smb: client: handle STATUS_STOPPED_ON_SYMLINK responses without a symlink target
Subsystems (2)
PATCH v2 1/1 smb: client: handle STATUS_STOPPED_ON_SYMLINK responses without a symlink target
The macOS built-in SMB server returns STATUS_STOPPED_ON_SYMLINK for a CREATE on a path whose final component is a symlink, but it does not include a Symbolic Link Error Response in the error data: both ErrorContextCount and ByteCount are zero, so the symlink target is not present in the response at all. Per [MS-SMB2] section 2.2.2 such a response should carry a valid Symbolic Link Error Response, so this is a server bug, but the target can still be retrieved with FSCTL_GET_REPARSE_POINT.
Frame from a capture against macOS 26.5.2 (build 25F84):
SMB2 hdr : Status=0x8000002d STATUS_STOPPED_ON_SYMLINK, Cmd=Create Error Rsp: StructureSize=0x0009
Error Context Count: 0
Byte Count: 0
Error Data: 00
symlink_data() cannot find a struct smb2_symlink_err_rsp in such a response and returns -EINVAL, which parse_create_response() propagates, so smb2_query_path_info() bails out at
if (rc || !data->reparse_point) goto out;
before it can retry with SMB2_OP_GET_REPARSE. stat(), readlink() and ls of any server-side symlink then fail with -EINVAL:
$ ls -la Config l????????? ? ? ? ? ? Config.json $ stat Config/Config.json stat: cannot statx 'Config/Config.json': Invalid argument
A 5.10 client resolves these symlinks correctly against the same server and share, so this is a regression for Apple SMB servers.
Handle it in several places:
- symlink_data() detects the empty response (ErrorContextCount and ByteCount both zero) and returns a distinct -ENODATA, so that "server did not send the target" can be told apart from a genuinely malformed response and only this case is worked around.
- parse_create_response() treats -ENODATA like STATUS_IO_REPARSE_TAG_NOT_HANDLED, which does not carry the target
either: leave the reparse tag unset and clear rc, so the existing SMB2_OP_GET_REPARSE path retrieves the target.
- smb2_query_path_info() only fixes up the symlink target type when the target is already known. SMB2_OP_GET_REPARSE sets data->reparse.tag but does not parse the target out of the reparse buffer; that happens later, in reparse_info_to_fattr(). Without this check smb2_fix_symlink_target_type() is called with a NULL target and returns -EIO. This could not happen with servers that send the target inline and therefore skip SMB2_OP_GET_REPARSE.
- smb2_open_file() maps -ENODATA to -EIO, matching STATUS_IO_REPARSE_TAG_NOT_HANDLED, so its callers retrieve the target with SMB2_OP_GET_REPARSE as well.
Tested on Debian 13, kernel 6.18.38 (armv7), against macOS 26.5.2:
symlinks now resolve, including relative, parent-traversing and directory symlinks, and reads through symlinks succeed.
---
v2: rebased onto cifs-2.6 for-next; no functional change. v1 was generated
against the 6.18.38 stable tree and did not apply to mainline/for-next.
fs/smb/client/smb2file.c | 21 +++++++++++++++++++++ fs/smb/client/smb2inode.c | 23 ++++++++++++++++++++--- 2 files changed, 41 insertions(+), 3 deletions(-)
fs/smb/client/smb2file.c +21 -0
@@ -30,6 +30,19 @@ static struct smb2_symlink_err_rsp *symlink_data(const struct kvec *iov)u8 *end = (u8 *)err + iov->iov_len;u32 len; +/* +* Per [MS-SMB2] section 2.2.2, a STATUS_STOPPED_ON_SYMLINK response has to +* carry a Symbolic Link Error Response, so ByteCount cannot be zero. Some +* servers (e.g. the macOS built-in SMB server) violate this and return an +* empty error response, with both ErrorContextCount and ByteCount set to +* zero, i.e. without the symlink target. Detect this and return -ENODATA +* so that callers can tell "server did not send the target" apart from a +* malformed response, and retrieve the target with FSCTL_GET_REPARSE_POINT +* instead. +*/ +if (!err->ErrorContextCount && !le32_to_cpu (err->ByteCount)) +return ERR_PTR (-ENODATA ); +if (err->ErrorContextCount) {struct smb2_error_context_rsp *p;
