Skip to contentSkip to conversation

ksmb: discard write access to the directory open

PATCH Applied Hobin Woo 5 participants 8 messages Jul 5, 2024 – Jul 8, 2024 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): CGME20240705032731epcas1p177d910a154ded37c459af1c2374a3571@epcas1p1.samsung.com

Hobin Woo

author
#1 lore raw

PATCH v2 1/1 ksmb: discard write access to the directory open

may_open() does not allow a directory to be opened with the write access. However, some writing flags set by client result in adding write access on server, making ksmbd incompatible with FUSE file system. Simply, let's discard the write access when opening a directory.

list_add corruption. next is NULL.

------------[ cut here ]------------
kernel BUG at lib/list_debug.c:26!
pc : __list_add_valid+0x88/0xbc
lr : __list_add_valid+0x88/0xbc
Call trace:
__list_add_valid+0x88/0xbc
fuse_finish_open+0x11c/0x170
fuse_open_common+0x284/0x5e8
fuse_dir_open+0x14/0x24
do_dentry_open+0x2a4/0x4e0
dentry_open+0x50/0x80
smb2_open+0xbe4/0x15a4
handle_ksmbd_work+0x478/0x5ec
process_one_work+0x1b4/0x448
worker_thread+0x25c/0x430
kthread+0x104/0x1d4
ret_from_fork+0x10/0x20

---
v2:
- Describe 'is_dir' in function parameters of 'smb2_create_open_flags'

 fs/smb/server/smb2pdu.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)
fs/smb/server/smb2pdu.c +11 -2
index e7e07891781b..7d26fdcebbf9 100644
@@ -2051,15 +2051,22 @@ int smb2_tree_connect(struct ksmbd_work *work)
  * @access:		file access flags
  * @disposition:	file disposition flags
  * @may_flags:		set with MAY_ flags
+ * @is_dir:		is creating open flags for directory
  *
  * Return:      file open flags
  */
 static int smb2_create_open_flags(bool file_present, __le32 access,
 				  __le32 disposition,
-				  int *may_flags)
+				  int *may_flags,
+				  bool is_dir)
 {
 	int oflags = O_NONBLOCK | O_LARGEFILE;
 
+	if (is_dir) {
+		access &= ~FILE_WRITE_DESIRE_ACCESS_LE;
+		ksmbd_debug(SMB, "Discard write access to a directory\n");
+	}
+
 	if (access & FILE_READ_DESIRED_ACCESS_LE &&
 	    access & FILE_WRITE_DESIRE_ACCESS_LE) {
 		oflags |= O_RDWR;
@@ -3167,7 +3174,9 @@ int smb2_open(struct ksmbd_work *work)
 
 	open_flags = smb2_create_open_flags(file_present, daccess,
 					    req->CreateDisposition,
-					    &may_flags);
+					    &may_flags,
+		req->CreateOptions & FILE_DIRECTORY_FILE_LE ||
+		(file_present && S_ISDIR(d_inode(path.dentry)->i_mode)));
 
 	if (!test_tree_conn_flag(tcon, KSMBD_TREE_CONN_FLAG_WRITABLE)) {
 		if (open_flags & (O_CREAT | O_TRUNC)) {
Signature · 1 line
2.43.0
maintainer Applied
#2 lore raw

2024년 7월 5일 (금) 오후 12:27, Hobin Woo <[email protected]>님이 작성:

Quoted 25 lines

may_open() does not allow a directory to be opened with the write access. However, some writing flags set by client result in adding write access on server, making ksmbd incompatible with FUSE file system. Simply, let's discard the write access when opening a directory.

list_add corruption. next is NULL.

------------[ cut here ]------------
kernel BUG at lib/list_debug.c:26!
pc : __list_add_valid+0x88/0xbc
lr : __list_add_valid+0x88/0xbc
Call trace:
__list_add_valid+0x88/0xbc
fuse_finish_open+0x11c/0x170
fuse_open_common+0x284/0x5e8
fuse_dir_open+0x14/0x24
do_dentry_open+0x2a4/0x4e0
dentry_open+0x50/0x80
smb2_open+0xbe4/0x15a4
handle_ksmbd_work+0x478/0x5ec
process_one_work+0x1b4/0x448
worker_thread+0x25c/0x430
kthread+0x104/0x1d4
ret_from_fork+0x10/0x20
Quoted 5 lines

---
v2:
- Describe 'is_dir' in function parameters of 'smb2_create_open_flags'

Applied it to #ksmbd-for-next-next.
Thanks for your patch!

