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);
137 * This function checks for proper alignment of input addr and len parameters.
139 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
141 if (len & ~HPAGE_MASK)
143 if (addr & ~HPAGE_MASK)
145 if (! (within_hugepage_low_range(addr, len)
146 || within_hugepage_high_range(addr, len)) )
151 struct slb_flush_info {
152 struct mm_struct *mm;
156 static void flush_low_segments(void *parm)
158 struct slb_flush_info *fi = parm;
161 BUILD_BUG_ON((sizeof(fi->newareas)*8) != NUM_LOW_AREAS);
163 if (current->active_mm != fi->mm)
166 /* Only need to do anything if this CPU is working in the same
167 * mm as the one which has changed */
169 /* update the paca copy of the context struct */
170 get_paca()->context = current->active_mm->context;
172 asm volatile("isync" : : : "memory");
173 for (i = 0; i < NUM_LOW_AREAS; i++) {
174 if (! (fi->newareas & (1U << i)))
176 asm volatile("slbie %0"
177 : : "r" ((i << SID_SHIFT) | SLBIE_C));
179 asm volatile("isync" : : : "memory");
182 static void flush_high_segments(void *parm)
184 struct slb_flush_info *fi = parm;
188 BUILD_BUG_ON((sizeof(fi->newareas)*8) != NUM_HIGH_AREAS);
190 if (current->active_mm != fi->mm)
193 /* Only need to do anything if this CPU is working in the same
194 * mm as the one which has changed */
196 /* update the paca copy of the context struct */
197 get_paca()->context = current->active_mm->context;
199 asm volatile("isync" : : : "memory");
200 for (i = 0; i < NUM_HIGH_AREAS; i++) {
201 if (! (fi->newareas & (1U << i)))
203 for (j = 0; j < (1UL << (HTLB_AREA_SHIFT-SID_SHIFT)); j++)
204 asm volatile("slbie %0"
205 :: "r" (((i << HTLB_AREA_SHIFT)
206 + (j << SID_SHIFT)) | SLBIE_C));
208 asm volatile("isync" : : : "memory");
211 static int prepare_low_area_for_htlb(struct mm_struct *mm, unsigned long area)
213 unsigned long start = area << SID_SHIFT;
214 unsigned long end = (area+1) << SID_SHIFT;
215 struct vm_area_struct *vma;
217 BUG_ON(area >= NUM_LOW_AREAS);
219 /* Check no VMAs are in the region */
220 vma = find_vma(mm, start);
221 if (vma && (vma->vm_start < end))
227 static int prepare_high_area_for_htlb(struct mm_struct *mm, unsigned long area)
229 unsigned long start = area << HTLB_AREA_SHIFT;
230 unsigned long end = (area+1) << HTLB_AREA_SHIFT;
231 struct vm_area_struct *vma;
233 BUG_ON(area >= NUM_HIGH_AREAS);
235 /* Hack, so that each addresses is controlled by exactly one
236 * of the high or low area bitmaps, the first high area starts
239 start = 0x100000000UL;
241 /* Check no VMAs are in the region */
242 vma = find_vma(mm, start);
243 if (vma && (vma->vm_start < end))
249 static int open_low_hpage_areas(struct mm_struct *mm, u16 newareas)
252 struct slb_flush_info fi;
254 BUILD_BUG_ON((sizeof(newareas)*8) != NUM_LOW_AREAS);
255 BUILD_BUG_ON((sizeof(mm->context.low_htlb_areas)*8) != NUM_LOW_AREAS);
257 newareas &= ~(mm->context.low_htlb_areas);
259 return 0; /* The segments we want are already open */
261 for (i = 0; i < NUM_LOW_AREAS; i++)
262 if ((1 << i) & newareas)
263 if (prepare_low_area_for_htlb(mm, i) != 0)
266 mm->context.low_htlb_areas |= newareas;
268 /* the context change must make it to memory before the flush,
269 * so that further SLB misses do the right thing. */
273 fi.newareas = newareas;
274 on_each_cpu(flush_low_segments, &fi, 0, 1);
279 static int open_high_hpage_areas(struct mm_struct *mm, u16 newareas)
