hugetlb: modular state for hugetlb page size
authorAndi Kleen <ak@suse.de>
Thu, 24 Jul 2008 04:27:41 +0000 (21:27 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 24 Jul 2008 17:47:17 +0000 (10:47 -0700)
The goal of this patchset is to support multiple hugetlb page sizes.  This
is achieved by introducing a new struct hstate structure, which
encapsulates the important hugetlb state and constants (eg.  huge page
size, number of huge pages currently allocated, etc).

The hstate structure is then passed around the code which requires these
fields, they will do the right thing regardless of the exact hstate they
are operating on.

This patch adds the hstate structure, with a single global instance of it
(default_hstate), and does the basic work of converting hugetlb to use the
hstate.

Future patches will add more hstate structures to allow for different
hugetlbfs mounts to have different page sizes.

[akpm@linux-foundation.org: coding-style fixes]
Acked-by: Adam Litke <agl@us.ibm.com>
Acked-by: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
19 files changed:
arch/ia64/mm/hugetlbpage.c
arch/powerpc/mm/hugetlbpage.c
arch/s390/mm/hugetlbpage.c
arch/sh/mm/hugetlbpage.c
arch/sparc64/mm/hugetlbpage.c
arch/x86/mm/hugetlbpage.c
fs/hugetlbfs/inode.c
include/asm-ia64/hugetlb.h
include/asm-powerpc/hugetlb.h
include/asm-s390/hugetlb.h
include/asm-sh/hugetlb.h
include/asm-sparc/hugetlb.h
include/asm-x86/hugetlb.h
include/linux/hugetlb.h
ipc/shm.c
mm/hugetlb.c
mm/memory.c
mm/mempolicy.c
mm/mmap.c

index cd49e2860eef3beea11d28fcd2f26dcf666072b7..6170f097d2554d60f2828302393172e57d3839b8 100644 (file)
@@ -24,7 +24,7 @@
 unsigned int hpage_shift=HPAGE_SHIFT_DEFAULT;
 
 pte_t *
-huge_pte_alloc (struct mm_struct *mm, unsigned long addr)
+huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
 {
        unsigned long taddr = htlbpage_to_page(addr);
        pgd_t *pgd;
@@ -75,7 +75,8 @@ int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
  * Don't actually need to do any preparation, but need to make sure
  * the address is in the right region.
  */
-int prepare_hugepage_range(unsigned long addr, unsigned long len)
+int prepare_hugepage_range(struct file *file,
+                       unsigned long addr, unsigned long len)
 {
        if (len & ~HPAGE_MASK)
                return -EINVAL;
@@ -149,7 +150,7 @@ unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, u
 
        /* Handle MAP_FIXED */
        if (flags & MAP_FIXED) {
-               if (prepare_hugepage_range(addr, len))
+               if (prepare_hugepage_range(file, addr, len))
                        return -EINVAL;
                return addr;
        }
index 1a96cc891cf5f718c4ed87484360f29789bc5c05..c94dc71af989277007c33ee3f804ee561d944c33 100644 (file)
@@ -128,7 +128,8 @@ pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
        return NULL;
 }
 
-pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
+pte_t *huge_pte_alloc(struct mm_struct *mm,
+                       unsigned long addr, unsigned long sz)
 {
        pgd_t *pg;
        pud_t *pu;
index f4b6124fdb75d1308d6b112081fab60cd039da9d..9162dc84f77f685c6f66383276660d6d7c049a71 100644 (file)
@@ -72,7 +72,8 @@ void arch_release_hugepage(struct page *page)
        page[1].index = 0;
 }
 
-pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
+pte_t *huge_pte_alloc(struct mm_struct *mm,
+                       unsigned long addr, unsigned long sz)
 {
        pgd_t *pgdp;
        pud_t *pudp;
index ae8c321d6e2a960e0e7fe1bd55ab5224e7a586cc..2f9dbe0ef4acd57144e8575919aadd0442ddc054 100644 (file)
@@ -22,7 +22,8 @@
 #include <asm/tlbflush.h>
 #include <asm/cacheflush.h>
 
-pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
+pte_t *huge_pte_alloc(struct mm_struct *mm,
+                       unsigned long addr, unsigned long sz)
 {
        pgd_t *pgd;
        pud_t *pud;
index ebefd2a14375b25104030cf04d649dc9b7da72fd..1307b23f6a7636e1d040a5a16232e283da8578fe 100644 (file)
@@ -175,7 +175,7 @@ hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
                return -ENOMEM;
 
        if (flags & MAP_FIXED) {
-               if (prepare_hugepage_range(addr, len))
+               if (prepare_hugepage_range(file, addr, len))
                        return -EINVAL;
                return addr;
        }
@@ -195,7 +195,8 @@ hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
                                pgoff, flags);
 }
 
-pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
+pte_t *huge_pte_alloc(struct mm_struct *mm,
+                       unsigned long addr, unsigned long sz)
 {
        pgd_t *pgd;
        pud_t *pud;
index 0b3d567e686df15c1b042deb388091ab78a6ae25..52476fde8996669d7eed87e7f96ea2984a3a308a 100644 (file)
@@ -124,7 +124,8 @@ int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
        return 1;
 }
 
-pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
+pte_t *huge_pte_alloc(struct mm_struct *mm,
+                       unsigned long addr, unsigned long sz)
 {
        pgd_t *pgd;
        pud_t *pud;
@@ -368,7 +369,7 @@ hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
                return -ENOMEM;
 
        if (flags & MAP_FIXED) {
-               if (prepare_hugepage_range(addr, len))
+               if (prepare_hugepage_range(file, addr, len))
                        return -EINVAL;
                return addr;
        }
index 428eff5b73f35a399717590e8f6b6b8492b6a239..516c581b53714aa350e24a08215b37c9a00b194c 100644 (file)
@@ -80,6 +80,7 @@ static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
        struct inode *inode = file->f_path.dentry->d_inode;
        loff_t len, vma_len;
        int ret;
+       struct hstate *h = hstate_file(file);
 
        /*
         * vma address alignment (but not the pgoff alignment) has
@@ -92,7 +93,7 @@ static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
        vma->vm_flags |= VM_HUGETLB | VM_RESERVED;
        vma->vm_ops = &hugetlb_vm_ops;
 
-       if (vma->vm_pgoff & ~(HPAGE_MASK >> PAGE_SHIFT))
+       if (vma->vm_pgoff & ~(huge_page_mask(h) >> PAGE_SHIFT))
                return -EINVAL;
 
        vma_len = (loff_t)(vma->vm_end - vma->vm_start);
@@ -104,8 +105,8 @@ static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
        len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 
        if (hugetlb_reserve_pages(inode,
-                               vma->vm_pgoff >> (HPAGE_SHIFT-PAGE_SHIFT),
-                               len >> HPAGE_SHIFT, vma))
+                               vma->vm_pgoff >> huge_page_order(h),
+                               len >> huge_page_shift(h), vma))
                goto out;
 
        ret = 0;
@@ -130,20 +131,21 @@ hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
        struct mm_struct *mm = current->mm;
        struct vm_area_struct *vma;
        unsigned long start_addr;
+       struct hstate *h = hstate_file(file);
 
