[PATCH] RPC: introduce client-side transport switch
Chuck Lever [Thu, 11 Aug 2005 20:25:23 +0000 (16:25 -0400)]
 Move the bulk of client-side socket-specific code into a separate source
 file, net/sunrpc/xprtsock.c.

 Test-plan:
 Millions of fsx operations.  Performance characterization such as "sio" or
 "iozone".  Destructive testing (unplugging the network temporarily, server
 reboots).  Connectathon with v2, v3, and v4.

 Version: Thu, 11 Aug 2005 16:03:38 -0400

Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

include/linux/sunrpc/xdr.h
include/linux/sunrpc/xprt.h
net/sunrpc/Makefile
net/sunrpc/clnt.c
net/sunrpc/sysctl.c
net/sunrpc/xdr.c
net/sunrpc/xprt.c
net/sunrpc/xprtsock.c [new file with mode: 0644]

index d8b7656..5da9687 100644 (file)
@@ -165,11 +165,6 @@ extern int csum_partial_copy_to_xdr(struct xdr_buf *, struct sk_buff *);
 extern ssize_t xdr_partial_copy_from_skb(struct xdr_buf *, unsigned int,
                skb_reader_t *, skb_read_actor_t);
 
-struct socket;
-struct sockaddr;
-extern int xdr_sendpages(struct socket *, struct sockaddr *, int,
-               struct xdr_buf *, unsigned int, int);
-
 extern int xdr_encode_word(struct xdr_buf *, int, u32);
 extern int xdr_decode_word(struct xdr_buf *, int, u32 *);
 
index e618c16..d82b47a 100644 (file)
@@ -59,7 +59,13 @@ extern unsigned int xprt_tcp_slot_table_entries;
  */
 #define RPC_REESTABLISH_TIMEOUT        (15*HZ)
 
-/* RPC call and reply header size as number of 32bit words (verifier
+/*
+ * RPC transport idle timeout.
+ */
+#define RPC_IDLE_DISCONNECT_TIMEOUT    (5*60*HZ)
+
+/*
+ * RPC call and reply header size as number of 32bit words (verifier
  * size computed separately)
  */
 #define RPC_CALLHDRSIZE                6
@@ -121,12 +127,19 @@ struct rpc_rqst {
 #define rq_svec                        rq_snd_buf.head
 #define rq_slen                        rq_snd_buf.len
 
-#define XPRT_LAST_FRAG         (1 << 0)
-#define XPRT_COPY_RECM         (1 << 1)
-#define XPRT_COPY_XID          (1 << 2)
-#define XPRT_COPY_DATA         (1 << 3)
+struct rpc_task;
+struct rpc_xprt;
+
+struct rpc_xprt_ops {
+       void            (*set_buffer_size)(struct rpc_xprt *xprt);
+       void            (*connect)(struct rpc_task *task);
+       int             (*send_request)(struct rpc_task *task);
+       void            (*close)(struct rpc_xprt *xprt);
+       void            (*destroy)(struct rpc_xprt *xprt);
+};
 
 struct rpc_xprt {
+       struct rpc_xprt_ops *   ops;            /* transport methods */
        struct socket *         sock;           /* BSD socket layer */
        struct sock *           inet;           /* INET layer */
 
@@ -199,14 +212,22 @@ struct rpc_xprt {
        wait_queue_head_t       cong_wait;
 };
 
+#define XPRT_LAST_FRAG         (1 << 0)
+#define XPRT_COPY_RECM         (1 << 1)
+#define XPRT_COPY_XID          (1 << 2)
+#define XPRT_COPY_DATA         (1 << 3)
+
 #ifdef __KERNEL__
 
 struct rpc_xprt *      xprt_create_proto(int proto, struct sockaddr_in *addr,
                                        struct rpc_timeout *toparms);
+void                   xprt_disconnect(struct rpc_xprt *);
 int                    xprt_destroy(struct rpc_xprt *);
 void                   xprt_set_timeout(struct rpc_timeout *, unsigned int,
                                        unsigned long);
-
+struct rpc_rqst *      xprt_lookup_rqst(struct rpc_xprt *, u32);
+void                   xprt_complete_rqst(struct rpc_xprt *,
+                                       struct rpc_rqst *, int);
 void                   xprt_reserve(struct rpc_task *);
 int                    xprt_prepare_transmit(struct rpc_task *);
 void                   xprt_transmit(struct rpc_task *);
@@ -214,7 +235,10 @@ void                       xprt_receive(struct rpc_task *);
 int                    xprt_adjust_timeout(struct rpc_rqst *req);
 void                   xprt_release(struct rpc_task *);
 void                   xprt_connect(struct rpc_task *);
-void                   xprt_sock_setbufsize(struct rpc_xprt *);
+int                    xs_setup_udp(struct rpc_xprt *,
+                                       struct rpc_timeout *);
+int                    xs_setup_tcp(struct rpc_xprt *,
+                                       struct rpc_timeout *);
 
 #define XPRT_LOCKED    0
 #define XPRT_CONNECT   1
index f0a9556..cdcab9c 100644 (file)
@@ -6,7 +6,7 @@
 obj-$(CONFIG_SUNRPC) += sunrpc.o
 obj-$(CONFIG_SUNRPC_GSS) += auth_gss/
 
-sunrpc-y := clnt.o xprt.o socklib.o sched.o \
+sunrpc-y := clnt.o xprt.o socklib.o xprtsock.o sched.o \
            auth.o auth_null.o auth_unix.o \
            svc.o svcsock.o svcauth.o svcauth_unix.o \
            pmap_clnt.o timer.o xdr.o \
index 2d3cf0a..ab50c3c 100644 (file)
@@ -525,8 +525,7 @@ rpc_setbufsize(struct rpc_clnt *clnt, unsigned int sndsize, unsigned int rcvsize
        xprt->rcvsize = 0;
        if (rcvsize)
                xprt->rcvsize = rcvsize + RPC_SLACK_SPACE;
-       if (xprt_connected(xprt))
-               xprt_sock_setbufsize(xprt);
+       xprt->ops->set_buffer_size(xprt);
 }
 
 /*
index 1b9616a..ef48326 100644 (file)
@@ -119,6 +119,9 @@ done:
        return 0;
 }
 
+unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
+unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
+
 static unsigned int min_slot_table_size = RPC_MIN_SLOT_TABLE;
 static unsigned int max_slot_table_size = RPC_MAX_SLOT_TABLE;
 
index 9cc12ae..32df433 100644 (file)
@@ -6,15 +6,12 @@
  * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
  */
 
+#include <linux/module.h>
 #include <linux/types.h>
-#include <linux/socket.h>
 #include <linux/string.h>
 #include <linux/kernel.h>
 #include <linux/pagemap.h>
 #include <linux/errno.h>
-#include <linux/in.h>
-#include <linux/net.h>
-#include <net/sock.h>
 #include <linux/sunrpc/xdr.h>
 #include <linux/sunrpc/msg_prot.h>
 
