)
222 /* 16MB hugepage */
223 return (pte_t *)pm;
224 else
225 hpdp = (hugepd_t *)pm;
226 }
227 }
228 if (!hpdp)
229 return NULL;
230
231 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
232
233 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
234 return NULL;
235
236 return hugepte_offset(hpdp, addr, pdshift);
237}
238
239#else
240
David Gibsona4fe3ce2009-10-26 19:24:31 +0000241pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
242{
243 pgd_t *pg;
244 pud_t *pu;
245 pmd_t *pm;
246 hugepd_t *hpdp = NULL;
247 unsigned pshift = __ffs(sz);
248 unsigned pdshift = PGDIR_SHIFT;
David Gibson0b264252008-09-05 11:49:54 +1000249
David Gibsona4fe3ce2009-10-26 19:24:31 +0000250 addr &= ~(sz-1);
251
252 pg = pgd_offset(mm, addr);
Becky Brucea1cd5412011-10-10 10:50:39 +0000253
254 if (pshift >= HUGEPD_PGD_SHIFT) {
David Gibsona4fe3ce2009-10-26 19:24:31 +0000255 hpdp = (hugepd_t *)pg;
256 } else {
257 pdshift = PUD_SHIFT;
258 pu = pud_alloc(mm, pg, addr);
Becky Brucea1cd5412011-10-10 10:50:39 +0000259 if (pshift >= HUGEPD_PUD_SHIFT) {
David Gibsona4fe3ce2009-10-26 19:24:31 +0000260 hpdp = (hugepd_t *)pu;
261 } else {
262 pdshift = PMD_SHIFT;
263 pm = pmd_alloc(mm, pu, addr);
264 hpdp = (hugepd_t *)pm;
265 }
266 }
267
268 if (!hpdp)
269 return NULL;
270
271 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
272
273 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
274 return NULL;
275
276 return hugepte_offset(hpdp, addr, pdshift);
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100277}
Aneesh Kumar K.Ve2b3d202013-04-28 09:37:30 +0000278#endif
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100279
Becky Bruce881fde12011-10-10 10:50:40 +0000280#ifdef CONFIG_PPC_FSL_BOOK3E
Jon Tollefson658013e2008-07-23 21:27:54 -0700281/* Build list of addresses of gigantic pages. This function is used in early
Anton Blanchard14ed7402014-09-17 22:15:34 +1000282 * boot before the buddy allocator is setup.
Jon Tollefson658013e2008-07-23 21:27:54 -0700283 */
Becky Bruce41151e72011-06-28 09:54:48 +0000284void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
285{
286 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
287 int i;
288
289 if (addr == 0)
290 return;
291
292 gpage_freearray[idx].nr_gpages = number_of_pages;
293
294 for (i = 0; i < number_of_pages; i++) {
295 gpage_freearray[idx].gpage_list[i] = addr;
296 addr += page_size;
297 }
298}
299
300/*
301 * Moves the gigantic page addresses from the temporary list to the
302 * huge_boot_pages list.
303 */
304int alloc_bootmem_huge_page(struct hstate *hstate)
305{
306 struct huge_bootmem_page *m;
Wanpeng Li2415cf12013-07-03 15:02:43 -0700307 int idx = shift_to_mmu_psize(huge_page_shift(hstate));
Becky Bruce41151e72011-06-28 09:54:48 +0000308 int nr_gpages = gpage_freearray[idx].nr_gpages;
309
310 if (nr_gpages == 0)
311 return 0;
312
313#ifdef CONFIG_HIGHMEM
314 /*
315 * If gpages can be in highmem we can't use the trick of storing the
316 * data structure in the page; allocate space for this
317 */
318 m = alloc_bootmem(sizeof(struct huge_bootmem_page));
319 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
320#else
321 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
322#endif
323
324 list_add(&m->list, &huge_boot_pages);
325 gpage_freearray[idx].nr_gpages = nr_gpages;
326 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
327 m->hstate = hstate;
328
329 return 1;
330}
331/*
332 * Scan the command line hugepagesz= options for gigantic pages; store those in
333 * a list that we use to allocate the memory once all options are parsed.
