Created attachment 19302 [details] log of a failing run Samba versions: 4.24.7 (Debian trixie-backports, 4.24.7+dfsg-1~bpo13+1) and 4.25.0 (Debian testing, 4.25.0+dfsg-1); also seen on 4.22.11 (Debian 13). Setup: a share with a directory on an ntfs-3g (FUSE) mount, e.g. path = /srv/data with ntfs-3g -o rw,windows_names,uid=1000,gid=1000 /dev/sdX1 /srv/data/ntfs Nothing special in smb.conf: "smb3 directory leases" and "vfs mkdir use tmp name" are at their defaults (auto, which resolves to yes). Client: Windows 11, SMB 3.1.1. The attached rig reproduces it in a Debian container with ntfs-3g mounted from an image file. Steps: from Windows, copy a small tree of new folders onto the ntfs-3g part of the share: C:\src\test\readme.txt C:\src\test\dir1\dir2\dir3\dir4\test.txt robocopy C:\src \\server\data\ntfs /E Result: creating the new folders fails with NT_STATUS_INSUFFICIENT_RESOURCES; robocopy reports ERROR 1450 (ERROR_NO_SYSTEM_RESOURCES) for "test", "dir1" and "dir2". The folder exists afterwards; robocopy's retry gets NT_STATUS_OBJECT_NAME_COLLISION and carries on. At the default log level smbd logs, once per failure: source3/locking/leases_db.c:150(leases_db_do_locked_fn) leases_db_do_locked_fn: ndr_push_struct_blob_failed: Character Conversion Error At log level 3 (attached) the lines before it are: convert_string_handle: Conversion error: Incomplete multibyte sequence (on other folders: convert_string_internal: Conversion error: Illegal multibyte sequence) ndr_push_charset: ndr_push_error(Character Conversion Error): Bad character conversion at librpc/ndr/ndr_string.c:1065 leases_db_do_locked_fn: ndr_push_struct_blob_failed: Character Conversion Error smbd_smb2_request_error_ex: idx[1] status[NT_STATUS_INSUFFICIENT_RESOURCES] || at source3/smbd/smb2_create.c:393 Every folder name is plain ASCII, so the string being pushed is not the folder's name. In one run at log level 5 the share mode record failed the same way instead (share_mode_data_ltdb_store: ndr_push_share_mode_data failed: Character Conversion Error). Expected: the folders are created without an error, as on a non-FUSE filesystem. It is timing-dependent: at the default log level it failed in every run (17 of 17), at log level 3 in 3 of 7, at log level 5 in 2 of 8, at log level 10 in none of 4. What isolates it (4.24.7, same container, client and tree): - share on ntfs-3g, defaults: fails - share on ext4, defaults: works - ntfs-3g with "vfs mkdir use tmp name = no": works - ntfs-3g with "smb3 directory leases = no": works Copying files into folders that already exist always works; only creating a folder fails. Analysis (from reading source3/smbd/open.c, not from a debugger): mkdir_internal() creates the directory under a temporary name and renames it into place with VFS_RENAME_HOW_NO_REPLACE. FUSE filesystems such as ntfs-3g do not support RENAME_NOREPLACE, so renameat2() returns EINVAL and the fallback runs: mkdirat() the real name, then rename the temporary directory over it. After that path the lease (and share mode) record of the new directory holds a string that is not valid in the unix charset although every name is ASCII, which suggests a pointer to freed or reused memory; that would also explain why more logging makes it rarer. Not confirmed with valgrind or ASan. Related: bug 15801 (same tmp-name mkdir code on OpenBSD, fixed by mapping "auto" to "no" on that platform only). Linux with a filesystem that lacks RENAME_NOREPLACE still takes the fallback. Workarounds: "vfs mkdir use tmp name = no" on the affected share, or "smb3 directory leases = no".
Created attachment 19303 [details] test setup with a generic share (data), user (smbtest) and a README