Skip to contentSkip to conversation

Sanitize pagetable handling for hugepages

PATCH New David Gibson 2 participants 4 messages Sep 4, 2009 PowerPC

lore
Share & tools
lore
#0 lore raw

cover Sanitize pagetable handling for hugepages

Currently, ordinary pages use one pagetable layout, and each different hugepage size uses a slightly different variant layout. A number of places which need to walk the pagetable must first check the slice map to see what the pagetable layout then handle the various different forms. New hardware, like Book3E is liable to introduce more possible variants.

This patch series, therefore, is designed to simplify the matter by limiting knowledge of the pagetable layout to only the allocation path. With this patch, ordinary pages are handled as ever, with a fixed 4 (or 3) level tree. All other variants branch off from some layer of that with a specially marked PGD/PUD/PMD pointer which also contains enough information to interpret the directories below that point. This means that things walking the pagetables (without allocating) don't need to look up the slice map, they can just step down the tree in the usual way, branching off to the "non-standard layout" path for hugepages, which uses the embdded information to interpret the tree from that point on.

This reduces the source size in a number of places, and means that newer variants on the pagetable layout to handle new hardware and new features will need to alter the existing code in less places.

I've used the libhugetlbfs testsuite to test these patches on a Power5+ machine, but they could do with some banging. In particular I don't have any suitable hardware to test 16G pages. So, think of this as the first draft of the series.

Signature · 4 lines
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson
#1 lore raw

PATCH 1/3 Make hpte_need_flush() correctly mask for multiple page sizes

Currently, hpte_need_flush() only correctly flushes the given address for normal pages. Callers for hugepages are required to mask the address themselves.

But hpte_nned_flush() already looks up the page sizes for its own reasons, so this is a rather silly imposition on the callers. This patch alters it to mask based on the pagesize it has looked up itself, and removes the awkward masking code in the hugepage caller.

---

---
 arch/powerpc/mm/hugetlbpage.c |    6 +-----
 arch/powerpc/mm/tlb_hash64.c  |    8 +++-----
 2 files changed, 4 insertions(+), 10 deletions(-)
arch/powerpc/mm/tlb_hash64.c +3 -5
Index: working-2.6/arch/powerpc/mm/tlb_hash64.c
===================================================================
@@ -53,11 +53,6 @@ void hpte_need_flush(struct mm_struct *m
 
 	i = batch->index;
 
-	/* We mask the address for the base page size. Huge pages will
-	 * have applied their own masking already
-	 */
-	addr &= PAGE_MASK;
-
 	/* Get page size (maybe move back to caller).
 	 *
 	 * NOTE: when using special 64K mappings in 4K environment like
@@ -75,6 +70,9 @@ void hpte_need_flush(struct mm_struct *m
 	} else
 		psize = pte_pagesize_index(mm, addr, pte);
 
+	/* Mask the address for the correct page size */
+	addr &= ~((1UL << mmu_psize_defs[psize].shift) - 1);
+
 	/* Build full vaddr */
 	if (!is_kernel_addr(addr)) {
 		ssize = user_segment_size(addr);
arch/powerpc/mm/hugetlbpage.c +1 -5
Index: working-2.6/arch/powerpc/mm/hugetlbpage.c
===================================================================
@@ -445,11 +445,7 @@ void set_huge_pte_at(struct mm_struct *m
 		 * necessary anymore if we make hpte_need_flush() get the
 		 * page size from the slices
 		 */
-		unsigned int psize = get_slice_psize(mm, addr);
-		unsigned int shift = mmu_psize_to_shift(psize);
-		unsigned long sz = ((1UL) << shift);
-		struct hstate *hstate = size_to_hstate(sz);
-		pte_update(mm, addr & hstate->mask, ptep, ~0UL, 1);
+		pte_update(mm, addr, ptep, ~0UL, 1);
 	}
 	*ptep = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
 }
#2 lore raw

PATCH 2/3 Cleanup management of kmem_caches for pagetables

In reply to David Gibson

Currently we have a fair bit of rather fiddly code to manage the various kmem_caches used to store page tables of various levels. We generally have two caches holding some combination of PGD, PUD and PMD tables, plus several more for the special hugepage pagetables.

This patch cleans this all up by taking a different approach. Rather than the caches being designated as for PUDs or for hugeptes for 16M pages, the caches are simply allocated to be a specific size. Thus sharing of caches between different types/levels of pagetables happens naturally. The pagetable size, where needed, is passed around encoded in the same way as {PGD,PUD,PMD}_INDEX_SIZE; that is n where the pagetable contains 2^n pointers.

