From: Wang Zhaolong Date: Thu, 3 Jul 2025 13:29:52 +0000 (+0800) Subject: smb: client: fix race condition in negotiate timeout by using more precise timing X-Git-Tag: v6.16-rc5~11^2~1 X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=266b5d02e14f3a0e07414e11f239397de0577a1d;p=thirdparty%2Fkernel%2Flinux.git smb: client: fix race condition in negotiate timeout by using more precise timing When the SMB server reboots and the client immediately accesses the mount point, a race condition can occur that causes operations to fail with "Host is down" error. Reproduction steps: # Mount SMB share mount -t cifs //192.168.245.109/TEST /mnt/ -o xxxx ls /mnt # Reboot server ssh root@192.168.245.109 reboot ssh root@192.168.245.109 /path/to/cifs_server_setup.sh ssh root@192.168.245.109 systemctl stop firewalld # Immediate access fails ls /mnt ls: cannot access '/mnt': Host is down # But works if there is a delay The issue is caused by a race condition between negotiate and reconnect. The 20-second negotiate timeout mechanism can interfere with the normal recovery process when both are triggered simultaneously. ls cifsd --------------------------------------------------- cifs_getattr cifs_revalidate_dentry cifs_get_inode_info cifs_get_fattr smb2_query_path_info smb2_compound_op SMB2_open_init smb2_reconnect cifs_negotiate_protocol smb2_negotiate cifs_send_recv smb_send_rqst wait_for_response cifs_demultiplex_thread cifs_read_from_socket cifs_readv_from_socket server_unresponsive cifs_reconnect __cifs_reconnect cifs_abort_connection mid->mid_state = MID_RETRY_NEEDED cifs_wake_up_task cifs_sync_mid_result // case MID_RETRY_NEEDED rc = -EAGAIN; // In smb2_negotiate() rc = -EHOSTDOWN; The server_unresponsive() timeout triggers cifs_reconnect(), which aborts ongoing mid requests and causes the ls command to receive -EAGAIN, leading to -EHOSTDOWN. Fix this by introducing a dedicated `neg_start` field to precisely tracks when the negotiate process begins. The timeout check now uses this accurate timestamp instead of `lstrp`, ensuring that: 1. Timeout is only triggered after negotiate has actually run for 20s 2. The mechanism doesn't interfere with concurrent recovery processes 3. Uninitialized timestamps (value 0) don't trigger false timeouts Fixes: 7ccc1465465d ("smb: client: fix hang in wait_for_response() for negproto") Signed-off-by: Wang Zhaolong Signed-off-by: Steve French --- diff --git a/fs/smb/client/cifsglob.h b/fs/smb/client/cifsglob.h index fdd95e5100cd..89160bc34d35 100644 --- a/fs/smb/client/cifsglob.h +++ b/fs/smb/client/cifsglob.h @@ -777,6 +777,7 @@ struct TCP_Server_Info { __le32 session_key_id; /* retrieved from negotiate response and send in session setup request */ struct session_key session_key; unsigned long lstrp; /* when we got last response from this server */ + unsigned long neg_start; /* when negotiate started (jiffies) */ struct cifs_secmech secmech; /* crypto sec mech functs, descriptors */ #define CIFS_NEGFLAVOR_UNENCAP 1 /* wct == 17, but no ext_sec */ #define CIFS_NEGFLAVOR_EXTENDED 2 /* wct == 17, ext_sec bit set */ diff --git a/fs/smb/client/connect.c b/fs/smb/client/connect.c index 484b677143fd..205f547ca49e 100644 --- a/fs/smb/client/connect.c +++ b/fs/smb/client/connect.c @@ -679,12 +679,12 @@ server_unresponsive(struct TCP_Server_Info *server) /* * If we're in the process of mounting a share or reconnecting a session * and the server abruptly shut down (e.g. socket wasn't closed, packet - * had been ACK'ed but no SMB response), don't wait longer than 20s to - * negotiate protocol. + * had been ACK'ed but no SMB response), don't wait longer than 20s from + * when negotiate actually started. */ spin_lock(&server->srv_lock); if (server->tcpStatus == CifsInNegotiate && - time_after(jiffies, server->lstrp + 20 * HZ)) { + time_after(jiffies, server->neg_start + 20 * HZ)) { spin_unlock(&server->srv_lock); cifs_reconnect(server, false); return true; @@ -4209,6 +4209,7 @@ retry: server->lstrp = jiffies; server->tcpStatus = CifsInNegotiate; + server->neg_start = jiffies; spin_unlock(&server->srv_lock); rc = server->ops->negotiate(xid, ses, server);