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);
57 } else if (IS_ENABLED(CONFIG_PPC_8xx)) {
60 new = pte_alloc_one(mm);
62 cachep = PGT_CACHE(pdshift - pshift);
67 if (!cachep && !new) {
68 WARN_ONCE(1, "No page table cache created for hugetlb tables");
73 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
75 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
76 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
82 * Make sure other cpus find the hugepd set only after a
83 * properly initialized page table is visible to them.
84 * For more details look for comment in __pte_alloc().
90 * We have multiple higher-level entries that point to the same
91 * actual pte location. Fill in each as we go and backtrack on error.
92 * We need all of these so the DTLB pgtable walk code can find the
93 * right higher-level entry without knowing if it's a hugepage or not.
95 for (i = 0; i < num_hugepd; i++, hpdp++) {
96 if (unlikely(!hugepd_none(*hpdp)))
98 hugepd_populate(hpdp, new, pshift);
100 /* If we bailed from the for loop early, an error occurred, clean up */
101 if (i < num_hugepd) {
102 for (i = i - 1 ; i >= 0; i--, hpdp--)
105 kmem_cache_free(cachep, new);
109 kmemleak_ignore(new);
116 * At this point we do the placement change only for BOOK3S 64. This would
117 * possibly work on other subarchs.
119 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
124 hugepd_t *hpdp = NULL;
125 unsigned pshift = __ffs(sz);
126 unsigned pdshift = PGDIR_SHIFT;
130 pg = pgd_offset(mm, addr);
132 #ifdef CONFIG_PPC_BOOK3S_64
133 if (pshift == PGDIR_SHIFT)
136 else if (pshift > PUD_SHIFT) {
138 * We need to use hugepd table
140 ptl = &mm->page_table_lock;
141 hpdp = (hugepd_t *)pg;
144 pu = pud_alloc(mm, pg, addr);
147 if (pshift == PUD_SHIFT)
149 else if (pshift > PMD_SHIFT) {
150 ptl = pud_lockptr(mm, pu);
151 hpdp = (hugepd_t *)pu;
154 pm = pmd_alloc(mm, pu, addr);
157 if (pshift == PMD_SHIFT)
161 ptl = pmd_lockptr(mm, pm);
162 hpdp = (hugepd_t *)pm;
167 if (pshift >= PGDIR_SHIFT) {
168 ptl = &mm->page_table_lock;
169 hpdp = (hugepd_t *)pg;
172 pu = pud_alloc(mm, pg, addr);
175 if (pshift >= PUD_SHIFT) {
176 ptl = pud_lockptr(mm, pu);
177 hpdp = (hugepd_t *)pu;
180 pm = pmd_alloc(mm, pu, addr);
183 ptl = pmd_lockptr(mm, pm);
184 hpdp = (hugepd_t *)pm;
191 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
193 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
194 pdshift, pshift, ptl))
197 return hugepte_offset(*hpdp, addr, pdshift);
200 #ifdef CONFIG_PPC_BOOK3S_64
202 * Tracks gpages after the device tree is scanned and before the
203 * huge_boot_pages list is ready on pseries.
205 #define MAX_NUMBER_GPAGES 1024
206 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
207 __initdata static unsigned nr_gpages;
210 * Build list of addresses of gigantic pages. This function is used in early
211 * boot before the buddy allocator is setup.