#3 lore raw

fixed typo in commit title and applied to ksmbd-for-next as well

Quoted 35 lines · Namjae Jeon

2024년 7월 5일 (금) 오후 12:27, Hobin Woo <[email protected]>님이 작성:

Quoted 25 lines

may_open() does not allow a directory to be opened with the write access. However, some writing flags set by client result in adding write access on server, making ksmbd incompatible with FUSE file system. Simply, let's discard the write access when opening a directory.

list_add corruption. next is NULL.

------------[ cut here ]------------
kernel BUG at lib/list_debug.c:26!
pc : __list_add_valid+0x88/0xbc
lr : __list_add_valid+0x88/0xbc
Call trace:
__list_add_valid+0x88/0xbc
fuse_finish_open+0x11c/0x170
fuse_open_common+0x284/0x5e8
fuse_dir_open+0x14/0x24
do_dentry_open+0x2a4/0x4e0
dentry_open+0x50/0x80
smb2_open+0xbe4/0x15a4
handle_ksmbd_work+0x478/0x5ec
process_one_work+0x1b4/0x448
worker_thread+0x25c/0x430
kthread+0x104/0x1d4
ret_from_fork+0x10/0x20
Quoted 5 lines

---
v2:
- Describe 'is_dir' in function parameters of 'smb2_create_open_flags'

Applied it to #ksmbd-for-next-next.
Thanks for your patch!

Signature · 3 lines
Thanks,

Steve
#4 lore raw
In reply to Hobin Woo
Quoted 4 lines · Hobin Woo

may_open() does not allow a directory to be opened with the write access. However, some writing flags set by client result in adding write access on server, making ksmbd incompatible with FUSE file system. Simply, let's discard the write access when opening a directory.

afair this should cause a failure like EACCESS or EISDIR and just be ignored. What does a Windows server return in this case, or Samba for that matter as it (hopefully) will behave like Windows.

-slow

maintainer
#5 lore raw

2024년 7월 5일 (금) 오후 8:54, Ralph Boehme <[email protected]>님이 작성:

Quoted 10 lines · Hobin Woo
Quoted 4 lines

may_open() does not allow a directory to be opened with the write access. However, some writing flags set by client result in adding write access on server, making ksmbd incompatible with FUSE file system. Simply, let's discard the write access when opening a directory.

afair this should cause a failure like EACCESS or EISDIR and just be ignored. What does a Windows server return in this case, or Samba for that matter as it (hopefully) will behave like Windows.

From what I've checked the packet dump, Samba doesn't return any errors in the same case.
Samba seems to open it with O_RDONLY if it is directory, So there is no problem, is it right?

Quoted 3 lines

-slow

#6 lore raw
Quoted 8 lines · Namjae Jeon

2024년 7월 5일 (금) 오후 8:54, Ralph Boehme <[email protected]>님이 작성:

Quoted 7 lines · Hobin Woo
Quoted 4 lines

may_open() does not allow a directory to be opened with the write access. However, some writing flags set by client result in adding write access on server, making ksmbd incompatible with FUSE file system. Simply, let's discard the write access when opening a directory.

Quoted 7 lines
Quoted 3 lines

afair this should cause a failure like EACCESS or EISDIR and just be ignored. What does a Windows server return in this case, or Samba for that matter as it (hopefully) will behave like Windows.

From what I've checked the packet dump, Samba doesn't return any
errors in the same case.
Samba seems to open it with O_RDONLY if it is directory, So there is no problem, is it right?

Hm, it seems my memory is playing tricks on me. Samba indeed forces O_RDONLY for directories in the servers and ignores the client requested access mask. Interestingly we don't seem to have any test for this case, at least I couldn't find any with a quick search. Quickly putting together a torture test shows Windows behaves the same. MS-FSA doesn't mention any such check should be done for directories as well. Getinfo on such a handle even returns the original unmodified client access mask, including the write access.

