2 * PPC Huge TLB Page Support for Kernel.
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
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>
25 #include <asm/setup.h>
26 #include <asm/hugetlb.h>
27 #include <asm/pte-walk.h>
29 bool hugetlb_disabled = false;
31 #define hugepd_none(hpd) (hpd_val(hpd) == 0)
33 #define PTE_T_ORDER (__builtin_ffs(sizeof(pte_t)) - __builtin_ffs(sizeof(void *)))
35 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
38 * Only called for hugetlbfs pages, hence can ignore THP and the
41 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
44 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
45 unsigned long address, unsigned int pdshift,
46 unsigned int pshift, spinlock_t *ptl)
48 struct kmem_cache *cachep;
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);
60 cachep = PGT_CACHE(pdshift - pshift);
64 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
66 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
67 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
73 * Make sure other cpus find the hugepd set only after a
74 * properly initialized page table is visible to them.
75 * For more details look for comment in __pte_alloc().
81 * We have multiple higher-level entries that point to the same
82 * actual pte location. Fill in each as we go and backtrack on error.
83 * We need all of these so the DTLB pgtable walk code can find the
84 * right higher-level entry without knowing if it's a hugepage or not.
86 for (i = 0; i < num_hugepd; i++, hpdp++) {
87 if (unlikely(!hugepd_none(*hpdp)))
89 hugepd_populate(hpdp, new, pshift);
91 /* If we bailed from the for loop early, an error occurred, clean up */
93 for (i = i - 1 ; i >= 0; i--, hpdp--)
95 kmem_cache_free(cachep, new);
104 * At this point we do the placement change only for BOOK3S 64. This would
105 * possibly work on other subarchs.
107 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
112 hugepd_t *hpdp = NULL;
113 unsigned pshift = __ffs(sz);
114 unsigned pdshift = PGDIR_SHIFT;
118 pg = pgd_offset(mm, addr);
120 #ifdef CONFIG_PPC_BOOK3S_64
121 if (pshift == PGDIR_SHIFT)
124 else if (pshift > PUD_SHIFT) {
126 * We need to use hugepd table
128 ptl = &mm->page_table_lock;
129 hpdp = (hugepd_t *)pg;
132 pu = pud_alloc(mm, pg, addr);
133 if (pshift == PUD_SHIFT)
135 else if (pshift > PMD_SHIFT) {
136 ptl = pud_lockptr(mm, pu);
137 hpdp = (hugepd_t *)pu;
140 pm = pmd_alloc(mm, pu, addr);
141 if (pshift == PMD_SHIFT)
145 ptl = pmd_lockptr(mm, pm);
146 hpdp = (hugepd_t *)pm;
151 if (pshift >= PGDIR_SHIFT) {
152 ptl = &mm->page_table_lock;
153 hpdp = (hugepd_t *)pg;
156 pu = pud_alloc(mm, pg, addr);
157 if (pshift >= PUD_SHIFT) {
158 ptl = pud_lockptr(mm, pu);
159 hpdp = (hugepd_t *)pu;
162 pm = pmd_alloc(mm, pu, addr);
163 ptl = pmd_lockptr(mm, pm);
164 hpdp = (hugepd_t *)pm;
171 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
173 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
174 pdshift, pshift, ptl))
177 return hugepte_offset(*hpdp, addr, pdshift);
180 #ifdef CONFIG_PPC_BOOK3S_64
182 * Tracks gpages after the device tree is scanned and before the
183 * huge_boot_pages list is ready on pseries.
185 #define MAX_NUMBER_GPAGES 1024
186 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
187 __initdata static unsigned nr_gpages;
190 * Build list of addresses of gigantic pages. This function is used in early
191 * boot before the buddy allocator is setup.
