Linux Audio

Check our new training course

Loading...
v6.8
  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/pgalloc.h>
 23#include <asm/tlb.h>
 24#include <asm/setup.h>
 25#include <asm/hugetlb.h>
 26#include <asm/pte-walk.h>
 27#include <asm/firmware.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_basic_t)) - \
 34			 __builtin_ffs(sizeof(void *)))
 35
 36pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 37{
 38	/*
 39	 * Only called for hugetlbfs pages, hence can ignore THP and the
 40	 * irq disabled walk.
 41	 */
 42	return __find_linux_pte(mm->pgd, addr, NULL, NULL);
 43}
 44
 45static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
 46			   unsigned long address, unsigned int pdshift,
 47			   unsigned int pshift, spinlock_t *ptl)
 48{
 49	struct kmem_cache *cachep;
 50	pte_t *new;
 51	int i;
 52	int num_hugepd;
 53
 54	if (pshift >= pdshift) {
 55		cachep = PGT_CACHE(PTE_T_ORDER);
 56		num_hugepd = 1 << (pshift - pdshift);
 57	} else {
 58		cachep = PGT_CACHE(pdshift - pshift);
 59		num_hugepd = 1;
 60	}
 61
 62	if (!cachep) {
 63		WARN_ONCE(1, "No page table cache created for hugetlb tables");
 64		return -ENOMEM;
 65	}
 66
 67	new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
 68
 69	BUG_ON(pshift > HUGEPD_SHIFT_MASK);
 70	BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
 71
 72	if (!new)
 73		return -ENOMEM;
 74
 75	/*
 76	 * Make sure other cpus find the hugepd set only after a
 77	 * properly initialized page table is visible to them.
 78	 * For more details look for comment in __pte_alloc().
 79	 */
 80	smp_wmb();
 81
 82	spin_lock(ptl);
 
 83	/*
 84	 * We have multiple higher-level entries that point to the same
 85	 * actual pte location.  Fill in each as we go and backtrack on error.
 86	 * We need all of these so the DTLB pgtable walk code can find the
 87	 * right higher-level entry without knowing if it's a hugepage or not.
 88	 */
 89	for (i = 0; i < num_hugepd; i++, hpdp++) {
 90		if (unlikely(!hugepd_none(*hpdp)))
 91			break;
 92		hugepd_populate(hpdp, new, pshift);
 
 
 
 
 
 
 
 
 
 
 
 
 93	}
 94	/* If we bailed from the for loop early, an error occurred, clean up */
 95	if (i < num_hugepd) {
 96		for (i = i - 1 ; i >= 0; i--, hpdp--)
 97			*hpdp = __hugepd(0);
 98		kmem_cache_free(cachep, new);
 99	} else {
100		kmemleak_ignore(new);
101	}
102	spin_unlock(ptl);
103	return 0;
104}
105
106/*
 
 
 
 
 
 
 
 
 
 
 
 
107 * At this point we do the placement change only for BOOK3S 64. This would
108 * possibly work on other subarchs.
109 */
110pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
111		      unsigned long addr, unsigned long sz)
112{
113	pgd_t *pg;
114	p4d_t *p4;
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	p4 = p4d_offset(pg, addr);
125
126#ifdef CONFIG_PPC_BOOK3S_64
127	if (pshift == PGDIR_SHIFT)
128		/* 16GB huge page */
129		return (pte_t *) p4;
130	else if (pshift > PUD_SHIFT) {
131		/*
132		 * We need to use hugepd table
133		 */
134		ptl = &mm->page_table_lock;
135		hpdp = (hugepd_t *)p4;
136	} else {
137		pdshift = PUD_SHIFT;
138		pu = pud_alloc(mm, p4, addr);
139		if (!pu)
140			return NULL;
141		if (pshift == PUD_SHIFT)
142			return (pte_t *)pu;
143		else if (pshift > PMD_SHIFT) {
144			ptl = pud_lockptr(mm, pu);
145			hpdp = (hugepd_t *)pu;
146		} else {
147			pdshift = PMD_SHIFT;
148			pm = pmd_alloc(mm, pu, addr);
149			if (!pm)
150				return NULL;
151			if (pshift == PMD_SHIFT)
152				/* 16MB hugepage */
153				return (pte_t *)pm;
154			else {
155				ptl = pmd_lockptr(mm, pm);
156				hpdp = (hugepd_t *)pm;
157			}
158		}
159	}
160#else
161	if (pshift >= PGDIR_SHIFT) {
162		ptl = &mm->page_table_lock;
163		hpdp = (hugepd_t *)p4;
164	} else {
165		pdshift = PUD_SHIFT;
166		pu = pud_alloc(mm, p4, addr);
167		if (!pu)
168			return NULL;
169		if (pshift >= PUD_SHIFT) {
170			ptl = pud_lockptr(mm, pu);
171			hpdp = (hugepd_t *)pu;
172		} else {
173			pdshift = PMD_SHIFT;
174			pm = pmd_alloc(mm, pu, addr);
175			if (!pm)
176				return NULL;
177			ptl = pmd_lockptr(mm, pm);
178			hpdp = (hugepd_t *)pm;
179		}
180	}
181#endif
182	if (!hpdp)
183		return NULL;
184
185	if (IS_ENABLED(CONFIG_PPC_8xx) && pshift < PMD_SHIFT)
186		return pte_alloc_huge(mm, (pmd_t *)hpdp, addr);
187
188	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
189
190	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
191						  pdshift, pshift, ptl))
192		return NULL;
193
194	return hugepte_offset(*hpdp, addr, pdshift);
195}
196
197#ifdef CONFIG_PPC_BOOK3S_64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
198/*
199 * Tracks gpages after the device tree is scanned and before the
200 * huge_boot_pages list is ready on pseries.
201 */
202#define MAX_NUMBER_GPAGES	1024
203__initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
204__initdata static unsigned nr_gpages;
 
 
205
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
206/*
207 * Build list of addresses of gigantic pages.  This function is used in early
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
208 * boot before the buddy allocator is setup.
209 */
210void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
211{
212	if (!addr)
213		return;
214	while (number_of_pages > 0) {
215		gpage_freearray[nr_gpages] = addr;
216		nr_gpages++;
217		number_of_pages--;
218		addr += page_size;
219	}
220}
221
222static int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
 
 
 