334 */
335
336unsigned long gpage_npages[MMU_PAGE_COUNT];
337
Paul Gortmaker89528122012-05-07 10:32:22 -0400338static int __init do_gpage_early_setup(char *param, char *val,
339 const char *unused)
Becky Bruce41151e72011-06-28 09:54:48 +0000340{
341 static phys_addr_t size;
342 unsigned long npages;
343
344 /*
345 * The hugepagesz and hugepages cmdline options are interleaved. We
346 * use the size variable to keep track of whether or not this was done
347 * properly and skip over instances where it is incorrect. Other
348 * command-line parsing code will issue warnings, so we don't need to.
349 *
350 */
351 if ((strcmp(param, "default_hugepagesz") == 0) ||
352 (strcmp(param, "hugepagesz") == 0)) {
353 size = memparse(val, NULL);
354 } else if (strcmp(param, "hugepages") == 0) {
355 if (size != 0) {
356 if (sscanf(val, "%lu", &npages) <= 0)
357 npages = 0;
358 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
359 size = 0;
360 }
361 }
362 return 0;
363}
364
365
366/*
367 * This function allocates physical space for pages that are larger than the
368 * buddy allocator can handle. We want to allocate these in highmem because
369 * the amount of lowmem is limited. This means that this function MUST be
370 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
371 * allocate to grab highmem.
372 */
373void __init reserve_hugetlb_gpages(void)
374{
375 static __initdata char cmdline[COMMAND_LINE_SIZE];
376 phys_addr_t size, base;
377 int i;
378
379 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
Pawel Moll026cee02012-03-26 12:50:51 +1030380 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
381 &do_gpage_early_setup);
Becky Bruce41151e72011-06-28 09:54:48 +0000382
383 /*
384 * Walk gpage list in reverse, allocating larger page sizes first.
385 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
386 * When we reach the point in the list where pages are no longer
387 * considered gpages, we're done.
388 */
389 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
390 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
391 continue;
392 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
393 break;
394
395 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
396 base = memblock_alloc_base(size * gpage_npages[i], size,
397 MEMBLOCK_ALLOC_ANYWHERE);
398 add_gpage(base, size, gpage_npages[i]);
399 }
400}
401
Becky Bruce881fde12011-10-10 10:50:40 +0000402#else /* !PPC_FSL_BOOK3E */
Becky Bruce41151e72011-06-28 09:54:48 +0000403
404/* Build list of addresses of gigantic pages. This function is used in early
Anton Blanchard14ed7402014-09-17 22:15:34 +1000405 * boot before the buddy allocator is setup.
Becky Bruce41151e72011-06-28 09:54:48 +0000406 */
407void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
Jon Tollefson658013e2008-07-23 21:27:54 -0700408{
409 if (!addr)
410 return;
411 while (number_of_pages > 0) {
412 gpage_freearray[nr_gpages] = addr;
413 nr_gpages++;
414 number_of_pages--;
415 addr += page_size;
416 }
417}
418
Jon Tollefsonec4b2c02008-07-23 21:27:53 -0700419/* Moves the gigantic page addresses from the temporary list to the
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700420 * huge_boot_pages list.
421 */
422int alloc_bootmem_huge_page(struct hstate *hstate)
Jon Tollefsonec4b2c02008-07-23 21:27:53 -0700423{
424 struct huge_bootmem_page *m;
425 if (nr_gpages == 0)
426 return 0;
427 m = phys_to_virt(gpage_freearray[--nr_gpages]);
428 gpage_freearray[nr_gpages] = 0;
429 list_add(&m->list, &huge_boot_pages);
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700430 m->hstate = hstate;
Jon Tollefsonec4b2c02008-07-23 21:27:53 -0700431 return 1;
432}
Becky Bruce41151e72011-06-28 09:54:48 +0000433#endif
Jon Tollefsonec4b2c02008-07-23 21:27:53 -0700434
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800435int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
436{
437 return 0;
438}
439
Becky Bruce881fde12011-10-10 10:50:40 +0000440#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000441#define HUGEPD_FREELIST_SIZE \