---
 arch/powerpc/include/asm/pgtable-ppc64.h |    1 
---
 arch/powerpc/include/asm/pgalloc-64.h    |   43 ++++++++++++-----------------
 arch/powerpc/include/asm/pgalloc.h       |   25 +++--------------
 arch/powerpc/include/asm/pgtable-ppc64.h |    1 
 arch/powerpc/mm/hugetlbpage.c            |   45 ++++++++-----------------------
 arch/powerpc/mm/init_64.c                |   42 ++++++++++++++--------------
 arch/powerpc/mm/pgtable.c                |   25 +++++++++++------
 6 files changed, 73 insertions(+), 108 deletions(-)
arch/powerpc/mm/init_64.c +21 -21
Index: working-2.6/arch/powerpc/mm/init_64.c
===================================================================
@@ -148,30 +148,30 @@ static void pmd_ctor(void *addr)
 	memset(addr, 0, PMD_TABLE_SIZE);
 }
 
-static const unsigned int pgtable_cache_size[2] = {
-	PGD_TABLE_SIZE, PMD_TABLE_SIZE
-};
-static const char *pgtable_cache_name[ARRAY_SIZE(pgtable_cache_size)] = {
-#ifdef CONFIG_PPC_64K_PAGES
-	"pgd_cache", "pmd_cache",
-#else
-	"pgd_cache", "pud_pmd_cache",
-#endif /* CONFIG_PPC_64K_PAGES */
-};
-
-#ifdef CONFIG_HUGETLB_PAGE
-/* Hugepages need an extra cache per hugepagesize, initialized in
- * hugetlbpage.c.  We can't put into the tables above, because HPAGE_SHIFT
- * is not compile time constant. */
-struct kmem_cache *pgtable_cache[ARRAY_SIZE(pgtable_cache_size)+MMU_PAGE_COUNT];
-#else
-struct kmem_cache *pgtable_cache[ARRAY_SIZE(pgtable_cache_size)];
-#endif
+struct kmem_cache *pgtable_cache[PGF_SHIFT_MASK];
+
+void pgtable_cache_add(unsigned shift, void (*ctor)(void *))
+{
+	char *name;
+	unsigned long table_size = sizeof(void *) << shift;
+	struct kmem_cache *new;
+
+	BUG_ON((shift < 1) || (shift > PGF_SHIFT_MASK));
+	if (PGT_CACHE(shift))
+		return; /* Already have a cache of this size */
+	name = kasprintf(GFP_KERNEL, "pgtable-2^%d", shift);
+	new = kmem_cache_create(name, table_size, table_size, 0, ctor);
+	PGT_CACHE(shift) = new;
+}
+
 
 void pgtable_cache_init(void)
 {
-	pgtable_cache[0] = kmem_cache_create(pgtable_cache_name[0], PGD_TABLE_SIZE, PGD_TABLE_SIZE, SLAB_PANIC, pgd_ctor);
-	pgtable_cache[1] = kmem_cache_create(pgtable_cache_name[1], PMD_TABLE_SIZE, PMD_TABLE_SIZE, SLAB_PANIC, pmd_ctor);
+	pgtable_cache_add(PGD_INDEX_SIZE, pgd_ctor);
+	pgtable_cache_add(PMD_INDEX_SIZE, pmd_ctor);
+	if (!PGT_CACHE(PGD_INDEX_SIZE) || !PGT_CACHE(PMD_INDEX_SIZE))
+		panic("Couldn't allocate pgtable caches");
+	BUG_ON(!PGT_CACHE(PUD_INDEX_SIZE));
 }
 
 #ifdef CONFIG_SPARSEMEM_VMEMMAP
arch/powerpc/include/asm/pgalloc-64.h +18 -25

90 lines not on this page · raw message

arch/powerpc/include/asm/pgalloc.h +5 -20
Index: working-2.6/arch/powerpc/include/asm/pgalloc.h
===================================================================
@@ -24,24 +24,12 @@ static inline void pte_free(struct mm_st
 	__free_page(ptepage);
 }
 
-typedef struct pgtable_free {
-	unsigned long val;
-} pgtable_free_t;
-
 /* This needs to be big enough to allow for MMU_PAGE_COUNT + 2 to be stored
  * and small enough to fit in the low bits of any naturally aligned page
  * table cache entry. Arbitrarily set to 0x1f, that should give us some
  * room to grow
  */
-#define PGF_CACHENUM_MASK	0x1f
-
-static inline pgtable_free_t pgtable_free_cache(void *p, int cachenum,
-						unsigned long mask)
-{
-	BUG_ON(cachenum > PGF_CACHENUM_MASK);
-
-	return (pgtable_free_t){.val = ((unsigned long) p & ~mask) | cachenum};
-}
+#define PGF_SHIFT_MASK		0xf
 
 #ifdef CONFIG_PPC64
 #include <asm/pgalloc-64.h>
@@ -50,12 +38,12 @@ static inline pgtable_free_t pgtable_fre
 #endif
 