@@ -177,103 +174,6 @@ xdr_inline_pages(struct xdr_buf *xdr, unsigned int offset,
 }
 
 
-int
-xdr_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen,
-               struct xdr_buf *xdr, unsigned int base, int msgflags)
-{
-       struct page **ppage = xdr->pages;
-       unsigned int len, pglen = xdr->page_len;
-       int err, ret = 0;
-       ssize_t (*sendpage)(struct socket *, struct page *, int, size_t, int);
-
-       len = xdr->head[0].iov_len;
-       if (base < len || (addr != NULL && base == 0)) {
-               struct kvec iov = {
-                       .iov_base = xdr->head[0].iov_base + base,
-                       .iov_len  = len - base,
-               };
-               struct msghdr msg = {
-                       .msg_name    = addr,
-                       .msg_namelen = addrlen,
-                       .msg_flags   = msgflags,
-               };
-               if (xdr->len > len)
-                       msg.msg_flags |= MSG_MORE;
-
-               if (iov.iov_len != 0)
-                       err = kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
-               else
-                       err = kernel_sendmsg(sock, &msg, NULL, 0, 0);
-               if (ret == 0)
-                       ret = err;
-               else if (err > 0)
-                       ret += err;
-               if (err != iov.iov_len)
-                       goto out;
-               base = 0;
-       } else
-               base -= len;
-
-       if (pglen == 0)
-               goto copy_tail;
-       if (base >= pglen) {
-               base -= pglen;
-               goto copy_tail;
-       }
-       if (base || xdr->page_base) {
-               pglen -= base;
-               base  += xdr->page_base;
-               ppage += base >> PAGE_CACHE_SHIFT;
-               base &= ~PAGE_CACHE_MASK;
-       }
-
-       sendpage = sock->ops->sendpage ? : sock_no_sendpage;
-       do {
-               int flags = msgflags;
-
-               len = PAGE_CACHE_SIZE;
-               if (base)
-                       len -= base;
-               if (pglen < len)
-                       len = pglen;
-
-               if (pglen != len || xdr->tail[0].iov_len != 0)
-                       flags |= MSG_MORE;
-
-               /* Hmm... We might be dealing with highmem pages */
-               if (PageHighMem(*ppage))
-                       sendpage = sock_no_sendpage;
-               err = sendpage(sock, *ppage, base, len, flags);
-               if (ret == 0)
-                       ret = err;
-               else if (err > 0)
-                       ret += err;
-               if (err != len)
-                       goto out;
-               base = 0;
-               ppage++;
-       } while ((pglen -= len) != 0);
-copy_tail:
-       len = xdr->tail[0].iov_len;
-       if (base < len) {
-               struct kvec iov = {
-                       .iov_base = xdr->tail[0].iov_base + base,
-                       .iov_len  = len - base,
-               };
-               struct msghdr msg = {
-                       .msg_flags   = msgflags,
-               };
-               err = kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
-               if (ret == 0)
-                       ret = err;
-               else if (err > 0)
-                       ret += err;
-       }
-out:
-       return ret;
-}
-
-
 /*
  * Helper routines for doing 'memmove' like operations on a struct xdr_buf
  *
index 67444f4..4342acf 100644 (file)
  *  tasks that rely on callbacks.
  *
  *  Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
- *
- *  TCP callback races fixes (C) 1998 Red Hat Software <alan@redhat.com>
- *  TCP send fixes (C) 1998 Red Hat Software <alan@redhat.com>
- *  TCP NFS related read + write fixes
- *   (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
- *
- *  Rewrite of larges part of the code in order to stabilize TCP stuff.
- *  Fix behaviour when socket buffer is full.
- *   (C) 1999 Trond Myklebust <trond.myklebust@fys.uio.no>
  */
 
+#include <linux/module.h>
+
 #include <linux/types.h>
-#include <linux/slab.h>
-#include <linux/capability.h>
-#include <linux/sched.h>
-#include <linux/errno.h>
-#include <linux/socket.h>
-#include <linux/in.h>
-#include <linux/net.h>
-#include <linux/mm.h>
-#include <linux/udp.h>
-#include <linux/tcp.h>
-#include <linux/sunrpc/clnt.h>
-#include <linux/file.h>
+#include <linux/interrupt.h>
 #include <linux/workqueue.h>
 #include <linux/random.h>
 
-#include <net/sock.h>
-#include <net/checksum.h>
-#include <net/udp.h>
-#include <net/tcp.h>
+#include <linux/sunrpc/clnt.h>
 
 /*
  * Local variables
 #endif
 
 #define XPRT_MAX_BACKOFF       (8)
-#define XPRT_IDLE_TIMEOUT      (5*60*HZ)
-#define XPRT_MAX_RESVPORT      (800)
 
 /*
  * Local functions
  */
 static void    xprt_request_init(struct rpc_task *, struct rpc_xprt *);
 static inline void     do_xprt_reserve(struct rpc_task *);
-static void    xprt_disconnect(struct rpc_xprt *);
 static void    xprt_connect_status(struct rpc_task *task);
-static struct rpc_xprt * xprt_setup(int proto, struct sockaddr_in *ap,
-                                               struct rpc_timeout *to);
-static struct socket *xprt_create_socket(struct rpc_xprt *, int, int);
-static void    xprt_bind_socket(struct rpc_xprt *, struct socket *);
 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
 
 static int     xprt_clear_backlog(struct rpc_xprt *xprt);
 
-#ifdef RPC_DEBUG_DATA
-/*
- * Print the buffer contents (first 128 bytes only--just enough for
- * diropres return).
- */
-static void
-xprt_pktdump(char *msg, u32 *packet, unsigned int count)
-{
-       u8      *buf = (u8 *) packet;
-       int     j;
-
-       dprintk("RPC:      %s\n", msg);
-       for (j = 0; j < count && j < 128; j += 4) {
-               if (!(j & 31)) {
-                       if (j)
-                               dprintk("\n");
-                       dprintk("0x%04x ", j);
-               }
-               dprintk("%02x%02x%02x%02x ",
-                       buf[j], buf[j+1], buf[j+2], buf[j+3]);
-       }
-       dprintk("\n");
-}
-#else
-static inline void
-xprt_pktdump(char *msg, u32 *packet, unsigned int count)
-{
-       /* NOP */
-}
-#endif
-
-/*
- * Look up RPC transport given an INET socket
- */
-static inline struct rpc_xprt *
-xprt_from_sock(struct sock *sk)
-{
-       return (struct rpc_xprt *) sk->sk_user_data;
-}
-
 /*
  * Serialize write access to sockets, in order to prevent different
  * requests from interfering with each other.
@@ -235,62 +167,6 @@ xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 }
 
