2 * PPC64 (POWER4) Huge TLB Page Support for Kernel.
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
6 * Based on the IA-32 version:
7 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
10 #include <linux/init.h>
13 #include <linux/hugetlb.h>
14 #include <linux/pagemap.h>
15 #include <linux/slab.h>
16 #include <linux/err.h>
17 #include <linux/sysctl.h>
19 #include <asm/pgalloc.h>
21 #include <asm/tlbflush.h>
22 #include <asm/mmu_context.h>
23 #include <asm/machdep.h>
24 #include <asm/cputable.h>
27 #define HPAGE_SHIFT_64K 16
28 #define HPAGE_SHIFT_16M 24
30 #define NUM_LOW_AREAS (0x100000000UL >> SID_SHIFT)
31 #define NUM_HIGH_AREAS (PGTABLE_RANGE >> HTLB_AREA_SHIFT)
33 unsigned int hugepte_shift;
34 #define PTRS_PER_HUGEPTE (1 << hugepte_shift)
35 #define HUGEPTE_TABLE_SIZE (sizeof(pte_t) << hugepte_shift)
37 #define HUGEPD_SHIFT (HPAGE_SHIFT + hugepte_shift)
38 #define HUGEPD_SIZE (1UL << HUGEPD_SHIFT)
39 #define HUGEPD_MASK (~(HUGEPD_SIZE-1))
41 #define huge_pgtable_cache (pgtable_cache[HUGEPTE_CACHE_NUM])
43 /* Flag to mark huge PD pointers. This means pmd_bad() and pud_bad()
44 * will choke on pointers to hugepte tables, which is handy for
45 * catching screwups early. */
48 typedef struct { unsigned long pd; } hugepd_t;
50 #define hugepd_none(hpd) ((hpd).pd == 0)
52 static inline pte_t *hugepd_page(hugepd_t hpd)
54 BUG_ON(!(hpd.pd & HUGEPD_OK));
55 return (pte_t *)(hpd.pd & ~HUGEPD_OK);
58 static inline pte_t *hugepte_offset(hugepd_t *hpdp, unsigned long addr)
60 unsigned long idx = ((addr >> HPAGE_SHIFT) & (PTRS_PER_HUGEPTE-1));
61 pte_t *dir = hugepd_page(*hpdp);
66 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
67 unsigned long address)
69 pte_t *new = kmem_cache_alloc(huge_pgtable_cache,
70 GFP_KERNEL|__GFP_REPEAT);
75 spin_lock(&mm->page_table_lock);
76 if (!hugepd_none(*hpdp))
77 kmem_cache_free(huge_pgtable_cache, new);
79 hpdp->pd = (unsigned long)new | HUGEPD_OK;
80 spin_unlock(&mm->page_table_lock);
84 /* Base page size affects how we walk hugetlb page tables */
85 #ifdef CONFIG_PPC_64K_PAGES
86 #define hpmd_offset(pud, addr) pmd_offset(pud, addr)
87 #define hpmd_alloc(mm, pud, addr) pmd_alloc(mm, pud, addr)
90 pmd_t *hpmd_offset(pud_t *pud, unsigned long addr)
92 if (HPAGE_SHIFT == HPAGE_SHIFT_64K)
93 return pmd_offset(pud, addr);
98 pmd_t *hpmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long addr)
100 if (HPAGE_SHIFT == HPAGE_SHIFT_64K)
101 return pmd_alloc(mm, pud, addr);
103 return (pmd_t *) pud;
107 /* Modelled after find_linux_pte() */
108 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
114 BUG_ON(get_slice_psize(mm, addr) != mmu_huge_psize);
118 pg = pgd_offset(mm, addr);
119 if (!pgd_none(*pg)) {
120 pu = pud_offset(pg, addr);
121 if (!pud_none(*pu)) {
122 pm = hpmd_offset(pu, addr);
124 return hugepte_offset((hugepd_t *)pm, addr);
131 pte_t *huge_pte_alloc(struct mm_struct *mm,
132 unsigned long addr, unsigned long sz)
137 hugepd_t *hpdp = NULL;
139 BUG_ON(get_slice_psize(mm, addr) != mmu_huge_psize);
143 pg = pgd_offset(mm, addr);
144 pu = pud_alloc(mm, pg, addr);
147 pm = hpmd_alloc(mm, pu, addr);
149 hpdp = (hugepd_t *)pm;
155 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr))
