[PATCH v2] fscrypt: try to avoid refing parent dentry in fscrypt_file_open

[PATCH v2] fscrypt: try to avoid refing parent dentry in fscrypt_file_open

[Mateusz Guzik]

Merely checking if the directory is encrypted happens for every open
when using ext4, at the moment refing and unrefing the parent, costing 2
atomics and serializing opens of different files.

The most common case of encryption not being used can be checked for
with RCU instead.

Sample result from open1_processes -t 20 ("Separate file open/close")
from will-it-scale on Sapphire Rapids (ops/s):
before:	12539898
after:	25575494 (+103%)

v2:
- add a comment justifying rcu usage, submitted by Eric Biggers
- whack spurious IS_ENCRYPTED check from the refed case

Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
---
 fs/crypto/hooks.c | 32 ++++++++++++++++++++++++++------
 1 file changed, 26 insertions(+), 6 deletions(-)

diff --git a/fs/crypto/hooks.c b/fs/crypto/hooks.c
index 104771c3d3f6..d8d5049b8fe1 100644
--- a/fs/crypto/hooks.c
+++ b/fs/crypto/hooks.c
@@ -30,21 +30,41 @@
 int fscrypt_file_open(struct inode *inode, struct file *filp)
 {
 	int err;
-	struct dentry *dir;
+	struct dentry *dentry, *dentry_parent;
+	struct inode *inode_parent;
 
 	err = fscrypt_require_key(inode);
 	if (err)
 		return err;
 
-	dir = dget_parent(file_dentry(filp));
-	if (IS_ENCRYPTED(d_inode(dir)) &&
-	    !fscrypt_has_permitted_context(d_inode(dir), inode)) {
+	dentry = file_dentry(filp);
+
+	/*
+	 * Getting a reference to the parent dentry is needed for the actual
+	 * encryption policy comparison, but it's expensive on multi-core
+	 * systems.  Since this function runs on unencrypted files too, start
+	 * with a lightweight RCU-mode check for the parent directory being
+	 * unencrypted (in which case it's fine for the child to be either
+	 * unencrypted, or encrypted with any policy).  Only continue on to the
+	 * full policy check if the parent directory is actually encrypted.
+	 */
+	rcu_read_lock();
+	dentry_parent = READ_ONCE(dentry->d_parent);
+	inode_parent = d_inode_rcu(dentry_parent);
+	if (inode_parent != NULL && !IS_ENCRYPTED(inode_parent)) {
+		rcu_read_unlock();
+		return 0;
+	}
+	rcu_read_unlock();
+
+	dentry_parent = dget_parent(dentry);
+	if (!fscrypt_has_permitted_context(d_inode(dentry_parent), inode)) {
 		fscrypt_warn(inode,
 			     "Inconsistent encryption context (parent directory: %lu)",
-			     d_inode(dir)->i_ino);
+			     d_inode(dentry_parent)->i_ino);
 		err = -EPERM;
 	}
-	dput(dir);
+	dput(dentry_parent);
 	return err;
 }
 EXPORT_SYMBOL_GPL(fscrypt_file_open);
-- 
2.39.2

Re: [PATCH v2] fscrypt: try to avoid refing parent dentry in fscrypt_file_open

[Eric Biggers]

On Wed, May 08, 2024 at 10:14:00AM +0200, Mateusz Guzik wrote:
> Merely checking if the directory is encrypted happens for every open
> when using ext4, at the moment refing and unrefing the parent, costing 2
> atomics and serializing opens of different files.
> 
> The most common case of encryption not being used can be checked for
> with RCU instead.
> 
> Sample result from open1_processes -t 20 ("Separate file open/close")
> from will-it-scale on Sapphire Rapids (ops/s):
> before:	12539898
> after:	25575494 (+103%)
> 
> v2:
> - add a comment justifying rcu usage, submitted by Eric Biggers
> - whack spurious IS_ENCRYPTED check from the refed case
> 
> Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
> ---
>  fs/crypto/hooks.c | 32 ++++++++++++++++++++++++++------
>  1 file changed, 26 insertions(+), 6 deletions(-)

Applied to https://git.kernel.org/pub/scm/fs/fscrypt/linux.git/log/?h=for-next

Thanks!

- Eric