442 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
443
444struct hugepd_freelist {
445 struct rcu_head rcu;
446 unsigned int index;
447 void *ptes[0];
448};
449
450static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
451
452static void hugepd_free_rcu_callback(struct rcu_head *head)
453{
454 struct hugepd_freelist *batch =
455 container_of(head, struct hugepd_freelist, rcu);
456 unsigned int i;
457
458 for (i = 0; i < batch->index; i++)
459 kmem_cache_free(hugepte_cache, batch->ptes[i]);
460
461 free_page((unsigned long)batch);
462}
463
464static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
465{
466 struct hugepd_freelist **batchp;
467
Christoph Lameter69111ba2014-10-21 15:23:25 -0500468 batchp = this_cpu_ptr(&hugepd_freelist_cur);
Becky Bruce41151e72011-06-28 09:54:48 +0000469
470 if (atomic_read(&tlb->mm->mm_users) < 2 ||
471 cpumask_equal(mm_cpumask(tlb->mm),
472 cpumask_of(smp_processor_id()))) {
473 kmem_cache_free(hugepte_cache, hugepte);
Tiejun Chen94b09d72014-01-20 16:39:34 +0800474 put_cpu_var(hugepd_freelist_cur);
Becky Bruce41151e72011-06-28 09:54:48 +0000475 return;
476 }
477
478 if (*batchp == NULL) {
479 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
480 (*batchp)->index = 0;
481 }
482
483 (*batchp)->ptes[(*batchp)->index++] = hugepte;
484 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
485 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
486 *batchp = NULL;
487 }
Tiejun Chen94b09d72014-01-20 16:39:34 +0800488 put_cpu_var(hugepd_freelist_cur);
Becky Bruce41151e72011-06-28 09:54:48 +0000489}
490#endif
491
David Gibsona4fe3ce2009-10-26 19:24:31 +0000492static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
493 unsigned long start, unsigned long end,
494 unsigned long floor, unsigned long ceiling)
David Gibsonf10a04c2006-04-28 15:02:51 +1000495{
496 pte_t *hugepte = hugepd_page(*hpdp);
Becky Bruce41151e72011-06-28 09:54:48 +0000497 int i;
498
David Gibsona4fe3ce2009-10-26 19:24:31 +0000499 unsigned long pdmask = ~((1UL << pdshift) - 1);
Becky Bruce41151e72011-06-28 09:54:48 +0000500 unsigned int num_hugepd = 1;
501
Becky Bruce881fde12011-10-10 10:50:40 +0000502#ifdef CONFIG_PPC_FSL_BOOK3E
503 /* Note: On fsl the hpdp may be the first of several */
Becky Bruce41151e72011-06-28 09:54:48 +0000504 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
Becky Bruce881fde12011-10-10 10:50:40 +0000505#else
506 unsigned int shift = hugepd_shift(*hpdp);
Becky Bruce41151e72011-06-28 09:54:48 +0000507#endif
David Gibsona4fe3ce2009-10-26 19:24:31 +0000508
509 start &= pdmask;
510 if (start < floor)
511 return;
512 if (ceiling) {
513 ceiling &= pdmask;
514 if (! ceiling)
515 return;
516 }
517 if (end - 1 > ceiling - 1)
518 return;
David Gibsonf10a04c2006-04-28 15:02:51 +1000519
Becky Bruce41151e72011-06-28 09:54:48 +0000520 for (i = 0; i < num_hugepd; i++, hpdp++)
521 hpdp->pd = 0;
522
David Gibsonf10a04c2006-04-28 15:02:51 +1000523 tlb->need_flush = 1;
Becky Bruce881fde12011-10-10 10:50:40 +0000524
525#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000526 hugepd_free(tlb, hugepte);
Becky Bruce881fde12011-10-10 10:50:40 +0000527#else
528 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
Becky Bruce41151e72011-06-28 09:54:48 +0000529#endif
David Gibsonf10a04c2006-04-28 15:02:51 +1000530}
531
David Gibsonf10a04c2006-04-28 15:02:51 +1000532static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
533 unsigned long addr, unsigned long end,
David Gibsona4fe3ce2009-10-26 19:24:31 +0000534 unsigned long floor, unsigned long ceiling)
David Gibsonf10a04c2006-04-28 15:02:51 +1000535{
536 pmd_t *pmd;
537 unsigned long next;
538 unsigned long start;
539
540 start = addr;
David Gibsonf10a04c2006-04-28 15:02:51 +1000541 do {
Becky Brucea1cd5412011-10-10 10:50:39 +0000542 pmd = pmd_offset(pud, addr);
David Gibsonf10a04c2006-04-28 15:02:51 +1000543 next = pmd_addr_end(addr, end);
Aneesh Kumar K.V8bbd9f02013-06-19 12:04:26 +0530544 if (!is_hugepd(pmd)) {
545 /*
546 * if it is not hugepd pointer, we should already find
547 * it cleared.
