Skip to contentSkip to conversation

fix filler_t callback type mismatches

PATCH New Christoph Hellwig 5 participants 12 messages May 1, 2019 – May 2, 2019

Subsystems (7)
lore
Share & tools
lore
#0 lore raw

Casting mapping->a_ops->readpage to filler_t causes an indirect call type mismatch with Control-Flow Integrity checking. This change fixes the mismatch in read_cache_page_gfp and read_mapping_page by adding using a NULL filler argument as an indication to call ->readpage directly, and by passing the right parameter callbacks in nfs and jffs2.

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/

#1 lore raw

PATCH 1/4 mm: fix an overly long line in read_cache_page

---
 mm/filemap.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
mm/filemap.c +2 -1
index d78f577baef2..a2fc59f56f50 100644
@@ -2956,7 +2956,8 @@ struct page *read_cache_page(struct address_space *mapping,
 				int (*filler)(void *, struct page *),
 				void *data)
 {
-	return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
+	return do_read_cache_page(mapping, index, filler, data,
+			mapping_gfp_mask(mapping));
 }
 EXPORT_SYMBOL(read_cache_page);
 
Signature · 1 line
2.20.1
#2 lore raw

PATCH 2/4 mm: don't cast ->readpage to filler_t for do_read_cache_page

In reply to Christoph Hellwig

We can just pass a NULL filler and do the right thing inside of do_read_cache_page based on the NULL parameter.

---
 include/linux/pagemap.h |  3 +--
 mm/filemap.c            | 10 ++++++----
 2 files changed, 7 insertions(+), 6 deletions(-)
include/linux/pagemap.h +1 -2
index bcf909d0de5f..f52c3a2074cd 100644
@@ -386,8 +386,7 @@ extern int read_cache_pages(struct address_space *mapping,
 static inline struct page *read_mapping_page(struct address_space *mapping,
 				pgoff_t index, void *data)
 {
-	filler_t *filler = (filler_t *)mapping->a_ops->readpage;
-	return read_cache_page(mapping, index, filler, data);
+	return read_cache_page(mapping, index, NULL, data);
 }
 
 /*
mm/filemap.c +6 -4
index a2fc59f56f50..51f5b02c299a 100644
@@ -2866,7 +2866,11 @@ static struct page *do_read_cache_page(struct address_space *mapping,
 		}
 
 filler:
-		err = filler(data, page);
+		if (filler)
+			err = filler(data, page);
+		else
+			err = mapping->a_ops->readpage(data, page);
+
 		if (err < 0) {
 			put_page(page);
 			return ERR_PTR(err);
@@ -2978,9 +2982,7 @@ struct page *read_cache_page_gfp(struct address_space *mapping,
 				pgoff_t index,
 				gfp_t gfp)
 {
-	filler_t *filler = (filler_t *)mapping->a_ops->readpage;
-
-	return do_read_cache_page(mapping, index, filler, NULL, gfp);
+	return do_read_cache_page(mapping, index, NULL, NULL, gfp);
 }
 EXPORT_SYMBOL(read_cache_page_gfp);
 
Signature · 1 line
2.20.1
#3 lore raw

PATCH 3/4 nfs: pass the correct prototype to read_cache_page

In reply to Christoph Hellwig

Fix the callbacks NFS passes to read_cache_page to actually have the proper type expected. Casting around function pointers can easily hide typing bugs, and defeats control flow protection.

---
 fs/nfs/dir.c     | 7 ++++---
 fs/nfs/symlink.c | 7 ++++---
 2 files changed, 8 insertions(+), 6 deletions(-)
fs/nfs/dir.c +4 -3
index a71d0b42d160..47d445bec8c9 100644
@@ -714,8 +714,9 @@ int nfs_readdir_xdr_to_array(nfs_readdir_descriptor_t *desc, struct page *page,
  * We only need to convert from xdr once so future lookups are much simpler
  */
 static
-int nfs_readdir_filler(nfs_readdir_descriptor_t *desc, struct page* page)
+int nfs_readdir_filler(void *data, struct page* page)
 {
+	nfs_readdir_descriptor_t *desc = data;
 	struct inode	*inode = file_inode(desc->file);
 	int ret;
 
@@ -762,8 +763,8 @@ void cache_page_release(nfs_readdir_descriptor_t *desc)
 static
 struct page *get_cache_page(nfs_readdir_descriptor_t *desc)
 {
-	return read_cache_page(desc->file->f_mapping,
-			desc->page_index, (filler_t *)nfs_readdir_filler, desc);
+	return read_cache_page(desc->file->f_mapping, desc->page_index,
+			nfs_readdir_filler, desc);
 }
 
 /*
fs/nfs/symlink.c +4 -3
index 06eb44b47885..25ba299fdac2 100644
@@ -26,8 +26,9 @@
  * and straight-forward than readdir caching.
  */
 