 #ifdef CONFIG_SMP
-extern void pgtable_free_tlb(struct mmu_gather *tlb, pgtable_free_t pgf);
+extern void pgtable_free_tlb(struct mmu_gather *tlb, void *table, unsigned shift);
 extern void pte_free_finish(void);
 #else /* CONFIG_SMP */
-static inline void pgtable_free_tlb(struct mmu_gather *tlb, pgtable_free_t pgf)
+static inline void pgtable_free_tlb(struct mmu_gather *tlb, void *table, unsigned shift)
 {
-	pgtable_free(pgf);
+	pgtable_free(table, shift);
 }
 static inline void pte_free_finish(void) { }
 #endif /* !CONFIG_SMP */
@@ -63,12 +51,9 @@ static inline void pte_free_finish(void)
 static inline void __pte_free_tlb(struct mmu_gather *tlb, struct page *ptepage,
 				  unsigned long address)
 {
-	pgtable_free_t pgf = pgtable_free_cache(page_address(ptepage),
-						PTE_NONCACHE_NUM,
-						PTE_TABLE_SIZE-1);
 	tlb_flush_pgtable(tlb, address);
 	pgtable_page_dtor(ptepage);
-	pgtable_free_tlb(tlb, pgf);
+	pgtable_free_tlb(tlb, page_address(ptepage), 0);
 }
 
 #endif /* __KERNEL__ */
arch/powerpc/mm/pgtable.c +16 -9
Index: working-2.6/arch/powerpc/mm/pgtable.c
===================================================================
@@ -47,12 +47,12 @@ struct pte_freelist_batch
 {
 	struct rcu_head	rcu;
 	unsigned int	index;
-	pgtable_free_t	tables[0];
+	unsigned long	tables[0];
 };
 
 #define PTE_FREELIST_SIZE \
 	((PAGE_SIZE - sizeof(struct pte_freelist_batch)) \
-	  / sizeof(pgtable_free_t))
+	  / sizeof(unsigned long))
 
 static void pte_free_smp_sync(void *arg)
 {
@@ -62,13 +62,13 @@ static void pte_free_smp_sync(void *arg)
 /* This is only called when we are critically out of memory
  * (and fail to get a page in pte_free_tlb).
  */
-static void pgtable_free_now(pgtable_free_t pgf)
+static void pgtable_free_now(void *table, unsigned shift)
 {
 	pte_freelist_forced_free++;
 
 	smp_call_function(pte_free_smp_sync, NULL, 1);
 
-	pgtable_free(pgf);
+	pgtable_free(table, shift);
 }
 
 static void pte_free_rcu_callback(struct rcu_head *head)
@@ -77,8 +77,12 @@ static void pte_free_rcu_callback(struct
 		container_of(head, struct pte_freelist_batch, rcu);
 	unsigned int i;
 
-	for (i = 0; i < batch->index; i++)
-		pgtable_free(batch->tables[i]);
+	for (i = 0; i < batch->index; i++) {
+		void *table = (void *)(batch->tables[i] & ~PGF_SHIFT_MASK);
+		unsigned shift = batch->tables[i] & PGF_SHIFT_MASK;
+
+		pgtable_free(table, shift);
+	}
 
 	free_page((unsigned long)batch);
 }
@@ -89,25 +93,28 @@ static void pte_free_submit(struct pte_f
 	call_rcu(&batch->rcu, pte_free_rcu_callback);
 }
 
-void pgtable_free_tlb(struct mmu_gather *tlb, pgtable_free_t pgf)
+void pgtable_free_tlb(struct mmu_gather *tlb, void *table, unsigned shift)
 {
 	/* This is safe since tlb_gather_mmu has disabled preemption */
 	struct pte_freelist_batch **batchp = &__get_cpu_var(pte_freelist_cur);
+	unsigned long pgf;
 
 	if (atomic_read(&tlb->mm->mm_users) < 2 ||
 	    cpumask_equal(mm_cpumask(tlb->mm), cpumask_of(smp_processor_id()))){
-		pgtable_free(pgf);
+		pgtable_free(table, shift);
 		return;
 	}
 
 	if (*batchp == NULL) {
 		*batchp = (struct pte_freelist_batch *)__get_free_page(GFP_ATOMIC);
 		if (*batchp == NULL) {
-			pgtable_free_now(pgf);
+			pgtable_free_now(table, shift);
 			return;
 		}
 		(*batchp)->index = 0;
 	}
+	BUG_ON(shift > (PGF_SHIFT_MASK + 1));
+	pgf = (unsigned long)table | (shift - 1);
 	(*batchp)->tables[(*batchp)->index++] = pgf;
 	if ((*batchp)->index == PTE_FREELIST_SIZE) {
 		pte_free_submit(*batchp);
arch/powerpc/mm/hugetlbpage.c +12 -33

86 lines not on this page · raw message

arch/powerpc/include/asm/pgtable-ppc64.h +1 -0
Index: working-2.6/arch/powerpc/include/asm/pgtable-ppc64.h
===================================================================
@@ -354,6 +354,7 @@ static inline void __ptep_set_access_fla
 #define pgoff_to_pte(off)	((pte_t) {((off) << PTE_RPN_SHIFT)|_PAGE_FILE})
 #define PTE_FILE_MAX_BITS	(BITS_PER_LONG - PTE_RPN_SHIFT)
 
+void pgtable_cache_add(unsigned shift, void (*ctor)(void *));
 void pgtable_cache_init(void);
 
 /*
#3 lore raw

PATCH 3/3 Allow more flexible layouts for hugepage pagetables

In reply to David Gibson

Currently each available hugepage size uses a slightly different pagetable layout: that is, the bottem level table of pointers to hugepages is a different size, and may branch off from the normal page tables at a different level. Every hugepage aware path that needs to walk the pagetables must therefore look up the hugepage size from the slice info first, and work out the correct way to walk the pagetables accordingly. Future hardware is likely to add more possible hugepage sizes, more layout options and more mess.