 /*
- * Write data to socket.
- */
-static inline int
-xprt_sendmsg(struct rpc_xprt *xprt, struct rpc_rqst *req)
-{
-       struct socket   *sock = xprt->sock;
-       struct xdr_buf  *xdr = &req->rq_snd_buf;
-       struct sockaddr *addr = NULL;
-       int addrlen = 0;
-       unsigned int    skip;
-       int             result;
-
-       if (!sock)
-               return -ENOTCONN;
-
-       xprt_pktdump("packet data:",
-                               req->rq_svec->iov_base,
-                               req->rq_svec->iov_len);
-
-       /* For UDP, we need to provide an address */
-       if (!xprt->stream) {
-               addr = (struct sockaddr *) &xprt->addr;
-               addrlen = sizeof(xprt->addr);
-       }
-       /* Dont repeat bytes */
-       skip = req->rq_bytes_sent;
-
-       clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
-       result = xdr_sendpages(sock, addr, addrlen, xdr, skip, MSG_DONTWAIT);
-
-       dprintk("RPC:      xprt_sendmsg(%d) = %d\n", xdr->len - skip, result);
-
-       if (result >= 0)
-               return result;
-
-       switch (result) {
-       case -ECONNREFUSED:
-               /* When the server has died, an ICMP port unreachable message
-                * prompts ECONNREFUSED.
-                */
-       case -EAGAIN:
-               break;
-       case -ECONNRESET:
-       case -ENOTCONN:
-       case -EPIPE:
-               /* connection broken */
-               if (xprt->stream)
-                       result = -ENOTCONN;
-               break;
-       default:
-               printk(KERN_NOTICE "RPC: sendmsg returned error %d\n", -result);
-       }
-       return result;
-}
-
-/*
  * Van Jacobson congestion avoidance. Check if the congestion window
  * overflowed. Put the task to sleep if this is the case.
  */
@@ -405,48 +281,20 @@ int xprt_adjust_timeout(struct rpc_rqst *req)
        return status;
 }
 
-/*
- * Close down a transport socket
- */
-static void
-xprt_close(struct rpc_xprt *xprt)
-{
-       struct socket   *sock = xprt->sock;
-       struct sock     *sk = xprt->inet;
-
-       if (!sk)
-               return;
-
-       write_lock_bh(&sk->sk_callback_lock);
-       xprt->inet = NULL;
-       xprt->sock = NULL;
-
-       sk->sk_user_data    = NULL;
-       sk->sk_data_ready   = xprt->old_data_ready;
-       sk->sk_state_change = xprt->old_state_change;
-       sk->sk_write_space  = xprt->old_write_space;
-       write_unlock_bh(&sk->sk_callback_lock);
-
-       sk->sk_no_check  = 0;
-
-       sock_release(sock);
-}
-
 static void
 xprt_socket_autoclose(void *args)
 {
        struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 
        xprt_disconnect(xprt);
-       xprt_close(xprt);
+       xprt->ops->close(xprt);
        xprt_release_write(xprt, NULL);
 }
 
 /*
  * Mark a transport as disconnected
  */
-static void
-xprt_disconnect(struct rpc_xprt *xprt)
+void xprt_disconnect(struct rpc_xprt *xprt)
 {
        dprintk("RPC:      disconnected transport %p\n", xprt);
        spin_lock_bh(&xprt->sock_lock);
@@ -479,57 +327,6 @@ out_abort:
        spin_unlock(&xprt->sock_lock);
 }
 
-static void xprt_socket_connect(void *args)
-{
-       struct rpc_xprt *xprt = (struct rpc_xprt *)args;
-       struct socket *sock = xprt->sock;
-       int status = -EIO;
-
-       if (xprt->shutdown || xprt->addr.sin_port == 0)
-               goto out;
-
-       /*
-        * Start by resetting any existing state
-        */
-       xprt_close(xprt);
-       sock = xprt_create_socket(xprt, xprt->prot, xprt->resvport);
-       if (sock == NULL) {
-               /* couldn't create socket or bind to reserved port;
-                * this is likely a permanent error, so cause an abort */
-               goto out;
-       }
-       xprt_bind_socket(xprt, sock);
-       xprt_sock_setbufsize(xprt);
-
-       status = 0;
-       if (!xprt->stream)
-               goto out;
-
-       /*
-        * Tell the socket layer to start connecting...
-        */
-       status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
-                       sizeof(xprt->addr), O_NONBLOCK);
-       dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
-                       xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
-       if (status < 0) {
-               switch (status) {
-                       case -EINPROGRESS:
-                       case -EALREADY:
-                               goto out_clear;
-               }
-       }
-out:
-       if (status < 0)
-               rpc_wake_up_status(&xprt->pending, status);
-       else
-               rpc_wake_up(&xprt->pending);
-out_clear:
-       smp_mb__before_clear_bit();
-       clear_bit(XPRT_CONNECTING, &xprt->sockstate);
-       smp_mb__after_clear_bit();
-}
-
 /*
  * Attempt to connect a TCP socket.
  *
@@ -552,30 +349,16 @@ void xprt_connect(struct rpc_task *task)
        if (!xprt_lock_write(xprt, task))
                return;
        if (xprt_connected(xprt))
-               goto out_write;
+               xprt_release_write(xprt, task);
+       else {
+               if (task->tk_rqstp)
+                       task->tk_rqstp->rq_bytes_sent = 0;
 
-       if (task->tk_rqstp)
-               task->tk_rqstp->rq_bytes_sent = 0;
-
-       task->tk_timeout = RPC_CONNECT_TIMEOUT;
-       rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
-       if (!test_and_set_bit(XPRT_CONNECTING, &xprt->sockstate)) {
-               /* Note: if we are here due to a dropped connection
-                *       we delay reconnecting by RPC_REESTABLISH_TIMEOUT/HZ
-                *       seconds
-                */
-               if (xprt->sock != NULL)
-                       schedule_delayed_work(&xprt->sock_connect,
-                                       RPC_REESTABLISH_TIMEOUT);
-               else {
-                       schedule_work(&xprt->sock_connect);
-                       if (!RPC_IS_ASYNC(task))
-                               flush_scheduled_work();
-               }
+               task->tk_timeout = RPC_CONNECT_TIMEOUT;
+               rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
+               xprt->ops->connect(task);
        }
        return;
- out_write:
-       xprt_release_write(xprt, task);
 }
 
 /*
@@ -624,8 +407,7 @@ xprt_connect_status(struct rpc_task *task)
 /*
  * Look up the RPC request corresponding to a reply, and then lock it.
  */
-static inline struct rpc_rqst *
-xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
+struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
 {
        struct list_head *pos;
        struct rpc_rqst *req = NULL;
@@ -644,8 +426,7 @@ xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
  * Complete reply received.
  * The TCP code relies on us to remove the request from xprt->pending.
  */
-static void
-xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
+void xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 {
        struct rpc_task *task = req->rq_task;
        struct rpc_clnt *clnt = task->tk_client;
@@ -692,409 +473,6 @@ xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 }
 