223{
224	struct huge_bootmem_page *m;
225	if (nr_gpages == 0)
226		return 0;
227	m = phys_to_virt(gpage_freearray[--nr_gpages]);
228	gpage_freearray[nr_gpages] = 0;
229	list_add(&m->list, &huge_boot_pages);
230	m->hstate = hstate;
231	return 1;
232}
233
234bool __init hugetlb_node_alloc_supported(void)
235{
236	return false;
237}
238#endif
239
240
241int __init alloc_bootmem_huge_page(struct hstate *h, int nid)
242{
243
244#ifdef CONFIG_PPC_BOOK3S_64
245	if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
246		return pseries_alloc_bootmem_huge_page(h);
247#endif
248	return __alloc_bootmem_huge_page(h, nid);
249}
250
251#ifndef CONFIG_PPC_BOOK3S_64
252#define HUGEPD_FREELIST_SIZE \
253	((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
254
255struct hugepd_freelist {
256	struct rcu_head	rcu;
257	unsigned int index;
258	void *ptes[];
259};
260
261static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
262
263static void hugepd_free_rcu_callback(struct rcu_head *head)
264{
265	struct hugepd_freelist *batch =
266		container_of(head, struct hugepd_freelist, rcu);
267	unsigned int i;
268
269	for (i = 0; i < batch->index; i++)
270		kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
271
272	free_page((unsigned long)batch);
273}
274
275static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
276{
277	struct hugepd_freelist **batchp;
278
279	batchp = &get_cpu_var(hugepd_freelist_cur);
280
281	if (atomic_read(&tlb->mm->mm_users) < 2 ||
282	    mm_is_thread_local(tlb->mm)) {
283		kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
 
284		put_cpu_var(hugepd_freelist_cur);
285		return;
286	}
287
288	if (*batchp == NULL) {
289		*batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
290		(*batchp)->index = 0;
291	}
292
293	(*batchp)->ptes[(*batchp)->index++] = hugepte;
294	if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
295		call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
296		*batchp = NULL;
297	}
298	put_cpu_var(hugepd_freelist_cur);
299}
300#else
301static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
302#endif
303
304/* Return true when the entry to be freed maps more than the area being freed */
305static bool range_is_outside_limits(unsigned long start, unsigned long end,
306				    unsigned long floor, unsigned long ceiling,
307				    unsigned long mask)
308{
309	if ((start & mask) < floor)
310		return true;
311	if (ceiling) {
312		ceiling &= mask;
313		if (!ceiling)
314			return true;
315	}
316	return end - 1 > ceiling - 1;
317}
318
319static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
320			      unsigned long start, unsigned long end,
321			      unsigned long floor, unsigned long ceiling)
322{
323	pte_t *hugepte = hugepd_page(*hpdp);
324	int i;
325
326	unsigned long pdmask = ~((1UL << pdshift) - 1);
327	unsigned int num_hugepd = 1;
328	unsigned int shift = hugepd_shift(*hpdp);
329
330	/* Note: On fsl the hpdp may be the first of several */
331	if (shift > pdshift)
332		num_hugepd = 1 << (shift - pdshift);
333
334	if (range_is_outside_limits(start, end, floor, ceiling, pdmask))
 
 
 
 
 
 
 
 
335		return;
336
337	for (i = 0; i < num_hugepd; i++, hpdp++)
338		*hpdp = __hugepd(0);
339
340	if (shift >= pdshift)
341		hugepd_free(tlb, hugepte);
342	else
343		pgtable_free_tlb(tlb, hugepte,
344				 get_hugepd_cache_index(pdshift - shift));
345}
346
347static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
348				   unsigned long addr, unsigned long end,
349				   unsigned long floor, unsigned long ceiling)
350{
351	pgtable_t token = pmd_pgtable(*pmd);
352
353	if (range_is_outside_limits(addr, end, floor, ceiling, PMD_MASK))
354		return;
355
356	pmd_clear(pmd);
357	pte_free_tlb(tlb, token, addr);
358	mm_dec_nr_ptes(tlb->mm);
359}
360
361static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
362				   unsigned long addr, unsigned long end,
363				   unsigned long floor, unsigned long ceiling)
364{
365	pmd_t *pmd;
366	unsigned long next;
367	unsigned long start;
368
369	start = addr;
370	do {
371		unsigned long more;
372
373		pmd = pmd_offset(pud, addr);
374		next = pmd_addr_end(addr, end);
375		if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
376			if (pmd_none_or_clear_bad(pmd))
377				continue;
378
379			/*
380			 * if it is not hugepd pointer, we should already find
381			 * it cleared.
382			 */
383			WARN_ON(!IS_ENABLED(CONFIG_PPC_8xx));
384
385			hugetlb_free_pte_range(tlb, pmd, addr, end, floor, ceiling);
386
387			continue;
388		}
389		/*
390		 * Increment next by the size of the huge mapping since
391		 * there may be more than one entry at this level for a
392		 * single hugepage, but all of them point to
393		 * the same kmem cache that holds the hugepte.
394		 */
395		more = addr + (1UL << hugepd_shift(*(hugepd_t *)pmd));
396		if (more > next)
397			next = more;
398
399		free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
400				  addr, next, floor, ceiling);
401	} while (addr = next, addr != end);
402
403	if (range_is_outside_limits(start, end, floor, ceiling, PUD_MASK))
 
 
 
 
 
 
 
 
404		return;
405
406	pmd = pmd_offset(pud, start & PUD_MASK);
407	pud_clear(pud);
408	pmd_free_tlb(tlb, pmd, start & PUD_MASK);
409	mm_dec_nr_pmds(tlb->mm);
410}
411
412static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
413				   unsigned long addr, unsigned long end,
414				   unsigned long floor, unsigned long ceiling)
415{
416	pud_t *pud;
417	unsigned long next;
418	unsigned long start;
419
420	start = addr;
421	do {
422		pud = pud_offset(p4d, addr);
423		next = pud_addr_end(addr, end);
424		if (!is_hugepd(__hugepd(pud_val(*pud)))) {
425			if (pud_none_or_clear_bad(pud))
426				continue;
427			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
428					       ceiling);
429		} else {
430			unsigned long more;
431			/*
432			 * Increment next by the size of the huge mapping since
433			 * there may be more than one entry at this level for a
434			 * single hugepage, but all of them point to
435			 * the same kmem cache that holds the hugepte.
436			 */
437			more = addr + (1UL << hugepd_shift(*(hugepd_t *)pud));
438			if (more > next)
439				next = more;
440
441			free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
442					  addr, next, floor, ceiling);
443		}
444	} while (addr = next, addr != end);
445
446	if (range_is_outside_limits(start, end, floor, ceiling, PGDIR_MASK))
 
