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/smp_lock.h>
16 #include <linux/slab.h>
17 #include <linux/err.h>
18 #include <linux/sysctl.h>
20 #include <asm/pgalloc.h>
22 #include <asm/tlbflush.h>
23 #include <asm/mmu_context.h>
24 #include <asm/machdep.h>
25 #include <asm/cputable.h>
28 #include <linux/sysctl.h>
30 #define NUM_LOW_AREAS (0x100000000UL >> SID_SHIFT)
31 #define NUM_HIGH_AREAS (PGTABLE_RANGE >> HTLB_AREA_SHIFT)
33 /* Modelled after find_linux_pte() */
34 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
41 BUG_ON(! in_hugepage_area(mm->context, addr));
45 pg = pgd_offset(mm, addr);
47 pu = pud_offset(pg, addr);
49 pm = pmd_offset(pu, addr);
50 #ifdef CONFIG_PPC_64K_PAGES
51 /* Currently, we use the normal PTE offset within full
52 * size PTE pages, thus our huge PTEs are scattered in
53 * the PTE page and we do waste some. We may change
54 * that in the future, but the current mecanism keeps
58 /* Note: pte_offset_* are all equivalent on
59 * ppc64 as we don't have HIGHMEM
61 pt = pte_offset_kernel(pm, addr);
64 #else /* CONFIG_PPC_64K_PAGES */
65 /* On 4k pages, we put huge PTEs in the PMD page */
68 #endif /* CONFIG_PPC_64K_PAGES */
75 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
82 BUG_ON(! in_hugepage_area(mm->context, addr));
86 pg = pgd_offset(mm, addr);
87 pu = pud_alloc(mm, pg, addr);
90 pm = pmd_alloc(mm, pu, addr);
92 #ifdef CONFIG_PPC_64K_PAGES
93 /* See comment in huge_pte_offset. Note that if we ever
94 * want to put the page size in the PMD, we would have
95 * to open code our own pte_alloc* function in order
96 * to populate and set the size atomically
98 pt = pte_alloc_map(mm, pm, addr);
99 #else /* CONFIG_PPC_64K_PAGES */
101 #endif /* CONFIG_PPC_64K_PAGES */
109 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
110 pte_t *ptep, pte_t pte)
112 if (pte_present(*ptep)) {
113 /* We open-code pte_clear because we need to pass the right
114 * argument to hpte_update (huge / !huge)
116 unsigned long old = pte_update(ptep, ~0UL);
117 if (old & _PAGE_HASHPTE)
118 hpte_update(mm, addr & HPAGE_MASK, ptep, old, 1);
121 *ptep = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
124 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
127 unsigned long old = pte_update(ptep, ~0UL);
129 if (old & _PAGE_HASHPTE)
130 hpte_update(mm, addr & HPAGE_MASK, ptep, old, 1);
136 struct slb_flush_info {
137 struct mm_struct *mm;
141 static void flush_low_segments(void *parm)
143 struct slb_flush_info *fi = parm;
146 BUILD_BUG_ON((sizeof(fi->newareas)*8) != NUM_LOW_AREAS);
148 if (current->active_mm != fi->mm)
151 /* Only need to do anything if this CPU is working in the same
152 * mm as the one which has changed */
154 /* update the paca copy of the context struct */
155 get_paca()->context = current->active_mm->context;
157 asm volatile("isync" : : : "memory");
158 for (i = 0; i < NUM_LOW_AREAS; i++) {
159 if (! (fi->newareas & (1U << i)))
161 asm volatile("slbie %0"
162 : : "r" ((i << SID_SHIFT) | SLBIE_C));
164 asm volatile("isync" : : : "memory");
167 static void flush_high_segments(void *parm)
169 struct slb_flush_info *fi = parm;
173 BUILD_BUG_ON((sizeof(fi->newareas)*8) != NUM_HIGH_AREAS);
175 if (current->active_mm != fi->mm)
178 /* Only need to do anything if this CPU is working in the same
179 * mm as the one which has changed */
181 /* update the paca copy of the context struct */
182 get_paca()->context = current->active_mm->context;
184 asm volatile("isync" : : : "memory");
185 for (i = 0; i < NUM_HIGH_AREAS; i++) {
186 if (! (fi->newareas & (1U << i)))
188 for (j = 0; j < (1UL << (HTLB_AREA_SHIFT-SID_SHIFT)); j++)
189 asm volatile("slbie %0"
190 :: "r" (((i << HTLB_AREA_SHIFT)
191 + (j << SID_SHIFT)) | SLBIE_C));
193 asm volatile("isync" : : : "memory");
196 static int prepare_low_area_for_htlb(struct mm_struct *mm, unsigned long area)
198 unsigned long start = area << SID_SHIFT;
199 unsigned long end = (area+1) << SID_SHIFT;
200 struct vm_area_struct *vma;
202 BUG_ON(area >= NUM_LOW_AREAS);