-       if (len & ~HPAGE_MASK)
+       if (len & ~huge_page_mask(h))
                return -EINVAL;
        if (len > TASK_SIZE)
                return -ENOMEM;
 
        if (flags & MAP_FIXED) {
-               if (prepare_hugepage_range(addr, len))
+               if (prepare_hugepage_range(file, addr, len))
                        return -EINVAL;
                return addr;
        }
 
        if (addr) {
-               addr = ALIGN(addr, HPAGE_SIZE);
+               addr = ALIGN(addr, huge_page_size(h));
                vma = find_vma(mm, addr);
                if (TASK_SIZE - len >= addr &&
                    (!vma || addr + len <= vma->vm_start))
@@ -156,7 +158,7 @@ hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
                start_addr = TASK_UNMAPPED_BASE;
 
 full_search:
-       addr = ALIGN(start_addr, HPAGE_SIZE);
+       addr = ALIGN(start_addr, huge_page_size(h));
 
        for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
                /* At this point:  (!vma || addr < vma->vm_end). */
@@ -174,7 +176,7 @@ full_search:
 
                if (!vma || addr + len <= vma->vm_start)
                        return addr;
-               addr = ALIGN(vma->vm_end, HPAGE_SIZE);
+               addr = ALIGN(vma->vm_end, huge_page_size(h));
        }
 }
 #endif
@@ -225,10 +227,11 @@ hugetlbfs_read_actor(struct page *page, unsigned long offset,
 static ssize_t hugetlbfs_read(struct file *filp, char __user *buf,
                              size_t len, loff_t *ppos)
 {
+       struct hstate *h = hstate_file(filp);
        struct address_space *mapping = filp->f_mapping;
        struct inode *inode = mapping->host;
-       unsigned long index = *ppos >> HPAGE_SHIFT;
-       unsigned long offset = *ppos & ~HPAGE_MASK;
+       unsigned long index = *ppos >> huge_page_shift(h);
+       unsigned long offset = *ppos & ~huge_page_mask(h);
        unsigned long end_index;
        loff_t isize;
        ssize_t retval = 0;
@@ -243,17 +246,17 @@ static ssize_t hugetlbfs_read(struct file *filp, char __user *buf,
        if (!isize)
                goto out;
 
-       end_index = (isize - 1) >> HPAGE_SHIFT;
+       end_index = (isize - 1) >> huge_page_shift(h);
        for (;;) {
                struct page *page;
-               int nr, ret;
+               unsigned long nr, ret;
 
                /* nr is the maximum number of bytes to copy from this page */
-               nr = HPAGE_SIZE;
+               nr = huge_page_size(h);
                if (index >= end_index) {
                        if (index > end_index)
                                goto out;
-                       nr = ((isize - 1) & ~HPAGE_MASK) + 1;
+                       nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
                        if (nr <= offset) {
                                goto out;
                        }
@@ -287,8 +290,8 @@ static ssize_t hugetlbfs_read(struct file *filp, char __user *buf,
                offset += ret;
                retval += ret;
                len -= ret;
-               index += offset >> HPAGE_SHIFT;
-               offset &= ~HPAGE_MASK;
+               index += offset >> huge_page_shift(h);
+               offset &= ~huge_page_mask(h);
 
                if (page)
                        page_cache_release(page);
@@ -298,7 +301,7 @@ static ssize_t hugetlbfs_read(struct file *filp, char __user *buf,
                        break;
        }
 out:
-       *ppos = ((loff_t)index << HPAGE_SHIFT) + offset;
+       *ppos = ((loff_t)index << huge_page_shift(h)) + offset;
        mutex_unlock(&inode->i_mutex);
        return retval;
 }
@@ -339,8 +342,9 @@ static void truncate_huge_page(struct page *page)
 
 static void truncate_hugepages(struct inode *inode, loff_t lstart)
 {
+       struct hstate *h = hstate_inode(inode);
        struct address_space *mapping = &inode->i_data;
-       const pgoff_t start = lstart >> HPAGE_SHIFT;
+       const pgoff_t start = lstart >> huge_page_shift(h);
        struct pagevec pvec;
        pgoff_t next;
        int i, freed = 0;
@@ -449,8 +453,9 @@ static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 {
        pgoff_t pgoff;
        struct address_space *mapping = inode->i_mapping;
+       struct hstate *h = hstate_inode(inode);
 
-       BUG_ON(offset & ~HPAGE_MASK);
+       BUG_ON(offset & ~huge_page_mask(h));
        pgoff = offset >> PAGE_SHIFT;
 
        i_size_write(inode, offset);
@@ -465,6 +470,7 @@ static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 {
        struct inode *inode = dentry->d_inode;
+       struct hstate *h = hstate_inode(inode);
        int error;
        unsigned int ia_valid = attr->ia_valid;
 
@@ -476,7 +482,7 @@ static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 
        if (ia_valid & ATTR_SIZE) {
                error = -EINVAL;
-               if (!(attr->ia_size & ~HPAGE_MASK))
+               if (!(attr->ia_size & ~huge_page_mask(h)))
                        error = hugetlb_vmtruncate(inode, attr->ia_size);
                if (error)
                        goto out;
@@ -610,9 +616,10 @@ static int hugetlbfs_set_page_dirty(struct page *page)
 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 {
        struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
+       struct hstate *h = hstate_inode(dentry->d_inode);
 
        buf->f_type = HUGETLBFS_MAGIC;
-       buf->f_bsize = HPAGE_SIZE;
+       buf->f_bsize = huge_page_size(h);
        if (sbinfo) {
                spin_lock(&sbinfo->stat_lock);
                /* If no limits set, just report 0 for max/free/used
@@ -942,7 +949,8 @@ struct file *hugetlb_file_setup(const char *name, size_t size)
                goto out_dentry;
 
        error = -ENOMEM;
-       if (hugetlb_reserve_pages(inode, 0, size >> HPAGE_SHIFT, NULL))
+       if (hugetlb_reserve_pages(inode, 0,
+                       size >> huge_page_shift(hstate_inode(inode)), NULL))
                goto out_inode;
 
        d_instantiate(dentry, inode);
index e9d1e5e2382dfd6bff1b65e61e254b263f02d067..da55c63728e0e3f5d2b269d8c6bcb1607bddee57 100644 (file)
@@ -8,7 +8,8 @@ void hugetlb_free_pgd_range(struct mmu_gather *tlb, unsigned long addr,
                            unsigned long end, unsigned long floor,
                            unsigned long ceiling);
 
-int prepare_hugepage_range(unsigned long addr, unsigned long len);
+int prepare_hugepage_range(struct file *file,
+                       unsigned long addr, unsigned long len);
 