 
 
 
 
 
 
 
447		return;
448
449	pud = pud_offset(p4d, start & PGDIR_MASK);
450	p4d_clear(p4d);
451	pud_free_tlb(tlb, pud, start & PGDIR_MASK);
452	mm_dec_nr_puds(tlb->mm);
453}
454
455/*
456 * This function frees user-level page tables of a process.
457 */
458void hugetlb_free_pgd_range(struct mmu_gather *tlb,
459			    unsigned long addr, unsigned long end,
460			    unsigned long floor, unsigned long ceiling)
461{
462	pgd_t *pgd;
463	p4d_t *p4d;
464	unsigned long next;
465
466	/*
467	 * Because there are a number of different possible pagetable
468	 * layouts for hugepage ranges, we limit knowledge of how
469	 * things should be laid out to the allocation path
470	 * (huge_pte_alloc(), above).  Everything else works out the
471	 * structure as it goes from information in the hugepd
472	 * pointers.  That means that we can't here use the
473	 * optimization used in the normal page free_pgd_range(), of
474	 * checking whether we're actually covering a large enough
475	 * range to have to do anything at the top level of the walk
476	 * instead of at the bottom.
477	 *
478	 * To make sense of this, you should probably go read the big
479	 * block comment at the top of the normal free_pgd_range(),
480	 * too.
481	 */
482
483	do {
484		next = pgd_addr_end(addr, end);
485		pgd = pgd_offset(tlb->mm, addr);
486		p4d = p4d_offset(pgd, addr);
487		if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
488			if (p4d_none_or_clear_bad(p4d))
489				continue;
490			hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling);
491		} else {
492			unsigned long more;
493			/*
494			 * Increment next by the size of the huge mapping since
495			 * there may be more than one entry at the pgd level
496			 * for a single hugepage, but all of them point to the
497			 * same kmem cache that holds the hugepte.
498			 */
499			more = addr + (1UL << hugepd_shift(*(hugepd_t *)pgd));
500			if (more > next)
501				next = more;
502
503			free_hugepd_range(tlb, (hugepd_t *)p4d, PGDIR_SHIFT,
504					  addr, next, floor, ceiling);
505		}
506	} while (addr = next, addr != end);
507}
508
509bool __init arch_hugetlb_valid_size(unsigned long size)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
510{
511	int shift = __ffs(size);
512	int mmu_psize;
 
513
514	/* Check that it is a page size supported by the hardware and
515	 * that it fits within pagetable and slice limits. */
516	if (size <= PAGE_SIZE || !is_power_of_2(size))
517		return false;
 
 
 
518
519	mmu_psize = check_and_get_huge_psize(shift);
520	if (mmu_psize < 0)
521		return false;
 
 
 
522
523	BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
524
525	return true;
 
 
 
 
 
 
 
526}
527
528static int __init add_huge_page_size(unsigned long long size)
529{
530	int shift = __ffs(size);
 
 
 
 
 
 
 
 
 
 
 
 
 
531
532	if (!arch_hugetlb_valid_size((unsigned long)size))
533		return -EINVAL;
534
 
 
 
 
 
 
535	hugetlb_add_hstate(shift - PAGE_SHIFT);
 
536	return 0;
537}
538
539static int __init hugetlbpage_init(void)
540{
541	bool configured = false;
542	int psize;
543
544	if (hugetlb_disabled) {
545		pr_info("HugeTLB support is disabled!\n");
546		return 0;
 
 
547	}
548
549	if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
550	    !mmu_has_feature(MMU_FTR_16M_PAGE))
551		return -ENODEV;
552
 
 
 
 
 
 
 
 
 
553	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
554		unsigned shift;
555		unsigned pdshift;
556
557		if (!mmu_psize_defs[psize].shift)
558			continue;
559
560		shift = mmu_psize_to_shift(psize);
561
562#ifdef CONFIG_PPC_BOOK3S_64
563		if (shift > PGDIR_SHIFT)
564			continue;
565		else if (shift > PUD_SHIFT)
566			pdshift = PGDIR_SHIFT;
567		else if (shift > PMD_SHIFT)
568			pdshift = PUD_SHIFT;
569		else
570			pdshift = PMD_SHIFT;
571#else
572		if (shift < PUD_SHIFT)
573			pdshift = PMD_SHIFT;
574		else if (shift < PGDIR_SHIFT)
575			pdshift = PUD_SHIFT;
576		else
577			pdshift = PGDIR_SHIFT;
578#endif
579
580		if (add_huge_page_size(1ULL << shift) < 0)
581			continue;
582		/*
583		 * if we have pdshift and shift value same, we don't
584		 * use pgt cache for hugepd.
585		 */
586		if (pdshift > shift) {
587			if (!IS_ENABLED(CONFIG_PPC_8xx))
588				pgtable_cache_add(pdshift - shift);
589		} else if (IS_ENABLED(CONFIG_PPC_E500) ||
590			   IS_ENABLED(CONFIG_PPC_8xx)) {
591			pgtable_cache_add(PTE_T_ORDER);
592		}
 
 
 
 
 
 
 
 
 
593
594		configured = true;
 
595	}
596
597	if (!configured)
598		pr_info("Failed to initialize. Disabling HugeTLB");
599
 
 
 
 
 
 
 
 
 
 
 
 
 
 
600	return 0;
601}
602
603arch_initcall(hugetlbpage_init);
604
605void __init gigantic_hugetlb_cma_reserve(void)
606{
607	unsigned long order = 0;
 
608
609	if (radix_enabled())
610		order = PUD_SHIFT - PAGE_SHIFT;
611	else if (!firmware_has_feature(FW_FEATURE_LPAR) && mmu_psize_defs[MMU_PAGE_16G].shift)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
612		/*
613		 * For pseries we do use ibm,expected#pages for reserving 16G pages.
 