204 /* Check no VMAs are in the region */
205 vma = find_vma(mm, start);
206 if (vma && (vma->vm_start < end))
212 static int prepare_high_area_for_htlb(struct mm_struct *mm, unsigned long area)
214 unsigned long start = area << HTLB_AREA_SHIFT;
215 unsigned long end = (area+1) << HTLB_AREA_SHIFT;
216 struct vm_area_struct *vma;
218 BUG_ON(area >= NUM_HIGH_AREAS);
220 /* Hack, so that each addresses is controlled by exactly one
221 * of the high or low area bitmaps, the first high area starts
224 start = 0x100000000UL;
226 /* Check no VMAs are in the region */
227 vma = find_vma(mm, start);
228 if (vma && (vma->vm_start < end))
234 static int open_low_hpage_areas(struct mm_struct *mm, u16 newareas)
237 struct slb_flush_info fi;
239 BUILD_BUG_ON((sizeof(newareas)*8) != NUM_LOW_AREAS);
240 BUILD_BUG_ON((sizeof(mm->context.low_htlb_areas)*8) != NUM_LOW_AREAS);
242 newareas &= ~(mm->context.low_htlb_areas);
244 return 0; /* The segments we want are already open */
246 for (i = 0; i < NUM_LOW_AREAS; i++)
247 if ((1 << i) & newareas)
248 if (prepare_low_area_for_htlb(mm, i) != 0)
251 mm->context.low_htlb_areas |= newareas;
253 /* the context change must make it to memory before the flush,
254 * so that further SLB misses do the right thing. */
258 fi.newareas = newareas;
259 on_each_cpu(flush_low_segments, &fi, 0, 1);
264 static int open_high_hpage_areas(struct mm_struct *mm, u16 newareas)
266 struct slb_flush_info fi;
269 BUILD_BUG_ON((sizeof(newareas)*8) != NUM_HIGH_AREAS);
270 BUILD_BUG_ON((sizeof(mm->context.high_htlb_areas)*8)
273 newareas &= ~(mm->context.high_htlb_areas);
275 return 0; /* The areas we want are already open */
277 for (i = 0; i < NUM_HIGH_AREAS; i++)
278 if ((1 << i) & newareas)
279 if (prepare_high_area_for_htlb(mm, i) != 0)
282 mm->context.high_htlb_areas |= newareas;
284 /* update the paca copy of the context struct */
285 get_paca()->context = mm->context;
287 /* the context change must make it to memory before the flush,
288 * so that further SLB misses do the right thing. */
292 fi.newareas = newareas;
293 on_each_cpu(flush_high_segments, &fi, 0, 1);
298 int prepare_hugepage_range(unsigned long addr, unsigned long len)
302 if ( (addr+len) < addr )
305 if (addr < 0x100000000UL)
306 err = open_low_hpage_areas(current->mm,
307 LOW_ESID_MASK(addr, len));
308 if ((addr + len) > 0x100000000UL)
309 err = open_high_hpage_areas(current->mm,
310 HTLB_AREA_MASK(addr, len));
312 printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
313 " failed (lowmask: 0x%04hx, highmask: 0x%04hx)\n",
315 LOW_ESID_MASK(addr, len), HTLB_AREA_MASK(addr, len));
323 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
328 if (! in_hugepage_area(mm->context, address))
329 return ERR_PTR(-EINVAL);
331 ptep = huge_pte_offset(mm, address);
332 page = pte_page(*ptep);
334 page += (address % HPAGE_SIZE) / PAGE_SIZE;
339 int pmd_huge(pmd_t pmd)
345 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
346 pmd_t *pmd, int write)
352 /* Because we have an exclusive hugepage region which lies within the
353 * normal user address space, we have to take special measures to make
354 * non-huge mmap()s evade the hugepage reserved regions. */
355 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
356 unsigned long len, unsigned long pgoff,
359 struct mm_struct *mm = current->mm;
360 struct vm_area_struct *vma;
361 unsigned long start_addr;
367 addr = PAGE_ALIGN(addr);
368 vma = find_vma(mm, addr);
369 if (((TASK_SIZE - len) >= addr)
370 && (!vma || (addr+len) <= vma->vm_start)
371 && !is_hugepage_only_range(mm, addr,len))
374 if (len > mm->cached_hole_size) {
375 start_addr = addr = mm->free_area_cache;
377 start_addr = addr = TASK_UNMAPPED_BASE;
378 mm->cached_hole_size = 0;
382 vma = find_vma(mm, addr);
383 while (TASK_SIZE - len >= addr) {
384 BUG_ON(vma && (addr >= vma->vm_end));
386 if (touches_hugepage_low_range(mm, addr, len)) {
387 addr = ALIGN(addr+1, 1<<SID_SHIFT);
388 vma = find_vma(mm, addr);
