Skip to contentSkip to conversation

Demand faulting for hugetlb

PATCH In review Adam Litke 1 participant 4 messages Sep 8, 2005

Subsystems (3)
lore
Share & tools
lore

Adam Litke

author
#0 lore raw

cover Demand faulting for hugetlb

Sending this out again after incorporating the feedback from Yanmin Zhang and Dave Hansen. Are we ready to go for -mm yet? The only patch which has seen any recent changes is #2 below so I assume people agree with #1 and #3 now :)

The three patches:
1) Remove a get_user_pages() optimization that is no longer valid for
demand faulting
2) Move fault logic from hugetlb_prefault() to hugetlb_pte_fault()
3) Apply a simple overcommit check so demand fault accounting behaves
in a manner in line with how prefault worked

Diffed against 2.6.13-git6

Signature · 2 lines
Adam Litke - (agl at us.ibm.com)
IBM Linux Technology Center
author
#1 lore raw

Initial Post (Thu, 18 Aug 2005)

In preparation for hugetlb demand faulting, remove this get_user_pages() optimization. Since huge pages will no longer be prefaulted, we can't assume that the huge ptes are established and hence, calling follow_hugetlb_page() is not valid.

With the follow_hugetlb_page() call removed, the normal code path will be triggered. follow_page() will either use follow_huge_addr() or follow_huge_pmd() to check for a previously faulted "page" to return. When this fails (ie. with demand faults), __handle_mm_fault() gets called which invokes the hugetlb_fault() handler to instantiate the huge page.

This patch doesn't make a lot of sense by itself, but I've broken it out to facilitate discussion on this specific element of the demand fault changes. While coding this up, I referenced previous discussion on this topic starting at http://lkml.org/lkml/2004/4/13/176 , which contains more opinions about the correctness of this approach.

Diffed against 2.6.13-git6

---
 memory.c |    5 -----
 1 files changed, 5 deletions(-)
mm/memory.c +0 -5
diff -upN reference/mm/memory.c current/mm/memory.c
@@ -949,11 +949,6 @@ int get_user_pages(struct task_struct *t
 				|| !(flags & vma->vm_flags))
 			return i ? : -EFAULT;
 
-		if (is_vm_hugetlb_page(vma)) {
-			i = follow_hugetlb_page(mm, vma, pages, vmas,
-						&start, &len, i);
-			continue;
-		}
 		spin_lock(&mm->page_table_lock);
 		do {
 			int write_access = write;
Signature · 2 lines
Adam Litke - (agl at us.ibm.com)
IBM Linux Technology Center
author
#2 lore raw
In reply to Adam Litke

Version 3 (Thu, 08 Aug 2005)
Organized logic in hugetlb_pte_fault() by breaking out
find_get_page/alloc_huge_page logic into separate function
Removed a few more paranoid checks <
Fixed tlb flushing in a race case < (thanks Yanmin Zhang)

Version 2 (Wed, 17 Aug 2005)
Removed spurious WARN_ON()
Patches added earlier in the series:
Check for p?d_none() in arch/i386/mm/hugetlbpage.c:huge_pte_offset() Move i386 stale pte check into huge_pte_alloc()

Initial Post (Fri, 05 Aug 2005)

Below is a patch to implement demand faulting for huge pages. The main motivation for changing from prefaulting to demand faulting is so that huge page memory areas can be allocated according to NUMA policy.

Thanks to consolidated hugetlb code, switching the behavior requires changing only one fault handler. The bulk of the patch just moves the logic from hugelb_prefault() to hugetlb_pte_fault().

Diffed against 2.6.13-git6

---
 fs/hugetlbfs/inode.c    |    6 -
 include/linux/hugetlb.h |    2 
 mm/hugetlb.c            |  154 +++++++++++++++++++++++++++++-------------------
 mm/memory.c             |    2 
 4 files changed, 98 insertions(+), 66 deletions(-)
fs/hugetlbfs/inode.c +1 -5
diff -upN reference/fs/hugetlbfs/inode.c current/fs/hugetlbfs/inode.c
@@ -48,7 +48,6 @@ int sysctl_hugetlb_shm_group;
 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 {
 	struct inode *inode = file->f_dentry->d_inode;
-	struct address_space *mapping = inode->i_mapping;
 	loff_t len, vma_len;
 	int ret;
 
@@ -79,10 +78,7 @@ static int hugetlbfs_file_mmap(struct fi
 	if (!(vma->vm_flags & VM_WRITE) && len > inode->i_size)
 		goto out;
 