 
614		 */
615		order = mmu_psize_to_shift(MMU_PAGE_16G) - PAGE_SHIFT;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
616
617	if (order) {
618		VM_WARN_ON(order <= MAX_PAGE_ORDER);
619		hugetlb_cma_reserve(order);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
620	}
 
 
621}
v4.10.11
   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#include <asm/hugetlb.h>
 
 
  25
  26#ifdef CONFIG_HUGETLB_PAGE
  27
  28#define PAGE_SHIFT_64K	16
  29#define PAGE_SHIFT_512K	19
  30#define PAGE_SHIFT_8M	23
  31#define PAGE_SHIFT_16M	24
  32#define PAGE_SHIFT_16G	34
  33
  34unsigned int HPAGE_SHIFT;
  35
  36/*
  37 * Tracks gpages after the device tree is scanned and before the
  38 * huge_boot_pages list is ready.  On non-Freescale implementations, this is
  39 * just used to track 16G pages and so is a single array.  FSL-based
  40 * implementations may have more than one gpage size, so we need multiple
  41 * arrays
  42 */
  43#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
  44#define MAX_NUMBER_GPAGES	128
  45struct psize_gpages {
  46	u64 gpage_list[MAX_NUMBER_GPAGES];
  47	unsigned int nr_gpages;
  48};
  49static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
  50#else
  51#define MAX_NUMBER_GPAGES	1024
  52static u64 gpage_freearray[MAX_NUMBER_GPAGES];
  53static unsigned nr_gpages;
  54#endif
  55
  56#define hugepd_none(hpd)	(hpd_val(hpd) == 0)
 
  57
  58pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  59{
  60	/* Only called for hugetlbfs pages, hence can ignore THP */
  61	return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL, NULL);
 
 
 
  62}
  63
  64static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
  65			   unsigned long address, unsigned pdshift, unsigned pshift)
 
  66{
  67	struct kmem_cache *cachep;
  68	pte_t *new;
  69	int i;
  70	int num_hugepd;
  71
  72	if (pshift >= pdshift) {
  73		cachep = hugepte_cache;
  74		num_hugepd = 1 << (pshift - pdshift);
  75	} else {
  76		cachep = PGT_CACHE(pdshift - pshift);
  77		num_hugepd = 1;
  78	}
  79
  80	new = kmem_cache_zalloc(cachep, GFP_KERNEL);
 
 
 
 
 
  81
  82	BUG_ON(pshift > HUGEPD_SHIFT_MASK);
  83	BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
  84
  85	if (! new)
  86		return -ENOMEM;
  87
  88	/*
  89	 * Make sure other cpus find the hugepd set only after a
  90	 * properly initialized page table is visible to them.
  91	 * For more details look for comment in __pte_alloc().
  92	 */
  93	smp_wmb();
  94
  95	spin_lock(&mm->page_table_lock);
  96
  97	/*
  98	 * We have multiple higher-level entries that point to the same
  99	 * actual pte location.  Fill in each as we go and backtrack on error.
 100	 * We need all of these so the DTLB pgtable walk code can find the
 101	 * right higher-level entry without knowing if it's a hugepage or not.
 102	 */
 103	for (i = 0; i < num_hugepd; i++, hpdp++) {
 104		if (unlikely(!hugepd_none(*hpdp)))
 105			break;
 106		else {
 107#ifdef CONFIG_PPC_BOOK3S_64
 108			*hpdp = __hugepd(__pa(new) |
 109					 (shift_to_mmu_psize(pshift) << 2));
 110#elif defined(CONFIG_PPC_8xx)
 111			*hpdp = __hugepd(__pa(new) |
 112					 (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
 113					  _PMD_PAGE_512K) | _PMD_PRESENT);
 114#else
 115			/* We use the old format for PPC_FSL_BOOK3E */
 116			*hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
 117#endif
 118		}
 119	}
 120	/* If we bailed from the for loop early, an error occurred, clean up */
 121	if (i < num_hugepd) {
 122		for (i = i - 1 ; i >= 0; i--, hpdp--)
 123			*hpdp = __hugepd(0);
 124		kmem_cache_free(cachep, new);
 
 
 125	}
 126	spin_unlock(&mm->page_table_lock);
 127	return 0;
 128}
 129
 130/*
 131 * These macros define how to determine which level of the page table holds
 132 * the hpdp.
 133 */
 134#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 135#define HUGEPD_PGD_SHIFT PGDIR_SHIFT
 136#define HUGEPD_PUD_SHIFT PUD_SHIFT
 137#else
 138#define HUGEPD_PGD_SHIFT PUD_SHIFT
 139#define HUGEPD_PUD_SHIFT PMD_SHIFT
 140#endif
 141
 142/*
 143 * At this point we do the placement change only for BOOK3S 64. This would
 144 * possibly work on other subarchs.
 145 */
 146pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 
 147{
 148	pgd_t *pg;
 
 149	pud_t *pu;
 150	pmd_t *pm;
 151	hugepd_t *hpdp = NULL;
 152	unsigned pshift = __ffs(sz);
 153	unsigned pdshift = PGDIR_SHIFT;
 
 154
 155	addr &= ~(sz-1);
 156	pg = pgd_offset(mm, addr);
 
 157
 158#ifdef CONFIG_PPC_BOOK3S_64
 159	if (pshift == PGDIR_SHIFT)
 160		/* 16GB huge page */
 161		return (pte_t *) pg;
 162	else if (pshift > PUD_SHIFT)
 163		/*
 164		 * We need to use hugepd table
 165		 */
 166		hpdp = (hugepd_t *)pg;
 167	else {
 
 168		pdshift = PUD_SHIFT;
 169		pu = pud_alloc(mm, pg, addr);
 
 
 170		if (pshift == PUD_SHIFT)
 171			return (pte_t *)pu;
 172		else if (pshift > PMD_SHIFT)
 
 173			hpdp = (hugepd_t *)pu;
 174		else {
 175			pdshift = PMD_SHIFT;
 176			pm = pmd_alloc(mm, pu, addr);
 
 
 177			if (pshift == PMD_SHIFT)
 178				/* 16MB hugepage */
 179				return (pte_t *)pm;
 180			else
 
 181				hpdp = (hugepd_t *)pm;
 
 182		}
 183	}
 184#else
 185	if (pshift >= HUGEPD_PGD_SHIFT) {
 186		hpdp = (hugepd_t *)pg;
 
 187	} else {
 188		pdshift = PUD_SHIFT;
 189		pu = pud_alloc(mm, pg, addr);
 190		if (pshift >= HUGEPD_PUD_SHIFT) {
 
 
 