-static int nfs_symlink_filler(struct inode *inode, struct page *page)
+static int nfs_symlink_filler(void *data, struct page *page)
 {
+	struct inode *inode = data;
 	int error;
 
 	error = NFS_PROTO(inode)->readlink(inode, page, 0, PAGE_SIZE);
@@ -65,8 +66,8 @@ static const char *nfs_get_link(struct dentry *dentry,
 		err = ERR_PTR(nfs_revalidate_mapping(inode, inode->i_mapping));
 		if (err)
 			return err;
-		page = read_cache_page(&inode->i_data, 0,
-					(filler_t *)nfs_symlink_filler, inode);
+		page = read_cache_page(&inode->i_data, 0, nfs_symlink_filler,
+				inode);
 		if (IS_ERR(page))
 			return ERR_CAST(page);
 	}
Signature · 1 line
2.20.1
#4 lore raw

PATCH 4/4 jffs2: pass the correct prototype to read_cache_page

In reply to Christoph Hellwig

Fix the callback jffs2 passes to read_cache_page to actually have the proper type expected. Casting around function pointers can easily hide typing bugs, and defeats control flow protection.

---
 fs/jffs2/file.c     | 4 ++--
 fs/jffs2/fs.c       | 2 +-
 fs/jffs2/os-linux.h | 2 +-
 3 files changed, 4 insertions(+), 4 deletions(-)
fs/jffs2/file.c +2 -2
index 7d8654a1472e..f8fb89b10227 100644
@@ -109,9 +109,9 @@ static int jffs2_do_readpage_nolock (struct inode *inode, struct page *pg)
 	return ret;
 }
 
-int jffs2_do_readpage_unlock(struct inode *inode, struct page *pg)
+int jffs2_do_readpage_unlock(void *data, struct page *pg)
 {
-	int ret = jffs2_do_readpage_nolock(inode, pg);
+	int ret = jffs2_do_readpage_nolock(data, pg);
 	unlock_page(pg);
 	return ret;
 }
fs/jffs2/fs.c +1 -1
index eab04eca95a3..7fbe8a7843b9 100644
@@ -686,7 +686,7 @@ unsigned char *jffs2_gc_fetch_page(struct jffs2_sb_info *c,
 	struct page *pg;
 
 	pg = read_cache_page(inode->i_mapping, offset >> PAGE_SHIFT,
-			     (void *)jffs2_do_readpage_unlock, inode);
+			     jffs2_do_readpage_unlock, inode);
 	if (IS_ERR(pg))
 		return (void *)pg;
 
fs/jffs2/os-linux.h +1 -1
index a2dbbb3f4c74..bd3d5f0ddc34 100644
@@ -155,7 +155,7 @@ extern const struct file_operations jffs2_file_operations;
 extern const struct inode_operations jffs2_file_inode_operations;
 extern const struct address_space_operations jffs2_file_address_operations;
 int jffs2_fsync(struct file *, loff_t, loff_t, int);
-int jffs2_do_readpage_unlock (struct inode *inode, struct page *pg);
+int jffs2_do_readpage_unlock(void *data, struct page *pg);
 
 /* ioctl.c */
 long jffs2_ioctl(struct file *, unsigned int, unsigned long);
Signature · 1 line
2.20.1
#5 lore raw
In reply to Christoph Hellwig
Quoted 7 lines · Christoph Hellwig

Casting mapping->a_ops->readpage to filler_t causes an indirect call type mismatch with Control-Flow Integrity checking. This change fixes the mismatch in read_cache_page_gfp and read_mapping_page by adding using a NULL filler argument as an indication to call ->readpage directly, and by passing the right parameter callbacks in nfs and jffs2.

Thanks, Christoph! This looks much cleaner.

I tested the patches on a kernel compiled with clang's -fsanitize=cfi and the fixes look good to me. However, you missed one more type mismatch in v9fs_vfs_readpages (fs/9p/vfs_addr.c). Could you please add that one to the series too?

Sami

#6 lore raw

[PATCH 5/4] 9p: pass the correct prototype to read_cache_page

In reply to Christoph Hellwig

Fix the callback 9p passes to read_cache_page to actually have the proper type expected. Casting around function pointers can easily hide typing bugs, and defeats control flow protection.

---
 fs/9p/vfs_addr.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
fs/9p/vfs_addr.c +4 -2
index 0bcbcc20f769..02e0fc51401e 100644
@@ -50,8 +50,9 @@
  * @page: structure to page
  *
  */
-static int v9fs_fid_readpage(struct p9_fid *fid, struct page *page)
+static int v9fs_fid_readpage(void *data, struct page *page)
 {
+	struct p9_fid *fid = data;
 	struct inode *inode = page->mapping->host;
 	struct bio_vec bvec = {.bv_page = page, .bv_len = PAGE_SIZE};
 	struct iov_iter to;
@@ -122,7 +123,8 @@ static int v9fs_vfs_readpages(struct file *filp, struct address_space *mapping,
 	if (ret == 0)
 		return ret;
 
-	ret = read_cache_pages(mapping, pages, (void *)v9fs_vfs_readpage, filp);
+	ret = read_cache_pages(mapping, pages, v9fs_fid_readpage,
+			filp->private_data);
 	p9_debug(P9_DEBUG_VFS, "  = %d\n", ret);
 	return ret;
 }
Signature · 1 line
2.20.1
Reviewed-by
#7 lore raw

Re: [PATCH 5/4] 9p: pass the correct prototype to read_cache_page

1) You need to pass "filp" rather than "filp->private_data" to read_cache_pages() in v9fs_fid_readpage().