281 struct slb_flush_info fi;
284 BUILD_BUG_ON((sizeof(newareas)*8) != NUM_HIGH_AREAS);
285 BUILD_BUG_ON((sizeof(mm->context.high_htlb_areas)*8)
288 newareas &= ~(mm->context.high_htlb_areas);
290 return 0; /* The areas we want are already open */
292 for (i = 0; i < NUM_HIGH_AREAS; i++)
293 if ((1 << i) & newareas)
294 if (prepare_high_area_for_htlb(mm, i) != 0)
297 mm->context.high_htlb_areas |= newareas;
299 /* update the paca copy of the context struct */
300 get_paca()->context = mm->context;
302 /* the context change must make it to memory before the flush,
303 * so that further SLB misses do the right thing. */
307 fi.newareas = newareas;
308 on_each_cpu(flush_high_segments, &fi, 0, 1);
313 int prepare_hugepage_range(unsigned long addr, unsigned long len)
317 if ( (addr+len) < addr )
320 if (addr < 0x100000000UL)
321 err = open_low_hpage_areas(current->mm,
322 LOW_ESID_MASK(addr, len));
323 if ((addr + len) > 0x100000000UL)
324 err = open_high_hpage_areas(current->mm,
325 HTLB_AREA_MASK(addr, len));
327 printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
328 " failed (lowmask: 0x%04hx, highmask: 0x%04hx)\n",
330 LOW_ESID_MASK(addr, len), HTLB_AREA_MASK(addr, len));
338 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
343 if (! in_hugepage_area(mm->context, address))
344 return ERR_PTR(-EINVAL);
346 ptep = huge_pte_offset(mm, address);
347 page = pte_page(*ptep);
349 page += (address % HPAGE_SIZE) / PAGE_SIZE;
354 int pmd_huge(pmd_t pmd)
360 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
361 pmd_t *pmd, int write)
367 /* Because we have an exclusive hugepage region which lies within the
368 * normal user address space, we have to take special measures to make
369 * non-huge mmap()s evade the hugepage reserved regions. */
370 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
371 unsigned long len, unsigned long pgoff,
374 struct mm_struct *mm = current->mm;
375 struct vm_area_struct *vma;
376 unsigned long start_addr;
382 addr = PAGE_ALIGN(addr);
383 vma = find_vma(mm, addr);
384 if (((TASK_SIZE - len) >= addr)
385 && (!vma || (addr+len) <= vma->vm_start)
386 && !is_hugepage_only_range(mm, addr,len))
389 if (len > mm->cached_hole_size) {
390 start_addr = addr = mm->free_area_cache;
392 start_addr = addr = TASK_UNMAPPED_BASE;
393 mm->cached_hole_size = 0;
397 vma = find_vma(mm, addr);
398 while (TASK_SIZE - len >= addr) {
399 BUG_ON(vma && (addr >= vma->vm_end));
401 if (touches_hugepage_low_range(mm, addr, len)) {
402 addr = ALIGN(addr+1, 1<<SID_SHIFT);
403 vma = find_vma(mm, addr);
406 if (touches_hugepage_high_range(mm, addr, len)) {
407 addr = ALIGN(addr+1, 1UL<<HTLB_AREA_SHIFT);
408 vma = find_vma(mm, addr);
411 if (!vma || addr + len <= vma->vm_start) {
413 * Remember the place where we stopped the search:
415 mm->free_area_cache = addr + len;
418 if (addr + mm->cached_hole_size < vma->vm_start)
419 mm->cached_hole_size = vma->vm_start - addr;
424 /* Make sure we didn't miss any holes */
425 if (start_addr != TASK_UNMAPPED_BASE) {
426 start_addr = addr = TASK_UNMAPPED_BASE;
427 mm->cached_hole_size = 0;
434 * This mmap-allocator allocates new areas top-down from below the
435 * stack's low limit (the base):
437 * Because we have an exclusive hugepage region which lies within the
438 * normal user address space, we have to take special measures to make
439 * non-huge mmap()s evade the hugepage reserved regions.