 static inline int is_hugepage_only_range(struct mm_struct *mm,
                                         unsigned long addr,
index 0a37aa5ecaa5bb86bea6e964fd2156ed2f926c27..ca37c4af27b1d872c1ae543df9ad9c50aaf7361c 100644 (file)
@@ -21,7 +21,8 @@ pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
  * If the arch doesn't supply something else, assume that hugepage
  * size aligned regions are ok without further preparation.
  */
-static inline int prepare_hugepage_range(unsigned long addr, unsigned long len)
+static inline int prepare_hugepage_range(struct file *file,
+                       unsigned long addr, unsigned long len)
 {
        if (len & ~HPAGE_MASK)
                return -EINVAL;
index 600a776f8f75121fb378ddda4e0aa82427ff7f14..670a1d1745d271e6e89a8246e6432c0c281b3d94 100644 (file)
@@ -22,7 +22,8 @@ void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
  * If the arch doesn't supply something else, assume that hugepage
  * size aligned regions are ok without further preparation.
  */
-static inline int prepare_hugepage_range(unsigned long addr, unsigned long len)
+static inline int prepare_hugepage_range(struct file *file,
+                       unsigned long addr, unsigned long len)
 {
        if (len & ~HPAGE_MASK)
                return -EINVAL;
index fb30018938c7d61e43c154907757beba98e42737..967068fb79ac483959ccd13436e1c14601133308 100644 (file)
@@ -14,7 +14,8 @@ static inline int is_hugepage_only_range(struct mm_struct *mm,
  * If the arch doesn't supply something else, assume that hugepage
  * size aligned regions are ok without further preparation.
  */
-static inline int prepare_hugepage_range(unsigned long addr, unsigned long len)
+static inline int prepare_hugepage_range(struct file *file,
+                       unsigned long addr, unsigned long len)
 {
        if (len & ~HPAGE_MASK)
                return -EINVAL;
index aeb92374ca3d663e6e0622fba55ffe93dbbcf20a..177061064ee602ed4dc0e8e956a4b682695c79f0 100644 (file)
@@ -22,7 +22,8 @@ static inline int is_hugepage_only_range(struct mm_struct *mm,
  * If the arch doesn't supply something else, assume that hugepage
  * size aligned regions are ok without further preparation.
  */
-static inline int prepare_hugepage_range(unsigned long addr, unsigned long len)
+static inline int prepare_hugepage_range(struct file *file,
+                       unsigned long addr, unsigned long len)
 {
        if (len & ~HPAGE_MASK)
                return -EINVAL;
index 7eed6e0883bf0ad6127e23088fe673924230c4e5..439a9acc132d10f77b469fc24b250f3acc2f51d9 100644 (file)
@@ -14,11 +14,13 @@ static inline int is_hugepage_only_range(struct mm_struct *mm,
  * If the arch doesn't supply something else, assume that hugepage
  * size aligned regions are ok without further preparation.
  */
-static inline int prepare_hugepage_range(unsigned long addr, unsigned long len)
+static inline int prepare_hugepage_range(struct file *file,
+                       unsigned long addr, unsigned long len)
 {
-       if (len & ~HPAGE_MASK)
+       struct hstate *h = hstate_file(file);
+       if (len & ~huge_page_mask(h))
                return -EINVAL;
-       if (addr & ~HPAGE_MASK)
+       if (addr & ~huge_page_mask(h))
                return -EINVAL;
        return 0;
 }
index abbc187193a1b9435b86daeb2462b4f20cc3b016..ad2271e11f9b0e731a317fef35dbee62945f0c1c 100644 (file)
@@ -8,7 +8,6 @@
 #include <linux/mempolicy.h>
 #include <linux/shm.h>
 #include <asm/tlbflush.h>
-#include <asm/hugetlb.h>
 
 struct ctl_table;
 
@@ -45,7 +44,8 @@ extern int sysctl_hugetlb_shm_group;
 
 /* arch callbacks */
 
-pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr);
+pte_t *huge_pte_alloc(struct mm_struct *mm,
+                       unsigned long addr, unsigned long sz);
 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr);
 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep);
 struct page *follow_huge_addr(struct mm_struct *mm, unsigned long address,
@@ -80,7 +80,7 @@ static inline unsigned long hugetlb_total_pages(void)
 #define hugetlb_report_meminfo(buf)            0
 #define hugetlb_report_node_meminfo(n, buf)    0
 #define follow_huge_pmd(mm, addr, pmd, write)  NULL
-#define prepare_hugepage_range(addr,len)       (-EINVAL)
+#define prepare_hugepage_range(file, addr, len)        (-EINVAL)
 #define pmd_huge(x)    0
 #define is_hugepage_only_range(mm, addr, len)  0
 #define hugetlb_free_pgd_range(tlb, addr, end, floor, ceiling) ({BUG(); 0; })
@@ -134,8 +134,6 @@ struct file *hugetlb_file_setup(const char *name, size_t);
 int hugetlb_get_quota(struct address_space *mapping, long delta);
 void hugetlb_put_quota(struct address_space *mapping, long delta);
 
-#define BLOCKS_PER_HUGEPAGE    (HPAGE_SIZE / 512)
-
 static inline int is_file_hugepages(struct file *file)
 {
        if (file->f_op == &hugetlbfs_file_operations)
@@ -164,4 +162,84 @@ unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
                                        unsigned long flags);
 #endif /* HAVE_ARCH_HUGETLB_UNMAPPED_AREA */
 
+#ifdef CONFIG_HUGETLB_PAGE
+
+/* Defines one hugetlb page size */
+struct hstate {
+       int hugetlb_next_nid;
+       unsigned int order;
+       unsigned long mask;
+       unsigned long max_huge_pages;
+       unsigned long nr_huge_pages;
+       unsigned long free_huge_pages;
+       unsigned long resv_huge_pages;
+       unsigned long surplus_huge_pages;
+       unsigned long nr_overcommit_huge_pages;
+       struct list_head hugepage_freelists[MAX_NUMNODES];
+       unsigned int nr_huge_pages_node[MAX_NUMNODES];
+       unsigned int free_huge_pages_node[MAX_NUMNODES];
+       unsigned int surplus_huge_pages_node[MAX_NUMNODES];
+};
+
+extern struct hstate default_hstate;
+
+static inline struct hstate *hstate_vma(struct vm_area_struct *vma)
+{
+       return &default_hstate;
+}
+
+static inline struct hstate *hstate_file(struct file *f)
+{
+       return &default_hstate;
+}
+
+static inline struct hstate *hstate_inode(struct inode *i)
+{
+       return &default_hstate;
+}
+
+static inline unsigned long huge_page_size(struct hstate *h)
+{
+       return (unsigned long)PAGE_SIZE << h->order;
+}
+
+static inline unsigned long huge_page_mask(struct hstate *h)
+{
+       return h->mask;
+}
+
+static inline unsigned int huge_page_order(struct hstate *h)
+{
+       return h->order;
+}
+
+static inline unsigned huge_page_shift(struct hstate *h)
+{
+       return h->order + PAGE_SHIFT;
+}
+
+static inline unsigned int pages_per_huge_page(struct hstate *h)
+{
+       return 1 << h->order;
+}
+
+static inline unsigned int blocks_per_huge_page(struct hstate *h)
+{
+       return huge_page_size(h) / 512;
+}
+
+#include <asm/hugetlb.h>
+
+#else
+struct hstate {};
+#define hstate_file(f) NULL
+#define hstate_vma(v) NULL
+#define hstate_inode(i) NULL
+#define huge_page_size(h) PAGE_SIZE
+#define huge_page_mask(h) PAGE_MASK
+#define huge_page_order(h) 0
+#define huge_page_shift(h) PAGE_SHIFT
+#define pages_per_huge_page(h) 1
+#endif
+
 #endif /* _LINUX_HUGETLB_H */
index 790240cd067f60860957951db601dfb2fc81e373..a726aebce7d7f9c679f468343aaf62a4354d805c 100644 (file)
--- a/ipc/shm.c
+++ b/ipc/shm.c
@@ -577,7 +577,8 @@ static void shm_get_stat(struct ipc_namespace *ns, unsigned long *rss,
 
                if (is_file_hugepages(shp->shm_file)) {
                        struct address_space *mapping = inode->i_mapping;
-                       *rss += (HPAGE_SIZE/PAGE_SIZE)*mapping->nrpages;
+                       struct hstate *h = hstate_file(shp->shm_file);
+                       *rss += pages_per_huge_page(h) * mapping->nrpages;
                } else {
                        struct shmem_inode_info *info = SHMEM_I(inode);
                        spin_lock(&info->lock);
index 32dff4290c66ede5e8fea935b7242e78b7d11928..0d8153e25f091fc04fdf7c2e4009e6161d454284 100644 (file)
 #include "internal.h"
 
 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
-static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
-static unsigned long surplus_huge_pages;
-static unsigned long nr_overcommit_huge_pages;
 unsigned long max_huge_pages;
 unsigned long sysctl_overcommit_huge_pages;
-static struct list_head hugepage_freelists[MAX_NUMNODES];
-static unsigned int nr_huge_pages_node[MAX_NUMNODES];
-static unsigned int free_huge_pages_node[MAX_NUMNODES];
-static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
 unsigned long hugepages_treat_as_movable;
-static int hugetlb_next_nid;
+
+struct hstate default_hstate;
 
