[CIFS] Reduce number of socket retries in large write path
CIFS in some heavy stress conditions cifs could get EAGAIN
repeatedly in smb_send2 which led to repeated retries and eventually
failure of large writes which could lead to data corruption.
There are three changes that were suggested by various network
developers:
1) convert cifs from non-blocking to blocking tcp sendmsg
(we left in the retry on failure)
2) change cifs to not set sendbuf and rcvbuf size for the socket
(let tcp autotune the buffer sizes since that works much better
in the TCP stack now)
3) if we have a partial frame sent in smb_send2, mark the tcp
session as invalid (close the socket and reconnect) so we do
not corrupt the remaining part of the SMB with the beginning
of the next SMB.
This does not appear to hurt performance measurably and has
been run in various scenarios, but it definately removes
a corruption that we were seeing in some high stress
test cases.
Acked-by: Shirish Pargaonkar <shirishp@us.ibm.com>
Signed-off-by: Steve French <sfrench@us.ibm.com>
diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c
index 71b7661..e9f9248 100644
--- a/fs/cifs/connect.c
+++ b/fs/cifs/connect.c
@@ -92,6 +92,8 @@
bool seal:1; /* request transport encryption on share */
bool nodfs:1; /* Do not request DFS, even if available */
bool local_lease:1; /* check leases only on local system, not remote */
+ bool noblocksnd:1;
+ bool noautotune:1;
unsigned int rsize;
unsigned int wsize;
unsigned int sockopt;
@@ -102,9 +104,11 @@
static int ipv4_connect(struct sockaddr_in *psin_server,
struct socket **csocket,
char *netb_name,
- char *server_netb_name);
+ char *server_netb_name,
+ bool noblocksnd,
+ bool nosndbuf); /* ipv6 never set sndbuf size */
static int ipv6_connect(struct sockaddr_in6 *psin_server,
- struct socket **csocket);
+ struct socket **csocket, bool noblocksnd);
/*
@@ -191,12 +195,13 @@
try_to_freeze();
if (server->protocolType == IPV6) {
rc = ipv6_connect(&server->addr.sockAddr6,
- &server->ssocket);
+ &server->ssocket, server->noautotune);
} else {
rc = ipv4_connect(&server->addr.sockAddr,
&server->ssocket,
server->workstation_RFC1001_name,
- server->server_RFC1001_name);
+ server->server_RFC1001_name,
+ server->noblocksnd, server->noautotune);
}
if (rc) {
cFYI(1, ("reconnect error %d", rc));
@@ -1192,6 +1197,10 @@
/* ignore */
} else if (strnicmp(data, "rw", 2) == 0) {
vol->rw = true;
+ } else if (strnicmp(data, "noblocksend", 11) == 0) {
+ vol->noblocksnd = 1;
+ } else if (strnicmp(data, "noautotune", 10) == 0) {
+ vol->noautotune = 1;
} else if ((strnicmp(data, "suid", 4) == 0) ||
(strnicmp(data, "nosuid", 6) == 0) ||
(strnicmp(data, "exec", 4) == 0) ||
@@ -1518,7 +1527,8 @@
static int
ipv4_connect(struct sockaddr_in *psin_server, struct socket **csocket,
- char *netbios_name, char *target_name)
+ char *netbios_name, char *target_name,
+ bool noblocksnd, bool noautotune)
{
int rc = 0;
int connected = 0;
@@ -1590,11 +1600,16 @@
(*csocket)->sk->sk_sndbuf,
(*csocket)->sk->sk_rcvbuf, (*csocket)->sk->sk_rcvtimeo));
(*csocket)->sk->sk_rcvtimeo = 7 * HZ;
+ if (!noblocksnd)
+ (*csocket)->sk->sk_sndtimeo = 3 * HZ;
+
/* make the bufsizes depend on wsize/rsize and max requests */
- if ((*csocket)->sk->sk_sndbuf < (200 * 1024))
- (*csocket)->sk->sk_sndbuf = 200 * 1024;
- if ((*csocket)->sk->sk_rcvbuf < (140 * 1024))
- (*csocket)->sk->sk_rcvbuf = 140 * 1024;
+ if (noautotune) {
+ if ((*csocket)->sk->sk_sndbuf < (200 * 1024))
+ (*csocket)->sk->sk_sndbuf = 200 * 1024;
+ if ((*csocket)->sk->sk_rcvbuf < (140 * 1024))
+ (*csocket)->sk->sk_rcvbuf = 140 * 1024;
+ }
/* send RFC1001 sessinit */
if (psin_server->sin_port == htons(RFC1001_PORT)) {
@@ -1631,7 +1646,7 @@
/* sizeof RFC1002_SESSION_REQUEST with no scope */
smb_buf->smb_buf_length = 0x81000044;
rc = smb_send(*csocket, smb_buf, 0x44,
- (struct sockaddr *)psin_server);
+ (struct sockaddr *)psin_server, noblocksnd);
kfree(ses_init_buf);
msleep(1); /* RFC1001 layer in at least one server
requires very short break before negprot
@@ -1651,7 +1666,8 @@
}
static int
-ipv6_connect(struct sockaddr_in6 *psin_server, struct socket **csocket)
+ipv6_connect(struct sockaddr_in6 *psin_server, struct socket **csocket,
+ bool noblocksnd)
{
int rc = 0;
int connected = 0;
@@ -1720,6 +1736,9 @@
the default. sock_setsockopt not used because it expects
user space buffer */
(*csocket)->sk->sk_rcvtimeo = 7 * HZ;
+ if (!noblocksnd)
+ (*csocket)->sk->sk_sndtimeo = 3 * HZ;
+
return rc;
}
@@ -1983,11 +2002,14 @@
cFYI(1, ("attempting ipv6 connect"));
/* BB should we allow ipv6 on port 139? */
/* other OS never observed in Wild doing 139 with v6 */
- rc = ipv6_connect(&sin_server6, &csocket);
+ rc = ipv6_connect(&sin_server6, &csocket,
+ volume_info.noblocksnd);
} else
rc = ipv4_connect(&sin_server, &csocket,
volume_info.source_rfc1001_name,
- volume_info.target_rfc1001_name);
+ volume_info.target_rfc1001_name,
+ volume_info.noblocksnd,
+ volume_info.noautotune);
if (rc < 0) {
cERROR(1, ("Error connecting to IPv4 socket. "
"Aborting operation"));
@@ -2002,6 +2024,8 @@
sock_release(csocket);
goto out;
} else {
+ srvTcp->noblocksnd = volume_info.noblocksnd;
+ srvTcp->noautotune = volume_info.noautotune;
memcpy(&srvTcp->addr.sockAddr, &sin_server,
sizeof(struct sockaddr_in));
atomic_set(&srvTcp->inFlight, 0);