 /*
- * Input handler for RPC replies. Called from a bottom half and hence
- * atomic.
- */
-static void
-udp_data_ready(struct sock *sk, int len)
-{
-       struct rpc_task *task;
-       struct rpc_xprt *xprt;
-       struct rpc_rqst *rovr;
-       struct sk_buff  *skb;
-       int err, repsize, copied;
-       u32 _xid, *xp;
-
-       read_lock(&sk->sk_callback_lock);
-       dprintk("RPC:      udp_data_ready...\n");
-       if (!(xprt = xprt_from_sock(sk))) {
-               printk("RPC:      udp_data_ready request not found!\n");
-               goto out;
-       }
-
-       dprintk("RPC:      udp_data_ready client %p\n", xprt);
-
-       if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
-               goto out;
-
-       if (xprt->shutdown)
-               goto dropit;
-
-       repsize = skb->len - sizeof(struct udphdr);
-       if (repsize < 4) {
-               printk("RPC: impossible RPC reply size %d!\n", repsize);
-               goto dropit;
-       }
-
-       /* Copy the XID from the skb... */
-       xp = skb_header_pointer(skb, sizeof(struct udphdr),
-                               sizeof(_xid), &_xid);
-       if (xp == NULL)
-               goto dropit;
-
-       /* Look up and lock the request corresponding to the given XID */
-       spin_lock(&xprt->sock_lock);
-       rovr = xprt_lookup_rqst(xprt, *xp);
-       if (!rovr)
-               goto out_unlock;
-       task = rovr->rq_task;
-
-       dprintk("RPC: %4d received reply\n", task->tk_pid);
-
-       if ((copied = rovr->rq_private_buf.buflen) > repsize)
-               copied = repsize;
-
-       /* Suck it into the iovec, verify checksum if not done by hw. */
-       if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
-               goto out_unlock;
-
-       /* Something worked... */
-       dst_confirm(skb->dst);
-
-       xprt_complete_rqst(xprt, rovr, copied);
-
- out_unlock:
-       spin_unlock(&xprt->sock_lock);
- dropit:
-       skb_free_datagram(sk, skb);
- out:
-       read_unlock(&sk->sk_callback_lock);
-}
-
-/*
- * Copy from an skb into memory and shrink the skb.
- */
-static inline size_t
-tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
-{
-       if (len > desc->count)
-               len = desc->count;
-       if (skb_copy_bits(desc->skb, desc->offset, p, len)) {
-               dprintk("RPC:      failed to copy %zu bytes from skb. %zu bytes remain\n",
-                               len, desc->count);
-               return 0;
-       }
-       desc->offset += len;
-       desc->count -= len;
-       dprintk("RPC:      copied %zu bytes from skb. %zu bytes remain\n",
-                       len, desc->count);
-       return len;
-}
-
-/*
- * TCP read fragment marker
- */
-static inline void
-tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
-{
-       size_t len, used;
-       char *p;
-
-       p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
-       len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
-       used = tcp_copy_data(desc, p, len);
-       xprt->tcp_offset += used;
-       if (used != len)
-               return;
-       xprt->tcp_reclen = ntohl(xprt->tcp_recm);
-       if (xprt->tcp_reclen & 0x80000000)
-               xprt->tcp_flags |= XPRT_LAST_FRAG;
-       else
-               xprt->tcp_flags &= ~XPRT_LAST_FRAG;
-       xprt->tcp_reclen &= 0x7fffffff;
-       xprt->tcp_flags &= ~XPRT_COPY_RECM;
-       xprt->tcp_offset = 0;
-       /* Sanity check of the record length */
-       if (xprt->tcp_reclen < 4) {
-               printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
-               xprt_disconnect(xprt);
-       }
-       dprintk("RPC:      reading TCP record fragment of length %d\n",
-                       xprt->tcp_reclen);
-}
-
-static void
-tcp_check_recm(struct rpc_xprt *xprt)
-{
-       dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u, tcp_flags = %lx\n",
-                       xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen, xprt->tcp_flags);
-       if (xprt->tcp_offset == xprt->tcp_reclen) {
-               xprt->tcp_flags |= XPRT_COPY_RECM;
-               xprt->tcp_offset = 0;
-               if (xprt->tcp_flags & XPRT_LAST_FRAG) {
-                       xprt->tcp_flags &= ~XPRT_COPY_DATA;
-                       xprt->tcp_flags |= XPRT_COPY_XID;
-                       xprt->tcp_copied = 0;
-               }
-       }
-}
-
-/*
- * TCP read xid
- */
-static inline void
-tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
-{
-       size_t len, used;
-       char *p;
-
-       len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
-       dprintk("RPC:      reading XID (%Zu bytes)\n", len);
-       p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
-       used = tcp_copy_data(desc, p, len);
-       xprt->tcp_offset += used;
-       if (used != len)
-               return;
-       xprt->tcp_flags &= ~XPRT_COPY_XID;
-       xprt->tcp_flags |= XPRT_COPY_DATA;
-       xprt->tcp_copied = 4;
-       dprintk("RPC:      reading reply for XID %08x\n",
-                                               ntohl(xprt->tcp_xid));
-       tcp_check_recm(xprt);
-}
-
-/*
- * TCP read and complete request
- */
-static inline void
-tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
-{
-       struct rpc_rqst *req;
-       struct xdr_buf *rcvbuf;
-       size_t len;
-       ssize_t r;
-
-       /* Find and lock the request corresponding to this xid */
-       spin_lock(&xprt->sock_lock);
-       req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
-       if (!req) {
-               xprt->tcp_flags &= ~XPRT_COPY_DATA;
-               dprintk("RPC:      XID %08x request not found!\n",
-                               ntohl(xprt->tcp_xid));
-               spin_unlock(&xprt->sock_lock);
-               return;
-       }
-
-       rcvbuf = &req->rq_private_buf;
-       len = desc->count;
-       if (len > xprt->tcp_reclen - xprt->tcp_offset) {
-               skb_reader_t my_desc;
-
-               len = xprt->tcp_reclen - xprt->tcp_offset;
-               memcpy(&my_desc, desc, sizeof(my_desc));
-               my_desc.count = len;
-               r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
-                                         &my_desc, tcp_copy_data);
-               desc->count -= r;
-               desc->offset += r;
-       } else
-               r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
-                                         desc, tcp_copy_data);
-
-       if (r > 0) {
-               xprt->tcp_copied += r;
-               xprt->tcp_offset += r;
-       }
-       if (r != len) {
-               /* Error when copying to the receive buffer,
-                * usually because we weren't able to allocate
-                * additional buffer pages. All we can do now
-                * is turn off XPRT_COPY_DATA, so the request
-                * will not receive any additional updates,