548 */
549 WARN_ON(!pmd_none_or_clear_bad(pmd));
David Gibsonf10a04c2006-04-28 15:02:51 +1000550 continue;
Aneesh Kumar K.V8bbd9f02013-06-19 12:04:26 +0530551 }
Becky Brucea1cd5412011-10-10 10:50:39 +0000552#ifdef CONFIG_PPC_FSL_BOOK3E
553 /*
554 * Increment next by the size of the huge mapping since
555 * there may be more than one entry at this level for a
556 * single hugepage, but all of them point to
557 * the same kmem cache that holds the hugepte.
558 */
559 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
560#endif
David Gibsona4fe3ce2009-10-26 19:24:31 +0000561 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
562 addr, next, floor, ceiling);
Becky Brucea1cd5412011-10-10 10:50:39 +0000563 } while (addr = next, addr != end);
David Gibsonf10a04c2006-04-28 15:02:51 +1000564
565 start &= PUD_MASK;
566 if (start < floor)
567 return;
568 if (ceiling) {
569 ceiling &= PUD_MASK;
570 if (!ceiling)
571 return;
572 }
573 if (end - 1 > ceiling - 1)
574 return;
575
576 pmd = pmd_offset(pud, start);
577 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000578 pmd_free_tlb(tlb, pmd, start);
David Gibsonf10a04c2006-04-28 15:02:51 +1000579}
David Gibsonf10a04c2006-04-28 15:02:51 +1000580
581static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
582 unsigned long addr, unsigned long end,
583 unsigned long floor, unsigned long ceiling)
584{
585 pud_t *pud;
586 unsigned long next;
587 unsigned long start;
588
589 start = addr;
David Gibsonf10a04c2006-04-28 15:02:51 +1000590 do {
Becky Brucea1cd5412011-10-10 10:50:39 +0000591 pud = pud_offset(pgd, addr);
David Gibsonf10a04c2006-04-28 15:02:51 +1000592 next = pud_addr_end(addr, end);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000593 if (!is_hugepd(pud)) {
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100594 if (pud_none_or_clear_bad(pud))
595 continue;
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700596 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
David Gibsona4fe3ce2009-10-26 19:24:31 +0000597 ceiling);
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100598 } else {
Becky Brucea1cd5412011-10-10 10:50:39 +0000599#ifdef CONFIG_PPC_FSL_BOOK3E
600 /*
601 * Increment next by the size of the huge mapping since
602 * there may be more than one entry at this level for a
603 * single hugepage, but all of them point to
604 * the same kmem cache that holds the hugepte.
605 */
606 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
607#endif
David Gibsona4fe3ce2009-10-26 19:24:31 +0000608 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
609 addr, next, floor, ceiling);
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100610 }
Becky Brucea1cd5412011-10-10 10:50:39 +0000611 } while (addr = next, addr != end);
David Gibsonf10a04c2006-04-28 15:02:51 +1000612
613 start &= PGDIR_MASK;
614 if (start < floor)
615 return;
616 if (ceiling) {
617 ceiling &= PGDIR_MASK;
618 if (!ceiling)
619 return;
620 }
621 if (end - 1 > ceiling - 1)
622 return;
623
624 pud = pud_offset(pgd, start);
625 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000626 pud_free_tlb(tlb, pud, start);
David Gibsonf10a04c2006-04-28 15:02:51 +1000627}
628
629/*
630 * This function frees user-level page tables of a process.
David Gibsonf10a04c2006-04-28 15:02:51 +1000631 */
Jan Beulich42b77722008-07-23 21:27:10 -0700632void hugetlb_free_pgd_range(struct mmu_gather *tlb,
David Gibsonf10a04c2006-04-28 15:02:51 +1000633 unsigned long addr, unsigned long end,
634 unsigned long floor, unsigned long ceiling)
635{
636 pgd_t *pgd;
637 unsigned long next;
David Gibsonf10a04c2006-04-28 15:02:51 +1000638
639 /*
David Gibsona4fe3ce2009-10-26 19:24:31 +0000640 * Because there are a number of different possible pagetable
641 * layouts for hugepage ranges, we limit knowledge of how
642 * things should be laid out to the allocation path
643 * (huge_pte_alloc(), above). Everything else works out the
644 * structure as it goes from information in the hugepd
645 * pointers. That means that we can't here use the
646 * optimization used in the normal page free_pgd_range(), of
647 * checking whether we're actually covering a large enough
648 * range to have to do anything at the top level of the walk
649 * instead of at the bottom.