442 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
443 const unsigned long len, const unsigned long pgoff,
444 const unsigned long flags)
446 struct vm_area_struct *vma, *prev_vma;
447 struct mm_struct *mm = current->mm;
448 unsigned long base = mm->mmap_base, addr = addr0;
449 unsigned long largest_hole = mm->cached_hole_size;
452 /* requested length too big for entire address space */
456 /* dont allow allocations above current base */
457 if (mm->free_area_cache > base)
458 mm->free_area_cache = base;
460 /* requesting a specific address */
462 addr = PAGE_ALIGN(addr);
463 vma = find_vma(mm, addr);
464 if (TASK_SIZE - len >= addr &&
465 (!vma || addr + len <= vma->vm_start)
466 && !is_hugepage_only_range(mm, addr,len))
470 if (len <= largest_hole) {
472 mm->free_area_cache = base;
475 /* make sure it can fit in the remaining address space */
476 if (mm->free_area_cache < len)
479 /* either no address requested or cant fit in requested address hole */
480 addr = (mm->free_area_cache - len) & PAGE_MASK;
483 if (touches_hugepage_low_range(mm, addr, len)) {
484 addr = (addr & ((~0) << SID_SHIFT)) - len;
485 goto hugepage_recheck;
486 } else if (touches_hugepage_high_range(mm, addr, len)) {
487 addr = (addr & ((~0UL) << HTLB_AREA_SHIFT)) - len;
488 goto hugepage_recheck;
492 * Lookup failure means no vma is above this address,
493 * i.e. return with success:
495 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
499 * new region fits between prev_vma->vm_end and
500 * vma->vm_start, use it:
502 if (addr+len <= vma->vm_start &&
503 (!prev_vma || (addr >= prev_vma->vm_end))) {
504 /* remember the address as a hint for next time */
505 mm->cached_hole_size = largest_hole;
506 return (mm->free_area_cache = addr);
508 /* pull free_area_cache down to the first hole */
509 if (mm->free_area_cache == vma->vm_end) {
510 mm->free_area_cache = vma->vm_start;
511 mm->cached_hole_size = largest_hole;
515 /* remember the largest hole we saw so far */
516 if (addr + largest_hole < vma->vm_start)
517 largest_hole = vma->vm_start - addr;
519 /* try just below the current vma->vm_start */
520 addr = vma->vm_start-len;
521 } while (len <= vma->vm_start);
525 * if hint left us with no space for the requested
526 * mapping then try again:
529 mm->free_area_cache = base;
535 * A failed mmap() very likely causes application failure,
536 * so fall back to the bottom-up function here. This scenario
537 * can happen with large stack limits and large mmap()
540 mm->free_area_cache = TASK_UNMAPPED_BASE;
541 mm->cached_hole_size = ~0UL;
542 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
544 * Restore the topdown base:
546 mm->free_area_cache = base;
547 mm->cached_hole_size = ~0UL;
552 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
554 unsigned long addr = 0;
555 struct vm_area_struct *vma;
557 vma = find_vma(current->mm, addr);
558 while (addr + len <= 0x100000000UL) {
559 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
561 if (! __within_hugepage_low_range(addr, len, segmask)) {
562 addr = ALIGN(addr+1, 1<<SID_SHIFT);
563 vma = find_vma(current->mm, addr);
567 if (!vma || (addr + len) <= vma->vm_start)
569 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
570 /* Depending on segmask this might not be a confirmed
571 * hugepage region, so the ALIGN could have skipped
573 vma = find_vma(current->mm, addr);
579 static unsigned long htlb_get_high_area(unsigned long len, u16 areamask)
581 unsigned long addr = 0x100000000UL;
582 struct vm_area_struct *vma;
584 vma = find_vma(current->mm, addr);
585 while (addr + len <= TASK_SIZE_USER64) {
586 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
588 if (! __within_hugepage_high_range(addr, len, areamask)) {
589 addr = ALIGN(addr+1, 1UL<<HTLB_AREA_SHIFT);
590 vma = find_vma(current->mm, addr);
594 if (!vma || (addr + len) <= vma->vm_start)
596 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
597 /* Depending on segmask this might not be a confirmed
598 * hugepage region, so the ALIGN could have skipped
600 vma = find_vma(current->mm, addr);
606 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
607 unsigned long len, unsigned long pgoff,
611 u16 areamask, curareas;
613 if (HPAGE_SHIFT == 0)
615 if (len & ~HPAGE_MASK)
618 if (!cpu_has_feature(CPU_FTR_16M_PAGE))
621 if (test_thread_flag(TIF_32BIT)) {
622 curareas = current->mm->context.low_htlb_areas;
624 /* First see if we can do the mapping in the existing
626 addr = htlb_get_low_area(len, curareas);
631 for (areamask = LOW_ESID_MASK(0x100000000UL-len, len);
632 ! lastshift; areamask >>=1) {
636 addr = htlb_get_low_area(len, curareas | areamask);
637 if ((addr != -ENOMEM)
638 && open_low_hpage_areas(current->mm, areamask) == 0)
642 curareas = current->mm->context.high_htlb_areas;
644 /* First see if we can do the mapping in the existing
646 addr = htlb_get_high_area(len, curareas);
651 for (areamask = HTLB_AREA_MASK(TASK_SIZE_USER64-len, len);
652 ! lastshift; areamask >>=1) {
656 addr = htlb_get_high_area(len, curareas | areamask);
657 if ((addr != -ENOMEM)
658 && open_high_hpage_areas(current->mm, areamask) == 0)
662 printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
668 * Called by asm hashtable.S for doing lazy icache flush
670 static unsigned int hash_huge_page_do_lazy_icache(unsigned long rflags,
676 if (!pfn_valid(pte_pfn(pte)))
679 page = pte_page(pte);
682 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
684 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++)
685 __flush_dcache_icache(page_address(page+i));
686 set_bit(PG_arch_1, &page->flags);
694 int hash_huge_page(struct mm_struct *mm, unsigned long access,
695 unsigned long ea, unsigned long vsid, int local,
699 unsigned long old_pte, new_pte;
700 unsigned long va, rflags, pa;
704 ptep = huge_pte_offset(mm, ea);
706 /* Search the Linux page table for a match with va */
707 va = (vsid << 28) | (ea & 0x0fffffff);
710 * If no pte found or not present, send the problem up to
713 if (unlikely(!ptep || pte_none(*ptep)))
717 * Check the user's access rights to the page. If access should be
718 * prevented then send the problem up to do_page_fault.