391 if (touches_hugepage_high_range(mm, addr, len)) {
392 addr = ALIGN(addr+1, 1UL<<HTLB_AREA_SHIFT);
393 vma = find_vma(mm, addr);
396 if (!vma || addr + len <= vma->vm_start) {
398 * Remember the place where we stopped the search:
400 mm->free_area_cache = addr + len;
403 if (addr + mm->cached_hole_size < vma->vm_start)
404 mm->cached_hole_size = vma->vm_start - addr;
409 /* Make sure we didn't miss any holes */
410 if (start_addr != TASK_UNMAPPED_BASE) {
411 start_addr = addr = TASK_UNMAPPED_BASE;
412 mm->cached_hole_size = 0;
419 * This mmap-allocator allocates new areas top-down from below the
420 * stack's low limit (the base):
422 * Because we have an exclusive hugepage region which lies within the
423 * normal user address space, we have to take special measures to make
424 * non-huge mmap()s evade the hugepage reserved regions.
427 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
428 const unsigned long len, const unsigned long pgoff,
429 const unsigned long flags)
431 struct vm_area_struct *vma, *prev_vma;
432 struct mm_struct *mm = current->mm;
433 unsigned long base = mm->mmap_base, addr = addr0;
434 unsigned long largest_hole = mm->cached_hole_size;
437 /* requested length too big for entire address space */
441 /* dont allow allocations above current base */
442 if (mm->free_area_cache > base)
443 mm->free_area_cache = base;
445 /* requesting a specific address */
447 addr = PAGE_ALIGN(addr);
448 vma = find_vma(mm, addr);
449 if (TASK_SIZE - len >= addr &&
450 (!vma || addr + len <= vma->vm_start)
451 && !is_hugepage_only_range(mm, addr,len))
455 if (len <= largest_hole) {
457 mm->free_area_cache = base;
460 /* make sure it can fit in the remaining address space */
461 if (mm->free_area_cache < len)
464 /* either no address requested or cant fit in requested address hole */
465 addr = (mm->free_area_cache - len) & PAGE_MASK;
468 if (touches_hugepage_low_range(mm, addr, len)) {
469 addr = (addr & ((~0) << SID_SHIFT)) - len;
470 goto hugepage_recheck;
471 } else if (touches_hugepage_high_range(mm, addr, len)) {
472 addr = (addr & ((~0UL) << HTLB_AREA_SHIFT)) - len;
473 goto hugepage_recheck;
477 * Lookup failure means no vma is above this address,
478 * i.e. return with success:
480 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
484 * new region fits between prev_vma->vm_end and
485 * vma->vm_start, use it:
487 if (addr+len <= vma->vm_start &&
488 (!prev_vma || (addr >= prev_vma->vm_end))) {
489 /* remember the address as a hint for next time */
490 mm->cached_hole_size = largest_hole;
491 return (mm->free_area_cache = addr);
493 /* pull free_area_cache down to the first hole */
494 if (mm->free_area_cache == vma->vm_end) {
495 mm->free_area_cache = vma->vm_start;
496 mm->cached_hole_size = largest_hole;
500 /* remember the largest hole we saw so far */
501 if (addr + largest_hole < vma->vm_start)
502 largest_hole = vma->vm_start - addr;
504 /* try just below the current vma->vm_start */
505 addr = vma->vm_start-len;
506 } while (len <= vma->vm_start);
510 * if hint left us with no space for the requested
511 * mapping then try again:
514 mm->free_area_cache = base;
520 * A failed mmap() very likely causes application failure,
521 * so fall back to the bottom-up function here. This scenario
522 * can happen with large stack limits and large mmap()
525 mm->free_area_cache = TASK_UNMAPPED_BASE;
526 mm->cached_hole_size = ~0UL;
527 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
529 * Restore the topdown base:
531 mm->free_area_cache = base;
532 mm->cached_hole_size = ~0UL;
537 static int htlb_check_hinted_area(unsigned long addr, unsigned long len)
539 struct vm_area_struct *vma;
541 vma = find_vma(current->mm, addr);
542 if (!vma || ((addr + len) <= vma->vm_start))
548 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
550 unsigned long addr = 0;
551 struct vm_area_struct *vma;
553 vma = find_vma(current->mm, addr);
554 while (addr + len <= 0x100000000UL) {
555 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
