Linux Audio

Check our new training course

Loading...
v5.4
  1/*
  2 * PPC Huge TLB Page Support for Kernel.
  3 *
  4 * Copyright (C) 2003 David Gibson, IBM Corporation.
  5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
  6 *
  7 * Based on the IA-32 version:
  8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
  9 */
 10
 11#include <linux/mm.h>
 12#include <linux/io.h>
 13#include <linux/slab.h>
 14#include <linux/hugetlb.h>
 15#include <linux/export.h>
 16#include <linux/of_fdt.h>
 17#include <linux/memblock.h>
 
 18#include <linux/moduleparam.h>
 19#include <linux/swap.h>
 20#include <linux/swapops.h>
 21#include <linux/kmemleak.h>
 22#include <asm/pgtable.h>
 23#include <asm/pgalloc.h>
 24#include <asm/tlb.h>
 25#include <asm/setup.h>
 26#include <asm/hugetlb.h>
 27#include <asm/pte-walk.h>
 28
 29bool hugetlb_disabled = false;
 
 
 30
 31#define hugepd_none(hpd)	(hpd_val(hpd) == 0)
 32
 33#define PTE_T_ORDER	(__builtin_ffs(sizeof(pte_t)) - __builtin_ffs(sizeof(void *)))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 34
 35pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 36{
 37	/*
 38	 * Only called for hugetlbfs pages, hence can ignore THP and the
 39	 * irq disabled walk.
 40	 */
 41	return __find_linux_pte(mm->pgd, addr, NULL, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 42}
 43
 44static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
 45			   unsigned long address, unsigned int pdshift,
 46			   unsigned int pshift, spinlock_t *ptl)
 47{
 48	struct kmem_cache *cachep;
 49	pte_t *new;
 50	int i;
 51	int num_hugepd;
 52
 53	if (pshift >= pdshift) {
 54		cachep = PGT_CACHE(PTE_T_ORDER);
 55		num_hugepd = 1 << (pshift - pdshift);
 56	} else if (IS_ENABLED(CONFIG_PPC_8xx)) {
 57		cachep = PGT_CACHE(PTE_INDEX_SIZE);
 58		num_hugepd = 1;
 59	} else {
 60		cachep = PGT_CACHE(pdshift - pshift);
 61		num_hugepd = 1;
 62	}
 63
 64	if (!cachep) {
 65		WARN_ONCE(1, "No page table cache created for hugetlb tables");
 66		return -ENOMEM;
 67	}
 68
 69	new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
 70
 71	BUG_ON(pshift > HUGEPD_SHIFT_MASK);
 72	BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
 73
 74	if (!new)
 75		return -ENOMEM;
 76
 77	/*
 78	 * Make sure other cpus find the hugepd set only after a
 79	 * properly initialized page table is visible to them.
 80	 * For more details look for comment in __pte_alloc().
 81	 */
 82	smp_wmb();
 83
 84	spin_lock(ptl);
 85	/*
 86	 * We have multiple higher-level entries that point to the same
 87	 * actual pte location.  Fill in each as we go and backtrack on error.
 88	 * We need all of these so the DTLB pgtable walk code can find the
 89	 * right higher-level entry without knowing if it's a hugepage or not.
 90	 */
 91	for (i = 0; i < num_hugepd; i++, hpdp++) {
 92		if (unlikely(!hugepd_none(*hpdp)))
 93			break;
 94		hugepd_populate(hpdp, new, pshift);
 
 95	}
 96	/* If we bailed from the for loop early, an error occurred, clean up */
 97	if (i < num_hugepd) {
 98		for (i = i - 1 ; i >= 0; i--, hpdp--)
 99			*hpdp = __hugepd(0);
100		kmem_cache_free(cachep, new);
101	} else {
102		kmemleak_ignore(new);
103	}
104	spin_unlock(ptl);
 
 
 
 
 
 
105	return 0;
106}
107
108/*
109 * At this point we do the placement change only for BOOK3S 64. This would
110 * possibly work on other subarchs.
111 */
 
 
 
 
 
 
 
 
112pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
113{
114	pgd_t *pg;
115	pud_t *pu;
116	pmd_t *pm;
117	hugepd_t *hpdp = NULL;
118	unsigned pshift = __ffs(sz);
119	unsigned pdshift = PGDIR_SHIFT;
120	spinlock_t *ptl;
121
122	addr &= ~(sz-1);
 