158 return hugepte_offset(hpdp, addr);
161 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
166 static void free_hugepte_range(struct mmu_gather *tlb, hugepd_t *hpdp)
168 pte_t *hugepte = hugepd_page(*hpdp);
172 pgtable_free_tlb(tlb, pgtable_free_cache(hugepte, HUGEPTE_CACHE_NUM,
176 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
177 unsigned long addr, unsigned long end,
178 unsigned long floor, unsigned long ceiling)
185 pmd = pmd_offset(pud, addr);
187 next = pmd_addr_end(addr, end);
190 free_hugepte_range(tlb, (hugepd_t *)pmd);
191 } while (pmd++, addr = next, addr != end);
201 if (end - 1 > ceiling - 1)
204 pmd = pmd_offset(pud, start);
206 pmd_free_tlb(tlb, pmd);
209 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
210 unsigned long addr, unsigned long end,
211 unsigned long floor, unsigned long ceiling)
218 pud = pud_offset(pgd, addr);
220 next = pud_addr_end(addr, end);
221 #ifdef CONFIG_PPC_64K_PAGES
222 if (pud_none_or_clear_bad(pud))
224 hugetlb_free_pmd_range(tlb, pud, addr, next, floor, ceiling);
226 if (HPAGE_SHIFT == HPAGE_SHIFT_64K) {
227 if (pud_none_or_clear_bad(pud))
229 hugetlb_free_pmd_range(tlb, pud, addr, next, floor, ceiling);
233 free_hugepte_range(tlb, (hugepd_t *)pud);
236 } while (pud++, addr = next, addr != end);
242 ceiling &= PGDIR_MASK;
246 if (end - 1 > ceiling - 1)
249 pud = pud_offset(pgd, start);
251 pud_free_tlb(tlb, pud);
255 * This function frees user-level page tables of a process.
257 * Must be called with pagetable lock held.
259 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
260 unsigned long addr, unsigned long end,
261 unsigned long floor, unsigned long ceiling)
268 * Comments below take from the normal free_pgd_range(). They
269 * apply here too. The tests against HUGEPD_MASK below are
270 * essential, because we *don't* test for this at the bottom
271 * level. Without them we'll attempt to free a hugepte table
272 * when we unmap just part of it, even if there are other
273 * active mappings using it.
275 * The next few lines have given us lots of grief...
277 * Why are we testing HUGEPD* at this top level? Because
278 * often there will be no work to do at all, and we'd prefer
279 * not to go all the way down to the bottom just to discover
282 * Why all these "- 1"s? Because 0 represents both the bottom
283 * of the address space and the top of it (using -1 for the
284 * top wouldn't help much: the masks would do the wrong thing).
285 * The rule is that addr 0 and floor 0 refer to the bottom of
286 * the address space, but end 0 and ceiling 0 refer to the top
287 * Comparisons need to use "end - 1" and "ceiling - 1" (though
288 * that end 0 case should be mythical).
290 * Wherever addr is brought up or ceiling brought down, we
291 * must be careful to reject "the opposite 0" before it
292 * confuses the subsequent tests. But what about where end is
293 * brought down by HUGEPD_SIZE below? no, end can't go down to
296 * Whereas we round start (addr) and ceiling down, by different
297 * masks at different levels, in order to test whether a table
298 * now has no other vmas using it, so can be freed, we don't
299 * bother to round floor or end up - the tests don't need that.