@@ -199,6 +212,14 @@ int smb2_open_file(const unsigned int xid, struct cifs_open_parms *oparms, rc =smb2_parse_symlink_response (oparms->cifs_sb, &err_iov, oparms->path, &data->symlink_target); +/* +* If smb2_parse_symlink_response returned -ENODATA then the +* symlink_target was not sent. Treat this as if the SMB2_open() +* failed with STATUS_IO_REPARSE_TAG_NOT_HANDLED status, which is +* indicated by the -EIO errno. +*/ +if (rc == -ENODATA ) + rc = -EIO ;if (!rc) {memset (&data->fi,0 ,sizeof (data->fi)); oparms->create_options |=OPEN_REPARSE_POINT ;
fs/smb/client/smb2inode.c +20 -3
@@ -792,9 +792,19 @@ static int parse_create_response(struct cifs_open_info_data *data, rc =smb2_parse_symlink_response (cifs_sb, iov, full_path, &data->symlink_target); -if (rc) +if (rc !=0 && rc != -ENODATA )return rc; - tag =IO_REPARSE_TAG_SYMLINK ; +/* +* -ENODATA means that the response was parsed but did not contain +* the symlink target at all (see symlink_data()). Treat it like +* STATUS_IO_REPARSE_TAG_NOT_HANDLED, which does not contain it +* either: leave the tag unset and clear rc, so that the caller +* retrieves the target with SMB2_OP_GET_REPARSE. +*/ +if (rc == -ENODATA ) + rc =0 ; +else + tag =IO_REPARSE_TAG_SYMLINK ; reparse_point = true;break ;case STATUS_SUCCESS :
@@ -987,7 +997,14 @@ int smb2_query_path_info(const unsigned int xid, rc = -EOPNOTSUPP ; } -if (data->reparse.tag ==IO_REPARSE_TAG_SYMLINK && !rc) { +/* +* If the symlink was already parsed in create response then it is needed to fix +* its type now (after the second call with OPEN_REPARSE_POINT which filled the +* data->fi.Attributes). If the symlink was not parsed in create response then +* the data->symlink_target was not filled yet and then the type will be fixed +* later after data->symlink_target is filled. +*/ +if (data->reparse.tag ==IO_REPARSE_TAG_SYMLINK && !rc && data->symlink_target) {bool directory =le32_to_cpu (data->fi.Attributes) &ATTR_DIRECTORY ; rc =smb2_fix_symlink_target_type (&data->symlink_target, directory, cifs_sb); }
Signature · 1 line
2.50.1 (Apple Git-155)
Steve French
Appliedmerged into cifs-2.6.git for-next and added Cc:stable
Quoted 160 lines · Carl Johnson
The macOS built-in SMB server returns
STATUS_STOPPED_ON_SYMLINKfor a CREATE on a path whose final component is a symlink, but it does not include a Symbolic Link Error Response in the error data: both ErrorContextCount and ByteCount are zero, so the symlink target is not present in the response at all. Per [MS-SMB2] section 2.2.2 such a response should carry a valid Symbolic Link Error Response, so this is a server bug, but the target can still be retrieved withFSCTL_GET_REPARSE_POINT.Frame from a capture against macOS 26.5.2 (build 25F84):
SMB2 hdr : Status=0x8000002d
STATUS_STOPPED_ON_SYMLINK, Cmd=Create Error Rsp: StructureSize=0x0009
Error Context Count: 0
Byte Count: 0
Error Data: 00
symlink_data()cannot find a structsmb2_symlink_err_rspin such a response and returns-EINVAL, whichparse_create_response()propagates, sosmb2_query_path_info()bails out atif (rc || !data->reparse_point) goto out;before it can retry with
SMB2_OP_GET_REPARSE.stat(),readlink()and ls of any server-side symlink then fail with-EINVAL:$ ls -la Config l????????? ? ? ? ? ? Config.json $ stat Config/Config.json stat: cannot statx 'Config/Config.json': Invalid argumentA 5.10 client resolves these symlinks correctly against the same server and share, so this is a regression for Apple SMB servers.
