Skip to contentSkip to conversation

ceph: truncate page cache when doing DIO in encrypted inodes

PATCH In review Luís Henriques 2 participants 3 messages Apr 1, 2022 Ceph client

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ceph: truncate page cache when doing DIO in encrypted inodes

When doing DIO on an encrypted node, we need to truncate the page cache in the range being written to, otherwise the cache will include invalid data.

---
 fs/ceph/file.c | 5 +++++
 1 file changed, 5 insertions(+)

This patch should fix generic/647 fstest when run with test_dummy_encryption.

fs/ceph/file.c +5 -0
index 5072570c2203..0f31c4d352a4 100644
@@ -1895,6 +1895,11 @@ ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
 		req->r_inode = inode;
 		req->r_mtime = mtime;
 
+		if (IS_ENCRYPTED(inode) && (iocb->ki_flags & IOCB_DIRECT))
+			truncate_inode_pages_range(
+				inode->i_mapping, write_pos,
+				PAGE_ALIGN(write_pos + write_len) - 1);
+
 		/* Set up the assertion */
 		if (rmw) {
 			/*
#1 lore raw
Quoted 16 lines · Luís Henriques

When doing DIO on an encrypted node, we need to truncate the page cache in the range being written to, otherwise the cache will include invalid data.

---
 fs/ceph/file.c | 5 +++++
 1 file changed, 5 insertions(+)

This patch should fix generic/647 fstest when run with test_dummy_encryption.

fs/ceph/file.c +0 -0
index 5072570c2203..0f31c4d352a4 100644
@@ -1895,6 +1895,11 @@ ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
 		req->r_inode = inode;

fs/ceph/file.c ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos, · line 1905

Quoted 10 lines
fs/ceph/file.c +5 -0
 		req->r_mtime = mtime;
 
+		if (IS_ENCRYPTED(inode) && (iocb->ki_flags & IOCB_DIRECT))
+			truncate_inode_pages_range(
+				inode->i_mapping, write_pos,
+				PAGE_ALIGN(write_pos + write_len) - 1);
+
 		/* Set up the assertion */
 		if (rmw) {
 			/*

Truncating the pagecache like this could cause dirty data to be discarded. I know we're planning to overwrite this range, but you are having to invalidate more than the written range here. We could potentially lose a write to that region.

Have you tried using something like invalidate_inode_pages2_range ? That's more of what we'd want here, as it's a bit more cautious about tossing out dirty pages. I see too that that is what ceph_direct_read_write calls in the write case as well.

Signature · 1 line
Jeff Layton <[email protected]>
#2 lore raw
Quoted 33 lines · Jeff Layton
Quoted 16 lines · Luís Henriques

When doing DIO on an encrypted node, we need to truncate the page cache in the range being written to, otherwise the cache will include invalid data.

---
 fs/ceph/file.c | 5 +++++
 1 file changed, 5 insertions(+)

This patch should fix generic/647 fstest when run with test_dummy_encryption.

fs/ceph/file.c +0 -0
index 5072570c2203..0f31c4d352a4 100644
@@ -1895,6 +1895,11 @@ ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
 		req->r_inode = inode;

fs/ceph/file.c ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos, · line 1905

Quoted 10 lines
fs/ceph/file.c +5 -0
 		req->r_mtime = mtime;
 
+		if (IS_ENCRYPTED(inode) && (iocb->ki_flags & IOCB_DIRECT))
+			truncate_inode_pages_range(
+				inode->i_mapping, write_pos,
+				PAGE_ALIGN(write_pos + write_len) - 1);
+
 		/* Set up the assertion */
 		if (rmw) {
 			/*

Truncating the pagecache like this could cause dirty data to be discarded. I know we're planning to overwrite this range, but you are having to invalidate more than the written range here. We could potentially lose a write to that region.

Quoted 4 lines

Have you tried using something like invalidate_inode_pages2_range ? That's more of what we'd want here, as it's a bit more cautious about tossing out dirty pages. I see too that that is what ceph_direct_read_write calls in the write case as well.

OK, let me try that instead. Yeah, I've used what the usual direct path was using (the truncate_inode_pages_range()), but if invalidate_inode_pages2_range works here I guess that's better. I'll test that and send out v2.

Cheers,

Signature · 1 line
Luís