123	pg = pgd_offset(mm, addr);
124
125#ifdef CONFIG_PPC_BOOK3S_64
126	if (pshift == PGDIR_SHIFT)
127		/* 16GB huge page */
128		return (pte_t *) pg;
129	else if (pshift > PUD_SHIFT) {
130		/*
131		 * We need to use hugepd table
132		 */
133		ptl = &mm->page_table_lock;
134		hpdp = (hugepd_t *)pg;
135	} else {
136		pdshift = PUD_SHIFT;
137		pu = pud_alloc(mm, pg, addr);
138		if (!pu)
139			return NULL;
140		if (pshift == PUD_SHIFT)
141			return (pte_t *)pu;
142		else if (pshift > PMD_SHIFT) {
143			ptl = pud_lockptr(mm, pu);
144			hpdp = (hugepd_t *)pu;
145		} else {
146			pdshift = PMD_SHIFT;
147			pm = pmd_alloc(mm, pu, addr);
148			if (!pm)
149				return NULL;
150			if (pshift == PMD_SHIFT)
151				/* 16MB hugepage */
152				return (pte_t *)pm;
153			else {
154				ptl = pmd_lockptr(mm, pm);
155				hpdp = (hugepd_t *)pm;
156			}
157		}
158	}
159#else
160	if (pshift >= PGDIR_SHIFT) {
161		ptl = &mm->page_table_lock;
162		hpdp = (hugepd_t *)pg;
163	} else {
164		pdshift = PUD_SHIFT;
165		pu = pud_alloc(mm, pg, addr);
166		if (!pu)
167			return NULL;
168		if (pshift >= PUD_SHIFT) {
169			ptl = pud_lockptr(mm, pu);
170			hpdp = (hugepd_t *)pu;
171		} else {
172			pdshift = PMD_SHIFT;
173			pm = pmd_alloc(mm, pu, addr);
174			if (!pm)
175				return NULL;
176			ptl = pmd_lockptr(mm, pm);
177			hpdp = (hugepd_t *)pm;
178		}
179	}
180#endif
181	if (!hpdp)
182		return NULL;
183
184	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
185
186	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
187						  pdshift, pshift, ptl))
188		return NULL;
189
190	return hugepte_offset(*hpdp, addr, pdshift);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
191}
192
193#ifdef CONFIG_PPC_BOOK3S_64
194/*
195 * Tracks gpages after the device tree is scanned and before the
196 * huge_boot_pages list is ready on pseries.
197 */
198#define MAX_NUMBER_GPAGES	1024
199__initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
200__initdata static unsigned nr_gpages;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
201
202/*
203 * Build list of addresses of gigantic pages.  This function is used in early
204 * boot before the buddy allocator is setup.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
205 */
206void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
207{
208	if (!addr)
209		return;
210	while (number_of_pages > 0) {
211		gpage_freearray[nr_gpages] = addr;
212		nr_gpages++;
213		number_of_pages--;
214		addr += page_size;
215	}
216}
217
218int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
 
 
 
219{
220	struct huge_bootmem_page *m;
221	if (nr_gpages == 0)
222		return 0;
223	m = phys_to_virt(gpage_freearray[--nr_gpages]);
224	gpage_freearray[nr_gpages] = 0;
225	list_add(&m->list, &huge_boot_pages);
226	m->hstate = hstate;
227	return 1;
228}
229#endif
230
231
232int __init alloc_bootmem_huge_page(struct hstate *h)
233{
234
235#ifdef CONFIG_PPC_BOOK3S_64
236	if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
237		return pseries_alloc_bootmem_huge_page(h);
238#endif
239	return __alloc_bootmem_huge_page(h);
240}
241
242#ifndef CONFIG_PPC_BOOK3S_64
243#define HUGEPD_FREELIST_SIZE \
244	((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
245
246struct hugepd_freelist {
247	struct rcu_head	rcu;
248	unsigned int index;
249	void *ptes[0];
250};
251
252static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
253
254static void hugepd_free_rcu_callback(struct rcu_head *head)
255{
256	struct hugepd_freelist *batch =
257		container_of(head, struct hugepd_freelist, rcu);
258	unsigned int i;
259
260	for (i = 0; i < batch->index; i++)
261		kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
262
263	free_page((unsigned long)batch);
264}
265
266static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
267{
268	struct hugepd_freelist **batchp;
269
270	batchp = &get_cpu_var(hugepd_freelist_cur);
271
272	if (atomic_read(&tlb->mm->mm_users) < 2 ||
273	    mm_is_thread_local(tlb->mm)) {
274		kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
275		put_cpu_var(hugepd_freelist_cur);
276		return;
277	}
278
279	if (*batchp == NULL) {
280		*batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
281		(*batchp)->index = 0;
282	}
283
284	(*batchp)->ptes[(*batchp)->index++] = hugepte;
285	if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
286		call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
287		*batchp = NULL;
288	}
289	put_cpu_var(hugepd_freelist_cur);
290}
291#else
292static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
293#endif
294
295static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
296			      unsigned long start, unsigned long end,
297			      unsigned long floor, unsigned long ceiling)
298{
299	pte_t *hugepte = hugepd_page(*hpdp);
300	int i;
301
302	unsigned long pdmask = ~((1UL << pdshift) - 1);
303	unsigned int num_hugepd = 1;
304	unsigned int shift = hugepd_shift(*hpdp);
305
 
306	/* Note: On fsl the hpdp may be the first of several */
307	if (shift > pdshift)
308		num_hugepd = 1 << (shift - pdshift);
 