The patched code passes "filp->private_data" as the "data" parameter to read_cache_pages(), which would generate a call to:

filler(data, page)

which would become a call to:

static int v9fs_vfs_readpage(struct file *filp, struct page *page)

{	
        return v9fs_fid_readpage(filp->private_data, page);
}

which would then effectively become:

    v9fs_fid_readpage(filp->private_data->private_data, page)

Which isn't correct; because data is a void *, no error is thrown when v9fs_vfs_readpages treats filp->private_data as if it is filp.

2) I'd also like to see an explicit comment in do_read_cache_page() along the lines of:

/*
* If a custom page filler was passed in use it, otherwise use the
* standard readpage() routine defined for the address_space. *
*/

3) Patch 5/4?

Otherwise it looks good.

Quoted 27 lines · Christoph Hellwig

Fix the callback 9p passes to read_cache_page to actually have the proper type expected. Casting around function pointers can easily hide typing bugs, and defeats control flow protection.

---
fs/9p/vfs_addr.c | 6 ++++--

1 file changed, 4 insertions(+), 2 deletions(-)
fs/9p/vfs_addr.c +2 -1
index 0bcbcc20f769..02e0fc51401e 100644
@@ -50,8 +50,9 @@
 * @page: structure to page
 *
 */
-static int v9fs_fid_readpage(struct p9_fid *fid, struct page *page)
+static int v9fs_fid_readpage(void *data, struct page *page)
 {
+	struct p9_fid *fid = data;
 	struct inode *inode = page->mapping->host;
 	struct bio_vec bvec = {.bv_page = page, .bv_len = PAGE_SIZE};
 	struct iov_iter to;

fs/9p/vfs_addr.c static int v9fs_vfs_readpages(struct file *filp, struct address_space *mapping, · line 130

Quoted 9 lines
fs/9p/vfs_addr.c +2 -1
@@ -122,7 +123,8 @@ static int v9fs_vfs_readpages(struct file *filp, struct address_space *mapping,
 	if (ret == 0)
 		return ret;
 
-	ret = read_cache_pages(mapping, pages, (void *)v9fs_vfs_readpage, filp);
+	ret = read_cache_pages(mapping, pages, v9fs_fid_readpage,
+			filp->private_data);
 	p9_debug(P9_DEBUG_VFS, "  = %d\n", ret);
 	return ret;
 }

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/

maintainer
#8 lore raw

Re: [PATCH 5/4] 9p: pass the correct prototype to read_cache_page

Quoted 1 line · William Kucharski

3) Patch 5/4?

That's a relatively common notation when an extra patch is needed to fix something after a series has been sent ;-)

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/

#9 lore raw

Re: [PATCH 5/4] 9p: pass the correct prototype to read_cache_page

In reply to William Kucharski
Quoted 2 lines · William Kucharski

1) You need to pass "filp" rather than "filp->private_data" to read_cache_pages() in v9fs_fid_readpage().

With this patch v9fs_fid_readpage takes a void pointer that must be a FID, and we pass the FID everywhere:

- v9fs_vfs_readpage passes filp->private_data
- v9fs_vfs_readpages passes filp->private_data through read_cache_pages
- v9fs_write_begin passes the local fid variable

Quoted 4 lines

The patched code passes "filp->private_data" as the "data" parameter to read_cache_pages(), which would generate a call to:

Quoted 8 lines

filler(data, page)

which would become a call to:

static int v9fs_vfs_readpage(struct file *filp, struct page *page)

{	
        return v9fs_fid_readpage(filp->private_data, page);
}

Except that we don't pass v9fs_vfs_readpage as the filler any more, we now pass v9fs_fid_readpage.

#10 lore raw

Re: [PATCH 5/4] 9p: pass the correct prototype to read_cache_page

Quoted 6 lines · Christoph Hellwig

Except that we don't pass v9fs_vfs_readpage as the filler any more, we now pass v9fs_fid_readpage.

True, so never mind. :-)

Reviewed-by
#11 lore raw
In reply to Christoph Hellwig
Quoted 6 lines · Christoph Hellwig

Casting mapping->a_ops->readpage to filler_t causes an indirect call type mismatch with Control-Flow Integrity checking. This change fixes the mismatch in read_cache_page_gfp and read_mapping_page by adding using a NULL filler argument as an indication to call ->readpage directly, and by passing the right parameter callbacks in nfs and jffs2.

Nice. This looks great; thanks for looking at this. For the series (including patch 5):

Signature · 1 line
Kees Cook