David Gibsonf10a04c2006-04-28 15:02:51 +1000650 *
David Gibsona4fe3ce2009-10-26 19:24:31 +0000651 * To make sense of this, you should probably go read the big
652 * block comment at the top of the normal free_pgd_range(),
653 * too.
David Gibsonf10a04c2006-04-28 15:02:51 +1000654 */
655
David Gibsonf10a04c2006-04-28 15:02:51 +1000656 do {
David Gibsonf10a04c2006-04-28 15:02:51 +1000657 next = pgd_addr_end(addr, end);
Becky Bruce41151e72011-06-28 09:54:48 +0000658 pgd = pgd_offset(tlb->mm, addr);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000659 if (!is_hugepd(pgd)) {
David Gibson0b264252008-09-05 11:49:54 +1000660 if (pgd_none_or_clear_bad(pgd))
661 continue;
662 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
663 } else {
Becky Bruce881fde12011-10-10 10:50:40 +0000664#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000665 /*
666 * Increment next by the size of the huge mapping since
Becky Bruce881fde12011-10-10 10:50:40 +0000667 * there may be more than one entry at the pgd level
668 * for a single hugepage, but all of them point to the
669 * same kmem cache that holds the hugepte.
Becky Bruce41151e72011-06-28 09:54:48 +0000670 */
671 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
672#endif
David Gibsona4fe3ce2009-10-26 19:24:31 +0000673 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
674 addr, next, floor, ceiling);
David Gibson0b264252008-09-05 11:49:54 +1000675 }
Becky Bruce41151e72011-06-28 09:54:48 +0000676 } while (addr = next, addr != end);
David Gibsone28f7fa2005-08-05 19:39:06 +1000677}
678
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679struct page *
680follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
681{
682 pte_t *ptep;
683 struct page *page;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000684 unsigned shift;
685 unsigned long mask;
Aneesh Kumar K.V12bc9f62013-06-20 14:30:18 +0530686 /*
687 * Transparent hugepages are handled by generic code. We can skip them
688 * here.
689 */
David Gibsona4fe3ce2009-10-26 19:24:31 +0000690 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700692 /* Verify it is a huge page else bail. */
Aneesh Kumar K.V12bc9f62013-06-20 14:30:18 +0530693 if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 return ERR_PTR(-EINVAL);
695
David Gibsona4fe3ce2009-10-26 19:24:31 +0000696 mask = (1UL << shift) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 page = pte_page(*ptep);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000698 if (page)
699 page += (address & mask) / PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700
701 return page;
702}
703
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704struct page *
705follow_huge_pmd(struct mm_struct *mm, unsigned long address,
706 pmd_t *pmd, int write)
707{
708 BUG();
709 return NULL;
710}
711
David Gibson39adfa52009-11-23 20:03:40 +0000712static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
713 unsigned long sz)
714{
715 unsigned long __boundary = (addr + sz) & ~(sz-1);
716 return (__boundary - 1 < end - 1) ? __boundary : end;
717}
718
David Gibsona4fe3ce2009-10-26 19:24:31 +0000719int gup_hugepd(hugepd_t *hugepd, unsigned pdshift,
720 unsigned long addr, unsigned long end,
721 int write, struct page **pages, int *nr)
722{
723 pte_t *ptep;
724 unsigned long sz = 1UL << hugepd_shift(*hugepd);
David Gibson39adfa52009-11-23 20:03:40 +0000725 unsigned long next;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000726
727 ptep = hugepte_offset(hugepd, addr, pdshift);
728 do {
David Gibson39adfa52009-11-23 20:03:40 +0000729 next = hugepte_addr_end(addr, end, sz);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000730 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
731 return 0;
David Gibson39adfa52009-11-23 20:03:40 +0000732 } while (ptep++, addr = next, addr != end);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000733
734 return 1;
735}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736
Becky Bruce76512952011-10-10 10:50:36 +0000737#ifdef CONFIG_PPC_MM_SLICES
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
739 unsigned long len, unsigned long pgoff,
740 unsigned long flags)
741{
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700742 struct hstate *hstate = hstate_file(file);
743 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
Brian King48f797d2008-12-04 04:07:54 +0000744
Michel Lespinasse34d07172013-04-29 11:53:52 -0700745 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746}
Becky Bruce76512952011-10-10 10:50:36 +0000747#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748
Mel Gorman33402892009-01-06 14:38:54 -0800749unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
750{
Paul Mackerras25c29f92011-09-20 19:58:10 +0000751#ifdef CONFIG_PPC_MM_SLICES