 
309
310	start &= pdmask;
311	if (start < floor)
312		return;
313	if (ceiling) {
314		ceiling &= pdmask;
315		if (! ceiling)
316			return;
317	}
318	if (end - 1 > ceiling - 1)
319		return;
320
321	for (i = 0; i < num_hugepd; i++, hpdp++)
322		*hpdp = __hugepd(0);
323
324	if (shift >= pdshift)
325		hugepd_free(tlb, hugepte);
326	else if (IS_ENABLED(CONFIG_PPC_8xx))
327		pgtable_free_tlb(tlb, hugepte,
328				 get_hugepd_cache_index(PTE_INDEX_SIZE));
329	else
330		pgtable_free_tlb(tlb, hugepte,
331				 get_hugepd_cache_index(pdshift - shift));
332}
333
334static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
335				   unsigned long addr, unsigned long end,
336				   unsigned long floor, unsigned long ceiling)
337{
338	pmd_t *pmd;
339	unsigned long next;
340	unsigned long start;
341
342	start = addr;
343	do {
344		unsigned long more;
345
346		pmd = pmd_offset(pud, addr);
347		next = pmd_addr_end(addr, end);
348		if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
349			/*
350			 * if it is not hugepd pointer, we should already find
351			 * it cleared.
352			 */
353			WARN_ON(!pmd_none_or_clear_bad(pmd));
354			continue;
355		}
356		/*
357		 * Increment next by the size of the huge mapping since
358		 * there may be more than one entry at this level for a
359		 * single hugepage, but all of them point to
360		 * the same kmem cache that holds the hugepte.
361		 */
362		more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
363		if (more > next)
364			next = more;
365
366		free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
367				  addr, next, floor, ceiling);
368	} while (addr = next, addr != end);
369
370	start &= PUD_MASK;
371	if (start < floor)
372		return;
373	if (ceiling) {
374		ceiling &= PUD_MASK;
375		if (!ceiling)
376			return;
377	}
378	if (end - 1 > ceiling - 1)
379		return;
380
381	pmd = pmd_offset(pud, start);
382	pud_clear(pud);
383	pmd_free_tlb(tlb, pmd, start);
384	mm_dec_nr_pmds(tlb->mm);
385}
386
387static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
388				   unsigned long addr, unsigned long end,
389				   unsigned long floor, unsigned long ceiling)
390{
391	pud_t *pud;
392	unsigned long next;
393	unsigned long start;
394
395	start = addr;
396	do {
397		pud = pud_offset(pgd, addr);
398		next = pud_addr_end(addr, end);
399		if (!is_hugepd(__hugepd(pud_val(*pud)))) {
400			if (pud_none_or_clear_bad(pud))
401				continue;
402			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
403					       ceiling);
404		} else {
405			unsigned long more;
406			/*
407			 * Increment next by the size of the huge mapping since
408			 * there may be more than one entry at this level for a
409			 * single hugepage, but all of them point to
410			 * the same kmem cache that holds the hugepte.
411			 */
412			more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
413			if (more > next)
414				next = more;
415
416			free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
417					  addr, next, floor, ceiling);
418		}
419	} while (addr = next, addr != end);
420
421	start &= PGDIR_MASK;
422	if (start < floor)
423		return;
424	if (ceiling) {
425		ceiling &= PGDIR_MASK;
426		if (!ceiling)
427			return;
428	}
429	if (end - 1 > ceiling - 1)
430		return;
431
432	pud = pud_offset(pgd, start);
433	pgd_clear(pgd);
434	pud_free_tlb(tlb, pud, start);
435	mm_dec_nr_puds(tlb->mm);
436}
437
438/*
439 * This function frees user-level page tables of a process.
 
 
440 */
441void hugetlb_free_pgd_range(struct mmu_gather *tlb,
442			    unsigned long addr, unsigned long end,
443			    unsigned long floor, unsigned long ceiling)
444{
445	pgd_t *pgd;
446	unsigned long next;
447
448	/*
449	 * Because there are a number of different possible pagetable
450	 * layouts for hugepage ranges, we limit knowledge of how
451	 * things should be laid out to the allocation path
452	 * (huge_pte_alloc(), above).  Everything else works out the
453	 * structure as it goes from information in the hugepd
454	 * pointers.  That means that we can't here use the
455	 * optimization used in the normal page free_pgd_range(), of
456	 * checking whether we're actually covering a large enough
457	 * range to have to do anything at the top level of the walk
458	 * instead of at the bottom.
459	 *
460	 * To make sense of this, you should probably go read the big
461	 * block comment at the top of the normal free_pgd_range(),
462	 * too.
463	 */
464
465	do {
466		next = pgd_addr_end(addr, end);
467		pgd = pgd_offset(tlb->mm, addr);
468		if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
469			if (pgd_none_or_clear_bad(pgd))
470				continue;
471			hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
472		} else {
473			unsigned long more;
474			/*
475			 * Increment next by the size of the huge mapping since
476			 * there may be more than one entry at the pgd level
477			 * for a single hugepage, but all of them point to the
478			 * same kmem cache that holds the hugepte.
479			 */
480			more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
481			if (more > next)
482				next = more;
483
484			free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
485					  addr, next, floor, ceiling);
486		}
487	} while (addr = next, addr != end);
488}
489
490struct page *follow_huge_pd(struct vm_area_struct *vma,
491			    unsigned long address, hugepd_t hpd,
492			    int flags, int pdshift)
493{
494	pte_t *ptep;
495	spinlock_t *ptl;
496	struct page *page = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
497	unsigned long mask;
498	int shift = hugepd_shift(hpd);
499	struct mm_struct *mm = vma->vm_mm;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
500
501retry:
502	/*
503	 * hugepage directory entries are protected by mm->page_table_lock
504	 * Use this instead of huge_pte_lockptr
505	 */
506	ptl = &mm->page_table_lock;
507	spin_lock(ptl);
508
509	ptep = hugepte_offset(hpd, address, pdshift);
510	if (pte_present(*ptep)) {
511		mask = (1UL << shift) - 1;
512		page = pte_page(*ptep);
513		page += ((address & mask) >> PAGE_SHIFT);
514		if (flags & FOLL_GET)
515			get_page(page);
516	} else {
517		if (is_hugetlb_entry_migration(*ptep)) {
518			spin_unlock(ptl);
519			__migration_entry_wait(mm, ptep, ptl);
520			goto retry;
521		}
522	}
523	spin_unlock(ptl);
524	return page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
525}
526
527#ifdef CONFIG_PPC_MM_SLICES
528unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
529					unsigned long len, unsigned long pgoff,
530					unsigned long flags)
531{
532	struct hstate *hstate = hstate_file(file);
533	int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
534
535#ifdef CONFIG_PPC_RADIX_MMU
536	if (radix_enabled())
537		return radix__hugetlb_get_unmapped_area(file, addr, len,
538						       pgoff, flags);
539#endif
540	return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
541}
542#endif
543
544unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
545{
546	/* With radix we don't use slice, so derive it from vma*/
547	if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
548		unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
549
550		return 1UL << mmu_psize_to_shift(psize);
551	}
552	return vma_kernel_pagesize(vma);
 