 /*
  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
@@ -203,11 +197,11 @@ static long region_count(struct list_head *head, long f, long t)
  * Convert the address within this vma to the page offset within
  * the mapping, in pagecache page units; huge pages here.
  */
-static pgoff_t vma_hugecache_offset(struct vm_area_struct *vma,
-                                       unsigned long address)
+static pgoff_t vma_hugecache_offset(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long address)
 {
-       return ((address - vma->vm_start) >> HPAGE_SHIFT) +
-                       (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
+       return ((address - vma->vm_start) >> huge_page_shift(h)) +
+                       (vma->vm_pgoff >> huge_page_order(h));
 }
 
 /*
@@ -309,20 +303,21 @@ static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 }
 
 /* Decrement the reserved pages in the hugepage pool by one */
-static void decrement_hugepage_resv_vma(struct vm_area_struct *vma)
+static void decrement_hugepage_resv_vma(struct hstate *h,
+                       struct vm_area_struct *vma)
 {
        if (vma->vm_flags & VM_NORESERVE)
                return;
 
        if (vma->vm_flags & VM_SHARED) {
                /* Shared mappings always use reserves */
-               resv_huge_pages--;
+               h->resv_huge_pages--;
        } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
                /*
                 * Only the process that called mmap() has reserves for
                 * private mappings.
                 */
-               resv_huge_pages--;
+               h->resv_huge_pages--;
        }
 }
 
@@ -344,12 +339,13 @@ static int vma_has_private_reserves(struct vm_area_struct *vma)
        return 1;
 }
 
-static void clear_huge_page(struct page *page, unsigned long addr)
+static void clear_huge_page(struct page *page,
+                       unsigned long addr, unsigned long sz)
 {
        int i;
 
        might_sleep();
-       for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
+       for (i = 0; i < sz/PAGE_SIZE; i++) {
                cond_resched();
                clear_user_highpage(page + i, addr + i * PAGE_SIZE);
        }
@@ -359,41 +355,43 @@ static void copy_huge_page(struct page *dst, struct page *src,
                           unsigned long addr, struct vm_area_struct *vma)
 {
        int i;
+       struct hstate *h = hstate_vma(vma);
 
        might_sleep();
-       for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
+       for (i = 0; i < pages_per_huge_page(h); i++) {
                cond_resched();
                copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
        }
 }
 
-static void enqueue_huge_page(struct page *page)
+static void enqueue_huge_page(struct hstate *h, struct page *page)
 {
        int nid = page_to_nid(page);
-       list_add(&page->lru, &hugepage_freelists[nid]);
-       free_huge_pages++;
-       free_huge_pages_node[nid]++;
+       list_add(&page->lru, &h->hugepage_freelists[nid]);
+       h->free_huge_pages++;
+       h->free_huge_pages_node[nid]++;
 }
 
-static struct page *dequeue_huge_page(void)
+static struct page *dequeue_huge_page(struct hstate *h)
 {
        int nid;
        struct page *page = NULL;
 
        for (nid = 0; nid < MAX_NUMNODES; ++nid) {
-               if (!list_empty(&hugepage_freelists[nid])) {
-                       page = list_entry(hugepage_freelists[nid].next,
+               if (!list_empty(&h->hugepage_freelists[nid])) {
+                       page = list_entry(h->hugepage_freelists[nid].next,
                                          struct page, lru);
                        list_del(&page->lru);
-                       free_huge_pages--;
-                       free_huge_pages_node[nid]--;
+                       h->free_huge_pages--;
+                       h->free_huge_pages_node[nid]--;
                        break;
                }
        }
        return page;
 }
 
-static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
+static struct page *dequeue_huge_page_vma(struct hstate *h,
+                               struct vm_area_struct *vma,
                                unsigned long address, int avoid_reserve)
 {
        int nid;
@@ -411,26 +409,26 @@ static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
         * not "stolen". The child may still get SIGKILLed
         */
        if (!vma_has_private_reserves(vma) &&
-                       free_huge_pages - resv_huge_pages == 0)
+                       h->free_huge_pages - h->resv_huge_pages == 0)
                return NULL;
 
        /* If reserves cannot be used, ensure enough pages are in the pool */
-       if (avoid_reserve && free_huge_pages - resv_huge_pages == 0)
+       if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
                return NULL;
 
        for_each_zone_zonelist_nodemask(zone, z, zonelist,
                                                MAX_NR_ZONES - 1, nodemask) {
                nid = zone_to_nid(zone);
                if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
-                   !list_empty(&hugepage_freelists[nid])) {
-                       page = list_entry(hugepage_freelists[nid].next,
+                   !list_empty(&h->hugepage_freelists[nid])) {
+                       page = list_entry(h->hugepage_freelists[nid].next,
                                          struct page, lru);
                        list_del(&page->lru);
-                       free_huge_pages--;
-                       free_huge_pages_node[nid]--;
+                       h->free_huge_pages--;
+                       h->free_huge_pages_node[nid]--;
 
                        if (!avoid_reserve)
-                               decrement_hugepage_resv_vma(vma);
+                               decrement_hugepage_resv_vma(h, vma);
 
                        break;
                }
@@ -439,12 +437,13 @@ static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
        return page;
 }
 
-static void update_and_free_page(struct page *page)
+static void update_and_free_page(struct hstate *h, struct page *page)
 {
        int i;
-       nr_huge_pages--;
-       nr_huge_pages_node[page_to_nid(page)]--;
-       for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
+
+       h->nr_huge_pages--;
+       h->nr_huge_pages_node[page_to_nid(page)]--;
+       for (i = 0; i < pages_per_huge_page(h); i++) {
                page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
                                1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
                                1 << PG_private | 1<< PG_writeback);
@@ -452,11 +451,16 @@ static void update_and_free_page(struct page *page)
        set_compound_page_dtor(page, NULL);
        set_page_refcounted(page);
        arch_release_hugepage(page);
-       __free_pages(page, HUGETLB_PAGE_ORDER);
+       __free_pages(page, huge_page_order(h));
 }
 
 static void free_huge_page(struct page *page)
 {
+       /*
+        * Can't pass hstate in here because it is called from the
+        * compound page destructor.
+        */
+       struct hstate *h = &default_hstate;
        int nid = page_to_nid(page);
        struct address_space *mapping;
 
@@ -466,12 +470,12 @@ static void free_huge_page(struct page *page)
        INIT_LIST_HEAD(&page->lru);
 
        spin_lock(&hugetlb_lock);
-       if (surplus_huge_pages_node[nid]) {
-               update_and_free_page(page);
-               surplus_huge_pages--;
-               surplus_huge_pages_node[nid]--;
+       if (h->surplus_huge_pages_node[nid]) {
+               update_and_free_page(h, page);
+               h->surplus_huge_pages--;
+               h->surplus_huge_pages_node[nid]--;
        } else {
-               enqueue_huge_page(page);
+               enqueue_huge_page(h, page);
        }
        spin_unlock(&hugetlb_lock);
        if (mapping)
@@ -483,7 +487,7 @@ static void free_huge_page(struct page *page)
  * balanced by operating on them in a round-robin fashion.
  * Returns 1 if an adjustment was made.
  */
-static int adjust_pool_surplus(int delta)
+static int adjust_pool_surplus(struct hstate *h, int delta)
 {
        static int prev_nid;
        int nid = prev_nid;
@@ -496,15 +500,15 @@ static int adjust_pool_surplus(int delta)
                        nid = first_node(node_online_map);
 
                /* To shrink on this node, there must be a surplus page */
-               if (delta < 0 && !surplus_huge_pages_node[nid])
+               if (delta < 0 && !h->surplus_huge_pages_node[nid])
                        continue;
                /* Surplus cannot exceed the total number of pages */
-               if (delta > 0 && surplus_huge_pages_node[nid] >=
-                                               nr_huge_pages_node[nid])
+               if (delta > 0 && h->surplus_huge_pages_node[nid] >=
+                                               h->nr_huge_pages_node[nid])
                        continue;
 