193 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
197 while (number_of_pages > 0) {
198 gpage_freearray[nr_gpages] = addr;
205 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
207 struct huge_bootmem_page *m;
210 m = phys_to_virt(gpage_freearray[--nr_gpages]);
211 gpage_freearray[nr_gpages] = 0;
212 list_add(&m->list, &huge_boot_pages);
219 int __init alloc_bootmem_huge_page(struct hstate *h)
222 #ifdef CONFIG_PPC_BOOK3S_64
223 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
224 return pseries_alloc_bootmem_huge_page(h);
226 return __alloc_bootmem_huge_page(h);
229 #ifndef CONFIG_PPC_BOOK3S_64
230 #define HUGEPD_FREELIST_SIZE \
231 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
233 struct hugepd_freelist {
239 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
241 static void hugepd_free_rcu_callback(struct rcu_head *head)
243 struct hugepd_freelist *batch =
244 container_of(head, struct hugepd_freelist, rcu);
247 for (i = 0; i < batch->index; i++)
248 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
250 free_page((unsigned long)batch);
253 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
255 struct hugepd_freelist **batchp;
257 batchp = &get_cpu_var(hugepd_freelist_cur);
259 if (atomic_read(&tlb->mm->mm_users) < 2 ||
260 mm_is_thread_local(tlb->mm)) {
261 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
262 put_cpu_var(hugepd_freelist_cur);
266 if (*batchp == NULL) {
267 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
268 (*batchp)->index = 0;
271 (*batchp)->ptes[(*batchp)->index++] = hugepte;
272 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
273 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
276 put_cpu_var(hugepd_freelist_cur);
279 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
282 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
283 unsigned long start, unsigned long end,
284 unsigned long floor, unsigned long ceiling)
286 pte_t *hugepte = hugepd_page(*hpdp);
289 unsigned long pdmask = ~((1UL << pdshift) - 1);
290 unsigned int num_hugepd = 1;
291 unsigned int shift = hugepd_shift(*hpdp);
293 /* Note: On fsl the hpdp may be the first of several */
295 num_hugepd = 1 << (shift - pdshift);
305 if (end - 1 > ceiling - 1)
308 for (i = 0; i < num_hugepd; i++, hpdp++)
311 if (shift >= pdshift)
312 hugepd_free(tlb, hugepte);
313 else if (IS_ENABLED(CONFIG_PPC_8xx))
314 pgtable_free_tlb(tlb, hugepte,
315 get_hugepd_cache_index(PTE_INDEX_SIZE));
317 pgtable_free_tlb(tlb, hugepte,
318 get_hugepd_cache_index(pdshift - shift));
321 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
322 unsigned long addr, unsigned long end,
323 unsigned long floor, unsigned long ceiling)
333 pmd = pmd_offset(pud, addr);
334 next = pmd_addr_end(addr, end);
335 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
337 * if it is not hugepd pointer, we should already find
340 WARN_ON(!pmd_none_or_clear_bad(pmd));
344 * Increment next by the size of the huge mapping since
345 * there may be more than one entry at this level for a
346 * single hugepage, but all of them point to
347 * the same kmem cache that holds the hugepte.
349 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
353 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
354 addr, next, floor, ceiling);
355 } while (addr = next, addr != end);
365 if (end - 1 > ceiling - 1)
368 pmd = pmd_offset(pud, start);
370 pmd_free_tlb(tlb, pmd, start);
371 mm_dec_nr_pmds(tlb->mm);
374 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
375 unsigned long addr, unsigned long end,
376 unsigned long floor, unsigned long ceiling)
384 pud = pud_offset(pgd, addr);
385 next = pud_addr_end(addr, end);
386 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
387 if (pud_none_or_clear_bad(pud))
389 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
394 * Increment next by the size of the huge mapping since
395 * there may be more than one entry at this level for a
396 * single hugepage, but all of them point to
397 * the same kmem cache that holds the hugepte.
399 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
403 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
404 addr, next, floor, ceiling);
406 } while (addr = next, addr != end);
412 ceiling &= PGDIR_MASK;
416 if (end - 1 > ceiling - 1)
419 pud = pud_offset(pgd, start);
421 pud_free_tlb(tlb, pud, start);
422 mm_dec_nr_puds(tlb->mm);
426 * This function frees user-level page tables of a process.
428 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
429 unsigned long addr, unsigned long end,
430 unsigned long floor, unsigned long ceiling)
436 * Because there are a number of different possible pagetable
437 * layouts for hugepage ranges, we limit knowledge of how
438 * things should be laid out to the allocation path
439 * (huge_pte_alloc(), above). Everything else works out the
440 * structure as it goes from information in the hugepd
441 * pointers. That means that we can't here use the
442 * optimization used in the normal page free_pgd_range(), of
443 * checking whether we're actually covering a large enough
444 * range to have to do anything at the top level of the walk
445 * instead of at the bottom.
