On 11/8/2022 11:33 AM, Shyam Prasad N wrote:
Came across a couple of race conditions between cifsd and i/o threads.
Fixed them here. Please review the changes.
This change should also ensure that binding session requests do not
race with the primary session request.
https://github.com/sprasad-microsoft/smb3-kernel-client/commit/c4ed4d985488640469acfc621bed5c3a55d67ac6.patch
Copy/pasting from that page:
@@ -4111,6 +4111,16 @@ cifs_setup_session(const unsigned int xid, struct
cifs_ses *ses,
else
scnprintf(ses->ip_addr, sizeof(ses->ip_addr), "%pI4", &addr->sin_addr);
+ /*
+ * if primary channel is still being setup,
+ * we cannot bind to it. try again after sometime
+ */
+ if (ses->ses_status == SES_IN_SETUP) {
+ spin_unlock(&ses->ses_lock);
+ msleep(10);
+ return -EAGAIN;
+ }
+
if (ses->ses_status != SES_GOOD &&
ses->ses_status != SES_NEW &&
ses->ses_status != SES_NEED_RECON) {
Sleeping for 10ms is really an ugly approach. Apart from making a
completely arbitrary timing choice, blocking the caller uninterruptibly
is pretty rude. And if the existing setup fails, this will blindly
retry it, immediately.
Can't this take a more formal queued approach, with a proper wakeup on
success, or bail out on interrupt, timeout or fail?
Tom.