-               surplus_huge_pages += delta;
-               surplus_huge_pages_node[nid] += delta;
+               h->surplus_huge_pages += delta;
+               h->surplus_huge_pages_node[nid] += delta;
                ret = 1;
                break;
        } while (nid != prev_nid);
@@ -513,46 +517,46 @@ static int adjust_pool_surplus(int delta)
        return ret;
 }
 
-static void prep_new_huge_page(struct page *page, int nid)
+static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
 {
        set_compound_page_dtor(page, free_huge_page);
        spin_lock(&hugetlb_lock);
-       nr_huge_pages++;
-       nr_huge_pages_node[nid]++;
+       h->nr_huge_pages++;
+       h->nr_huge_pages_node[nid]++;
        spin_unlock(&hugetlb_lock);
        put_page(page); /* free it into the hugepage allocator */
 }
 
-static struct page *alloc_fresh_huge_page_node(int nid)
+static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 {
        struct page *page;
 
        page = alloc_pages_node(nid,
                htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
                                                __GFP_REPEAT|__GFP_NOWARN,
-               HUGETLB_PAGE_ORDER);
+               huge_page_order(h));
        if (page) {
                if (arch_prepare_hugepage(page)) {
                        __free_pages(page, HUGETLB_PAGE_ORDER);
                        return NULL;
                }
-               prep_new_huge_page(page, nid);
+               prep_new_huge_page(h, page, nid);
        }
 
        return page;
 }
 
-static int alloc_fresh_huge_page(void)
+static int alloc_fresh_huge_page(struct hstate *h)
 {
        struct page *page;
        int start_nid;
        int next_nid;
        int ret = 0;
 
-       start_nid = hugetlb_next_nid;
+       start_nid = h->hugetlb_next_nid;
 
        do {
-               page = alloc_fresh_huge_page_node(hugetlb_next_nid);
+               page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
                if (page)
                        ret = 1;
                /*
@@ -566,11 +570,11 @@ static int alloc_fresh_huge_page(void)
                 * if we just successfully allocated a hugepage so that
                 * the next caller gets hugepages on the next node.
                 */
-               next_nid = next_node(hugetlb_next_nid, node_online_map);
+               next_nid = next_node(h->hugetlb_next_nid, node_online_map);
                if (next_nid == MAX_NUMNODES)
                        next_nid = first_node(node_online_map);
-               hugetlb_next_nid = next_nid;
-       } while (!page && hugetlb_next_nid != start_nid);
+               h->hugetlb_next_nid = next_nid;
+       } while (!page && h->hugetlb_next_nid != start_nid);
 
        if (ret)
                count_vm_event(HTLB_BUDDY_PGALLOC);
@@ -580,8 +584,8 @@ static int alloc_fresh_huge_page(void)
        return ret;
 }
 
-static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
-                                               unsigned long address)
+static struct page *alloc_buddy_huge_page(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long address)
 {
        struct page *page;
        unsigned int nid;
@@ -610,18 +614,18 @@ static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
         * per-node value is checked there.
         */
        spin_lock(&hugetlb_lock);
-       if (surplus_huge_pages >= nr_overcommit_huge_pages) {
+       if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
                spin_unlock(&hugetlb_lock);
                return NULL;
        } else {
-               nr_huge_pages++;
-               surplus_huge_pages++;
+               h->nr_huge_pages++;
+               h->surplus_huge_pages++;
        }
        spin_unlock(&hugetlb_lock);
 
        page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
                                        __GFP_REPEAT|__GFP_NOWARN,
-                                       HUGETLB_PAGE_ORDER);
+                                       huge_page_order(h));
 
        spin_lock(&hugetlb_lock);
        if (page) {
@@ -636,12 +640,12 @@ static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
                /*
                 * We incremented the global counters already
                 */
-               nr_huge_pages_node[nid]++;
-               surplus_huge_pages_node[nid]++;
+               h->nr_huge_pages_node[nid]++;
+               h->surplus_huge_pages_node[nid]++;
                __count_vm_event(HTLB_BUDDY_PGALLOC);
        } else {
-               nr_huge_pages--;
-               surplus_huge_pages--;
+               h->nr_huge_pages--;
+               h->surplus_huge_pages--;
                __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
        }
        spin_unlock(&hugetlb_lock);
@@ -653,16 +657,16 @@ static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
  * Increase the hugetlb pool such that it can accomodate a reservation
  * of size 'delta'.
  */
-static int gather_surplus_pages(int delta)
+static int gather_surplus_pages(struct hstate *h, int delta)
 {
        struct list_head surplus_list;
        struct page *page, *tmp;
        int ret, i;
        int needed, allocated;
 
-       needed = (resv_huge_pages + delta) - free_huge_pages;
+       needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
        if (needed <= 0) {
-               resv_huge_pages += delta;
+               h->resv_huge_pages += delta;
                return 0;
        }
 
@@ -673,7 +677,7 @@ static int gather_surplus_pages(int delta)
 retry:
        spin_unlock(&hugetlb_lock);
        for (i = 0; i < needed; i++) {
-               page = alloc_buddy_huge_page(NULL, 0);
+               page = alloc_buddy_huge_page(h, NULL, 0);
                if (!page) {
                        /*
                         * We were not able to allocate enough pages to
@@ -694,7 +698,8 @@ retry:
         * because either resv_huge_pages or free_huge_pages may have changed.
         */
        spin_lock(&hugetlb_lock);
-       needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
+       needed = (h->resv_huge_pages + delta) -
+                       (h->free_huge_pages + allocated);
        if (needed > 0)
                goto retry;
 
@@ -707,7 +712,7 @@ retry:
         * before they are reserved.
         */
        needed += allocated;
-       resv_huge_pages += delta;
+       h->resv_huge_pages += delta;
        ret = 0;
 free:
        /* Free the needed pages to the hugetlb pool */
@@ -715,7 +720,7 @@ free:
                if ((--needed) < 0)
                        break;
                list_del(&page->lru);
-               enqueue_huge_page(page);
+               enqueue_huge_page(h, page);
        }
 
        /* Free unnecessary surplus pages to the buddy allocator */
@@ -743,7 +748,8 @@ free:
  * allocated to satisfy the reservation must be explicitly freed if they were
  * never used.
  */
-static void return_unused_surplus_pages(unsigned long unused_resv_pages)
+static void return_unused_surplus_pages(struct hstate *h,
+                                       unsigned long unused_resv_pages)
 {
        static int nid = -1;
        struct page *page;
@@ -758,27 +764,27 @@ static void return_unused_surplus_pages(unsigned long unused_resv_pages)
        unsigned long remaining_iterations = num_online_nodes();
 
        /* Uncommit the reservation */
-       resv_huge_pages -= unused_resv_pages;
+       h->resv_huge_pages -= unused_resv_pages;
 
-       nr_pages = min(unused_resv_pages, surplus_huge_pages);
+       nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
 
        while (remaining_iterations-- && nr_pages) {
                nid = next_node(nid, node_online_map);
                if (nid == MAX_NUMNODES)
                        nid = first_node(node_online_map);
 
-               if (!surplus_huge_pages_node[nid])
+               if (!h->surplus_huge_pages_node[nid])
                        continue;
 