 
 
 
 
 
 
 
 
 
 
553}
554
555static int __init add_huge_page_size(unsigned long long size)
556{
557	int shift = __ffs(size);
558	int mmu_psize;
559
560	/* Check that it is a page size supported by the hardware and
561	 * that it fits within pagetable and slice limits. */
562	if (size <= PAGE_SIZE || !is_power_of_2(size))
 
 
 
 
 
 
 
 
 
563		return -EINVAL;
564
565	mmu_psize = check_and_get_huge_psize(shift);
566	if (mmu_psize < 0)
 
 
 
567		return -EINVAL;
 
568
569	BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
570
571	/* Return if huge page size has already been setup */
572	if (size_to_hstate(size))
573		return 0;
574
575	hugetlb_add_hstate(shift - PAGE_SHIFT);
576
577	return 0;
578}
579
580static int __init hugepage_setup_sz(char *str)
581{
582	unsigned long long size;
583
584	size = memparse(str, &str);
585
586	if (add_huge_page_size(size) != 0) {
587		hugetlb_bad_size();
588		pr_err("Invalid huge page size specified(%llu)\n", size);
589	}
590
591	return 1;
592}
593__setup("hugepagesz=", hugepage_setup_sz);
594
 
 
595static int __init hugetlbpage_init(void)
596{
597	bool configured = false;
598	int psize;
599
600	if (hugetlb_disabled) {
601		pr_info("HugeTLB support is disabled!\n");
602		return 0;
 
 
 
 
 
 
 
 
 
603	}
604
605	if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
606	    !mmu_has_feature(MMU_FTR_16M_PAGE))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
607		return -ENODEV;
608
609	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
610		unsigned shift;
611		unsigned pdshift;
612
613		if (!mmu_psize_defs[psize].shift)
614			continue;
615
616		shift = mmu_psize_to_shift(psize);
617
618#ifdef CONFIG_PPC_BOOK3S_64
619		if (shift > PGDIR_SHIFT)
620			continue;
621		else if (shift > PUD_SHIFT)
622			pdshift = PGDIR_SHIFT;
623		else if (shift > PMD_SHIFT)
624			pdshift = PUD_SHIFT;
625		else
626			pdshift = PMD_SHIFT;
627#else
628		if (shift < PUD_SHIFT)
629			pdshift = PMD_SHIFT;
630		else if (shift < PGDIR_SHIFT)
631			pdshift = PUD_SHIFT;
632		else
633			pdshift = PGDIR_SHIFT;
634#endif
635
636		if (add_huge_page_size(1ULL << shift) < 0)
637			continue;
638		/*
639		 * if we have pdshift and shift value same, we don't
640		 * use pgt cache for hugepd.
641		 */
642		if (pdshift > shift && IS_ENABLED(CONFIG_PPC_8xx))
643			pgtable_cache_add(PTE_INDEX_SIZE);
644		else if (pdshift > shift)
645			pgtable_cache_add(pdshift - shift);
646		else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) || IS_ENABLED(CONFIG_PPC_8xx))
647			pgtable_cache_add(PTE_T_ORDER);
648
649		configured = true;
650	}
651
652	if (configured) {
653		if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
654			hugetlbpage_init_default();
655	} else
656		pr_info("Failed to initialize. Disabling HugeTLB");
 
 
657
658	return 0;
659}
660
661arch_initcall(hugetlbpage_init);
662
663void flush_dcache_icache_hugepage(struct page *page)
664{
665	int i;
666	void *start;
667
668	BUG_ON(!PageCompound(page));
669
670	for (i = 0; i < compound_nr(page); i++) {
671		if (!PageHighMem(page)) {
672			__flush_dcache_icache(page_address(page+i));
673		} else {
674			start = kmap_atomic(page+i);
675			__flush_dcache_icache(start);
676			kunmap_atomic(start);
677		}
678	}
679}
v3.5.6
  1/*
  2 * PPC Huge TLB Page Support for Kernel.
  3 *
  4 * Copyright (C) 2003 David Gibson, IBM Corporation.
  5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
  6 *
  7 * Based on the IA-32 version:
  8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
  9 */
 10
 11#include <linux/mm.h>
 12#include <linux/io.h>
 13#include <linux/slab.h>
 14#include <linux/hugetlb.h>
 15#include <linux/export.h>
 16#include <linux/of_fdt.h>
 17#include <linux/memblock.h>
 18#include <linux/bootmem.h>
 19#include <linux/moduleparam.h>
 
 
 
 20#include <asm/pgtable.h>
 21#include <asm/pgalloc.h>
 22#include <asm/tlb.h>
 23#include <asm/setup.h>
 