-                * and time out.
-                * Any remaining data from this record will
-                * be discarded.
-                */
-               xprt->tcp_flags &= ~XPRT_COPY_DATA;
-               dprintk("RPC:      XID %08x truncated request\n",
-                               ntohl(xprt->tcp_xid));
-               dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
-                               xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
-               goto out;
-       }
-
-       dprintk("RPC:      XID %08x read %Zd bytes\n",
-                       ntohl(xprt->tcp_xid), r);
-       dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
-                       xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
-
-       if (xprt->tcp_copied == req->rq_private_buf.buflen)
-               xprt->tcp_flags &= ~XPRT_COPY_DATA;
-       else if (xprt->tcp_offset == xprt->tcp_reclen) {
-               if (xprt->tcp_flags & XPRT_LAST_FRAG)
-                       xprt->tcp_flags &= ~XPRT_COPY_DATA;
-       }
-
-out:
-       if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
-               dprintk("RPC: %4d received reply complete\n",
-                               req->rq_task->tk_pid);
-               xprt_complete_rqst(xprt, req, xprt->tcp_copied);
-       }
-       spin_unlock(&xprt->sock_lock);
-       tcp_check_recm(xprt);
-}
-
-/*
- * TCP discard extra bytes from a short read
- */
-static inline void
-tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
-{
-       size_t len;
-
-       len = xprt->tcp_reclen - xprt->tcp_offset;
-       if (len > desc->count)
-               len = desc->count;
-       desc->count -= len;
-       desc->offset += len;
-       xprt->tcp_offset += len;
-       dprintk("RPC:      discarded %Zu bytes\n", len);
-       tcp_check_recm(xprt);
-}
-
-/*
- * TCP record receive routine
- * We first have to grab the record marker, then the XID, then the data.
- */
-static int
-tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
-               unsigned int offset, size_t len)
-{
-       struct rpc_xprt *xprt = rd_desc->arg.data;
-       skb_reader_t desc = {
-               .skb    = skb,
-               .offset = offset,
-               .count  = len,
-               .csum   = 0
-               };
-
-       dprintk("RPC:      tcp_data_recv\n");
-       do {
-               /* Read in a new fragment marker if necessary */
-               /* Can we ever really expect to get completely empty fragments? */
-               if (xprt->tcp_flags & XPRT_COPY_RECM) {
-                       tcp_read_fraghdr(xprt, &desc);
-                       continue;
-               }
-               /* Read in the xid if necessary */
-               if (xprt->tcp_flags & XPRT_COPY_XID) {
-                       tcp_read_xid(xprt, &desc);
-                       continue;
-               }
-               /* Read in the request data */
-               if (xprt->tcp_flags & XPRT_COPY_DATA) {
-                       tcp_read_request(xprt, &desc);
-                       continue;
-               }
-               /* Skip over any trailing bytes on short reads */
-               tcp_read_discard(xprt, &desc);
-       } while (desc.count);
-       dprintk("RPC:      tcp_data_recv done\n");
-       return len - desc.count;
-}
-
-static void tcp_data_ready(struct sock *sk, int bytes)
-{
-       struct rpc_xprt *xprt;
-       read_descriptor_t rd_desc;
-
-       read_lock(&sk->sk_callback_lock);
-       dprintk("RPC:      tcp_data_ready...\n");
-       if (!(xprt = xprt_from_sock(sk))) {
-               printk("RPC:      tcp_data_ready socket info not found!\n");
-               goto out;
-       }
-       if (xprt->shutdown)
-               goto out;
-
-       /* We use rd_desc to pass struct xprt to tcp_data_recv */
-       rd_desc.arg.data = xprt;
-       rd_desc.count = 65536;
-       tcp_read_sock(sk, &rd_desc, tcp_data_recv);
-out:
-       read_unlock(&sk->sk_callback_lock);
-}
-
-static void
-tcp_state_change(struct sock *sk)
-{
-       struct rpc_xprt *xprt;
-
-       read_lock(&sk->sk_callback_lock);
-       if (!(xprt = xprt_from_sock(sk)))
-               goto out;
-       dprintk("RPC:      tcp_state_change client %p...\n", xprt);
-       dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
-                               sk->sk_state, xprt_connected(xprt),
-                               sock_flag(sk, SOCK_DEAD),
-                               sock_flag(sk, SOCK_ZAPPED));
-
-       switch (sk->sk_state) {
-       case TCP_ESTABLISHED:
-               spin_lock_bh(&xprt->sock_lock);
-               if (!xprt_test_and_set_connected(xprt)) {
-                       /* Reset TCP record info */
-                       xprt->tcp_offset = 0;
-                       xprt->tcp_reclen = 0;
-                       xprt->tcp_copied = 0;
-                       xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
-                       rpc_wake_up(&xprt->pending);
-               }
-               spin_unlock_bh(&xprt->sock_lock);
-               break;
-       case TCP_SYN_SENT:
-       case TCP_SYN_RECV:
-               break;
-       default:
-               xprt_disconnect(xprt);
-               break;
-       }
- out:
-       read_unlock(&sk->sk_callback_lock);
-}
-
-/*
- * Called when more output buffer space is available for this socket.
- * We try not to wake our writers until they can make "significant"
- * progress, otherwise we'll waste resources thrashing sock_sendmsg
- * with a bunch of small requests.
- */
-static void
-xprt_write_space(struct sock *sk)
-{
-       struct rpc_xprt *xprt;
-       struct socket   *sock;
-
-       read_lock(&sk->sk_callback_lock);
-       if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
-               goto out;
-       if (xprt->shutdown)
-               goto out;
-
-       /* Wait until we have enough socket memory */
-       if (xprt->stream) {
-               /* from net/core/stream.c:sk_stream_write_space */
-               if (sk_stream_wspace(sk) < sk_stream_min_wspace(sk))
-                       goto out;
-       } else {
-               /* from net/core/sock.c:sock_def_write_space */
-               if (!sock_writeable(sk))
-                       goto out;
-       }
-
-       if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
-               goto out;
-
-       spin_lock_bh(&xprt->sock_lock);
-       if (xprt->snd_task)
-               rpc_wake_up_task(xprt->snd_task);
-       spin_unlock_bh(&xprt->sock_lock);
-out:
-       read_unlock(&sk->sk_callback_lock);
-}
-
-/*
  * RPC receive timeout handler.
  */
 static void
@@ -1161,19 +539,10 @@ xprt_transmit(struct rpc_task *task)
        struct rpc_clnt *clnt = task->tk_client;
        struct rpc_rqst *req = task->tk_rqstp;
        struct rpc_xprt *xprt = req->rq_xprt;
-       int status, retry = 0;
-
+       int status;
 
        dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 
-       /* set up everything as needed. */
-       /* Write the record marker */
-       if (xprt->stream) {
-               u32     *marker = req->rq_svec[0].iov_base;
-
-               *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
-       }
-
        smp_rmb();
        if (!req->rq_received) {
                if (list_empty(&req->rq_list)) {
@@ -1191,41 +560,9 @@ xprt_transmit(struct rpc_task *task)
        } else if (!req->rq_bytes_sent)
                return;
 
-       /* Continue transmitting the packet/record. We must be careful
-        * to cope with writespace callbacks arriving _after_ we have
-        * called xprt_sendmsg().
-        */
-       while (1) {
-               req->rq_xtime = jiffies;
-               status = xprt_sendmsg(xprt, req);
-
-               if (status < 0)
-                       break;
-
-               if (xprt->stream) {
-                       req->rq_bytes_sent += status;
-
-                       /* If we've sent the entire packet, immediately
-                        * reset the count of bytes sent. */
-                       if (req->rq_bytes_sent >= req->rq_slen) {
-                               req->rq_bytes_sent = 0;
-                               goto out_receive;
-                       }
-               } else {
-                       if (status >= req->rq_slen)
-                               goto out_receive;
-                       status = -EAGAIN;
-                       break;
-               }
-
-               dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
-                               task->tk_pid, req->rq_slen - req->rq_bytes_sent,
-                               req->rq_slen);
-
-               status = -EAGAIN;
-               if (retry++ > 50)
-                       break;
-       }
+       status = xprt->ops->send_request(task);
+       if (!status)
+               goto out_receive;
 
        /* Note: at this point, task->tk_sleeping has not yet been set,
         *       hence there is no danger of the waking up task being put on
@@ -1234,26 +571,10 @@ xprt_transmit(struct rpc_task *task)
        task->tk_status = status;
 
        switch (status) {
-       case -EAGAIN:
-               if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
-                       /* Protect against races with xprt_write_space */
-                       spin_lock_bh(&xprt->sock_lock);
-                       /* Don't race with disconnect */
-                       if (!xprt_connected(xprt))
-                               task->tk_status = -ENOTCONN;
-                       else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
-                               task->tk_timeout = req->rq_timeout;
-                               rpc_sleep_on(&xprt->pending, task, NULL, NULL);
-                       }
-                       spin_unlock_bh(&xprt->sock_lock);
-                       return;
-               }
-               /* Keep holding the socket if it is blocked */
-               rpc_delay(task, HZ>>4);
-               return;
        case -ECONNREFUSED:
                task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
                rpc_sleep_on(&xprt->sending, task, NULL, NULL);
+       case -EAGAIN:
        case -ENOTCONN:
                return;
        default:
@@ -1367,7 +688,8 @@ xprt_release(struct rpc_task *task)
                list_del(&req->rq_list);
        xprt->last_used = jiffies;
        if (list_empty(&xprt->recv) && !xprt->shutdown)
-               mod_timer(&xprt->timer, xprt->last_used + XPRT_IDLE_TIMEOUT);
+               mod_timer(&xprt->timer,
+                               xprt->last_used + RPC_IDLE_DISCONNECT_TIMEOUT);
        spin_unlock_bh(&xprt->sock_lock);
        task->tk_rqstp = NULL;
        memset(req, 0, sizeof(*req));   /* mark unused */
@@ -1381,18 +703,6 @@ xprt_release(struct rpc_task *task)
 }
 
 /*
- * Set default timeout parameters
- */
-static void
-xprt_default_timeout(struct rpc_timeout *to, int proto)
-{
-       if (proto == IPPROTO_UDP)
-               xprt_set_timeout(to, 5,  5 * HZ);
-       else
-               xprt_set_timeout(to, 2, 60 * HZ);
-}
-
-/*
  * Set constant timeout
  */
 void
@@ -1405,68 +715,51 @@ xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
        to->to_exponential = 0;
 }
 
-unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
-unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
-
 /*
  * Initialize an RPC client
  */
 static struct rpc_xprt *
 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
 {
+       int result;
        struct rpc_xprt *xprt;
-       unsigned int entries;
-       size_t slot_table_size;
        struct rpc_rqst *req;
 
-       dprintk("RPC:      setting up %s transport...\n",
-                               proto == IPPROTO_UDP? "UDP" : "TCP");
-
-       entries = (proto == IPPROTO_TCP)?
-               xprt_tcp_slot_table_entries : xprt_udp_slot_table_entries;
-
        if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
                return ERR_PTR(-ENOMEM);
        memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
-       xprt->max_reqs = entries;
-       slot_table_size = entries * sizeof(xprt->slot[0]);
-       xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
-       if (xprt->slot == NULL) {
-               kfree(xprt);
-               return ERR_PTR(-ENOMEM);
-       }
-       memset(xprt->slot, 0, slot_table_size);
 
        xprt->addr = *ap;
-       xprt->prot = proto;
-       xprt->stream = (proto == IPPROTO_TCP)? 1 : 0;
-       if (xprt->stream) {
-               xprt->cwnd = RPC_MAXCWND(xprt);
-               xprt->nocong = 1;
-               xprt->max_payload = (1U << 31) - 1;
-       } else {
-               xprt->cwnd = RPC_INITCWND;
-               xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
+
+       switch (proto) {
+       case IPPROTO_UDP:
+               result = xs_setup_udp(xprt, to);
+               break;
+       case IPPROTO_TCP:
+               result = xs_setup_tcp(xprt, to);
+               break;
+       default:
+               printk(KERN_ERR "RPC: unrecognized transport protocol: %d\n",
+                               proto);
+               result = -EIO;
+               break;
+       }
+       if (result) {
+               kfree(xprt);
+               return ERR_PTR(result);
        }
+
        spin_lock_init(&xprt->sock_lock);
        spin_lock_init(&xprt->xprt_lock);
        init_waitqueue_head(&xprt->cong_wait);
 
        INIT_LIST_HEAD(&xprt->free);
        INIT_LIST_HEAD(&xprt->recv);
-       INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
        INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
        init_timer(&xprt->timer);
        xprt->timer.function = xprt_init_autodisconnect;
        xprt->timer.data = (unsigned long) xprt;
        xprt->last_used = jiffies;
-       xprt->port = XPRT_MAX_RESVPORT;
-
-       /* Set timeout parameters */
-       if (to) {
-               xprt->timeout = *to;
-       } else
-               xprt_default_timeout(&xprt->timeout, xprt->prot);
 
        rpc_init_wait_queue(&xprt->pending, "xprt_pending");
        rpc_init_wait_queue(&xprt->sending, "xprt_sending");
@@ -1474,14 +767,11 @@ xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
        rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
 
        /* initialize free list */
-       for (req = &xprt->slot[entries-1]; req >= &xprt->slot[0]; req--)
+       for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
                list_add(&req->rq_list, &xprt->free);
 
        xprt_init_xid(xprt);
 
-       /* Check whether we want to use a reserved port */
-       xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
-
        dprintk("RPC:      created transport %p with %u slots\n", xprt,
                        xprt->max_reqs);
        