-	ret = hugetlb_prefault(mapping, vma);
-	if (ret)
-		goto out;
-
+	ret = 0;
 	if (inode->i_size < len)
 		inode->i_size = len;
 out:
include/linux/hugetlb.h +2 -0
diff -upN reference/include/linux/hugetlb.h current/include/linux/hugetlb.h
@@ -25,6 +25,8 @@ int is_hugepage_mem_enough(size_t);
 unsigned long hugetlb_total_pages(void);
 struct page *alloc_huge_page(void);
 void free_huge_page(struct page *);
+int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct * vma,
+			unsigned long address, int write_access);
 
 extern unsigned long max_huge_pages;
 extern const unsigned long hugetlb_zero, hugetlb_infinity;
mm/hugetlb.c +94 -60

179 lines not on this page · raw message

mm/memory.c +1 -1
diff -upN reference/mm/memory.c current/mm/memory.c
@@ -2041,7 +2041,7 @@ int __handle_mm_fault(struct mm_struct *
 	inc_page_state(pgfault);
 
 	if (is_vm_hugetlb_page(vma))
-		return VM_FAULT_SIGBUS;	/* mapping truncation does this. */
+		return hugetlb_fault(mm, vma, address, write_access);
 
 	/*
 	 * We need the page table lock to synchronize with kswapd
Signature · 2 lines
Adam Litke - (agl at us.ibm.com)
IBM Linux Technology Center
author
#3 lore raw
In reply to Adam Litke

Initial Post (Thu, 18 Aug 2005)

Basic overcommit checking for hugetlb_file_map() based on an implementation used with demand faulting in SLES9.

Since demand faulting can't guarantee the availability of pages at mmap time, this patch implements a basic sanity check to ensure that the number of huge pages required to satisfy the mmap are currently available. Despite the obvious race, I think it is a good start on doing proper accounting. I'd like to work towards an accounting system that mimics the semantics of normal pages (especially for the MAP_PRIVATE/COW case). That work is underway and builds on what this patch starts.

Huge page shared memory segments are simpler and still maintain their commit on shmget semantics.

Diffed against 2.6.13-git6

---
 inode.c |   47 +++++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 47 insertions(+)
fs/hugetlbfs/inode.c +47 -0
diff -upN reference/fs/hugetlbfs/inode.c current/fs/hugetlbfs/inode.c
@@ -45,9 +45,51 @@ static struct backing_dev_info hugetlbfs
 
 int sysctl_hugetlb_shm_group;
 
+static void huge_pagevec_release(struct pagevec *pvec);
+
+unsigned long
+huge_pages_needed(struct address_space *mapping, struct vm_area_struct *vma)
+{
+	int i;
+	struct pagevec pvec;
+	unsigned long start = vma->vm_start;
+	unsigned long end = vma->vm_end;
+	unsigned long hugepages = (end - start) >> HPAGE_SHIFT;
+	pgoff_t next = vma->vm_pgoff;
+	pgoff_t endpg = next + ((end - start) >> PAGE_SHIFT);
+	struct inode *inode = vma->vm_file->f_dentry->d_inode;
+
+	/*
+	 * Shared memory segments are accounted for at shget time,
+	 * not at shmat (when the mapping is actually created) so 
+	 * check here if the memory has already been accounted for.
+	 */
+	if (inode->i_blocks != 0)
+		return 0;
+
+	pagevec_init(&pvec, 0);
+	while (next < endpg) {
+		if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE))
+			break;
+		for (i = 0; i < pagevec_count(&pvec); i++) {
+			struct page *page = pvec.pages[i];
+			if (page->index > next)
+				next = page->index;
+			if (page->index >= endpg)
+				break;
+			next++;
+			hugepages--;
+		}
+		huge_pagevec_release(&pvec);
+	}
+	return hugepages << HPAGE_SHIFT;
+}
+
 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 {
 	struct inode *inode = file->f_dentry->d_inode;
+	struct address_space *mapping = inode->i_mapping;
+	unsigned long bytes;
 	loff_t len, vma_len;
 	int ret;
 
@@ -66,6 +108,10 @@ static int hugetlbfs_file_mmap(struct fi
 	if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
 		return -EINVAL;
 
+	bytes = huge_pages_needed(mapping, vma);
+	if (!is_hugepage_mem_enough(bytes))
+		return -ENOMEM;
+
 	vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 
 	down(&inode->i_sem);
@@ -794,6 +840,7 @@ struct file *hugetlb_zero_setup(size_t s
 	d_instantiate(dentry, inode);
 	inode->i_size = size;
 	inode->i_nlink = 0;
+	inode->i_blocks = 1;
 	file->f_vfsmnt = mntget(hugetlbfs_vfsmount);
 	file->f_dentry = dentry;
 	file->f_mapping = inode->i_mapping;
Signature · 2 lines
Adam Litke - (agl at us.ibm.com)
IBM Linux Technology Center