309 ceiling &= HUGEPD_MASK;
313 if (end - 1 > ceiling - 1)
319 pgd = pgd_offset(tlb->mm, addr);
321 BUG_ON(get_slice_psize(tlb->mm, addr) != mmu_huge_psize);
322 next = pgd_addr_end(addr, end);
323 if (pgd_none_or_clear_bad(pgd))
325 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
326 } while (pgd++, addr = next, addr != end);
329 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
330 pte_t *ptep, pte_t pte)
332 if (pte_present(*ptep)) {
333 /* We open-code pte_clear because we need to pass the right
334 * argument to hpte_need_flush (huge / !huge). Might not be
335 * necessary anymore if we make hpte_need_flush() get the
336 * page size from the slices
338 pte_update(mm, addr & HPAGE_MASK, ptep, ~0UL, 1);
340 *ptep = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
343 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
346 unsigned long old = pte_update(mm, addr, ptep, ~0UL, 1);
351 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
356 if (get_slice_psize(mm, address) != mmu_huge_psize)
357 return ERR_PTR(-EINVAL);
359 ptep = huge_pte_offset(mm, address);
360 page = pte_page(*ptep);
362 page += (address % HPAGE_SIZE) / PAGE_SIZE;
367 int pmd_huge(pmd_t pmd)
373 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
374 pmd_t *pmd, int write)
381 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
382 unsigned long len, unsigned long pgoff,
385 return slice_get_unmapped_area(addr, len, flags,
386 mmu_huge_psize, 1, 0);
390 * Called by asm hashtable.S for doing lazy icache flush
392 static unsigned int hash_huge_page_do_lazy_icache(unsigned long rflags,
398 if (!pfn_valid(pte_pfn(pte)))
401 page = pte_page(pte);
404 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
406 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++)
407 __flush_dcache_icache(page_address(page+i));
408 set_bit(PG_arch_1, &page->flags);
416 int hash_huge_page(struct mm_struct *mm, unsigned long access,
417 unsigned long ea, unsigned long vsid, int local,
421 unsigned long old_pte, new_pte;
422 unsigned long va, rflags, pa;
425 int ssize = user_segment_size(ea);
427 ptep = huge_pte_offset(mm, ea);
429 /* Search the Linux page table for a match with va */
430 va = hpt_va(ea, vsid, ssize);
433 * If no pte found or not present, send the problem up to
436 if (unlikely(!ptep || pte_none(*ptep)))
440 * Check the user's access rights to the page. If access should be
441 * prevented then send the problem up to do_page_fault.
443 if (unlikely(access & ~pte_val(*ptep)))
446 * At this point, we have a pte (old_pte) which can be used to build
447 * or update an HPTE. There are 2 cases:
449 * 1. There is a valid (present) pte with no associated HPTE (this is
450 * the most common case)
451 * 2. There is a valid (present) pte with an associated HPTE. The
452 * current values of the pp bits in the HPTE prevent access
453 * because we are doing software DIRTY bit management and the
454 * page is currently not DIRTY.