 191			hpdp = (hugepd_t *)pu;
 192		} else {
 193			pdshift = PMD_SHIFT;
 194			pm = pmd_alloc(mm, pu, addr);
 
 
 
 195			hpdp = (hugepd_t *)pm;
 196		}
 197	}
 198#endif
 199	if (!hpdp)
 200		return NULL;
 201
 
 
 
 202	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
 203
 204	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
 
 205		return NULL;
 206
 207	return hugepte_offset(*hpdp, addr, pdshift);
 208}
 209
 210#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 211/* Build list of addresses of gigantic pages.  This function is used in early
 212 * boot before the buddy allocator is setup.
 213 */
 214void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 215{
 216	unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
 217	int i;
 218
 219	if (addr == 0)
 220		return;
 221
 222	gpage_freearray[idx].nr_gpages = number_of_pages;
 223
 224	for (i = 0; i < number_of_pages; i++) {
 225		gpage_freearray[idx].gpage_list[i] = addr;
 226		addr += page_size;
 227	}
 228}
 229
 230/*
 231 * Moves the gigantic page addresses from the temporary list to the
 232 * huge_boot_pages list.
 233 */
 234int alloc_bootmem_huge_page(struct hstate *hstate)
 235{
 236	struct huge_bootmem_page *m;
 237	int idx = shift_to_mmu_psize(huge_page_shift(hstate));
 238	int nr_gpages = gpage_freearray[idx].nr_gpages;
 239
 240	if (nr_gpages == 0)
 241		return 0;
 242
 243#ifdef CONFIG_HIGHMEM
 244	/*
 245	 * If gpages can be in highmem we can't use the trick of storing the
 246	 * data structure in the page; allocate space for this
 247	 */
 248	m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
 249	m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
 250#else
 251	m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
 252#endif
 253
 254	list_add(&m->list, &huge_boot_pages);
 255	gpage_freearray[idx].nr_gpages = nr_gpages;
 256	gpage_freearray[idx].gpage_list[nr_gpages] = 0;
 257	m->hstate = hstate;
 258
 259	return 1;
 260}
 261/*
 262 * Scan the command line hugepagesz= options for gigantic pages; store those in
 263 * a list that we use to allocate the memory once all options are parsed.
 264 */
 265
 266unsigned long gpage_npages[MMU_PAGE_COUNT];
 267
 268static int __init do_gpage_early_setup(char *param, char *val,
 269				       const char *unused, void *arg)
 270{
 271	static phys_addr_t size;
 272	unsigned long npages;
 273
 274	/*
 275	 * The hugepagesz and hugepages cmdline options are interleaved.  We
 276	 * use the size variable to keep track of whether or not this was done
 277	 * properly and skip over instances where it is incorrect.  Other
 278	 * command-line parsing code will issue warnings, so we don't need to.
 279	 *
 280	 */
 281	if ((strcmp(param, "default_hugepagesz") == 0) ||
 282	    (strcmp(param, "hugepagesz") == 0)) {
 283		size = memparse(val, NULL);
 284	} else if (strcmp(param, "hugepages") == 0) {
 285		if (size != 0) {
 286			if (sscanf(val, "%lu", &npages) <= 0)
 287				npages = 0;
 288			if (npages > MAX_NUMBER_GPAGES) {
 289				pr_warn("MMU: %lu pages requested for page "
 290#ifdef CONFIG_PHYS_ADDR_T_64BIT
 291					"size %llu KB, limiting to "
 292#else
 293					"size %u KB, limiting to "
 294#endif
 295					__stringify(MAX_NUMBER_GPAGES) "\n",
 296					npages, size / 1024);
 297				npages = MAX_NUMBER_GPAGES;
 298			}
 299			gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
 300			size = 0;
 301		}
 302	}
 303	return 0;
 304}
 305
 306
 307/*
 308 * This function allocates physical space for pages that are larger than the
 309 * buddy allocator can handle.  We want to allocate these in highmem because
 310 * the amount of lowmem is limited.  This means that this function MUST be
 311 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
 312 * allocate to grab highmem.
 313 */
 314void __init reserve_hugetlb_gpages(void)
 315{
 316	static __initdata char cmdline[COMMAND_LINE_SIZE];
 317	phys_addr_t size, base;
 318	int i;
 319
 320	strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
 321	parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
 322			NULL, &do_gpage_early_setup);
 323
 324	/*
 325	 * Walk gpage list in reverse, allocating larger page sizes first.
 326	 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
 327	 * When we reach the point in the list where pages are no longer
 328	 * considered gpages, we're done.
 329	 */
 330	for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
 331		if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
 332			continue;
 333		else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
 334			break;
 335
 336		size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
 337		base = memblock_alloc_base(size * gpage_npages[i], size,
 338					   MEMBLOCK_ALLOC_ANYWHERE);
 339		add_gpage(base, size, gpage_npages[i]);
 340	}
 341}
 342
 343#else /* !PPC_FSL_BOOK3E */
 344
 345/* Build list of addresses of gigantic pages.  This function is used in early
 346 * boot before the buddy allocator is setup.
 347 */
 348void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
 349{
 350	if (!addr)
 351		return;
 352	while (number_of_pages > 0) {
 353		gpage_freearray[nr_gpages] = addr;
 354		nr_gpages++;
 355		number_of_pages--;
 356		addr += page_size;
 357	}
 358}
 359
 360/* Moves the gigantic page addresses from the temporary list to the
 361 * huge_boot_pages list.
 362 */
 363int alloc_bootmem_huge_page(struct hstate *hstate)
 364{
 365	struct huge_bootmem_page *m;
 366	if (nr_gpages == 0)
 367		return 0;
 368	m = phys_to_virt(gpage_freearray[--nr_gpages]);
 369	gpage_freearray[nr_gpages] = 0;
 370	list_add(&m->list, &huge_boot_pages);
 371	m->hstate = hstate;
 372	return 1;
 373}
 
 
 
 
 
 374#endif
 375
 376#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 
 
 
 
 
 
 
 
 
 