 
 24
 25#define PAGE_SHIFT_64K	16
 26#define PAGE_SHIFT_16M	24
 27#define PAGE_SHIFT_16G	34
 28
 29unsigned int HPAGE_SHIFT;
 30
 31/*
 32 * Tracks gpages after the device tree is scanned and before the
 33 * huge_boot_pages list is ready.  On non-Freescale implementations, this is
 34 * just used to track 16G pages and so is a single array.  FSL-based
 35 * implementations may have more than one gpage size, so we need multiple
 36 * arrays
 37 */
 38#ifdef CONFIG_PPC_FSL_BOOK3E
 39#define MAX_NUMBER_GPAGES	128
 40struct psize_gpages {
 41	u64 gpage_list[MAX_NUMBER_GPAGES];
 42	unsigned int nr_gpages;
 43};
 44static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
 45#else
 46#define MAX_NUMBER_GPAGES	1024
 47static u64 gpage_freearray[MAX_NUMBER_GPAGES];
 48static unsigned nr_gpages;
 49#endif
 50
 51static inline int shift_to_mmu_psize(unsigned int shift)
 52{
 53	int psize;
 54
 55	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize)
 56		if (mmu_psize_defs[psize].shift == shift)
 57			return psize;
 58	return -1;
 59}
 60
 61static inline unsigned int mmu_psize_to_shift(unsigned int mmu_psize)
 62{
 63	if (mmu_psize_defs[mmu_psize].shift)
 64		return mmu_psize_defs[mmu_psize].shift;
 65	BUG();
 66}
 67
 68#define hugepd_none(hpd)	((hpd).pd == 0)
 69
 70pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
 71{
 72	pgd_t *pg;
 73	pud_t *pu;
 74	pmd_t *pm;
 75	hugepd_t *hpdp = NULL;
 76	unsigned pdshift = PGDIR_SHIFT;
 77
 78	if (shift)
 79		*shift = 0;
 80
 81	pg = pgdir + pgd_index(ea);
 82	if (is_hugepd(pg)) {
 83		hpdp = (hugepd_t *)pg;
 84	} else if (!pgd_none(*pg)) {
 85		pdshift = PUD_SHIFT;
 86		pu = pud_offset(pg, ea);
 87		if (is_hugepd(pu))
 88			hpdp = (hugepd_t *)pu;
 89		else if (!pud_none(*pu)) {
 90			pdshift = PMD_SHIFT;
 91			pm = pmd_offset(pu, ea);
 92			if (is_hugepd(pm))
 93				hpdp = (hugepd_t *)pm;
 94			else if (!pmd_none(*pm)) {
 95				return pte_offset_kernel(pm, ea);
 96			}
 97		}
 98	}
 99
100	if (!hpdp)
101		return NULL;
102
103	if (shift)
104		*shift = hugepd_shift(*hpdp);
105	return hugepte_offset(hpdp, ea, pdshift);
106}
107EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte);
108
109pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
110{
111	return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
112}
113
114static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
115			   unsigned long address, unsigned pdshift, unsigned pshift)
 
116{
117	struct kmem_cache *cachep;
118	pte_t *new;
 
 
119
120#ifdef CONFIG_PPC_FSL_BOOK3E
121	int i;
122	int num_hugepd = 1 << (pshift - pdshift);
123	cachep = hugepte_cache;
124#else
125	cachep = PGT_CACHE(pdshift - pshift);
126#endif
 
 
 
 
 
 
 
 
127
128	new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
129
130	BUG_ON(pshift > HUGEPD_SHIFT_MASK);
131	BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
132
133	if (! new)
134		return -ENOMEM;
135
136	spin_lock(&mm->page_table_lock);
137#ifdef CONFIG_PPC_FSL_BOOK3E
 
 
 
 
 
 
138	/*
139	 * We have multiple higher-level entries that point to the same
140	 * actual pte location.  Fill in each as we go and backtrack on error.
141	 * We need all of these so the DTLB pgtable walk code can find the
142	 * right higher-level entry without knowing if it's a hugepage or not.
143	 */
144	for (i = 0; i < num_hugepd; i++, hpdp++) {
145		if (unlikely(!hugepd_none(*hpdp)))
146			break;
147		else
148			hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
149	}
150	/* If we bailed from the for loop early, an error occurred, clean up */
151	if (i < num_hugepd) {
152		for (i = i - 1 ; i >= 0; i--, hpdp--)
153			hpdp->pd = 0;
154		kmem_cache_free(cachep, new);
 
 
155	}
156#else
157	if (!hugepd_none(*hpdp))
158		kmem_cache_free(cachep, new);
159	else
160		hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
161#endif
162	spin_unlock(&mm->page_table_lock);
163	return 0;
164}
165
166/*
167 * These macros define how to determine which level of the page table holds
168 * the hpdp.
169 */
170#ifdef CONFIG_PPC_FSL_BOOK3E
171#define HUGEPD_PGD_SHIFT PGDIR_SHIFT
172#define HUGEPD_PUD_SHIFT PUD_SHIFT
173#else
174#define HUGEPD_PGD_SHIFT PUD_SHIFT
175#define HUGEPD_PUD_SHIFT PMD_SHIFT
176#endif
177
178pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
179{
180	pgd_t *pg;
181	pud_t *pu;
182	pmd_t *pm;
183	hugepd_t *hpdp = NULL;
184	unsigned pshift = __ffs(sz);
185	unsigned pdshift = PGDIR_SHIFT;
 
186
187	addr &= ~(sz-1);
188
189	pg = pgd_offset(mm, addr);
190
191	if (pshift >= HUGEPD_PGD_SHIFT) {
 
 
 
 
 
 
 
 
192		hpdp = (hugepd_t *)pg;
193	} else {
194		pdshift = PUD_SHIFT;
195		pu = pud_alloc(mm, pg, addr);
196		if (pshift >= HUGEPD_PUD_SHIFT) {
 
 
 
 
 
197			hpdp = (hugepd_t *)pu;
198		} else {
199			pdshift = PMD_SHIFT;
200			pm = pmd_alloc(mm, pu, addr);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
201			hpdp = (hugepd_t *)pm;
202		}
203	}
204
205	if (!hpdp)
206		return NULL;
207
208	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
209
210	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
 
211		return NULL;
212
213	return hugepte_offset(hpdp, addr, pdshift);
214}
215
216#ifdef CONFIG_PPC_FSL_BOOK3E
217/* Build list of addresses of gigantic pages.  This function is used in early
218 * boot before the buddy or bootmem allocator is setup.
219 */
220void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
221{
222	unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
223	int i;
224
225	if (addr == 0)
226		return;
227
228	gpage_freearray[idx].nr_gpages = number_of_pages;
229
230	for (i = 0; i < number_of_pages; i++) {
231		gpage_freearray[idx].gpage_list[i] = addr;
232		addr += page_size;
233	}
234}
235
 
