[Linux-cachefs] [RFC PATCH 37/53] netfs: Support decryption on ubuffered/DIO read

David Howells dhowells at redhat.com
Fri Oct 13 16:04:06 UTC 2023


Support unbuffered and direct I/O reads from an encrypted file.  This may
require making a larger read than is required into a bounce buffer and
copying out the required bits.  We don't decrypt in-place in the user
buffer lest userspace interfere and muck up the decryption.

Signed-off-by: David Howells <dhowells at redhat.com>
cc: Jeff Layton <jlayton at kernel.org>
cc: linux-cachefs at redhat.com
cc: linux-fsdevel at vger.kernel.org
cc: linux-mm at kvack.org
---
 fs/netfs/direct_read.c | 10 ++++++++++
 fs/netfs/internal.h    | 17 +++++++++++++++++
 2 files changed, 27 insertions(+)

diff --git a/fs/netfs/direct_read.c b/fs/netfs/direct_read.c
index 52ad8fa66dd5..158719b56900 100644
--- a/fs/netfs/direct_read.c
+++ b/fs/netfs/direct_read.c
@@ -181,6 +181,16 @@ static ssize_t netfs_unbuffered_read_iter_locked(struct kiocb *iocb, struct iov_
 		iov_iter_advance(iter, orig_count);
 	}
 
+	/* If we're going to do decryption or decompression, we're going to
+	 * need a bounce buffer - and if the data is misaligned for the crypto
+	 * algorithm, we decrypt in place and then copy.
+	 */
+	if (test_bit(NETFS_RREQ_CONTENT_ENCRYPTION, &rreq->flags)) {
+		if (!netfs_is_crypto_aligned(rreq, iter))
+			__set_bit(NETFS_RREQ_CRYPT_IN_PLACE, &rreq->flags);
+		__set_bit(NETFS_RREQ_USE_BOUNCE_BUFFER, &rreq->flags);
+	}
+
 	/* If we're going to use a bounce buffer, we need to set it up.  We
 	 * will then need to pad the request out to the minimum block size.
 	 */
diff --git a/fs/netfs/internal.h b/fs/netfs/internal.h
index 8dc68a75d6cd..7dd37d3aff3f 100644
--- a/fs/netfs/internal.h
+++ b/fs/netfs/internal.h
@@ -196,6 +196,23 @@ static inline void netfs_put_group_many(struct netfs_group *netfs_group, int nr)
 		netfs_group->free(netfs_group);
 }
 
+/*
+ * Check to see if a buffer aligns with the crypto unit block size.  If it
+ * doesn't the crypto layer is going to copy all the data - in which case
+ * relying on the crypto op for a free copy is pointless.
+ */
+static inline bool netfs_is_crypto_aligned(struct netfs_io_request *rreq,
+					   struct iov_iter *iter)
+{
+	struct netfs_inode *ctx = netfs_inode(rreq->inode);
+	unsigned long align, mask = (1UL << ctx->min_bshift) - 1;
+
+	if (!ctx->min_bshift)
+		return true;
+	align = iov_iter_alignment(iter);
+	return (align & mask) == 0;
+}
+
 /*****************************************************************************/
 /*
  * debug tracing


More information about the Linux-cachefs mailing list