 
 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		put_cpu_var(hugepd_freelist_cur);
 411		return;
 412	}
 413
 414	if (*batchp == NULL) {
 415		*batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
 416		(*batchp)->index = 0;
 417	}
 418
 419	(*batchp)->ptes[(*batchp)->index++] = hugepte;
 420	if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
 421		call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
 422		*batchp = NULL;
 423	}
 424	put_cpu_var(hugepd_freelist_cur);
 425}
 426#else
 427static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
 428#endif
 429
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 430static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
 431			      unsigned long start, unsigned long end,
 432			      unsigned long floor, unsigned long ceiling)
 433{
 434	pte_t *hugepte = hugepd_page(*hpdp);
 435	int i;
 436
 437	unsigned long pdmask = ~((1UL << pdshift) - 1);
 438	unsigned int num_hugepd = 1;
 439	unsigned int shift = hugepd_shift(*hpdp);
 440
 441	/* Note: On fsl the hpdp may be the first of several */
 442	if (shift > pdshift)
 443		num_hugepd = 1 << (shift - pdshift);
 444
 445	start &= pdmask;
 446	if (start < floor)
 447		return;
 448	if (ceiling) {
 449		ceiling &= pdmask;
 450		if (! ceiling)
 451			return;
 452	}
 453	if (end - 1 > ceiling - 1)
 454		return;
 455
 456	for (i = 0; i < num_hugepd; i++, hpdp++)
 457		*hpdp = __hugepd(0);
 458
 459	if (shift >= pdshift)
 460		hugepd_free(tlb, hugepte);
 461	else
 462		pgtable_free_tlb(tlb, hugepte, pdshift - shift);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 463}
 464
 465static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 466				   unsigned long addr, unsigned long end,
 467				   unsigned long floor, unsigned long ceiling)
 468{
 469	pmd_t *pmd;
 470	unsigned long next;
 471	unsigned long start;
 472
 473	start = addr;
 474	do {
 475		unsigned long more;
 476
 477		pmd = pmd_offset(pud, addr);
 478		next = pmd_addr_end(addr, end);
 479		if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
 
 
 
 480			/*
 481			 * if it is not hugepd pointer, we should already find
 482			 * it cleared.
 483			 */
 484			WARN_ON(!pmd_none_or_clear_bad(pmd));
 
 
 
 485			continue;
 486		}
 487		/*
 488		 * Increment next by the size of the huge mapping since
 489		 * there may be more than one entry at this level for a
 490		 * single hugepage, but all of them point to
 491		 * the same kmem cache that holds the hugepte.
 492		 */
 493		more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
 494		if (more > next)
 495			next = more;
 496
 497		free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
 498				  addr, next, floor, ceiling);
 499	} while (addr = next, addr != end);
 500
 501	start &= PUD_MASK;
 502	if (start < floor)
 503		return;
 504	if (ceiling) {
 505		ceiling &= PUD_MASK;
 506		if (!ceiling)
 507			return;
 508	}
 509	if (end - 1 > ceiling - 1)
 510		return;
 511
 512	pmd = pmd_offset(pud, start);
 513	pud_clear(pud);
 514	pmd_free_tlb(tlb, pmd, start);
 515	mm_dec_nr_pmds(tlb->mm);
 516}
 517
 518static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 519				   unsigned long addr, unsigned long end,
 520				   unsigned long floor, unsigned long ceiling)
 521{
 522	pud_t *pud;
 523	unsigned long next;
 524	unsigned long start;
 525
 526	start = addr;
 527	do {
 528		pud = pud_offset(pgd, addr);
 529		next = pud_addr_end(addr, end);
 530		if (!is_hugepd(__hugepd(pud_val(*pud)))) {
 531			if (pud_none_or_clear_bad(pud))
 532				continue;
 533			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
 534					       ceiling);
 535		} else {
 536			unsigned long more;
 537			/*
 538			 * Increment next by the size of the huge mapping since
 539			 * there may be more than one entry at this level for a
 540			 * single hugepage, but all of them point to
 541			 * the same kmem cache that holds the hugepte.
 542			 */
 543			more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
 544			if (more > next)
 545				next = more;
 546
 547			free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
 548					  addr, next, floor, ceiling);
 549		}
 550	} while (addr = next, addr != end);
 551
 552	start &= PGDIR_MASK;
 553	if (start < floor)
 554		return;
 555	if (ceiling) {
 556		ceiling &= PGDIR_MASK;
 557		if (!ceiling)
 558			return;
 559	}
 560	if (end - 1 > ceiling - 1)
 561		return;
 562
 563	pud = pud_offset(pgd, start);
 564	pgd_clear(pgd);
 565	pud_free_tlb(tlb, pud, start);
 
 566}
 567
 568/*
 569 * This function frees user-level page tables of a process.
 570 */
 571void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 572			    unsigned long addr, unsigned long end,
 573			    unsigned long floor, unsigned long ceiling)
 574{
 575	pgd_t *pgd;
 
 576	unsigned long next;
 577
 578	/*
 579	 * Because there are a number of different possible pagetable
 580	 * layouts for hugepage ranges, we limit knowledge of how
 581	 * things should be laid out to the allocation path
 582	 * (huge_pte_alloc(), above).  Everything else works out the
 583	 * structure as it goes from information in the hugepd
 584	 * pointers.  That means that we can't here use the
 585	 * optimization used in the normal page free_pgd_range(), of
 586	 * checking whether we're actually covering a large enough
 587	 * range to have to do anything at the top level of the walk
 588	 * instead of at the bottom.
 589	 *
 590	 * To make sense of this, you should probably go read the big
 591	 * block comment at the top of the normal free_pgd_range(),
 592	 * too.
 593	 */
 594
 595	do {
 596		next = pgd_addr_end(addr, end);
 597		pgd = pgd_offset(tlb->mm, addr);
 