236/*
237 * Moves the gigantic page addresses from the temporary list to the
238 * huge_boot_pages list.
239 */
240int alloc_bootmem_huge_page(struct hstate *hstate)
241{
242	struct huge_bootmem_page *m;
243	int idx = shift_to_mmu_psize(hstate->order + PAGE_SHIFT);
244	int nr_gpages = gpage_freearray[idx].nr_gpages;
245
246	if (nr_gpages == 0)
247		return 0;
248
249#ifdef CONFIG_HIGHMEM
250	/*
251	 * If gpages can be in highmem we can't use the trick of storing the
252	 * data structure in the page; allocate space for this
253	 */
254	m = alloc_bootmem(sizeof(struct huge_bootmem_page));
255	m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
256#else
257	m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
258#endif
259
260	list_add(&m->list, &huge_boot_pages);
261	gpage_freearray[idx].nr_gpages = nr_gpages;
262	gpage_freearray[idx].gpage_list[nr_gpages] = 0;
263	m->hstate = hstate;
264
265	return 1;
266}
267/*
268 * Scan the command line hugepagesz= options for gigantic pages; store those in
269 * a list that we use to allocate the memory once all options are parsed.
270 */
271
272unsigned long gpage_npages[MMU_PAGE_COUNT];
273
274static int __init do_gpage_early_setup(char *param, char *val,
275				       const char *unused)
276{
277	static phys_addr_t size;
278	unsigned long npages;
279
280	/*
281	 * The hugepagesz and hugepages cmdline options are interleaved.  We
282	 * use the size variable to keep track of whether or not this was done
283	 * properly and skip over instances where it is incorrect.  Other
284	 * command-line parsing code will issue warnings, so we don't need to.
285	 *
286	 */
287	if ((strcmp(param, "default_hugepagesz") == 0) ||
288	    (strcmp(param, "hugepagesz") == 0)) {
289		size = memparse(val, NULL);
290	} else if (strcmp(param, "hugepages") == 0) {
291		if (size != 0) {
292			if (sscanf(val, "%lu", &npages) <= 0)
293				npages = 0;
294			gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
295			size = 0;
296		}
297	}
298	return 0;
299}
300
301
302/*
303 * This function allocates physical space for pages that are larger than the
304 * buddy allocator can handle.  We want to allocate these in highmem because
305 * the amount of lowmem is limited.  This means that this function MUST be
306 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
307 * allocate to grab highmem.
308 */
309void __init reserve_hugetlb_gpages(void)
310{
311	static __initdata char cmdline[COMMAND_LINE_SIZE];
312	phys_addr_t size, base;
313	int i;
314
315	strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
316	parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
317			&do_gpage_early_setup);
318
319	/*
320	 * Walk gpage list in reverse, allocating larger page sizes first.
321	 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
322	 * When we reach the point in the list where pages are no longer
323	 * considered gpages, we're done.
324	 */
325	for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
326		if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
327			continue;
328		else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
329			break;
330
331		size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
332		base = memblock_alloc_base(size * gpage_npages[i], size,
333					   MEMBLOCK_ALLOC_ANYWHERE);
334		add_gpage(base, size, gpage_npages[i]);
335	}
336}
337
338#else /* !PPC_FSL_BOOK3E */
339
340/* Build list of addresses of gigantic pages.  This function is used in early
341 * boot before the buddy or bootmem allocator is setup.
342 */
343void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
344{
345	if (!addr)
346		return;
347	while (number_of_pages > 0) {
348		gpage_freearray[nr_gpages] = addr;
349		nr_gpages++;
350		number_of_pages--;
351		addr += page_size;
352	}
353}
354
355/* Moves the gigantic page addresses from the temporary list to the
356 * huge_boot_pages list.
357 */
358int alloc_bootmem_huge_page(struct hstate *hstate)
359{
360	struct huge_bootmem_page *m;
361	if (nr_gpages == 0)
362		return 0;
363	m = phys_to_virt(gpage_freearray[--nr_gpages]);
364	gpage_freearray[nr_gpages] = 0;
365	list_add(&m->list, &huge_boot_pages);
366	m->hstate = hstate;
367	return 1;
368}
369#endif
370
371int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 
372{
373	return 0;
 
 
 
 
 
374}
375
376#ifdef CONFIG_PPC_FSL_BOOK3E
377#define HUGEPD_FREELIST_SIZE \
378	((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
379
380struct hugepd_freelist {
381	struct rcu_head	rcu;
382	unsigned int index;
383	void *ptes[0];
384};
385
386static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
387
388static void hugepd_free_rcu_callback(struct rcu_head *head)
389{
390	struct hugepd_freelist *batch =
391		container_of(head, struct hugepd_freelist, rcu);
392	unsigned int i;
393
394	for (i = 0; i < batch->index; i++)
395		kmem_cache_free(hugepte_cache, batch->ptes[i]);
396
397	free_page((unsigned long)batch);
398}
399
400static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
401{
402	struct hugepd_freelist **batchp;
403
404	batchp = &__get_cpu_var(hugepd_freelist_cur);
405
406	if (atomic_read(&tlb->mm->mm_users) < 2 ||
407	    cpumask_equal(mm_cpumask(tlb->mm),
408			  cpumask_of(smp_processor_id()))) {
409		kmem_cache_free(hugepte_cache, hugepte);
410		return;
411	}
412
413	if (*batchp == NULL) {
414		*batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
415		(*batchp)->index = 0;
416	}
417
418	(*batchp)->ptes[(*batchp)->index++] = hugepte;
419	if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
420		call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
421		*batchp = NULL;
422	}
 