This patch, therefore reworks the handling of hugepage pagetables to reduce this complexity. In the new scheme, instead of having to consult the slice mask, pagetable walking code can check a flag in the PGD/PUD/PMD entries to see where to branch off to hugepage pagetables, and the entry also contains the information (eseentially hugepage shift) necessary to then interpret that table without recourse to the slice mask. This scheme can be extended neatly to handle multiple levels of self-describing "special" hugepage pagetables, although for now we assume only one level exists.

This approach means that only the pagetable allocation path needs to know how the pagetables should be set out. All other (hugepage) pagetable walking paths can just interpret the structure as they go.

There already was a flag bit in PGD/PUD/PMD entries for hugepage directory pointers, but it was only used for debug. We alter that flag bit to instead be a 0 in the MSB to indicate a hugepage pagetable pointer (normally it would be 1 since the pointer lies in the linear mapping). This means that asm pagetable walking can test for (and punt on) hugepage pointers with the same test that checks for unpopulated page directory entries (beq becomes bge), since hugepage pointers will always be positive, and normal pointers always negative.

While we're at it, we get rid of the confusing (and grep defeating) #defining of hugepte_shift to be the same thing as mmu_huge_psizes.

---
 arch/powerpc/include/asm/hugetlb.h |    9 
 arch/powerpc/mm/hugetlbpage.c      |  396 +++++++++++++++----------------------
 arch/powerpc/mm/init_64.c          |   11 -
 3 files changed, 181 insertions(+), 235 deletions(-)
arch/powerpc/mm/hugetlbpage.c +162 -234
Index: working-2.6/arch/powerpc/mm/hugetlbpage.c
===================================================================
@@ -42,23 +42,9 @@ static unsigned nr_gpages;
  */
 unsigned int mmu_huge_psizes[MMU_PAGE_COUNT] = { }; /* initialize all to 0 */
 
-#define hugepte_shift			mmu_huge_psizes
-#define HUGEPTE_INDEX_SIZE(psize)	(mmu_huge_psizes[(psize)])
-#define PTRS_PER_HUGEPTE(psize)		(1 << mmu_huge_psizes[psize])
-
-#define HUGEPD_SHIFT(psize)		(mmu_psize_to_shift(psize) \
-					 + HUGEPTE_INDEX_SIZE(psize))
-#define HUGEPD_SIZE(psize)		(1UL << HUGEPD_SHIFT(psize))
-#define HUGEPD_MASK(psize)		(~(HUGEPD_SIZE(psize)-1))
-
 /* Flag to mark huge PD pointers.  This means pmd_bad() and pud_bad()
  * will choke on pointers to hugepte tables, which is handy for
  * catching screwups early. */
-#define HUGEPD_OK	0x1
-
-typedef struct { unsigned long pd; } hugepd_t;
-
-#define hugepd_none(hpd)	((hpd).pd == 0)
 
 static inline int shift_to_mmu_psize(unsigned int shift)
 {
@@ -82,71 +68,127 @@ static inline unsigned int mmu_psize_to_
 	BUG();
 }
 
+#define hugepd_none(hpd)	((hpd).pd == 0)
+
 static inline pte_t *hugepd_page(hugepd_t hpd)
 {
-	BUG_ON(!(hpd.pd & HUGEPD_OK));
-	return (pte_t *)(hpd.pd & ~HUGEPD_OK);
+	BUG_ON(!hugepd_ok(hpd));
+	return (pte_t *)((hpd.pd & ~HUGEPD_SHIFT_MASK) | 0xc000000000000000);
+}
+
+static inline unsigned int hugepd_shift(hugepd_t hpd)
+{
+	return hpd.pd & HUGEPD_SHIFT_MASK;
 }
 
-static inline pte_t *hugepte_offset(hugepd_t *hpdp, unsigned long addr,
-				    struct hstate *hstate)
+static inline pte_t *hugepte_offset(hugepd_t *hpdp, unsigned long addr, unsigned pdshift)
 {
-	unsigned int shift = huge_page_shift(hstate);
-	int psize = shift_to_mmu_psize(shift);
-	unsigned long idx = ((addr >> shift) & (PTRS_PER_HUGEPTE(psize)-1));
+	unsigned long idx = (addr & ((1UL << pdshift) - 1)) >> hugepd_shift(*hpdp);
 	pte_t *dir = hugepd_page(*hpdp);
 