-               if (!list_empty(&hugepage_freelists[nid])) {
-                       page = list_entry(hugepage_freelists[nid].next,
+               if (!list_empty(&h->hugepage_freelists[nid])) {
+                       page = list_entry(h->hugepage_freelists[nid].next,
                                          struct page, lru);
                        list_del(&page->lru);
-                       update_and_free_page(page);
-                       free_huge_pages--;
-                       free_huge_pages_node[nid]--;
-                       surplus_huge_pages--;
-                       surplus_huge_pages_node[nid]--;
+                       update_and_free_page(h, page);
+                       h->free_huge_pages--;
+                       h->free_huge_pages_node[nid]--;
+                       h->surplus_huge_pages--;
+                       h->surplus_huge_pages_node[nid]--;
                        nr_pages--;
                        remaining_iterations = num_online_nodes();
                }
@@ -794,13 +800,14 @@ static void return_unused_surplus_pages(unsigned long unused_resv_pages)
  * an instantiated the change should be committed via vma_commit_reservation.
  * No action is required on failure.
  */
-static int vma_needs_reservation(struct vm_area_struct *vma, unsigned long addr)
+static int vma_needs_reservation(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long addr)
 {
        struct address_space *mapping = vma->vm_file->f_mapping;
        struct inode *inode = mapping->host;
 
        if (vma->vm_flags & VM_SHARED) {
-               pgoff_t idx = vma_hugecache_offset(vma, addr);
+               pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                return region_chg(&inode->i_mapping->private_list,
                                                        idx, idx + 1);
 
@@ -809,7 +816,7 @@ static int vma_needs_reservation(struct vm_area_struct *vma, unsigned long addr)
 
        } else  {
                int err;
-               pgoff_t idx = vma_hugecache_offset(vma, addr);
+               pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                struct resv_map *reservations = vma_resv_map(vma);
 
                err = region_chg(&reservations->regions, idx, idx + 1);
@@ -818,18 +825,18 @@ static int vma_needs_reservation(struct vm_area_struct *vma, unsigned long addr)
                return 0;
        }
 }
-static void vma_commit_reservation(struct vm_area_struct *vma,
-                                                       unsigned long addr)
+static void vma_commit_reservation(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long addr)
 {
        struct address_space *mapping = vma->vm_file->f_mapping;
        struct inode *inode = mapping->host;
 
        if (vma->vm_flags & VM_SHARED) {
-               pgoff_t idx = vma_hugecache_offset(vma, addr);
+               pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                region_add(&inode->i_mapping->private_list, idx, idx + 1);
 
        } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
-               pgoff_t idx = vma_hugecache_offset(vma, addr);
+               pgoff_t idx = vma_hugecache_offset(h, vma, addr);
                struct resv_map *reservations = vma_resv_map(vma);
 
                /* Mark this page used in the map. */
@@ -840,6 +847,7 @@ static void vma_commit_reservation(struct vm_area_struct *vma,
 static struct page *alloc_huge_page(struct vm_area_struct *vma,
                                    unsigned long addr, int avoid_reserve)
 {
+       struct hstate *h = hstate_vma(vma);
        struct page *page;
        struct address_space *mapping = vma->vm_file->f_mapping;
        struct inode *inode = mapping->host;
@@ -852,7 +860,7 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
         * MAP_NORESERVE mappings may also need pages and quota allocated
         * if no reserve mapping overlaps.
         */
-       chg = vma_needs_reservation(vma, addr);
+       chg = vma_needs_reservation(h, vma, addr);
        if (chg < 0)
                return ERR_PTR(chg);
        if (chg)
@@ -860,11 +868,11 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
                        return ERR_PTR(-ENOSPC);
 
        spin_lock(&hugetlb_lock);
-       page = dequeue_huge_page_vma(vma, addr, avoid_reserve);
+       page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
        spin_unlock(&hugetlb_lock);
 
        if (!page) {
-               page = alloc_buddy_huge_page(vma, addr);
+               page = alloc_buddy_huge_page(h, vma, addr);
                if (!page) {
                        hugetlb_put_quota(inode->i_mapping, chg);
                        return ERR_PTR(-VM_FAULT_OOM);
@@ -874,7 +882,7 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
        set_page_refcounted(page);
        set_page_private(page, (unsigned long) mapping);
 
-       vma_commit_reservation(vma, addr);
+       vma_commit_reservation(h, vma, addr);
 
        return page;
 }
@@ -882,21 +890,28 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
 static int __init hugetlb_init(void)
 {
        unsigned long i;
+       struct hstate *h = &default_hstate;
 
        if (HPAGE_SHIFT == 0)
                return 0;
 
+       if (!h->order) {
+               h->order = HPAGE_SHIFT - PAGE_SHIFT;
+               h->mask = HPAGE_MASK;
+       }
+
        for (i = 0; i < MAX_NUMNODES; ++i)
-               INIT_LIST_HEAD(&hugepage_freelists[i]);
+               INIT_LIST_HEAD(&h->hugepage_freelists[i]);
 
-       hugetlb_next_nid = first_node(node_online_map);
+       h->hugetlb_next_nid = first_node(node_online_map);
 
        for (i = 0; i < max_huge_pages; ++i) {
-               if (!alloc_fresh_huge_page())
+               if (!alloc_fresh_huge_page(h))
                        break;
        }
-       max_huge_pages = free_huge_pages = nr_huge_pages = i;
-       printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
+       max_huge_pages = h->free_huge_pages = h->nr_huge_pages = i;
+       printk(KERN_INFO "Total HugeTLB memory allocated, %ld\n",
+                       h->free_huge_pages);
        return 0;
 }
 module_init(hugetlb_init);
@@ -922,34 +937,36 @@ static unsigned int cpuset_mems_nr(unsigned int *array)
 
 #ifdef CONFIG_SYSCTL
 #ifdef CONFIG_HIGHMEM
-static void try_to_free_low(unsigned long count)
+static void try_to_free_low(struct hstate *h, unsigned long count)
 {
        int i;
 
        for (i = 0; i < MAX_NUMNODES; ++i) {
                struct page *page, *next;
-               list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
-                       if (count >= nr_huge_pages)
+               struct list_head *freel = &h->hugepage_freelists[i];
+               list_for_each_entry_safe(page, next, freel, lru) {
+                       if (count >= h->nr_huge_pages)
                                return;
                        if (PageHighMem(page))
                                continue;
                        list_del(&page->lru);
                        update_and_free_page(page);
-                       free_huge_pages--;
-                       free_huge_pages_node[page_to_nid(page)]--;
+                       h->free_huge_pages--;
+                       h->free_huge_pages_node[page_to_nid(page)]--;
                }
        }
 }
 #else
-static inline void try_to_free_low(unsigned long count)
+static inline void try_to_free_low(struct hstate *h, unsigned long count)
 {
 }
 #endif
 
-#define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
+#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
 static unsigned long set_max_huge_pages(unsigned long count)
 {
        unsigned long min_count, ret;
+       struct hstate *h = &default_hstate;
 
        /*
         * Increase the pool size
@@ -963,19 +980,19 @@ static unsigned long set_max_huge_pages(unsigned long count)
         * within all the constraints specified by the sysctls.
         */
        spin_lock(&hugetlb_lock);
-       while (surplus_huge_pages && count > persistent_huge_pages) {
-               if (!adjust_pool_surplus(-1))
+       while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
+               if (!adjust_pool_surplus(h, -1))
                        break;
        }
 