 598		if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
 599			if (pgd_none_or_clear_bad(pgd))
 600				continue;
 601			hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 602		} else {
 603			unsigned long more;
 604			/*
 605			 * Increment next by the size of the huge mapping since
 606			 * there may be more than one entry at the pgd level
 607			 * for a single hugepage, but all of them point to the
 608			 * same kmem cache that holds the hugepte.
 609			 */
 610			more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
 611			if (more > next)
 612				next = more;
 613
 614			free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
 615					  addr, next, floor, ceiling);
 616		}
 617	} while (addr = next, addr != end);
 618}
 619
 620/*
 621 * We are holding mmap_sem, so a parallel huge page collapse cannot run.
 622 * To prevent hugepage split, disable irq.
 623 */
 624struct page *
 625follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 626{
 627	bool is_thp;
 628	pte_t *ptep, pte;
 629	unsigned shift;
 630	unsigned long mask, flags;
 631	struct page *page = ERR_PTR(-EINVAL);
 632
 633	local_irq_save(flags);
 634	ptep = find_linux_pte_or_hugepte(mm->pgd, address, &is_thp, &shift);
 635	if (!ptep)
 636		goto no_page;
 637	pte = READ_ONCE(*ptep);
 638	/*
 639	 * Verify it is a huge page else bail.
 640	 * Transparent hugepages are handled by generic code. We can skip them
 641	 * here.
 642	 */
 643	if (!shift || is_thp)
 644		goto no_page;
 645
 646	if (!pte_present(pte)) {
 647		page = NULL;
 648		goto no_page;
 649	}
 650	mask = (1UL << shift) - 1;
 651	page = pte_page(pte);
 652	if (page)
 653		page += (address & mask) / PAGE_SIZE;
 654
 655no_page:
 656	local_irq_restore(flags);
 657	return page;
 658}
 659
 660struct page *
 661follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 662		pmd_t *pmd, int write)
 663{
 664	BUG();
 665	return NULL;
 666}
 667
 668struct page *
 669follow_huge_pud(struct mm_struct *mm, unsigned long address,
 670		pud_t *pud, int write)
 671{
 672	BUG();
 673	return NULL;
 674}
 675
 676static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
 677				      unsigned long sz)
 678{
 679	unsigned long __boundary = (addr + sz) & ~(sz-1);
 680	return (__boundary - 1 < end - 1) ? __boundary : end;
 681}
 682
 683int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
 684		unsigned long end, int write, struct page **pages, int *nr)
 685{
 686	pte_t *ptep;
 687	unsigned long sz = 1UL << hugepd_shift(hugepd);
 688	unsigned long next;
 689
 690	ptep = hugepte_offset(hugepd, addr, pdshift);
 691	do {
 692		next = hugepte_addr_end(addr, end, sz);
 693		if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
 694			return 0;
 695	} while (ptep++, addr = next, addr != end);
 696
 697	return 1;
 698}
 699
 700#ifdef CONFIG_PPC_MM_SLICES
 701unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 702					unsigned long len, unsigned long pgoff,
 703					unsigned long flags)
 704{
 705	struct hstate *hstate = hstate_file(file);
 706	int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
 707
 708	if (radix_enabled())
 709		return radix__hugetlb_get_unmapped_area(file, addr, len,
 710						       pgoff, flags);
 711	return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
 712}
 713#endif
 714
 715unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 716{
 717#ifdef CONFIG_PPC_MM_SLICES
 718	unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
 719	/* With radix we don't use slice, so derive it from vma*/
 720	if (!radix_enabled())
 721		return 1UL << mmu_psize_to_shift(psize);
 722#endif
 723	if (!is_vm_hugetlb_page(vma))
 724		return PAGE_SIZE;
 725
 726	return huge_page_size(hstate_vma(vma));
 727}
 728
 729static inline bool is_power_of_4(unsigned long x)
 730{
 731	if (is_power_of_2(x))
 732		return (__ilog2(x) % 2) ? false : true;
 733	return false;
 734}
 735
 736static int __init add_huge_page_size(unsigned long long size)
 737{
 738	int shift = __ffs(size);
 739	int mmu_psize;
 740
 741	/* Check that it is a page size supported by the hardware and
 742	 * that it fits within pagetable and slice limits. */
 743	if (size <= PAGE_SIZE)
 744		return -EINVAL;
 745#if defined(CONFIG_PPC_FSL_BOOK3E)
 746	if (!is_power_of_4(size))
 747		return -EINVAL;
 748#elif !defined(CONFIG_PPC_8xx)
 749	if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
 750		return -EINVAL;
 751#endif
 752
 753	if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
 754		return -EINVAL;
 755
 756	BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
 757
 758	/* Return if huge page size has already been setup */
 759	if (size_to_hstate(size))
 760		return 0;
 761
 762	hugetlb_add_hstate(shift - PAGE_SHIFT);
 763
 764	return 0;
 765}
 766
 767static int __init hugepage_setup_sz(char *str)
 768{
 769	unsigned long long size;
 
 770
 771	size = memparse(str, &str);
 772
 773	if (add_huge_page_size(size) != 0) {
 774		hugetlb_bad_size();
 775		pr_err("Invalid huge page size specified(%llu)\n", size);
 776	}
 777
 778	return 1;
 779}
 780__setup("hugepagesz=", hugepage_setup_sz);
 781
 782struct kmem_cache *hugepte_cache;
 783static int __init hugetlbpage_init(void)
 784{
 785	int psize;
 786
 787#if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
 788	if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
 789		return -ENODEV;
 790#endif
 791	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 792		unsigned shift;
 793		unsigned pdshift;
 794
 795		if (!mmu_psize_defs[psize].shift)
 796			continue;
 797
 798		shift = mmu_psize_to_shift(psize);
 799
 800		if (add_huge_page_size(1ULL << shift) < 0)
 
 801			continue;
 802
 803		if (shift < HUGEPD_PUD_SHIFT)
 
 
 
 
 
 
 804			pdshift = PMD_SHIFT;
 805		else if (shift < HUGEPD_PGD_SHIFT)
 806			pdshift = PUD_SHIFT;
 807		else
 808			pdshift = PGDIR_SHIFT;
 
 
 