Mel Gorman33402892009-01-06 14:38:54 -0800752 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
753
754 return 1UL << mmu_psize_to_shift(psize);
Becky Bruce41151e72011-06-28 09:54:48 +0000755#else
756 if (!is_vm_hugetlb_page(vma))
757 return PAGE_SIZE;
758
759 return huge_page_size(hstate_vma(vma));
760#endif
761}
762
763static inline bool is_power_of_4(unsigned long x)
764{
765 if (is_power_of_2(x))
766 return (__ilog2(x) % 2) ? false : true;
767 return false;
Mel Gorman33402892009-01-06 14:38:54 -0800768}
769
David Gibsond1837cb2009-10-26 19:24:31 +0000770static int __init add_huge_page_size(unsigned long long size)
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100771{
David Gibsond1837cb2009-10-26 19:24:31 +0000772 int shift = __ffs(size);
773 int mmu_psize;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000774
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100775 /* Check that it is a page size supported by the hardware and
David Gibsond1837cb2009-10-26 19:24:31 +0000776 * that it fits within pagetable and slice limits. */
Becky Bruce41151e72011-06-28 09:54:48 +0000777#ifdef CONFIG_PPC_FSL_BOOK3E
778 if ((size < PAGE_SIZE) || !is_power_of_4(size))
779 return -EINVAL;
780#else
David Gibsond1837cb2009-10-26 19:24:31 +0000781 if (!is_power_of_2(size)
782 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
783 return -EINVAL;
Becky Bruce41151e72011-06-28 09:54:48 +0000784#endif
Jon Tollefson91224342008-07-23 21:27:55 -0700785
David Gibsond1837cb2009-10-26 19:24:31 +0000786 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
787 return -EINVAL;
788
789#ifdef CONFIG_SPU_FS_64K_LS
790 /* Disable support for 64K huge pages when 64K SPU local store
791 * support is enabled as the current implementation conflicts.
792 */
793 if (shift == PAGE_SHIFT_64K)
794 return -EINVAL;
795#endif /* CONFIG_SPU_FS_64K_LS */
796
797 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
798
799 /* Return if huge page size has already been setup */
800 if (size_to_hstate(size))
801 return 0;
802
803 hugetlb_add_hstate(shift - PAGE_SHIFT);
804
805 return 0;
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100806}
807
808static int __init hugepage_setup_sz(char *str)
809{
810 unsigned long long size;
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100811
812 size = memparse(str, &str);
813
David Gibsond1837cb2009-10-26 19:24:31 +0000814 if (add_huge_page_size(size) != 0)
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100815 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
816
817 return 1;
818}
819__setup("hugepagesz=", hugepage_setup_sz);
820
Becky Bruce881fde12011-10-10 10:50:40 +0000821#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000822struct kmem_cache *hugepte_cache;
823static int __init hugetlbpage_init(void)
824{
825 int psize;
826
827 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
828 unsigned shift;
829
830 if (!mmu_psize_defs[psize].shift)
831 continue;
832
833 shift = mmu_psize_to_shift(psize);
834
835 /* Don't treat normal page sizes as huge... */
836 if (shift != PAGE_SHIFT)
837 if (add_huge_page_size(1ULL << shift) < 0)
838 continue;
839 }
840
841 /*
842 * Create a kmem cache for hugeptes. The bottom bits in the pte have
843 * size information encoded in them, so align them to allow this
844 */
845 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
846 HUGEPD_SHIFT_MASK + 1, 0, NULL);
847 if (hugepte_cache == NULL)
848 panic("%s: Unable to create kmem cache for hugeptes\n",
849 __func__);
850
851 /* Default hpage size = 4M */
852 if (mmu_psize_defs[MMU_PAGE_4M].shift)
853 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
854 else
855 panic("%s: Unable to set default huge page size\n", __func__);
856
857
858 return 0;
859}
860#else
David Gibsonf10a04c2006-04-28 15:02:51 +1000861static int __init hugetlbpage_init(void)
862{
David Gibsona4fe3ce2009-10-26 19:24:31 +0000863 int psize;
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700864
Matt Evans44ae3ab2011-04-06 19:48:50 +0000865 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
David Gibsonf10a04c2006-04-28 15:02:51 +1000866 return -ENODEV;
Benjamin Herrenschmidt00df4382008-07-28 16:13:18 +1000867
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700868 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
David Gibsond1837cb2009-10-26 19:24:31 +0000869 unsigned shift;
870 unsigned pdshift;
871
872 if (!mmu_psize_defs[psize].shift)
873 continue;
874
875 shift = mmu_psize_to_shift(psize);
876
877 if (add_huge_page_size(1ULL << shift) < 0)
878 continue;
879
880 if (shift < PMD_SHIFT)
881 pdshift = PMD_SHIFT;
882 else if (shift < PUD_SHIFT)
883 pdshift = PUD_SHIFT;
884 else
885 pdshift = PGDIR_SHIFT;
Aneesh Kumar K.Ve2b3d202013-04-28 09:37:30 +0000886 /*
887 * if we have pdshift and shift value same, we don't
888 * use pgt cache for hugepd.