-       while (count > persistent_huge_pages) {
+       while (count > persistent_huge_pages(h)) {
                /*
                 * If this allocation races such that we no longer need the
                 * page, free_huge_page will handle it by freeing the page
                 * and reducing the surplus.
                 */
                spin_unlock(&hugetlb_lock);
-               ret = alloc_fresh_huge_page();
+               ret = alloc_fresh_huge_page(h);
                spin_lock(&hugetlb_lock);
                if (!ret)
                        goto out;
@@ -997,21 +1014,21 @@ static unsigned long set_max_huge_pages(unsigned long count)
         * and won't grow the pool anywhere else. Not until one of the
         * sysctls are changed, or the surplus pages go out of use.
         */
-       min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
+       min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
        min_count = max(count, min_count);
-       try_to_free_low(min_count);
-       while (min_count < persistent_huge_pages) {
-               struct page *page = dequeue_huge_page();
+       try_to_free_low(h, min_count);
+       while (min_count < persistent_huge_pages(h)) {
+               struct page *page = dequeue_huge_page(h);
                if (!page)
                        break;
-               update_and_free_page(page);
+               update_and_free_page(h, page);
        }
-       while (count < persistent_huge_pages) {
-               if (!adjust_pool_surplus(1))
+       while (count < persistent_huge_pages(h)) {
+               if (!adjust_pool_surplus(h, 1))
                        break;
        }
 out:
-       ret = persistent_huge_pages;
+       ret = persistent_huge_pages(h);
        spin_unlock(&hugetlb_lock);
        return ret;
 }
@@ -1041,9 +1058,10 @@ int hugetlb_overcommit_handler(struct ctl_table *table, int write,
                        struct file *file, void __user *buffer,
                        size_t *length, loff_t *ppos)
 {
+       struct hstate *h = &default_hstate;
        proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
        spin_lock(&hugetlb_lock);
-       nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
+       h->nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
        spin_unlock(&hugetlb_lock);
        return 0;
 }
@@ -1052,37 +1070,40 @@ int hugetlb_overcommit_handler(struct ctl_table *table, int write,
 
 int hugetlb_report_meminfo(char *buf)
 {
+       struct hstate *h = &default_hstate;
        return sprintf(buf,
                        "HugePages_Total: %5lu\n"
                        "HugePages_Free:  %5lu\n"
                        "HugePages_Rsvd:  %5lu\n"
                        "HugePages_Surp:  %5lu\n"
                        "Hugepagesize:    %5lu kB\n",
-                       nr_huge_pages,
-                       free_huge_pages,
-                       resv_huge_pages,
-                       surplus_huge_pages,
-                       HPAGE_SIZE/1024);
+                       h->nr_huge_pages,
+                       h->free_huge_pages,
+                       h->resv_huge_pages,
+                       h->surplus_huge_pages,
+                       1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
 }
 
 int hugetlb_report_node_meminfo(int nid, char *buf)
 {
+       struct hstate *h = &default_hstate;
        return sprintf(buf,
                "Node %d HugePages_Total: %5u\n"
                "Node %d HugePages_Free:  %5u\n"
                "Node %d HugePages_Surp:  %5u\n",
-               nid, nr_huge_pages_node[nid],
-               nid, free_huge_pages_node[nid],
-               nid, surplus_huge_pages_node[nid]);
+               nid, h->nr_huge_pages_node[nid],
+               nid, h->free_huge_pages_node[nid],
+               nid, h->surplus_huge_pages_node[nid]);
 }
 
 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 unsigned long hugetlb_total_pages(void)
 {
-       return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
+       struct hstate *h = &default_hstate;
+       return h->nr_huge_pages * pages_per_huge_page(h);
 }
 
-static int hugetlb_acct_memory(long delta)
+static int hugetlb_acct_memory(struct hstate *h, long delta)
 {
        int ret = -ENOMEM;
 
@@ -1105,18 +1126,18 @@ static int hugetlb_acct_memory(long delta)
         * semantics that cpuset has.
         */
        if (delta > 0) {
-               if (gather_surplus_pages(delta) < 0)
+               if (gather_surplus_pages(h, delta) < 0)
                        goto out;
 
-               if (delta > cpuset_mems_nr(free_huge_pages_node)) {
-                       return_unused_surplus_pages(delta);
+               if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
+                       return_unused_surplus_pages(h, delta);
                        goto out;
                }
        }
 
        ret = 0;
        if (delta < 0)
-               return_unused_surplus_pages((unsigned long) -delta);
+               return_unused_surplus_pages(h, (unsigned long) -delta);
 
 out:
        spin_unlock(&hugetlb_lock);
@@ -1141,14 +1162,15 @@ static void hugetlb_vm_op_open(struct vm_area_struct *vma)
 
 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
 {
+       struct hstate *h = hstate_vma(vma);
        struct resv_map *reservations = vma_resv_map(vma);
        unsigned long reserve;
        unsigned long start;
        unsigned long end;
 
        if (reservations) {
-               start = vma_hugecache_offset(vma, vma->vm_start);
-               end = vma_hugecache_offset(vma, vma->vm_end);
+               start = vma_hugecache_offset(h, vma, vma->vm_start);
+               end = vma_hugecache_offset(h, vma, vma->vm_end);
 
                reserve = (end - start) -
                        region_count(&reservations->regions, start, end);
@@ -1156,7 +1178,7 @@ static void hugetlb_vm_op_close(struct vm_area_struct *vma)
                kref_put(&reservations->refs, resv_map_release);
 
                if (reserve)
-                       hugetlb_acct_memory(-reserve);
+                       hugetlb_acct_memory(h, -reserve);
        }
 }
 
@@ -1214,14 +1236,16 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
        struct page *ptepage;
        unsigned long addr;
        int cow;
+       struct hstate *h = hstate_vma(vma);
+       unsigned long sz = huge_page_size(h);
 
        cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 
-       for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
+       for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
                src_pte = huge_pte_offset(src, addr);
                if (!src_pte)
                        continue;
-               dst_pte = huge_pte_alloc(dst, addr);
+               dst_pte = huge_pte_alloc(dst, addr, sz);
                if (!dst_pte)
                        goto nomem;
 
@@ -1257,6 +1281,9 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
        pte_t pte;
        struct page *page;
        struct page *tmp;
+       struct hstate *h = hstate_vma(vma);
+       unsigned long sz = huge_page_size(h);
+
        /*
         * A page gathering list, protected by per file i_mmap_lock. The
         * lock is used to avoid list corruption from multiple unmapping
@@ -1265,11 +1292,11 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
        LIST_HEAD(page_list);
 
        WARN_ON(!is_vm_hugetlb_page(vma));
-       BUG_ON(start & ~HPAGE_MASK);
-       BUG_ON(end & ~HPAGE_MASK);
+       BUG_ON(start & ~huge_page_mask(h));
+       BUG_ON(end & ~huge_page_mask(h));
 
        spin_lock(&mm->page_table_lock);
-       for (address = start; address < end; address += HPAGE_SIZE) {
+       for (address = start; address < end; address += sz) {
                ptep = huge_pte_offset(mm, address);
                if (!ptep)
                        continue;
@@ -1383,6 +1410,7 @@ static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
                        unsigned long address, pte_t *ptep, pte_t pte,
                        struct page *pagecache_page)
 {
+       struct hstate *h = hstate_vma(vma);
        struct page *old_page, *new_page;
        int avoidcopy;
        int outside_reserve = 0;
@@ -1443,7 +1471,7 @@ retry_avoidcopy:
        __SetPageUptodate(new_page);
        spin_lock(&mm->page_table_lock);
 
-       ptep = huge_pte_offset(mm, address & HPAGE_MASK);
+       ptep = huge_pte_offset(mm, address & huge_page_mask(h));
        if (likely(pte_same(huge_ptep_get(ptep), pte))) {
                /* Break COW */
                huge_ptep_clear_flush(vma, address, ptep);
@@ -1458,14 +1486,14 @@ retry_avoidcopy:
 }
 
 /* Return the pagecache page at a given address within a VMA */
-static struct page *hugetlbfs_pagecache_page(struct vm_area_struct *vma,
-                       unsigned long address)
+static struct page *hugetlbfs_pagecache_page(struct hstate *h,
+                       struct vm_area_struct *vma, unsigned long address)
 {
        struct address_space *mapping;
        pgoff_t idx;
 
        mapping = vma->vm_file->f_mapping;
-       idx = vma_hugecache_offset(vma, address);
+       idx = vma_hugecache_offset(h, vma, address);
 
        return find_lock_page(mapping, idx);
 }
@@ -1473,6 +1501,7 @@ static struct page *hugetlbfs_pagecache_page(struct vm_area_struct *vma,
 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
                        unsigned long address, pte_t *ptep, int write_access)
 {
+       struct hstate *h = hstate_vma(vma);
        int ret = VM_FAULT_SIGBUS;
        pgoff_t idx;
        unsigned long size;
@@ -1493,7 +1522,7 @@ static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
        }
 
        mapping = vma->vm_file->f_mapping;
-       idx = vma_hugecache_offset(vma, address);
+       idx = vma_hugecache_offset(h, vma, address);
 