447 * To make sense of this, you should probably go read the big
448 * block comment at the top of the normal free_pgd_range(),
453 next = pgd_addr_end(addr, end);
454 pgd = pgd_offset(tlb->mm, addr);
455 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
456 if (pgd_none_or_clear_bad(pgd))
458 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
462 * Increment next by the size of the huge mapping since
463 * there may be more than one entry at the pgd level
464 * for a single hugepage, but all of them point to the
465 * same kmem cache that holds the hugepte.
467 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
471 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
472 addr, next, floor, ceiling);
474 } while (addr = next, addr != end);
477 struct page *follow_huge_pd(struct vm_area_struct *vma,
478 unsigned long address, hugepd_t hpd,
479 int flags, int pdshift)
483 struct page *page = NULL;
485 int shift = hugepd_shift(hpd);
486 struct mm_struct *mm = vma->vm_mm;
490 * hugepage directory entries are protected by mm->page_table_lock
491 * Use this instead of huge_pte_lockptr
493 ptl = &mm->page_table_lock;
496 ptep = hugepte_offset(hpd, address, pdshift);
497 if (pte_present(*ptep)) {
498 mask = (1UL << shift) - 1;
499 page = pte_page(*ptep);
500 page += ((address & mask) >> PAGE_SHIFT);
501 if (flags & FOLL_GET)
504 if (is_hugetlb_entry_migration(*ptep)) {
506 __migration_entry_wait(mm, ptep, ptl);
514 #ifdef CONFIG_PPC_MM_SLICES
515 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
516 unsigned long len, unsigned long pgoff,
519 struct hstate *hstate = hstate_file(file);
520 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
522 #ifdef CONFIG_PPC_RADIX_MMU
524 return radix__hugetlb_get_unmapped_area(file, addr, len,
527 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
531 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
533 /* With radix we don't use slice, so derive it from vma*/
534 if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
535 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
537 return 1UL << mmu_psize_to_shift(psize);
539 return vma_kernel_pagesize(vma);
542 static int __init add_huge_page_size(unsigned long long size)
544 int shift = __ffs(size);
547 /* Check that it is a page size supported by the hardware and
548 * that it fits within pagetable and slice limits. */
549 if (size <= PAGE_SIZE || !is_power_of_2(size))
552 mmu_psize = check_and_get_huge_psize(shift);
556 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
558 /* Return if huge page size has already been setup */
559 if (size_to_hstate(size))
562 hugetlb_add_hstate(shift - PAGE_SHIFT);
567 static int __init hugepage_setup_sz(char *str)
569 unsigned long long size;
571 size = memparse(str, &str);
573 if (add_huge_page_size(size) != 0) {
575 pr_err("Invalid huge page size specified(%llu)\n", size);
580 __setup("hugepagesz=", hugepage_setup_sz);
582 static int __init hugetlbpage_init(void)
586 if (hugetlb_disabled) {
587 pr_info("HugeTLB support is disabled!\n");
591 if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
592 !mmu_has_feature(MMU_FTR_16M_PAGE))
595 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
599 if (!mmu_psize_defs[psize].shift)
602 shift = mmu_psize_to_shift(psize);
604 #ifdef CONFIG_PPC_BOOK3S_64
605 if (shift > PGDIR_SHIFT)
607 else if (shift > PUD_SHIFT)
608 pdshift = PGDIR_SHIFT;
609 else if (shift > PMD_SHIFT)
614 if (shift < PUD_SHIFT)
616 else if (shift < PGDIR_SHIFT)
619 pdshift = PGDIR_SHIFT;
622 if (add_huge_page_size(1ULL << shift) < 0)
625 * if we have pdshift and shift value same, we don't
626 * use pgt cache for hugepd.
628 if (pdshift > shift && IS_ENABLED(CONFIG_PPC_8xx))
629 pgtable_cache_add(PTE_INDEX_SIZE);
630 else if (pdshift > shift)
631 pgtable_cache_add(pdshift - shift);
632 else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) || IS_ENABLED(CONFIG_PPC_8xx))
633 pgtable_cache_add(PTE_T_ORDER);
636 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
637 hugetlbpage_init_default();
642 arch_initcall(hugetlbpage_init);
644 void flush_dcache_icache_hugepage(struct page *page)
649 BUG_ON(!PageCompound(page));
651 for (i = 0; i < (1UL << compound_order(page)); i++) {
652 if (!PageHighMem(page)) {
653 __flush_dcache_icache(page_address(page+i));
655 start = kmap_atomic(page+i);
656 __flush_dcache_icache(start);
657 kunmap_atomic(start);