557 if (! __within_hugepage_low_range(addr, len, segmask)) {
558 addr = ALIGN(addr+1, 1<<SID_SHIFT);
559 vma = find_vma(current->mm, addr);
563 if (!vma || (addr + len) <= vma->vm_start)
565 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
566 /* Depending on segmask this might not be a confirmed
567 * hugepage region, so the ALIGN could have skipped
569 vma = find_vma(current->mm, addr);
575 static unsigned long htlb_get_high_area(unsigned long len, u16 areamask)
577 unsigned long addr = 0x100000000UL;
578 struct vm_area_struct *vma;
580 vma = find_vma(current->mm, addr);
581 while (addr + len <= TASK_SIZE_USER64) {
582 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
584 if (! __within_hugepage_high_range(addr, len, areamask)) {
585 addr = ALIGN(addr+1, 1UL<<HTLB_AREA_SHIFT);
586 vma = find_vma(current->mm, addr);
590 if (!vma || (addr + len) <= vma->vm_start)
592 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
593 /* Depending on segmask this might not be a confirmed
594 * hugepage region, so the ALIGN could have skipped
596 vma = find_vma(current->mm, addr);
602 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
603 unsigned long len, unsigned long pgoff,
607 u16 areamask, curareas;
609 if (HPAGE_SHIFT == 0)
611 if (len & ~HPAGE_MASK)
614 if (!cpu_has_feature(CPU_FTR_16M_PAGE))
617 /* Paranoia, caller should have dealt with this */
618 BUG_ON((addr + len) < addr);
620 if (test_thread_flag(TIF_32BIT)) {
621 /* Paranoia, caller should have dealt with this */
622 BUG_ON((addr + len) > 0x100000000UL);
624 curareas = current->mm->context.low_htlb_areas;
626 /* First see if we can use the hint address */
627 if (addr && (htlb_check_hinted_area(addr, len) == 0)) {
628 areamask = LOW_ESID_MASK(addr, len);
629 if (open_low_hpage_areas(current->mm, areamask) == 0)
633 /* Next see if we can map in the existing low areas */
634 addr = htlb_get_low_area(len, curareas);
638 /* Finally go looking for areas to open */
640 for (areamask = LOW_ESID_MASK(0x100000000UL-len, len);
641 ! lastshift; areamask >>=1) {
645 addr = htlb_get_low_area(len, curareas | areamask);
646 if ((addr != -ENOMEM)
647 && open_low_hpage_areas(current->mm, areamask) == 0)
651 curareas = current->mm->context.high_htlb_areas;
653 /* First see if we can use the hint address */
654 /* We discourage 64-bit processes from doing hugepage
655 * mappings below 4GB (must use MAP_FIXED) */
656 if ((addr >= 0x100000000UL)
657 && (htlb_check_hinted_area(addr, len) == 0)) {
658 areamask = HTLB_AREA_MASK(addr, len);
659 if (open_high_hpage_areas(current->mm, areamask) == 0)
663 /* Next see if we can map in the existing high areas */
664 addr = htlb_get_high_area(len, curareas);
668 /* Finally go looking for areas to open */
670 for (areamask = HTLB_AREA_MASK(TASK_SIZE_USER64-len, len);
671 ! lastshift; areamask >>=1) {
675 addr = htlb_get_high_area(len, curareas | areamask);
676 if ((addr != -ENOMEM)
677 && open_high_hpage_areas(current->mm, areamask) == 0)
681 printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
687 * Called by asm hashtable.S for doing lazy icache flush
689 static unsigned int hash_huge_page_do_lazy_icache(unsigned long rflags,
695 if (!pfn_valid(pte_pfn(pte)))
698 page = pte_page(pte);
701 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
703 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++)
704 __flush_dcache_icache(page_address(page+i));
705 set_bit(PG_arch_1, &page->flags);
713 int hash_huge_page(struct mm_struct *mm, unsigned long access,
714 unsigned long ea, unsigned long vsid, int local,
718 unsigned long old_pte, new_pte;
719 unsigned long va, rflags, pa;
723 ptep = huge_pte_offset(mm, ea);
725 /* Search the Linux page table for a match with va */
726 va = (vsid << 28) | (ea & 0x0fffffff);
729 * If no pte found or not present, send the problem up to
732 if (unlikely(!ptep || pte_none(*ptep)))
736 * Check the user's access rights to the page. If access should be
737 * prevented then send the problem up to do_page_fault.