Sorry for the noise! :)

-slow

reviewer
#7 lore raw
Quoted 17 lines · Ralph Boehme
Quoted 8 lines · Namjae Jeon

2024년 7월 5일 (금) 오후 8:54, Ralph Boehme <[email protected]>님이 작성:

Quoted 7 lines · Hobin Woo
Quoted 4 lines

may_open() does not allow a directory to be opened with the write access. However, some writing flags set by client result in adding write access on server, making ksmbd incompatible with FUSE file system. Simply, let's discard the write access when opening a directory.

Quoted 7 lines
Quoted 3 lines

afair this should cause a failure like EACCESS or EISDIR and just be ignored. What does a Windows server return in this case, or Samba for that matter as it (hopefully) will behave like Windows.

 From what I've checked the packet dump, Samba doesn't return any errors in the same case.
Samba seems to open it with O_RDONLY if it is directory, So there is no problem, is it right?

Quoted 5 lines

Hm, it seems my memory is playing tricks on me. Samba indeed forces O_RDONLY for directories in the servers and ignores the client requested access mask. Interestingly we don't seem to have any test for this case, at least I couldn't find any with a quick search. Quickly putting together a torture test shows Windows behaves the same. MS-FSA doesn't mention any such check should be done for directories as well. Getinfo on such a handle even returns the original unmodified client access mask, including the write access.

Sorry for the noise! :)

-slow

I would ask to see that the SMB3 protocol test suite results are not impacted by this change, and ideally the various Linux/XFS tests as well. I don't see any indication these were run?

The other thing I want to point out is that the crash reported in the commit message is a FUSE failure. Why is that not a bug, and why does the message not justify why the "fix" is in ksmbd?

Tom.

author
#8 lore raw
Quoted 16 lines · Ralph Boehme
Quoted 7 lines · Namjae Jeon
Quoted 6 lines

2024년 7월 5일 (금) 오후 8:54, Ralph Boehme <[email protected]>님이 작성:

Quoted 5 lines · Hobin Woo
Quoted 3 lines

may_open() does not allow a directory to be opened with the write access.
However, some writing flags set by client result in adding write

access

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 3 lines

on server, making ksmbd incompatible with FUSE file system. Simply, let's
discard the write access when opening a directory.

afair this should cause a failure like EACCESS or EISDIR and just be ignored. What does a Windows server return in this case, or Samba for that matter as it (hopefully) will behave like Windows.

Quoted 14 lines
Quoted 7 lines
Quoted 4 lines

From what I've checked the packet dump, Samba doesn't return any
errors in the same case.
Samba seems to open it with O_RDONLY if it is directory, So there is no problem, is it right?

Hm, it seems my memory is playing tricks on me. Samba indeed forces O_RDONLY for directories in the servers and ignores the client

requested

Quoted 6 lines

access mask. Interestingly we don't seem to have any test for this case, at least I couldn't find any with a quick search. Quickly putting together a torture test shows Windows behaves the same. MS-FSA doesn't mention any such check should be done for directories as well. Getinfo on such a handle even returns the original unmodified client access mask, including the write access.

Right. That is why I fixed ksmd not FUSE.

Quoted 5 lines
Quoted 4 lines

Sorry for the noise! :)

-slow

Quoted 7 lines

I would ask to see that the SMB3 protocol test suite results are not impacted by this change, and ideally the various Linux/XFS tests as well. I don't see any indication these were run?

The other thing I want to point out is that the crash reported in the commit message is a FUSE failure. Why is that not a bug, and why does the message not justify why the "fix" is in ksmbd?

As you pointed out, if FUSE had additional checks for this, then the issue wouldn't have occurred in the first place. However, I believe it is not the fundamental fix since FUSE operates correctly under the sanity checks of VFS.

Widely used file systems may perform such checks by themselves or might not have the functionality related to this. But if any new functionality relevant to this were added, we can expect similar problems to happen.

Hobin

Quoted 2 lines

Tom.