 	return dir + idx;
 }
 
+static pte_t *huge_pte_offset_and_shift(struct mm_struct *mm,
+					unsigned long addr, unsigned *shift)
+{
+	pgd_t *pg;
+	pud_t *pu;
+	pmd_t *pm;
+	hugepd_t *hpdp = NULL;
+	unsigned pdshift = PGDIR_SHIFT;
+
+	if (shift)
+		*shift = 0;
+
+	pg = pgd_offset(mm, addr);
+	if (is_hugepd(pg)) {
+		hpdp = (hugepd_t *)pg;
+	} else if (!pgd_none(*pg)) {
+		pdshift = PUD_SHIFT;
+		pu = pud_offset(pg, addr);
+		if (is_hugepd(pu))
+			hpdp = (hugepd_t *)pu;
+		else if (!pud_none(*pu)) {
+			pdshift = PMD_SHIFT;
+			pm = pmd_offset(pu, addr);
+			if (is_hugepd(pm))
+				hpdp = (hugepd_t *)pm;
+			else if (!pmd_none(*pm)) {
+				return pte_offset_map(pm, addr);
+			}
+		}
+	}
+
+	if (!hpdp)
+		return NULL;
+
+	if (shift)
+		*shift = hugepd_shift(*hpdp);
+	return hugepte_offset(hpdp, addr, pdshift);
+}
+
+pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
+{
+	return huge_pte_offset_and_shift(mm, addr, NULL);
+}
+
 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
-			   unsigned long address, unsigned int psize)
+			   unsigned long address, unsigned pdshift, unsigned pshift)
 {
-	pte_t *new = kmem_cache_zalloc(PGT_CACHE(hugepte_shift[psize]),
+	pte_t *new = kmem_cache_zalloc(PGT_CACHE(pdshift - pshift),
 				       GFP_KERNEL|__GFP_REPEAT);
 
+	BUG_ON(pshift > HUGEPD_SHIFT_MASK);
+	BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
+
 	if (! new)
 		return -ENOMEM;
 
 	spin_lock(&mm->page_table_lock);
 	if (!hugepd_none(*hpdp))
-		kmem_cache_free(PGT_CACHE(hugepte_shift[psize]), new);
+		kmem_cache_free(PGT_CACHE(pdshift - pshift), new);
 	else
-		hpdp->pd = (unsigned long)new | HUGEPD_OK;
+		hpdp->pd = ((unsigned long)new & ~0x8000000000000000) | pshift;
 	spin_unlock(&mm->page_table_lock);
 	return 0;
 }
 
-
-static pud_t *hpud_offset(pgd_t *pgd, unsigned long addr, struct hstate *hstate)
-{
-	if (huge_page_shift(hstate) < PUD_SHIFT)
-		return pud_offset(pgd, addr);
-	else
-		return (pud_t *) pgd;
-}
-static pud_t *hpud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long addr,
-			 struct hstate *hstate)
+pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 {
-	if (huge_page_shift(hstate) < PUD_SHIFT)
-		return pud_alloc(mm, pgd, addr);
-	else
-		return (pud_t *) pgd;
-}
-static pmd_t *hpmd_offset(pud_t *pud, unsigned long addr, struct hstate *hstate)
-{
-	if (huge_page_shift(hstate) < PMD_SHIFT)
-		return pmd_offset(pud, addr);
-	else
-		return (pmd_t *) pud;
-}
-static pmd_t *hpmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long addr,
-			 struct hstate *hstate)
-{
-	if (huge_page_shift(hstate) < PMD_SHIFT)
-		return pmd_alloc(mm, pud, addr);
-	else
-		return (pmd_t *) pud;
+	pgd_t *pg;
+	pud_t *pu;
+	pmd_t *pm;
+	hugepd_t *hpdp = NULL;
+	unsigned pshift = __ffs(sz);
+	unsigned pdshift = PGDIR_SHIFT;
+
+	addr &= ~(sz-1);
+
+	pg = pgd_offset(mm, addr);
+	if (pshift >= PUD_SHIFT) {
+		hpdp = (hugepd_t *)pg;
+	} else {
+		pdshift = PUD_SHIFT;
+		pu = pud_alloc(mm, pg, addr);
+		if (pshift >= PMD_SHIFT) {
+			hpdp = (hugepd_t *)pu;
+		} else {
+			pdshift = PMD_SHIFT;
+			pm = pmd_alloc(mm, pu, addr);
+			hpdp = (hugepd_t *)pm;
+		}
+	}
+
+	if (!hpdp)
+		return NULL;
+
+	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
+
+	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
+		return NULL;
+
+	return hugepte_offset(hpdp, addr, pdshift);
 }
 
 /* Build list of addresses of gigantic pages.  This function is used in early
@@ -180,92 +222,38 @@ int alloc_bootmem_huge_page(struct hstat
 	return 1;
 }
 