213 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
217 while (number_of_pages > 0) {
218 gpage_freearray[nr_gpages] = addr;
225 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
227 struct huge_bootmem_page *m;
230 m = phys_to_virt(gpage_freearray[--nr_gpages]);
231 gpage_freearray[nr_gpages] = 0;
232 list_add(&m->list, &huge_boot_pages);
239 int __init alloc_bootmem_huge_page(struct hstate *h)
242 #ifdef CONFIG_PPC_BOOK3S_64
243 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
244 return pseries_alloc_bootmem_huge_page(h);
246 return __alloc_bootmem_huge_page(h);
249 #ifndef CONFIG_PPC_BOOK3S_64
250 #define HUGEPD_FREELIST_SIZE \
251 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
253 struct hugepd_freelist {
259 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
261 static void hugepd_free_rcu_callback(struct rcu_head *head)
263 struct hugepd_freelist *batch =
264 container_of(head, struct hugepd_freelist, rcu);
267 for (i = 0; i < batch->index; i++)
268 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
270 free_page((unsigned long)batch);
273 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
275 struct hugepd_freelist **batchp;
277 batchp = &get_cpu_var(hugepd_freelist_cur);
279 if (atomic_read(&tlb->mm->mm_users) < 2 ||
280 mm_is_thread_local(tlb->mm)) {
281 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
282 put_cpu_var(hugepd_freelist_cur);
286 if (*batchp == NULL) {
287 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
288 (*batchp)->index = 0;
291 (*batchp)->ptes[(*batchp)->index++] = hugepte;
292 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
293 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
296 put_cpu_var(hugepd_freelist_cur);
299 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
302 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
303 unsigned long start, unsigned long end,
304 unsigned long floor, unsigned long ceiling)
306 pte_t *hugepte = hugepd_page(*hpdp);
309 unsigned long pdmask = ~((1UL << pdshift) - 1);
310 unsigned int num_hugepd = 1;
311 unsigned int shift = hugepd_shift(*hpdp);
313 /* Note: On fsl the hpdp may be the first of several */
315 num_hugepd = 1 << (shift - pdshift);
325 if (end - 1 > ceiling - 1)
328 for (i = 0; i < num_hugepd; i++, hpdp++)
331 if (shift >= pdshift)
332 hugepd_free(tlb, hugepte);
333 else if (IS_ENABLED(CONFIG_PPC_8xx))
334 pgtable_free_tlb(tlb, hugepte, 0);
336 pgtable_free_tlb(tlb, hugepte,
337 get_hugepd_cache_index(pdshift - shift));
340 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
341 unsigned long addr, unsigned long end,
342 unsigned long floor, unsigned long ceiling)
352 pmd = pmd_offset(pud, addr);
353 next = pmd_addr_end(addr, end);
354 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
356 * if it is not hugepd pointer, we should already find
359 WARN_ON(!pmd_none_or_clear_bad(pmd));
363 * Increment next by the size of the huge mapping since
364 * there may be more than one entry at this level for a
365 * single hugepage, but all of them point to
366 * the same kmem cache that holds the hugepte.
368 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
372 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
373 addr, next, floor, ceiling);
374 } while (addr = next, addr != end);
384 if (end - 1 > ceiling - 1)
387 pmd = pmd_offset(pud, start);
389 pmd_free_tlb(tlb, pmd, start);
390 mm_dec_nr_pmds(tlb->mm);
393 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
394 unsigned long addr, unsigned long end,
395 unsigned long floor, unsigned long ceiling)
403 pud = pud_offset(pgd, addr);
404 next = pud_addr_end(addr, end);
405 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
406 if (pud_none_or_clear_bad(pud))
408 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
413 * Increment next by the size of the huge mapping since
414 * there may be more than one entry at this level for a
415 * single hugepage, but all of them point to
416 * the same kmem cache that holds the hugepte.
418 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
422 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
423 addr, next, floor, ceiling);
425 } while (addr = next, addr != end);
431 ceiling &= PGDIR_MASK;
435 if (end - 1 > ceiling - 1)
438 pud = pud_offset(pgd, start);
440 pud_free_tlb(tlb, pud, start);
441 mm_dec_nr_puds(tlb->mm);
445 * This function frees user-level page tables of a process.
447 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
448 unsigned long addr, unsigned long end,
449 unsigned long floor, unsigned long ceiling)
455 * Because there are a number of different possible pagetable
456 * layouts for hugepage ranges, we limit knowledge of how
457 * things should be laid out to the allocation path
458 * (huge_pte_alloc(), above). Everything else works out the
459 * structure as it goes from information in the hugepd
460 * pointers. That means that we can't here use the
461 * optimization used in the normal page free_pgd_range(), of
462 * checking whether we're actually covering a large enough
463 * range to have to do anything at the top level of the walk
464 * instead of at the bottom.
466 * To make sense of this, you should probably go read the big
467 * block comment at the top of the normal free_pgd_range(),
472 next = pgd_addr_end(addr, end);
473 pgd = pgd_offset(tlb->mm, addr);
474 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
475 if (pgd_none_or_clear_bad(pgd))
477 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
481 * Increment next by the size of the huge mapping since
482 * there may be more than one entry at the pgd level
483 * for a single hugepage, but all of them point to the
484 * same kmem cache that holds the hugepte.