459 old_pte = pte_val(*ptep);
460 if (old_pte & _PAGE_BUSY)
462 new_pte = old_pte | _PAGE_BUSY | _PAGE_ACCESSED;
463 } while(old_pte != __cmpxchg_u64((unsigned long *)ptep,
466 rflags = 0x2 | (!(new_pte & _PAGE_RW));
467 /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
468 rflags |= ((new_pte & _PAGE_EXEC) ? 0 : HPTE_R_N);
469 if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
470 /* No CPU has hugepages but lacks no execute, so we
471 * don't need to worry about that case */
472 rflags = hash_huge_page_do_lazy_icache(rflags, __pte(old_pte),
475 /* Check if pte already has an hpte (case 2) */
476 if (unlikely(old_pte & _PAGE_HASHPTE)) {
477 /* There MIGHT be an HPTE for this pte */
478 unsigned long hash, slot;
480 hash = hpt_hash(va, HPAGE_SHIFT, ssize);
481 if (old_pte & _PAGE_F_SECOND)
483 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
484 slot += (old_pte & _PAGE_F_GIX) >> 12;
486 if (ppc_md.hpte_updatepp(slot, rflags, va, mmu_huge_psize,
488 old_pte &= ~_PAGE_HPTEFLAGS;
491 if (likely(!(old_pte & _PAGE_HASHPTE))) {
492 unsigned long hash = hpt_hash(va, HPAGE_SHIFT, ssize);
493 unsigned long hpte_group;
495 pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
498 hpte_group = ((hash & htab_hash_mask) *
499 HPTES_PER_GROUP) & ~0x7UL;
501 /* clear HPTE slot informations in new PTE */
502 #ifdef CONFIG_PPC_64K_PAGES
503 new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HPTE_SUB0;
505 new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HASHPTE;
507 /* Add in WIMG bits */
508 rflags |= (new_pte & (_PAGE_WRITETHRU | _PAGE_NO_CACHE |
509 _PAGE_COHERENT | _PAGE_GUARDED));
511 /* Insert into the hash table, primary slot */
512 slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags, 0,
513 mmu_huge_psize, ssize);
515 /* Primary is full, try the secondary */
516 if (unlikely(slot == -1)) {
517 hpte_group = ((~hash & htab_hash_mask) *
518 HPTES_PER_GROUP) & ~0x7UL;
519 slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags,
521 mmu_huge_psize, ssize);
524 hpte_group = ((hash & htab_hash_mask) *
525 HPTES_PER_GROUP)&~0x7UL;
527 ppc_md.hpte_remove(hpte_group);
532 if (unlikely(slot == -2))
533 panic("hash_huge_page: pte_insert failed\n");
535 new_pte |= (slot << 12) & (_PAGE_F_SECOND | _PAGE_F_GIX);
539 * No need to use ldarx/stdcx here
541 *ptep = __pte(new_pte & ~_PAGE_BUSY);
549 void set_huge_psize(int psize)
551 /* Check that it is a page size supported by the hardware and
552 * that it fits within pagetable limits. */
553 if (mmu_psize_defs[psize].shift && mmu_psize_defs[psize].shift < SID_SHIFT &&
554 (mmu_psize_defs[psize].shift > MIN_HUGEPTE_SHIFT ||
555 mmu_psize_defs[psize].shift == HPAGE_SHIFT_64K)) {
556 HPAGE_SHIFT = mmu_psize_defs[psize].shift;
557 mmu_huge_psize = psize;
558 #ifdef CONFIG_PPC_64K_PAGES
559 hugepte_shift = (PMD_SHIFT-HPAGE_SHIFT);
561 if (HPAGE_SHIFT == HPAGE_SHIFT_64K)
562 hugepte_shift = (PMD_SHIFT-HPAGE_SHIFT);
564 hugepte_shift = (PUD_SHIFT-HPAGE_SHIFT);
571 static int __init hugepage_setup_sz(char *str)
573 unsigned long long size;
577 size = memparse(str, &str);
581 #ifndef CONFIG_PPC_64K_PAGES
582 case HPAGE_SHIFT_64K:
583 mmu_psize = MMU_PAGE_64K;
586 case HPAGE_SHIFT_16M:
587 mmu_psize = MMU_PAGE_16M;
591 if (mmu_psize >=0 && mmu_psize_defs[mmu_psize].shift)
592 set_huge_psize(mmu_psize);
594 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
598 __setup("hugepagesz=", hugepage_setup_sz);
600 static void zero_ctor(struct kmem_cache *cache, void *addr)
602 memset(addr, 0, kmem_cache_size(cache));
605 static int __init hugetlbpage_init(void)
607 if (!cpu_has_feature(CPU_FTR_16M_PAGE))
610 huge_pgtable_cache = kmem_cache_create("hugepte_cache",
615 if (! huge_pgtable_cache)
616 panic("hugetlbpage_init(): could not create hugepte cache\n");
621 module_init(hugetlbpage_init);