739 if (unlikely(access & ~pte_val(*ptep)))
742 * At this point, we have a pte (old_pte) which can be used to build
743 * or update an HPTE. There are 2 cases:
745 * 1. There is a valid (present) pte with no associated HPTE (this is
746 * the most common case)
747 * 2. There is a valid (present) pte with an associated HPTE. The
748 * current values of the pp bits in the HPTE prevent access
749 * because we are doing software DIRTY bit management and the
750 * page is currently not DIRTY.
755 old_pte = pte_val(*ptep);
756 if (old_pte & _PAGE_BUSY)
758 new_pte = old_pte | _PAGE_BUSY |
759 _PAGE_ACCESSED | _PAGE_HASHPTE;
760 } while(old_pte != __cmpxchg_u64((unsigned long *)ptep,
763 rflags = 0x2 | (!(new_pte & _PAGE_RW));
764 /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
765 rflags |= ((new_pte & _PAGE_EXEC) ? 0 : HPTE_R_N);
766 if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
767 /* No CPU has hugepages but lacks no execute, so we
768 * don't need to worry about that case */
769 rflags = hash_huge_page_do_lazy_icache(rflags, __pte(old_pte),
772 /* Check if pte already has an hpte (case 2) */
773 if (unlikely(old_pte & _PAGE_HASHPTE)) {
774 /* There MIGHT be an HPTE for this pte */
775 unsigned long hash, slot;
777 hash = hpt_hash(va, HPAGE_SHIFT);
778 if (old_pte & _PAGE_F_SECOND)
780 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
781 slot += (old_pte & _PAGE_F_GIX) >> 12;
783 if (ppc_md.hpte_updatepp(slot, rflags, va, mmu_huge_psize,
785 old_pte &= ~_PAGE_HPTEFLAGS;
788 if (likely(!(old_pte & _PAGE_HASHPTE))) {
789 unsigned long hash = hpt_hash(va, HPAGE_SHIFT);
790 unsigned long hpte_group;
792 pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
795 hpte_group = ((hash & htab_hash_mask) *
796 HPTES_PER_GROUP) & ~0x7UL;
798 /* clear HPTE slot informations in new PTE */
799 new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HASHPTE;
801 /* Add in WIMG bits */
802 /* XXX We should store these in the pte */
803 /* --BenH: I think they are ... */
804 rflags |= _PAGE_COHERENT;
806 /* Insert into the hash table, primary slot */
807 slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags, 0,
810 /* Primary is full, try the secondary */
811 if (unlikely(slot == -1)) {
812 new_pte |= _PAGE_F_SECOND;
813 hpte_group = ((~hash & htab_hash_mask) *
814 HPTES_PER_GROUP) & ~0x7UL;
815 slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags,
820 hpte_group = ((hash & htab_hash_mask) *
821 HPTES_PER_GROUP)&~0x7UL;
823 ppc_md.hpte_remove(hpte_group);
828 if (unlikely(slot == -2))
829 panic("hash_huge_page: pte_insert failed\n");
831 new_pte |= (slot << 12) & _PAGE_F_GIX;
835 * No need to use ldarx/stdcx here
837 *ptep = __pte(new_pte & ~_PAGE_BUSY);