486 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
490 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
491 addr, next, floor, ceiling);
493 } while (addr = next, addr != end);
496 struct page *follow_huge_pd(struct vm_area_struct *vma,
497 unsigned long address, hugepd_t hpd,
498 int flags, int pdshift)
502 struct page *page = NULL;
504 int shift = hugepd_shift(hpd);
505 struct mm_struct *mm = vma->vm_mm;
509 * hugepage directory entries are protected by mm->page_table_lock
510 * Use this instead of huge_pte_lockptr
512 ptl = &mm->page_table_lock;
515 ptep = hugepte_offset(hpd, address, pdshift);
516 if (pte_present(*ptep)) {
517 mask = (1UL << shift) - 1;
518 page = pte_page(*ptep);
519 page += ((address & mask) >> PAGE_SHIFT);
520 if (flags & FOLL_GET)
523 if (is_hugetlb_entry_migration(*ptep)) {
525 __migration_entry_wait(mm, ptep, ptl);
533 #ifdef CONFIG_PPC_MM_SLICES
534 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
535 unsigned long len, unsigned long pgoff,
538 struct hstate *hstate = hstate_file(file);
539 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
541 #ifdef CONFIG_PPC_RADIX_MMU
543 return radix__hugetlb_get_unmapped_area(file, addr, len,
546 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
550 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
552 /* With radix we don't use slice, so derive it from vma*/
553 if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
554 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
556 return 1UL << mmu_psize_to_shift(psize);
558 return vma_kernel_pagesize(vma);
561 static int __init add_huge_page_size(unsigned long long size)
563 int shift = __ffs(size);
566 /* Check that it is a page size supported by the hardware and
567 * that it fits within pagetable and slice limits. */
568 if (size <= PAGE_SIZE || !is_power_of_2(size))
571 mmu_psize = check_and_get_huge_psize(shift);
575 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
577 /* Return if huge page size has already been setup */
578 if (size_to_hstate(size))
581 hugetlb_add_hstate(shift - PAGE_SHIFT);
586 static int __init hugepage_setup_sz(char *str)
588 unsigned long long size;
590 size = memparse(str, &str);
592 if (add_huge_page_size(size) != 0) {
594 pr_err("Invalid huge page size specified(%llu)\n", size);
599 __setup("hugepagesz=", hugepage_setup_sz);
601 static int __init hugetlbpage_init(void)
603 bool configured = false;
606 if (hugetlb_disabled) {
607 pr_info("HugeTLB support is disabled!\n");
611 if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
612 !mmu_has_feature(MMU_FTR_16M_PAGE))
615 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
619 if (!mmu_psize_defs[psize].shift)
622 shift = mmu_psize_to_shift(psize);
624 #ifdef CONFIG_PPC_BOOK3S_64
625 if (shift > PGDIR_SHIFT)
627 else if (shift > PUD_SHIFT)
628 pdshift = PGDIR_SHIFT;
629 else if (shift > PMD_SHIFT)
634 if (shift < PUD_SHIFT)
636 else if (shift < PGDIR_SHIFT)
639 pdshift = PGDIR_SHIFT;
642 if (add_huge_page_size(1ULL << shift) < 0)
645 * if we have pdshift and shift value same, we don't
646 * use pgt cache for hugepd.
648 if (pdshift > shift) {
649 if (!IS_ENABLED(CONFIG_PPC_8xx))
650 pgtable_cache_add(pdshift - shift);
651 } else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) ||
652 IS_ENABLED(CONFIG_PPC_8xx)) {
653 pgtable_cache_add(PTE_T_ORDER);
660 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
661 hugetlbpage_init_default();
663 pr_info("Failed to initialize. Disabling HugeTLB");
668 arch_initcall(hugetlbpage_init);
670 void flush_dcache_icache_hugepage(struct page *page)
675 BUG_ON(!PageCompound(page));
677 for (i = 0; i < compound_nr(page); i++) {
678 if (!PageHighMem(page)) {
679 __flush_dcache_icache(page_address(page+i));
681 start = kmap_atomic(page+i);
682 __flush_dcache_icache(start);
683 kunmap_atomic(start);