423}
 
 
424#endif
425
426static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
427			      unsigned long start, unsigned long end,
428			      unsigned long floor, unsigned long ceiling)
429{
430	pte_t *hugepte = hugepd_page(*hpdp);
431	int i;
432
433	unsigned long pdmask = ~((1UL << pdshift) - 1);
434	unsigned int num_hugepd = 1;
 
435
436#ifdef CONFIG_PPC_FSL_BOOK3E
437	/* Note: On fsl the hpdp may be the first of several */
438	num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
439#else
440	unsigned int shift = hugepd_shift(*hpdp);
441#endif
442
443	start &= pdmask;
444	if (start < floor)
445		return;
446	if (ceiling) {
447		ceiling &= pdmask;
448		if (! ceiling)
449			return;
450	}
451	if (end - 1 > ceiling - 1)
452		return;
453
454	for (i = 0; i < num_hugepd; i++, hpdp++)
455		hpdp->pd = 0;
456
457	tlb->need_flush = 1;
458
459#ifdef CONFIG_PPC_FSL_BOOK3E
460	hugepd_free(tlb, hugepte);
461#else
462	pgtable_free_tlb(tlb, hugepte, pdshift - shift);
463#endif
 
464}
465
466static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
467				   unsigned long addr, unsigned long end,
468				   unsigned long floor, unsigned long ceiling)
469{
470	pmd_t *pmd;
471	unsigned long next;
472	unsigned long start;
473
474	start = addr;
475	do {
 
 
476		pmd = pmd_offset(pud, addr);
477		next = pmd_addr_end(addr, end);
478		if (pmd_none(*pmd))
 
 
 
 
 
479			continue;
480#ifdef CONFIG_PPC_FSL_BOOK3E
481		/*
482		 * Increment next by the size of the huge mapping since
483		 * there may be more than one entry at this level for a
484		 * single hugepage, but all of them point to
485		 * the same kmem cache that holds the hugepte.
486		 */
487		next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
488#endif
 
 
489		free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
490				  addr, next, floor, ceiling);
491	} while (addr = next, addr != end);
492
493	start &= PUD_MASK;
494	if (start < floor)
495		return;
496	if (ceiling) {
497		ceiling &= PUD_MASK;
498		if (!ceiling)
499			return;
500	}
501	if (end - 1 > ceiling - 1)
502		return;
503
504	pmd = pmd_offset(pud, start);
505	pud_clear(pud);
506	pmd_free_tlb(tlb, pmd, start);
 
507}
508
509static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
510				   unsigned long addr, unsigned long end,
511				   unsigned long floor, unsigned long ceiling)
512{
513	pud_t *pud;
514	unsigned long next;
515	unsigned long start;
516
517	start = addr;
518	do {
519		pud = pud_offset(pgd, addr);
520		next = pud_addr_end(addr, end);
521		if (!is_hugepd(pud)) {
522			if (pud_none_or_clear_bad(pud))
523				continue;
524			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
525					       ceiling);
526		} else {
527#ifdef CONFIG_PPC_FSL_BOOK3E
528			/*
529			 * Increment next by the size of the huge mapping since
530			 * there may be more than one entry at this level for a
531			 * single hugepage, but all of them point to
532			 * the same kmem cache that holds the hugepte.
533			 */
534			next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
535#endif
 
 
536			free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
537					  addr, next, floor, ceiling);
538		}
539	} while (addr = next, addr != end);
540
541	start &= PGDIR_MASK;
542	if (start < floor)
543		return;
544	if (ceiling) {
545		ceiling &= PGDIR_MASK;
546		if (!ceiling)
547			return;
548	}
549	if (end - 1 > ceiling - 1)
550		return;
551
552	pud = pud_offset(pgd, start);
553	pgd_clear(pgd);
554	pud_free_tlb(tlb, pud, start);
 
555}
556
557/*
558 * This function frees user-level page tables of a process.
559 *
560 * Must be called with pagetable lock held.
561 */
562void hugetlb_free_pgd_range(struct mmu_gather *tlb,
563			    unsigned long addr, unsigned long end,
564			    unsigned long floor, unsigned long ceiling)
565{
566	pgd_t *pgd;
567	unsigned long next;
568
569	/*
570	 * Because there are a number of different possible pagetable
571	 * layouts for hugepage ranges, we limit knowledge of how
572	 * things should be laid out to the allocation path
573	 * (huge_pte_alloc(), above).  Everything else works out the
574	 * structure as it goes from information in the hugepd
575	 * pointers.  That means that we can't here use the
576	 * optimization used in the normal page free_pgd_range(), of
577	 * checking whether we're actually covering a large enough
578	 * range to have to do anything at the top level of the walk
579	 * instead of at the bottom.
580	 *
581	 * To make sense of this, you should probably go read the big
582	 * block comment at the top of the normal free_pgd_range(),
583	 * too.
584	 */
585
586	do {
587		next = pgd_addr_end(addr, end);
588		pgd = pgd_offset(tlb->mm, addr);
589		if (!is_hugepd(pgd)) {
590			if (pgd_none_or_clear_bad(pgd))
591				continue;
592			hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
593		} else {
594#ifdef CONFIG_PPC_FSL_BOOK3E
595			/*
596			 * Increment next by the size of the huge mapping since
597			 * there may be more than one entry at the pgd level
598			 * for a single hugepage, but all of them point to the
599			 * same kmem cache that holds the hugepte.
600			 */
601			next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
602#endif
 