@@ -1489,120 +779,6 @@ xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
 }
 
 /*
- * Bind to a reserved port
- */
-static inline int xprt_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
-{
-       struct sockaddr_in myaddr = {
-               .sin_family = AF_INET,
-       };
-       int             err, port;
-
-       /* Were we already bound to a given port? Try to reuse it */
-       port = xprt->port;
-       do {
-               myaddr.sin_port = htons(port);
-               err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
-                                               sizeof(myaddr));
-               if (err == 0) {
-                       xprt->port = port;
-                       return 0;
-               }
-               if (--port == 0)
-                       port = XPRT_MAX_RESVPORT;
-       } while (err == -EADDRINUSE && port != xprt->port);
-
-       printk("RPC: Can't bind to reserved port (%d).\n", -err);
-       return err;
-}
-
-static void
-xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
-{
-       struct sock     *sk = sock->sk;
-
-       if (xprt->inet)
-               return;
-
-       write_lock_bh(&sk->sk_callback_lock);
-       sk->sk_user_data = xprt;
-       xprt->old_data_ready = sk->sk_data_ready;
-       xprt->old_state_change = sk->sk_state_change;
-       xprt->old_write_space = sk->sk_write_space;
-       if (xprt->prot == IPPROTO_UDP) {
-               sk->sk_data_ready = udp_data_ready;
-               sk->sk_no_check = UDP_CSUM_NORCV;
-               xprt_set_connected(xprt);
-       } else {
-               tcp_sk(sk)->nonagle = 1;        /* disable Nagle's algorithm */
-               sk->sk_data_ready = tcp_data_ready;
-               sk->sk_state_change = tcp_state_change;
-               xprt_clear_connected(xprt);
-       }
-       sk->sk_write_space = xprt_write_space;
-
-       /* Reset to new socket */
-       xprt->sock = sock;
-       xprt->inet = sk;
-       write_unlock_bh(&sk->sk_callback_lock);
-
-       return;
-}
-
-/*
- * Set socket buffer length
- */
-void
-xprt_sock_setbufsize(struct rpc_xprt *xprt)
-{
-       struct sock *sk = xprt->inet;
-
-       if (xprt->stream)
-               return;
-       if (xprt->rcvsize) {
-               sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
-               sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
-       }
-       if (xprt->sndsize) {
-               sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
-               sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
-               sk->sk_write_space(sk);
-       }
-}
-
-/*
- * Datastream sockets are created here, but xprt_connect will create
- * and connect stream sockets.
- */
-static struct socket * xprt_create_socket(struct rpc_xprt *xprt, int proto, int resvport)
-{
-       struct socket   *sock;
-       int             type, err;
-
-       dprintk("RPC:      xprt_create_socket(%s %d)\n",
-                          (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
-
-       type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
-
-       if ((err = sock_create_kern(PF_INET, type, proto, &sock)) < 0) {
-               printk("RPC: can't create socket (%d).\n", -err);
-               return NULL;
-       }
-
-       /* If the caller has the capability, bind to a reserved port */
-       if (resvport && xprt_bindresvport(xprt, sock) < 0) {
-               printk("RPC: can't bind to reserved port.\n");
-               goto failed;
-       }
-
-       return sock;
-
-failed:
-       sock_release(sock);
-       return NULL;
-}
-
-/*
  * Create an RPC client transport given the protocol and peer address.
  */
 struct rpc_xprt *
@@ -1631,10 +807,6 @@ xprt_shutdown(struct rpc_xprt *xprt)
        rpc_wake_up(&xprt->backlog);
        wake_up(&xprt->cong_wait);
        del_timer_sync(&xprt->timer);
-
-       /* synchronously wait for connect worker to finish */
-       cancel_delayed_work(&xprt->sock_connect);
-       flush_scheduled_work();
 }
 
 /*
@@ -1655,9 +827,7 @@ xprt_destroy(struct rpc_xprt *xprt)
 {
        dprintk("RPC:      destroying transport %p\n", xprt);
        xprt_shutdown(xprt);
-       xprt_disconnect(xprt);
-       xprt_close(xprt);
-       kfree(xprt->slot);
+       xprt->ops->destroy(xprt);
        kfree(xprt);
 