720 if (unlikely(access & ~pte_val(*ptep)))
723 * At this point, we have a pte (old_pte) which can be used to build
724 * or update an HPTE. There are 2 cases:
726 * 1. There is a valid (present) pte with no associated HPTE (this is
727 * the most common case)
728 * 2. There is a valid (present) pte with an associated HPTE. The
729 * current values of the pp bits in the HPTE prevent access
730 * because we are doing software DIRTY bit management and the
731 * page is currently not DIRTY.
736 old_pte = pte_val(*ptep);
737 if (old_pte & _PAGE_BUSY)
739 new_pte = old_pte | _PAGE_BUSY |
740 _PAGE_ACCESSED | _PAGE_HASHPTE;
741 } while(old_pte != __cmpxchg_u64((unsigned long *)ptep,
744 rflags = 0x2 | (!(new_pte & _PAGE_RW));
745 /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
746 rflags |= ((new_pte & _PAGE_EXEC) ? 0 : HPTE_R_N);
747 if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
748 /* No CPU has hugepages but lacks no execute, so we
749 * don't need to worry about that case */
750 rflags = hash_huge_page_do_lazy_icache(rflags, __pte(old_pte),
753 /* Check if pte already has an hpte (case 2) */
754 if (unlikely(old_pte & _PAGE_HASHPTE)) {
755 /* There MIGHT be an HPTE for this pte */
756 unsigned long hash, slot;
758 hash = hpt_hash(va, HPAGE_SHIFT);
759 if (old_pte & _PAGE_F_SECOND)
761 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
762 slot += (old_pte & _PAGE_F_GIX) >> 12;
764 if (ppc_md.hpte_updatepp(slot, rflags, va, mmu_huge_psize,
766 old_pte &= ~_PAGE_HPTEFLAGS;
769 if (likely(!(old_pte & _PAGE_HASHPTE))) {
770 unsigned long hash = hpt_hash(va, HPAGE_SHIFT);
771 unsigned long hpte_group;
773 pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
776 hpte_group = ((hash & htab_hash_mask) *
777 HPTES_PER_GROUP) & ~0x7UL;
779 /* clear HPTE slot informations in new PTE */
780 new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HASHPTE;
782 /* Add in WIMG bits */
783 /* XXX We should store these in the pte */
784 /* --BenH: I think they are ... */
785 rflags |= _PAGE_COHERENT;
787 /* Insert into the hash table, primary slot */
788 slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags, 0,
791 /* Primary is full, try the secondary */
792 if (unlikely(slot == -1)) {
793 new_pte |= _PAGE_F_SECOND;
794 hpte_group = ((~hash & htab_hash_mask) *
795 HPTES_PER_GROUP) & ~0x7UL;
796 slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags,
801 hpte_group = ((hash & htab_hash_mask) *
802 HPTES_PER_GROUP)&~0x7UL;
804 ppc_md.hpte_remove(hpte_group);
809 if (unlikely(slot == -2))
810 panic("hash_huge_page: pte_insert failed\n");
812 new_pte |= (slot << 12) & _PAGE_F_GIX;
816 * No need to use ldarx/stdcx here
818 *ptep = __pte(new_pte & ~_PAGE_BUSY);