 
 809		/*
 810		 * if we have pdshift and shift value same, we don't
 811		 * use pgt cache for hugepd.
 812		 */
 813		if (pdshift > shift)
 814			pgtable_cache_add(pdshift - shift, NULL);
 815#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 816		else if (!hugepte_cache) {
 817			/*
 818			 * Create a kmem cache for hugeptes.  The bottom bits in
 819			 * the pte have size information encoded in them, so
 820			 * align them to allow this
 821			 */
 822			hugepte_cache = kmem_cache_create("hugepte-cache",
 823							  sizeof(pte_t),
 824							  HUGEPD_SHIFT_MASK + 1,
 825							  0, NULL);
 826			if (hugepte_cache == NULL)
 827				panic("%s: Unable to create kmem cache "
 828				      "for hugeptes\n", __func__);
 829
 830		}
 831#endif
 832	}
 833
 834#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
 835	/* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
 836	if (mmu_psize_defs[MMU_PAGE_4M].shift)
 837		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
 838	else if (mmu_psize_defs[MMU_PAGE_512K].shift)
 839		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
 840#else
 841	/* Set default large page size. Currently, we pick 16M or 1M
 842	 * depending on what is available
 843	 */
 844	if (mmu_psize_defs[MMU_PAGE_16M].shift)
 845		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
 846	else if (mmu_psize_defs[MMU_PAGE_1M].shift)
 847		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
 848	else if (mmu_psize_defs[MMU_PAGE_2M].shift)
 849		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
 850#endif
 851	return 0;
 852}
 853
 854arch_initcall(hugetlbpage_init);
 855
 856void flush_dcache_icache_hugepage(struct page *page)
 857{
 858	int i;
 859	void *start;
 860
 861	BUG_ON(!PageCompound(page));
 862
 863	for (i = 0; i < (1UL << compound_order(page)); i++) {
 864		if (!PageHighMem(page)) {
 865			__flush_dcache_icache(page_address(page+i));
 866		} else {
 867			start = kmap_atomic(page+i);
 868			__flush_dcache_icache(start);
 869			kunmap_atomic(start);
 870		}
 871	}
 872}
 873
 874#endif /* CONFIG_HUGETLB_PAGE */
 875
 876/*
 877 * We have 4 cases for pgds and pmds:
 878 * (1) invalid (all zeroes)
 879 * (2) pointer to next table, as normal; bottom 6 bits == 0
 880 * (3) leaf pte for huge page _PAGE_PTE set
 881 * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
 882 *
 883 * So long as we atomically load page table pointers we are safe against teardown,
 884 * we can follow the address down to the the page and take a ref on it.
 885 * This function need to be called with interrupts disabled. We use this variant
 886 * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
 887 */
 888
 889pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
 890				   bool *is_thp, unsigned *shift)
 891{
 892	pgd_t pgd, *pgdp;
 893	pud_t pud, *pudp;
 894	pmd_t pmd, *pmdp;
 895	pte_t *ret_pte;
 896	hugepd_t *hpdp = NULL;
 897	unsigned pdshift = PGDIR_SHIFT;
 898
 899	if (shift)
 900		*shift = 0;
 901
 902	if (is_thp)
 903		*is_thp = false;
 904
 905	pgdp = pgdir + pgd_index(ea);
 906	pgd  = READ_ONCE(*pgdp);
 907	/*
 908	 * Always operate on the local stack value. This make sure the
 909	 * value don't get updated by a parallel THP split/collapse,
 910	 * page fault or a page unmap. The return pte_t * is still not
 911	 * stable. So should be checked there for above conditions.
 912	 */
 913	if (pgd_none(pgd))
 914		return NULL;
 915	else if (pgd_huge(pgd)) {
 916		ret_pte = (pte_t *) pgdp;
 917		goto out;
 918	} else if (is_hugepd(__hugepd(pgd_val(pgd))))
 919		hpdp = (hugepd_t *)&pgd;
 920	else {
 921		/*
 922		 * Even if we end up with an unmap, the pgtable will not
 923		 * be freed, because we do an rcu free and here we are
 924		 * irq disabled
 925		 */
 926		pdshift = PUD_SHIFT;
 927		pudp = pud_offset(&pgd, ea);
 928		pud  = READ_ONCE(*pudp);
 929
 930		if (pud_none(pud))
 931			return NULL;
 932		else if (pud_huge(pud)) {
 933			ret_pte = (pte_t *) pudp;
 934			goto out;
 935		} else if (is_hugepd(__hugepd(pud_val(pud))))
 936			hpdp = (hugepd_t *)&pud;
 937		else {
 938			pdshift = PMD_SHIFT;
 939			pmdp = pmd_offset(&pud, ea);
 940			pmd  = READ_ONCE(*pmdp);
 941			/*
 942			 * A hugepage collapse is captured by pmd_none, because
 943			 * it mark the pmd none and do a hpte invalidate.
 944			 */
 945			if (pmd_none(pmd))
 946				return NULL;
 947
 948			if (pmd_trans_huge(pmd)) {
 949				if (is_thp)
 950					*is_thp = true;
 951				ret_pte = (pte_t *) pmdp;
 952				goto out;
 953			}
 954
 955			if (pmd_huge(pmd)) {
 956				ret_pte = (pte_t *) pmdp;
 957				goto out;
 958			} else if (is_hugepd(__hugepd(pmd_val(pmd))))
 959				hpdp = (hugepd_t *)&pmd;
 960			else
 961				return pte_offset_kernel(&pmd, ea);
 962		}
 963	}
 964	if (!hpdp)
 965		return NULL;
 966
 967	ret_pte = hugepte_offset(*hpdp, ea, pdshift);
 968	pdshift = hugepd_shift(*hpdp);
 969out:
 970	if (shift)
 971		*shift = pdshift;
 972	return ret_pte;
 973}
 974EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
 975
 976int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
 977		unsigned long end, int write, struct page **pages, int *nr)
 978{
 979	unsigned long mask;
 980	unsigned long pte_end;
 981	struct page *head, *page;
 982	pte_t pte;
 983	int refs;
 984
 985	pte_end = (addr + sz) & ~(sz-1);
 986	if (pte_end < end)
 987		end = pte_end;
 988
 989	pte = READ_ONCE(*ptep);
 990	mask = _PAGE_PRESENT | _PAGE_READ;
 991
 992	/*
 993	 * On some CPUs like the 8xx, _PAGE_RW hence _PAGE_WRITE is defined
 994	 * as 0 and _PAGE_RO has to be set when a page is not writable
 995	 */
 996	if (write)
 997		mask |= _PAGE_WRITE;
 998	else
 999		mask |= _PAGE_RO;
1000
1001	if ((pte_val(pte) & mask) != mask)
1002		return 0;
1003
1004	/* hugepages are never "special" */
1005	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1006
1007	refs = 0;
1008	head = pte_page(pte);
1009
1010	page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1011	do {
1012		VM_BUG_ON(compound_head(page) != head);
1013		pages[*nr] = page;
1014		(*nr)++;
1015		page++;
1016		refs++;
1017	} while (addr += PAGE_SIZE, addr != end);
1018
1019	if (!page_cache_add_speculative(head, refs)) {
1020		*nr -= refs;
1021		return 0;
1022	}
1023
1024	if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1025		/* Could be optimized better */
1026		*nr -= refs;
1027		while (refs--)
1028			put_page(head);
1029		return 0;
1030	}
1031
1032	return 1;
1033}