| Message ID | 20180117172200.3221-1-aaptel@suse.com |
|---|---|
| Headers | show |
| Series | Make IPC a regular tcon and fix SMB2 domain-based DFS | expand |
Very Nice. I like that it also gets rid of a lot of magic tcon==NULL. Reviewed-by: Ronnie Sahlberg <lsahlber@redhat.com> ----- Original Message ----- From: "Aurelien Aptel" <aaptel@suse.com> To: linux-cifs@vger.kernel.org Cc: smfrench@gmail.com, "Aurelien Aptel" <aaptel@suse.com> Sent: Thursday, 18 January, 2018 4:21:57 AM Subject: [PATCH v1 0/3] Make IPC a regular tcon and fix SMB2 domain-based DFS Hi, The DFS code for smb2 doesn't work [1] when cifs.ko connects to a domain-based DFS where you have an extra level on indirection (so 3 machines: AD, nameserver and the final storage). * In SMB1 code, the DFS request is made without the tcon object. It works directly on the session. * In SMB2 code, we use SMB2_ioctl() which works on a tcon. For domain-based DFS UNC \\DOMAIN\namespace\link, cifs.ko first tries to connect to \\DOMAIN\namespace. DOMAIN doesn't have a "namespace" share so no valid tcon object is created. * SMB1 doesn't care as it's working on the session object and just use the ses->ipc_tid number to make the DFS request. * SMB2 code on the other hand fails as SMB2_ioctl() needs a tcon object (even if you pass use_ipc=true) because it uses the tcon you pass and simply overwrites the packet tid with ipc_tid. There is no valid tcon object at that time since DOMAIN rejected the "namespace" tcon request. So to solve this I thought I would get rid of ses->ipc_tid and make a real IPC tcon, that would be added to the session tcon_list and be managed like any other tcon (almost). This patch series cleans some definitions and makes the IPC connection a regular tcon owned by the session. It removes the use_ipc param from SMB2_ioct() (you can pass ses->tcon_ipc as the tcon param). I've made the choice of having the IPC tcon completely associated to the session instead of any other tcon. Its refcount is not really used as it's always created as soon as we have a valid session and always destroy when the session is destroyed and no mount points directly refer to it. All patches compile when applied sequentially, no warnings with checkpatch.pl, no warnings with make C=1. I've tested this against Windows Server 2016 vms with smb1, 2, 3 (with and without encryption) and it seems to be working properly. I've also quickly tested the reconnection code path but it probably needs some work still (making sure DFS works after a reconnection) as there are multiple ways to trigger the reconnection (I've been using qemu monitor "set_link xyz off" which is analogous to unplugging the network cable IIUC). Please take a look, review, test :) 1: https://bugzilla.samba.org/show_bug.cgi?id=12917 Aurelien Aptel (3): CIFS: set SERVER_NAME_LENGTH to serverName actual size CIFS: make IPC a regular tcon CIFS: use tcon_ipc instead of use_ipc parameter from SMB2_ioctl fs/cifs/cifsglob.h | 14 ++--- fs/cifs/cifssmb.c | 5 +- fs/cifs/connect.c | 159 ++++++++++++++++++++++++++++++++++++++++------------ fs/cifs/inode.c | 2 +- fs/cifs/smb2file.c | 2 +- fs/cifs/smb2ops.c | 50 +++++++---------- fs/cifs/smb2pdu.c | 36 ++---------- fs/cifs/smb2proto.h | 3 +- 8 files changed, 160 insertions(+), 111 deletions(-)