-
-/* Modelled after find_linux_pte() */
-pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
-{
-	pgd_t *pg;
-	pud_t *pu;
-	pmd_t *pm;
-
-	unsigned int psize;
-	unsigned int shift;
-	unsigned long sz;
-	struct hstate *hstate;
-	psize = get_slice_psize(mm, addr);
-	shift = mmu_psize_to_shift(psize);
-	sz = ((1UL) << shift);
-	hstate = size_to_hstate(sz);
-
-	addr &= hstate->mask;
-
-	pg = pgd_offset(mm, addr);
-	if (!pgd_none(*pg)) {
-		pu = hpud_offset(pg, addr, hstate);
-		if (!pud_none(*pu)) {
-			pm = hpmd_offset(pu, addr, hstate);
-			if (!pmd_none(*pm))
-				return hugepte_offset((hugepd_t *)pm, addr,
-						      hstate);
-		}
-	}
-
-	return NULL;
-}
-
-pte_t *huge_pte_alloc(struct mm_struct *mm,
-			unsigned long addr, unsigned long sz)
-{
-	pgd_t *pg;
-	pud_t *pu;
-	pmd_t *pm;
-	hugepd_t *hpdp = NULL;
-	struct hstate *hstate;
-	unsigned int psize;
-	hstate = size_to_hstate(sz);
-
-	psize = get_slice_psize(mm, addr);
-	BUG_ON(!mmu_huge_psizes[psize]);
-
-	addr &= hstate->mask;
-
-	pg = pgd_offset(mm, addr);
-	pu = hpud_alloc(mm, pg, addr, hstate);
-
-	if (pu) {
-		pm = hpmd_alloc(mm, pu, addr, hstate);
-		if (pm)
-			hpdp = (hugepd_t *)pm;
-	}
-
-	if (! hpdp)
-		return NULL;
-
-	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, psize))
-		return NULL;
-
-	return hugepte_offset(hpdp, addr, hstate);
-}
-
 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 {
 	return 0;
 }
 
-static void free_hugepte_range(struct mmu_gather *tlb, hugepd_t *hpdp,
-			       unsigned int psize)
+static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
+			      unsigned long start, unsigned long end,
+			      unsigned long floor, unsigned long ceiling)
 {
 	pte_t *hugepte = hugepd_page(*hpdp);
+	unsigned shift = hugepd_shift(*hpdp);
+	unsigned long pdmask = ~((1UL << pdshift) - 1);
+
+	start &= pdmask;
+	if (start < floor)
+		return;
+	if (ceiling) {
+		ceiling &= pdmask;
+		if (! ceiling)
+			return;
+	}
+	if (end - 1 > ceiling - 1)
+		return;
 
 	hpdp->pd = 0;
 	tlb->need_flush = 1;
-	pgtable_free_tlb(tlb, hugepte, hugepte_shift[psize]);
+	pgtable_free_tlb(tlb, hugepte, pdshift - shift);
 }
 
 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 				   unsigned long addr, unsigned long end,
-				   unsigned long floor, unsigned long ceiling,
-				   unsigned int psize)
+				   unsigned long floor, unsigned long ceiling)
 {
 	pmd_t *pmd;
 	unsigned long next;
@@ -277,7 +265,8 @@ static void hugetlb_free_pmd_range(struc
 		next = pmd_addr_end(addr, end);
 		if (pmd_none(*pmd))
 			continue;
-		free_hugepte_range(tlb, (hugepd_t *)pmd, psize);
+		free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
+				  addr, next, floor, ceiling);
 	} while (pmd++, addr = next, addr != end);
 
 	start &= PUD_MASK;
@@ -303,23 +292,19 @@ static void hugetlb_free_pud_range(struc
 	pud_t *pud;
 	unsigned long next;
 	unsigned long start;
-	unsigned int shift;
-	unsigned int psize = get_slice_psize(tlb->mm, addr);
-	shift = mmu_psize_to_shift(psize);
 
 	start = addr;
 	pud = pud_offset(pgd, addr);
 	do {
 		next = pud_addr_end(addr, end);
-		if (shift < PMD_SHIFT) {
+		if (!is_hugepd(pud)) {
 			if (pud_none_or_clear_bad(pud))
 				continue;
 			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
-					       ceiling, psize);
+					       ceiling);
 		} else {
-			if (pud_none(*pud))
-				continue;
-			free_hugepte_range(tlb, (hugepd_t *)pud, psize);
+			free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
+					  addr, next, floor, ceiling);
 		}
 	} while (pud++, addr = next, addr != end);
 
@@ -350,74 +335,34 @@ void hugetlb_free_pgd_range(struct mmu_g
 {
 	pgd_t *pgd;
 	unsigned long next;
-	unsigned long start;
 