 
603			free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
604					  addr, next, floor, ceiling);
605		}
606	} while (addr = next, addr != end);
607}
608
609struct page *
610follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 
611{
612	pte_t *ptep;
613	struct page *page;
614	unsigned shift;
615	unsigned long mask;
616
617	ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
618
619	/* Verify it is a huge page else bail. */
620	if (!ptep || !shift)
621		return ERR_PTR(-EINVAL);
622
623	mask = (1UL << shift) - 1;
624	page = pte_page(*ptep);
625	if (page)
626		page += (address & mask) / PAGE_SIZE;
627
628	return page;
629}
630
631int pmd_huge(pmd_t pmd)
632{
633	return 0;
634}
635
636int pud_huge(pud_t pud)
637{
638	return 0;
639}
640
641struct page *
642follow_huge_pmd(struct mm_struct *mm, unsigned long address,
643		pmd_t *pmd, int write)
644{
645	BUG();
646	return NULL;
647}
648
649static noinline int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
650		       unsigned long end, int write, struct page **pages, int *nr)
651{
652	unsigned long mask;
653	unsigned long pte_end;
654	struct page *head, *page, *tail;
655	pte_t pte;
656	int refs;
657
658	pte_end = (addr + sz) & ~(sz-1);
659	if (pte_end < end)
660		end = pte_end;
661
662	pte = *ptep;
663	mask = _PAGE_PRESENT | _PAGE_USER;
664	if (write)
665		mask |= _PAGE_RW;
666
667	if ((pte_val(pte) & mask) != mask)
668		return 0;
669
670	/* hugepages are never "special" */
671	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
672
673	refs = 0;
674	head = pte_page(pte);
675
676	page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
677	tail = page;
678	do {
679		VM_BUG_ON(compound_head(page) != head);
680		pages[*nr] = page;
681		(*nr)++;
682		page++;
683		refs++;
684	} while (addr += PAGE_SIZE, addr != end);
685
686	if (!page_cache_add_speculative(head, refs)) {
687		*nr -= refs;
688		return 0;
689	}
690
691	if (unlikely(pte_val(pte) != pte_val(*ptep))) {
692		/* Could be optimized better */
693		*nr -= refs;
694		while (refs--)
695			put_page(head);
696		return 0;
697	}
698
 
699	/*
700	 * Any tail page need their mapcount reference taken before we
701	 * return.
702	 */
703	while (refs--) {
704		if (PageTail(tail))
705			get_huge_page_tail(tail);
706		tail++;
 
 
 
 
 
 
 
 
 
 
 
 
707	}
708
709	return 1;
710}
711
712static 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
719int 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);
725	unsigned long next;
726
727	ptep = hugepte_offset(hugepd, addr, pdshift);
728	do {
729		next = hugepte_addr_end(addr, end, sz);
730		if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
731			return 0;
732	} while (ptep++, addr = next, addr != end);
733
734	return 1;
735}
736
737#ifdef CONFIG_PPC_MM_SLICES
738unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
739					unsigned long len, unsigned long pgoff,
740					unsigned long flags)
741{
742	struct hstate *hstate = hstate_file(file);
743	int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
744
745	return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1, 0);
 
 
 
 
 
746}
747#endif
748
749unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
750{
751#ifdef CONFIG_PPC_MM_SLICES
752	unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 
753
754	return 1UL << mmu_psize_to_shift(psize);
755#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;
768}
769
770static int __init add_huge_page_size(unsigned long long size)
771{
772	int shift = __ffs(size);
773	int mmu_psize;
774
775	/* Check that it is a page size supported by the hardware and
776	 * that it fits within pagetable and slice limits. */
777#ifdef CONFIG_PPC_FSL_BOOK3E
778	if ((size < PAGE_SIZE) || !is_power_of_4(size))
779		return -EINVAL;
780#else
781	if (!is_power_of_2(size)
782	    || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
783		return -EINVAL;
784#endif
785
786	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;
806}
807
808static int __init hugepage_setup_sz(char *str)
809{
810	unsigned long long size;
811
812	size = memparse(str, &str);
813
814	if (add_huge_page_size(size) != 0)
815		printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
 
 
816
817	return 1;
818}
819__setup("hugepagesz=", hugepage_setup_sz);
820
821#ifdef CONFIG_PPC_FSL_BOOK3E
822struct 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
861static int __init hugetlbpage_init(void)
862{
863	int psize;
864
865	if (!mmu_has_feature(MMU_FTR_16M_PAGE))
866		return -ENODEV;
867
868	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
869		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;
 
886
887		pgtable_cache_add(pdshift - shift, NULL);
888		if (!PGT_CACHE(pdshift - shift))
889			panic("hugetlbpage_init(): could not create "
890			      "pgtable cache for %d bit pagesize\n", shift);
 
 
 
 
 
 
 
 
 
 
891	}
892
893	/* Set default large page size. Currently, we pick 16M or 1M
894	 * depending on what is available
895	 */
896	if (mmu_psize_defs[MMU_PAGE_16M].shift)
897		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
898	else if (mmu_psize_defs[MMU_PAGE_1M].shift)
899		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
900
901	return 0;
902}
903#endif
904module_init(hugetlbpage_init);
905
906void flush_dcache_icache_hugepage(struct page *page)
907{
908	int i;
909	void *start;
910
911	BUG_ON(!PageCompound(page));
912
913	for (i = 0; i < (1UL << compound_order(page)); i++) {
914		if (!PageHighMem(page)) {
915			__flush_dcache_icache(page_address(page+i));
916		} else {
917			start = kmap_atomic(page+i);
918			__flush_dcache_icache(start);
919			kunmap_atomic(start);
920		}
921	}
922}