        /*
         * Use page lock to guard against racing truncation
@@ -1502,7 +1531,7 @@ static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 retry:
        page = find_lock_page(mapping, idx);
        if (!page) {
-               size = i_size_read(mapping->host) >> HPAGE_SHIFT;
+               size = i_size_read(mapping->host) >> huge_page_shift(h);
                if (idx >= size)
                        goto out;
                page = alloc_huge_page(vma, address, 0);
@@ -1510,7 +1539,7 @@ retry:
                        ret = -PTR_ERR(page);
                        goto out;
                }
-               clear_huge_page(page, address);
+               clear_huge_page(page, address, huge_page_size(h));
                __SetPageUptodate(page);
 
                if (vma->vm_flags & VM_SHARED) {
@@ -1526,14 +1555,14 @@ retry:
                        }
 
                        spin_lock(&inode->i_lock);
-                       inode->i_blocks += BLOCKS_PER_HUGEPAGE;
+                       inode->i_blocks += blocks_per_huge_page(h);
                        spin_unlock(&inode->i_lock);
                } else
                        lock_page(page);
        }
 
        spin_lock(&mm->page_table_lock);
-       size = i_size_read(mapping->host) >> HPAGE_SHIFT;
+       size = i_size_read(mapping->host) >> huge_page_shift(h);
        if (idx >= size)
                goto backout;
 
@@ -1569,8 +1598,9 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        pte_t entry;
        int ret;
        static DEFINE_MUTEX(hugetlb_instantiation_mutex);
+       struct hstate *h = hstate_vma(vma);
 
-       ptep = huge_pte_alloc(mm, address);
+       ptep = huge_pte_alloc(mm, address, huge_page_size(h));
        if (!ptep)
                return VM_FAULT_OOM;
 
@@ -1594,7 +1624,7 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        if (likely(pte_same(entry, huge_ptep_get(ptep))))
                if (write_access && !pte_write(entry)) {
                        struct page *page;
-                       page = hugetlbfs_pagecache_page(vma, address);
+                       page = hugetlbfs_pagecache_page(h, vma, address);
                        ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
                        if (page) {
                                unlock_page(page);
@@ -1615,6 +1645,7 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
        unsigned long pfn_offset;
        unsigned long vaddr = *position;
        int remainder = *length;
+       struct hstate *h = hstate_vma(vma);
 
        spin_lock(&mm->page_table_lock);
        while (vaddr < vma->vm_end && remainder) {
@@ -1626,7 +1657,7 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                 * each hugepage.  We have to make * sure we get the
                 * first, for the page indexing below to work.
                 */
-               pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
+               pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
 
                if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
                    (write && !pte_write(huge_ptep_get(pte)))) {
@@ -1644,7 +1675,7 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                        break;
                }
 
-               pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
+               pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
                page = pte_page(huge_ptep_get(pte));
 same_page:
                if (pages) {
@@ -1660,7 +1691,7 @@ same_page:
                --remainder;
                ++i;
                if (vaddr < vma->vm_end && remainder &&
-                               pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
+                               pfn_offset < pages_per_huge_page(h)) {
                        /*
                         * We use pfn_offset to avoid touching the pageframes
                         * of this compound page.
@@ -1682,13 +1713,14 @@ void hugetlb_change_protection(struct vm_area_struct *vma,
        unsigned long start = address;
        pte_t *ptep;
        pte_t pte;
+       struct hstate *h = hstate_vma(vma);
 
        BUG_ON(address >= end);
        flush_cache_range(vma, address, end);
 
        spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
        spin_lock(&mm->page_table_lock);
-       for (; address < end; address += HPAGE_SIZE) {
+       for (; address < end; address += huge_page_size(h)) {
                ptep = huge_pte_offset(mm, address);
                if (!ptep)
                        continue;
@@ -1711,6 +1743,7 @@ int hugetlb_reserve_pages(struct inode *inode,
                                        struct vm_area_struct *vma)
 {
        long ret, chg;
+       struct hstate *h = hstate_inode(inode);
 
        if (vma && vma->vm_flags & VM_NORESERVE)
                return 0;
@@ -1739,7 +1772,7 @@ int hugetlb_reserve_pages(struct inode *inode,
 
        if (hugetlb_get_quota(inode->i_mapping, chg))
                return -ENOSPC;
-       ret = hugetlb_acct_memory(chg);
+       ret = hugetlb_acct_memory(h, chg);
        if (ret < 0) {
                hugetlb_put_quota(inode->i_mapping, chg);
                return ret;
@@ -1751,12 +1784,13 @@ int hugetlb_reserve_pages(struct inode *inode,
 
 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
 {
+       struct hstate *h = hstate_inode(inode);
        long chg = region_truncate(&inode->i_mapping->private_list, offset);
 
        spin_lock(&inode->i_lock);
-       inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
+       inode->i_blocks -= blocks_per_huge_page(h);
        spin_unlock(&inode->i_lock);
 
        hugetlb_put_quota(inode->i_mapping, (chg - freed));
-       hugetlb_acct_memory(-(chg - freed));
+       hugetlb_acct_memory(h, -(chg - freed));
 }
index 72932489a082c313998a0cebb1ce62cdd40c769a..c1c1d6d8c22bba3db72f9614fa8659c3002e0445 100644 (file)
@@ -903,7 +903,7 @@ unsigned long unmap_vmas(struct mmu_gather **tlbp,
                        if (unlikely(is_vm_hugetlb_page(vma))) {
                                unmap_hugepage_range(vma, start, end, NULL);
                                zap_work -= (end - start) /
-                                               (HPAGE_SIZE / PAGE_SIZE);
+                                       pages_per_huge_page(hstate_vma(vma));
                                start = end;
                        } else
                                start = unmap_page_range(*tlbp, vma,
index c94e58b192c3c9913f4f4ab57860b6ac9469f839..e550bec205826bd63f46082d6c5aac7bc69aa23e 100644 (file)
@@ -1481,7 +1481,7 @@ struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
 
        if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
                zl = node_zonelist(interleave_nid(*mpol, vma, addr,
-                                               HPAGE_SHIFT), gfp_flags);
+                               huge_page_shift(hstate_vma(vma))), gfp_flags);
        } else {
                zl = policy_zonelist(gfp_flags, *mpol);
                if ((*mpol)->mode == MPOL_BIND)
@@ -2220,9 +2220,12 @@ static void check_huge_range(struct vm_area_struct *vma,
 {
        unsigned long addr;
        struct page *page;
+       struct hstate *h = hstate_vma(vma);
+       unsigned long sz = huge_page_size(h);
 
-       for (addr = start; addr < end; addr += HPAGE_SIZE) {
-               pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
+       for (addr = start; addr < end; addr += sz) {
+               pte_t *ptep = huge_pte_offset(vma->vm_mm,
+                                               addr & huge_page_mask(h));
                pte_t pte;
 
                if (!ptep)
index 57d3b6097deb38e3f73d3319845681207b86c293..5e0cc99e9cd57704a1019dc2c6764d895ad3d756 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -1812,7 +1812,8 @@ int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
        struct mempolicy *pol;
        struct vm_area_struct *new;
 
-       if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
+       if (is_vm_hugetlb_page(vma) && (addr &
+                                       ~(huge_page_mask(hstate_vma(vma)))))
                return -EINVAL;
 
        if (mm->map_count >= sysctl_max_map_count)