889 */
890 if (pdshift != shift) {
891 pgtable_cache_add(pdshift - shift, NULL);
892 if (!PGT_CACHE(pdshift - shift))
893 panic("hugetlbpage_init(): could not create "
894 "pgtable cache for %d bit pagesize\n", shift);
895 }
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700896 }
David Gibsonf10a04c2006-04-28 15:02:51 +1000897
David Gibsond1837cb2009-10-26 19:24:31 +0000898 /* Set default large page size. Currently, we pick 16M or 1M
899 * depending on what is available
900 */
901 if (mmu_psize_defs[MMU_PAGE_16M].shift)
902 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
903 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
904 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
905
David Gibsonf10a04c2006-04-28 15:02:51 +1000906 return 0;
907}
Becky Bruce41151e72011-06-28 09:54:48 +0000908#endif
David Gibsonf10a04c2006-04-28 15:02:51 +1000909module_init(hugetlbpage_init);
David Gibson0895ecd2009-10-26 19:24:31 +0000910
911void flush_dcache_icache_hugepage(struct page *page)
912{
913 int i;
Becky Bruce41151e72011-06-28 09:54:48 +0000914 void *start;
David Gibson0895ecd2009-10-26 19:24:31 +0000915
916 BUG_ON(!PageCompound(page));
917
Becky Bruce41151e72011-06-28 09:54:48 +0000918 for (i = 0; i < (1UL << compound_order(page)); i++) {
919 if (!PageHighMem(page)) {
920 __flush_dcache_icache(page_address(page+i));
921 } else {
Cong Wang2480b202011-11-25 23:14:16 +0800922 start = kmap_atomic(page+i);
Becky Bruce41151e72011-06-28 09:54:48 +0000923 __flush_dcache_icache(start);
Cong Wang2480b202011-11-25 23:14:16 +0800924 kunmap_atomic(start);
Becky Bruce41151e72011-06-28 09:54:48 +0000925 }
926 }
David Gibson0895ecd2009-10-26 19:24:31 +0000927}
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530928
929#endif /* CONFIG_HUGETLB_PAGE */
930
931/*
932 * We have 4 cases for pgds and pmds:
933 * (1) invalid (all zeroes)
934 * (2) pointer to next table, as normal; bottom 6 bits == 0
935 * (3) leaf pte for huge page, bottom two bits != 00
936 * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530937 *
938 * So long as we atomically load page table pointers we are safe against teardown,
939 * we can follow the address down to the the page and take a ref on it.
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530940 */
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530941
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530942pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
943{
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530944 pgd_t pgd, *pgdp;
945 pud_t pud, *pudp;
946 pmd_t pmd, *pmdp;
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530947 pte_t *ret_pte;
948 hugepd_t *hpdp = NULL;
949 unsigned pdshift = PGDIR_SHIFT;
950
951 if (shift)
952 *shift = 0;
953
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530954 pgdp = pgdir + pgd_index(ea);
955 pgd = ACCESS_ONCE(*pgdp);
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +0530956 /*
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530957 * Always operate on the local stack value. This make sure the
958 * value don't get updated by a parallel THP split/collapse,
959 * page fault or a page unmap. The return pte_t * is still not
960 * stable. So should be checked there for above conditions.