Handle it in several places:
-
symlink_data()detects the empty response (ErrorContextCount and ByteCount both zero) and returns a distinct-ENODATA, so that "server did not send the target" can be told apart from a genuinely malformed response and only this case is worked around.-
parse_create_response()treats-ENODATAlikeSTATUS_IO_REPARSE_TAG_NOT_HANDLED, which does not carry the target
either: leave the reparse tag unset and clear rc, so the existingSMB2_OP_GET_REPARSEpath retrieves the target.-
smb2_query_path_info()only fixes up the symlink target type when the target is already known.SMB2_OP_GET_REPARSEsetsdata->reparse.tagbut does not parse the target out of the reparse buffer; that happens later, inreparse_info_to_fattr(). Without this checksmb2_fix_symlink_target_type()is called with a NULL target and returns-EIO. This could not happen with servers that send the target inline and therefore skipSMB2_OP_GET_REPARSE.-
smb2_open_file()maps-ENODATAto-EIO, matchingSTATUS_IO_REPARSE_TAG_NOT_HANDLED, so its callers retrieve the target withSMB2_OP_GET_REPARSEas well.Tested on Debian 13, kernel 6.18.38 (armv7), against macOS 26.5.2:
symlinks now resolve, including relative, parent-traversing and directory symlinks, and reads through symlinks succeed.Co-developed-by Pali Rohár <[email protected]>Signed-off-by Pali Rohár <[email protected]>Signed-off-by Carl Johnson <[email protected]>---
v2: rebased onto cifs-2.6 for-next; no functional change. v1 was generated
against the 6.18.38 stable tree and did not apply to mainline/for-next.fs/smb/client/smb2file.c | 21 +++++++++++++++++++++ fs/smb/client/smb2inode.c | 23 ++++++++++++++++++++--- 2 files changed, 41 insertions(+), 3 deletions(-)fs/smb/client/smb2file.c +21
-0@@ -30,6 +30,19 @@ static struct smb2_symlink_err_rsp *symlink_data(const struct kvec *iov)u8 *end = (u8 *)err + iov->iov_len;u32 len; +/* +* Per [MS-SMB2] section 2.2.2, a STATUS_STOPPED_ON_SYMLINK response has to +* carry a Symbolic Link Error Response, so ByteCount cannot be zero. Some +* servers (e.g. the macOS built-in SMB server) violate this and return an +* empty error response, with both ErrorContextCount and ByteCount set to +* zero, i.e. without the symlink target. Detect this and return -ENODATA +* so that callers can tell "server did not send the target" apart from a +* malformed response, and retrieve the target with FSCTL_GET_REPARSE_POINT +* instead. +*/ +if (!err->ErrorContextCount && !le32_to_cpu (err->ByteCount)) +return ERR_PTR (-ENODATA ); +if (err->ErrorContextCount) {struct smb2_error_context_rsp *p;@@ -199,6 +212,14 @@ int smb2_open_file(const unsigned int xid, struct cifs_open_parms *oparms, rc =smb2_parse_symlink_response (oparms->cifs_sb, &err_iov, oparms->path, &data->symlink_target); +/* +* If smb2_parse_symlink_response returned -ENODATA then the +* symlink_target was not sent. Treat this as if the SMB2_open() +* failed with STATUS_IO_REPARSE_TAG_NOT_HANDLED status, which is +* indicated by the -EIO errno. +*/ +if (rc == -ENODATA ) + rc = -EIO ;if (!rc) {memset (&data->fi,0 ,sizeof (data->fi)); oparms->create_options |=OPEN_REPARSE_POINT ;fs/smb/client/smb2inode.c +20
-4@@ -792,9 +792,19 @@ static int parse_create_response(struct cifs_open_info_data *data, rc =smb2_parse_symlink_response (cifs_sb, iov, full_path, &data->symlink_target); -if (rc) +if (rc !=0 && rc != -ENODATA )return rc; - tag =IO_REPARSE_TAG_SYMLINK ; +/* +* -ENODATA means that the response was parsed but did not contain +* the symlink target at all (see symlink_data()). Treat it like +* STATUS_IO_REPARSE_TAG_NOT_HANDLED, which does not contain it +* either: leave the tag unset and clear rc, so that the caller +* retrieves the target with SMB2_OP_GET_REPARSE. +*/ +if (rc == -ENODATA ) + rc =0 ; +else + tag =IO_REPARSE_TAG_SYMLINK ; reparse_point = true;break ;case STATUS_SUCCESS :@@ -987,7 +997,14 @@ int smb2_query_path_info(const unsigned int xid, rc = -EOPNOTSUPP ; } -if (data->reparse.tag ==IO_REPARSE_TAG_SYMLINK && !rc) { +/* +* If the symlink was already parsed in create response then it is needed to fix +* its type now (after the second call with OPEN_REPARSE_POINT which filled the +* data->fi.Attributes). If the symlink was not parsed in create response then +* the data->symlink_target was not filled yet and then the type will be fixed +* later after data->symlink_target is filled. +*/ +if (data->reparse.tag ==IO_REPARSE_TAG_SYMLINK && !rc && data->symlink_target) {bool directory =le32_to_cpu (data->fi.Attributes) &ATTR_DIRECTORY ; rc =smb2_fix_symlink_target_type (&data->symlink_target, directory, cifs_sb); } --2.50.1 (Apple Git-155)
Signature · 3 lines
Thanks, Steve