 	/*
-	 * Comments below take from the normal free_pgd_range().  They
-	 * apply here too.  The tests against HUGEPD_MASK below are
-	 * essential, because we *don't* test for this at the bottom
-	 * level.  Without them we'll attempt to free a hugepte table
-	 * when we unmap just part of it, even if there are other
-	 * active mappings using it.
-	 *
-	 * The next few lines have given us lots of grief...
+	 * Because there are a number of different possible pagetable
+	 * layouts for hugepage ranges, we limit knowledge of how
+	 * things should be laid out to the allocation path
+	 * (huge_pte_alloc(), above).  Everything else works out the
+	 * structure as it goes from information in the hugepd
+	 * pointers.  That means that we can't here use the
+	 * optimization used in the normal page free_pgd_range(), of
+	 * checking whether we're actually covering a large enough
+	 * range to have to do anything at the top level of the walk
+	 * instead of at the bottom.
 	 *
-	 * Why are we testing HUGEPD* at this top level?  Because
-	 * often there will be no work to do at all, and we'd prefer
-	 * not to go all the way down to the bottom just to discover
-	 * that.
-	 *
-	 * Why all these "- 1"s?  Because 0 represents both the bottom
-	 * of the address space and the top of it (using -1 for the
-	 * top wouldn't help much: the masks would do the wrong thing).
-	 * The rule is that addr 0 and floor 0 refer to the bottom of
-	 * the address space, but end 0 and ceiling 0 refer to the top
-	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
-	 * that end 0 case should be mythical).
-	 *
-	 * Wherever addr is brought up or ceiling brought down, we
-	 * must be careful to reject "the opposite 0" before it
-	 * confuses the subsequent tests.  But what about where end is
-	 * brought down by HUGEPD_SIZE below? no, end can't go down to
-	 * 0 there.
-	 *
-	 * Whereas we round start (addr) and ceiling down, by different
-	 * masks at different levels, in order to test whether a table
-	 * now has no other vmas using it, so can be freed, we don't
-	 * bother to round floor or end up - the tests don't need that.
+	 * To make sense of this, you should probably go read the big
+	 * block comment at the top of the normal free_pgd_range(),
+	 * too.
 	 */
-	unsigned int psize = get_slice_psize(tlb->mm, addr);
-
-	addr &= HUGEPD_MASK(psize);
-	if (addr < floor) {
-		addr += HUGEPD_SIZE(psize);
-		if (!addr)
-			return;
-	}
-	if (ceiling) {
-		ceiling &= HUGEPD_MASK(psize);
-		if (!ceiling)
-			return;
-	}
-	if (end - 1 > ceiling - 1)
-		end -= HUGEPD_SIZE(psize);
-	if (addr > end - 1)
-		return;
 
-	start = addr;
 	pgd = pgd_offset(tlb->mm, addr);
 	do {
-		psize = get_slice_psize(tlb->mm, addr);
-		BUG_ON(!mmu_huge_psizes[psize]);
 		next = pgd_addr_end(addr, end);
-		if (mmu_psize_to_shift(psize) < PUD_SHIFT) {
+		if (!is_hugepd(pgd)) {
 			if (pgd_none_or_clear_bad(pgd))
 				continue;
 			hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 		} else {
-			if (pgd_none(*pgd))
-				continue;
-			free_hugepte_range(tlb, (hugepd_t *)pgd, psize);
+			free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
+					  addr, next, floor, ceiling);
 		}
 	} while (pgd++, addr = next, addr != end);
 }
@@ -448,19 +393,19 @@ follow_huge_addr(struct mm_struct *mm, u
 {
 	pte_t *ptep;
 	struct page *page;
-	unsigned int mmu_psize = get_slice_psize(mm, address);
+	unsigned shift;
+	unsigned long mask;
+
+	ptep = huge_pte_offset_and_shift(mm, address, &shift);
 
 	/* Verify it is a huge page else bail. */
-	if (!mmu_huge_psizes[mmu_psize])
+	if (!ptep || !shift)
 		return ERR_PTR(-EINVAL);
 
-	ptep = huge_pte_offset(mm, address);
+	mask = (1UL << shift) - 1;
 	page = pte_page(*ptep);
-	if (page) {
-		unsigned int shift = mmu_psize_to_shift(mmu_psize);
-		unsigned long sz = ((1UL) << shift);
-		page += (address % sz) / PAGE_SIZE;
-	}
+	if (page)
+		page += (address & mask) / PAGE_SIZE;
 
 	return page;
 }
@@ -541,21 +486,18 @@ int hash_huge_page(struct mm_struct *mm,
 	int err = 1;
 	int ssize = user_segment_size(ea);
 	unsigned int mmu_psize;
-	int shift;
+	unsigned shift;
 	mmu_psize = get_slice_psize(mm, ea);
 
-	if (!mmu_huge_psizes[mmu_psize])
-		goto out;
-	ptep = huge_pte_offset(mm, ea);
-
+	ptep = huge_pte_offset_and_shift(mm, ea, &shift);
 	/* Search the Linux page table for a match with va */
 	va = hpt_va(ea, vsid, ssize);
 
 	/*
-	 * If no pte found or not present, send the problem up to
-	 * do_page_fault
+	 * If no pte found or not present, or it's not a hugepage pte,
+	 * send the problem up to do_page_fault
 	 */
-	if (unlikely(!ptep || pte_none(*ptep)))
+	if (unlikely(!ptep || pte_none(*ptep) || !shift))
 		goto out;
 