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +0530961 */
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530962 if (pgd_none(pgd))
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +0530963 return NULL;
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530964 else if (pgd_huge(pgd)) {
965 ret_pte = (pte_t *) pgdp;
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530966 goto out;
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530967 } else if (is_hugepd(&pgd))
968 hpdp = (hugepd_t *)&pgd;
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +0530969 else {
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530970 /*
971 * Even if we end up with an unmap, the pgtable will not
972 * be freed, because we do an rcu free and here we are
973 * irq disabled
974 */
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530975 pdshift = PUD_SHIFT;
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530976 pudp = pud_offset(&pgd, ea);
977 pud = ACCESS_ONCE(*pudp);
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530978
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530979 if (pud_none(pud))
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +0530980 return NULL;
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530981 else if (pud_huge(pud)) {
982 ret_pte = (pte_t *) pudp;
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530983 goto out;
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530984 } else if (is_hugepd(&pud))
985 hpdp = (hugepd_t *)&pud;
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +0530986 else {
Aneesh Kumar K.V29409992013-06-20 14:30:16 +0530987 pdshift = PMD_SHIFT;
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530988 pmdp = pmd_offset(&pud, ea);
989 pmd = ACCESS_ONCE(*pmdp);
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +0530990 /*
991 * A hugepage collapse is captured by pmd_none, because
992 * it mark the pmd none and do a hpte invalidate.
993 *
994 * A hugepage split is captured by pmd_trans_splitting
995 * because we mark the pmd trans splitting and do a
996 * hpte invalidate
997 *
998 */
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +0530999 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +05301000 return NULL;
Aneesh Kumar K.V29409992013-06-20 14:30:16 +05301001
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +05301002 if (pmd_huge(pmd) || pmd_large(pmd)) {
1003 ret_pte = (pte_t *) pmdp;
Aneesh Kumar K.V29409992013-06-20 14:30:16 +05301004 goto out;
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +05301005 } else if (is_hugepd(&pmd))
1006 hpdp = (hugepd_t *)&pmd;
Aneesh Kumar K.Vac52ae42013-06-20 14:30:17 +05301007 else
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +05301008 return pte_offset_kernel(&pmd, ea);
Aneesh Kumar K.V29409992013-06-20 14:30:16 +05301009 }
1010 }
1011 if (!hpdp)
1012 return NULL;
1013
1014 ret_pte = hugepte_offset(hpdp, ea, pdshift);
1015 pdshift = hugepd_shift(*hpdp);
1016out:
1017 if (shift)
1018 *shift = pdshift;
1019 return ret_pte;
1020}
1021EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte);
1022
1023int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1024 unsigned long end, int write, struct page **pages, int *nr)
1025{
1026 unsigned long mask;
1027 unsigned long pte_end;
1028 struct page *head, *page, *tail;
1029 pte_t pte;
1030 int refs;
1031
1032 pte_end = (addr + sz) & ~(sz-1);
1033 if (pte_end < end)
1034 end = pte_end;
1035
Aneesh Kumar K.V7888b4d2013-06-20 14:30:23 +05301036 pte = ACCESS_ONCE(*ptep);
Aneesh Kumar K.V29409992013-06-20 14:30:16 +05301037 mask = _PAGE_PRESENT | _PAGE_USER;
1038 if (write)
1039 mask |= _PAGE_RW;
1040
1041 if ((pte_val(pte) & mask) != mask)
1042 return 0;
1043
Aneesh Kumar K.V0ac52dd2013-06-20 14:30:22 +05301044#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1045 /*
1046 * check for splitting here
1047 */
1048 if (pmd_trans_splitting(pte_pmd(pte)))
1049 return 0;
1050#endif
1051
Aneesh Kumar K.V29409992013-06-20 14:30:16 +05301052 /* hugepages are never "special" */
1053 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1054
1055 refs = 0;
1056 head = pte_page(pte);
1057
1058 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1059 tail = page;
1060 do {
1061 VM_BUG_ON(compound_head(page) != head);
1062 pages[*nr] = page;
1063 (*nr)++;
1064 page++;
1065 refs++;
1066 } while (addr += PAGE_SIZE, addr != end);
1067
1068 if (!page_cache_add_speculative(head, refs)) {
1069 *nr -= refs;
1070 return 0;
1071 }
1072
1073 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1074 /* Could be optimized better */
1075 *nr -= refs;
1076 while (refs--)
1077 put_page(head);
1078 return 0;
1079 }
1080
1081 /*
1082 * Any tail page need their mapcount reference taken before we
1083 * return.
1084 */
1085 while (refs--) {
1086 if (PageTail(tail))
1087 get_huge_page_tail(tail);
1088 tail++;
1089 }
1090
1091 return 1;
1092}