 	/* 
@@ -588,7 +530,6 @@ int hash_huge_page(struct mm_struct *mm,
 	rflags = 0x2 | (!(new_pte & _PAGE_RW));
  	/* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
 	rflags |= ((new_pte & _PAGE_EXEC) ? 0 : HPTE_R_N);
-	shift = mmu_psize_to_shift(mmu_psize);
 	sz = ((1UL) << shift);
 	if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
 		/* No CPU has hugepages but lacks no execute, so we
@@ -672,6 +613,8 @@ repeat:
 
 static void __init set_huge_psize(int psize)
 {
+	unsigned pdshift;
+
 	/* Check that it is a page size supported by the hardware and
 	 * that it fits within pagetable limits. */
 	if (mmu_psize_defs[psize].shift &&
@@ -686,29 +629,14 @@ static void __init set_huge_psize(int ps
 			return;
 		hugetlb_add_hstate(mmu_psize_defs[psize].shift - PAGE_SHIFT);
 
-		switch (mmu_psize_defs[psize].shift) {
-		case PAGE_SHIFT_64K:
-		    /* We only allow 64k hpages with 4k base page,
-		     * which was checked above, and always put them
-		     * at the PMD */
-		    hugepte_shift[psize] = PMD_SHIFT;
-		    break;
-		case PAGE_SHIFT_16M:
-		    /* 16M pages can be at two different levels
-		     * of pagestables based on base page size */
-		    if (PAGE_SHIFT == PAGE_SHIFT_64K)
-			    hugepte_shift[psize] = PMD_SHIFT;
-		    else /* 4k base page */
-			    hugepte_shift[psize] = PUD_SHIFT;
-		    break;
-		case PAGE_SHIFT_16G:
-		    /* 16G pages are always at PGD level */
-		    hugepte_shift[psize] = PGDIR_SHIFT;
-		    break;
-		}
-		hugepte_shift[psize] -= mmu_psize_defs[psize].shift;
-	} else
-		hugepte_shift[psize] = 0;
+		if (mmu_psize_defs[psize].shift < PMD_SHIFT)
+			pdshift = PMD_SHIFT;
+		else if (mmu_psize_defs[psize].shift < PUD_SHIFT)
+			pdshift = PUD_SHIFT;
+		else
+			pdshift = PGDIR_SHIFT;
+		mmu_huge_psizes[psize] = pdshift - mmu_psize_defs[psize].shift;
+	}
 }
 
 static int __init hugepage_setup_sz(char *str)
@@ -732,7 +660,7 @@ __setup("hugepagesz=", hugepage_setup_sz
 
 static int __init hugetlbpage_init(void)
 {
-	unsigned int psize;
+	int psize;
 
 	if (!cpu_has_feature(CPU_FTR_16M_PAGE))
 		return -ENODEV;
@@ -753,8 +681,8 @@ static int __init hugetlbpage_init(void)
 
 	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
 		if (mmu_huge_psizes[psize]) {
-			pgtable_cache_add(hugepte_shift[psize], NULL);
-			if (!PGT_CACHE(hugepte_shift[psize]))
+			pgtable_cache_add(mmu_huge_psizes[psize], NULL);
+			if (!PGT_CACHE(mmu_huge_psizes[psize]))
 				panic("hugetlbpage_init(): could not create "
 				      "pgtable cache for %d bit pagesize\n",
 				      mmu_psize_to_shift(psize));
arch/powerpc/include/asm/hugetlb.h +9 -0
Index: working-2.6/arch/powerpc/include/asm/hugetlb.h
===================================================================
@@ -3,6 +3,15 @@
 
 #include <asm/page.h>
 
+typedef struct { signed long pd; } hugepd_t;
+
+static inline int hugepd_ok(hugepd_t hpd)
+{
+	return (hpd.pd > 0);
+}
+
+#define is_hugepd(pdep)               (hugepd_ok(*((hugepd_t *)(pdep))))
+#define HUGEPD_SHIFT_MASK     0x3f
 
 int is_hugepage_only_range(struct mm_struct *mm, unsigned long addr,
 			   unsigned long len);
arch/powerpc/mm/init_64.c +10 -1
Index: working-2.6/arch/powerpc/mm/init_64.c
===================================================================
@@ -41,6 +41,7 @@
 #include <linux/module.h>
 #include <linux/poison.h>
 #include <linux/lmb.h>
+#include <linux/hugetlb.h>
 
 #include <asm/pgalloc.h>
 #include <asm/page.h>
@@ -154,13 +155,21 @@ void pgtable_cache_add(unsigned shift, v
 {
 	char *name;
 	unsigned long table_size = sizeof(void *) << shift;
+	unsigned long align = table_size;
 	struct kmem_cache *new;
 
 	BUG_ON((shift < 1) || (shift > PGF_SHIFT_MASK));
+#ifdef CONFIG_HUGETLB_PAGE
+	/* We use low bits in hugepage dir pointers to store index
+	 * size information.  Table alignment must be big enough to
+	 * fit it. */
+	align = max_t(unsigned long, align, HUGEPD_SHIFT_MASK + 1);
+#endif
+
 	if (PGT_CACHE(shift))
 		return; /* Already have a cache of this size */
 	name = kasprintf(GFP_KERNEL, "pgtable-2^%d", shift);
-	new = kmem_cache_create(name, table_size, table_size, 0, ctor);
+	new = kmem_cache_create(name, table_size, align, 0, ctor);
 	PGT_CACHE(shift) = new;
 }