        return 0;
diff --git a/net/sunrpc/xprtsock.c b/net/sunrpc/xprtsock.c
new file mode 100644 (file)
index 0000000..fa1180a
--- /dev/null
@@ -0,0 +1,1021 @@
+/*
+ * linux/net/sunrpc/xprtsock.c
+ *
+ * Client-side transport implementation for sockets.
+ *
+ * TCP callback races fixes (C) 1998 Red Hat Software <alan@redhat.com>
+ * TCP send fixes (C) 1998 Red Hat Software <alan@redhat.com>
+ * TCP NFS related read + write fixes
+ *  (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
+ *
+ * Rewrite of larges part of the code in order to stabilize TCP stuff.
+ * Fix behaviour when socket buffer is full.
+ *  (C) 1999 Trond Myklebust <trond.myklebust@fys.uio.no>
+ */
+
+#include <linux/types.h>
+#include <linux/slab.h>
+#include <linux/capability.h>
+#include <linux/sched.h>
+#include <linux/pagemap.h>
+#include <linux/errno.h>
+#include <linux/socket.h>
+#include <linux/in.h>
+#include <linux/net.h>
+#include <linux/mm.h>
+#include <linux/udp.h>
+#include <linux/tcp.h>
+#include <linux/sunrpc/clnt.h>
+#include <linux/file.h>
+
+#include <net/sock.h>
+#include <net/checksum.h>
+#include <net/udp.h>
+#include <net/tcp.h>
+
+#ifdef RPC_DEBUG
+# undef  RPC_DEBUG_DATA
+# define RPCDBG_FACILITY       RPCDBG_XPRT
+#endif
+
+#define XPRT_MAX_RESVPORT      (800)
+
+#ifdef RPC_DEBUG_DATA
+/*
+ * Print the buffer contents (first 128 bytes only--just enough for
+ * diropres return).
+ */
+static void
+xprt_pktdump(char *msg, u32 *packet, unsigned int count)
+{
+       u8      *buf = (u8 *) packet;
+       int     j;
+
+       dprintk("RPC:      %s\n", msg);
+       for (j = 0; j < count && j < 128; j += 4) {
+               if (!(j & 31)) {
+                       if (j)
+                               dprintk("\n");
+                       dprintk("0x%04x ", j);
+               }
+               dprintk("%02x%02x%02x%02x ",
+                       buf[j], buf[j+1], buf[j+2], buf[j+3]);
+       }
+       dprintk("\n");
+}
+#else
+static inline void
+xprt_pktdump(char *msg, u32 *packet, unsigned int count)
+{
+       /* NOP */
+}
+#endif
+
+/*
+ * Look up RPC transport given an INET socket
+ */
+static inline struct rpc_xprt *
+xprt_from_sock(struct sock *sk)
+{
+       return (struct rpc_xprt *) sk->sk_user_data;
+}
+
+static int
+xdr_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen,
+               struct xdr_buf *xdr, unsigned int base, int msgflags)
+{
+       struct page **ppage = xdr->pages;
+       unsigned int len, pglen = xdr->page_len;
+       int err, ret = 0;
+       ssize_t (*sendpage)(struct socket *, struct page *, int, size_t, int);
+
+       len = xdr->head[0].iov_len;
+       if (base < len || (addr != NULL && base == 0)) {
+               struct kvec iov = {
+                       .iov_base = xdr->head[0].iov_base + base,
+                       .iov_len  = len - base,
+               };
+               struct msghdr msg = {
+                       .msg_name    = addr,
+                       .msg_namelen = addrlen,
+                       .msg_flags   = msgflags,
+               };
+               if (xdr->len > len)
+                       msg.msg_flags |= MSG_MORE;
+
+               if (iov.iov_len != 0)
+                       err = kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
+               else
+                       err = kernel_sendmsg(sock, &msg, NULL, 0, 0);
+               if (ret == 0)
+                       ret = err;
+               else if (err > 0)
+                       ret += err;
+               if (err != iov.iov_len)
+                       goto out;
+               base = 0;
+       } else
+               base -= len;
+
+       if (pglen == 0)
+               goto copy_tail;
+       if (base >= pglen) {
+               base -= pglen;
+               goto copy_tail;
+       }
+       if (base || xdr->page_base) {
+               pglen -= base;
+               base  += xdr->page_base;
+               ppage += base >> PAGE_CACHE_SHIFT;
+               base &= ~PAGE_CACHE_MASK;
+       }
+
+       sendpage = sock->ops->sendpage ? : sock_no_sendpage;
+       do {
+               int flags = msgflags;
+
+               len = PAGE_CACHE_SIZE;
+               if (base)
+                       len -= base;
+               if (pglen < len)
+                       len = pglen;
+
+               if (pglen != len || xdr->tail[0].iov_len != 0)
+                       flags |= MSG_MORE;
+
+               /* Hmm... We might be dealing with highmem pages */
+               if (PageHighMem(*ppage))
+                       sendpage = sock_no_sendpage;
+               err = sendpage(sock, *ppage, base, len, flags);
+               if (ret == 0)
+                       ret = err;
+               else if (err > 0)
+                       ret += err;
+               if (err != len)
+                       goto out;
+               base = 0;
+               ppage++;
+       } while ((pglen -= len) != 0);
+copy_tail:
+       len = xdr->tail[0].iov_len;
+       if (base < len) {
+               struct kvec iov = {
+                       .iov_base = xdr->tail[0].iov_base + base,
+                       .iov_len  = len - base,
+               };
+               struct msghdr msg = {
+                       .msg_flags   = msgflags,
+               };
+               err = kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
+               if (ret == 0)
+                       ret = err;
+               else if (err > 0)
+                       ret += err;
+       }
+out:
+       return ret;
+}
+
+/*
+ * Write data to socket.
+ */
+static inline int
+xprt_sendmsg(struct rpc_xprt *xprt, struct rpc_rqst *req)
+{
+       struct socket   *sock = xprt->sock;
+       struct xdr_buf  *xdr = &req->rq_snd_buf;
+       struct sockaddr *addr = NULL;
+       int addrlen = 0;
+       unsigned int    skip;
+       int             result;
+
+       if (!sock)
+               return -ENOTCONN;
+
+       xprt_pktdump("packet data:",
+                               req->rq_svec->iov_base,
+                               req->rq_svec->iov_len);
+
+       /* For UDP, we need to provide an address */
+       if (!xprt->stream) {
+               addr = (struct sockaddr *) &xprt->addr;
+               addrlen = sizeof(xprt->addr);
+       }
+       /* Dont repeat bytes */
+       skip = req->rq_bytes_sent;
+
+       clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
+       result = xdr_sendpages(sock, addr, addrlen, xdr, skip, MSG_DONTWAIT);
+
+       dprintk("RPC:      xprt_sendmsg(%d) = %d\n", xdr->len - skip, result);
+
+       if (result >= 0)
+               return result;
+
+       switch (result) {
+       case -ECONNREFUSED:
+               /* When the server has died, an ICMP port unreachable message
+                * prompts ECONNREFUSED.
+                */
+       case -EAGAIN:
+               break;
+       case -ECONNRESET:
+       case -ENOTCONN:
+       case -EPIPE:
+               /* connection broken */
+               if (xprt->stream)
+                       result = -ENOTCONN;
+               break;
+       default:
+               printk(KERN_NOTICE "RPC: sendmsg returned error %d\n", -result);
+       }
+       return result;
+}
+
+static int
+xprt_send_request(struct rpc_task *task)
+{
+       struct rpc_rqst *req = task->tk_rqstp;
+       struct rpc_xprt *xprt = req->rq_xprt;
+       int status, retry = 0;
+
+       /* set up everything as needed. */
+       /* Write the record marker */
+       if (xprt->stream) {
+               u32     *marker = req->rq_svec[0].iov_base;
+
+               *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
+       }
+
+       /* Continue transmitting the packet/record. We must be careful
+        * to cope with writespace callbacks arriving _after_ we have
+        * called xprt_sendmsg().
+        */
+       while (1) {
+               req->rq_xtime = jiffies;
+               status = xprt_sendmsg(xprt, req);
+
+               if (status < 0)
+                       break;
+
+               if (xprt->stream) {
+                       req->rq_bytes_sent += status;
+
+                       /* If we've sent the entire packet, immediately
+                        * reset the count of bytes sent. */
+                       if (req->rq_bytes_sent >= req->rq_slen) {
+                               req->rq_bytes_sent = 0;
+                               return 0;
+                       }
+               } else {
+                       if (status >= req->rq_slen)
+                               return 0;
+                       status = -EAGAIN;
+                       break;
+               }
+
+               dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
+                               task->tk_pid, req->rq_slen - req->rq_bytes_sent,
+                               req->rq_slen);
+
+               status = -EAGAIN;
+               if (retry++ > 50)
+                       break;
+       }
+
+       if (status == -EAGAIN) {
+               if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
+                       /* Protect against races with xprt_write_space */
+                       spin_lock_bh(&xprt->sock_lock);
+                       /* Don't race with disconnect */
+                       if (!xprt_connected(xprt))
+                               task->tk_status = -ENOTCONN;
+                       else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
+                               task->tk_timeout = req->rq_timeout;
+                               rpc_sleep_on(&xprt->pending, task, NULL, NULL);
+                       }
+                       spin_unlock_bh(&xprt->sock_lock);
+                       return status;
+               }
+               /* Keep holding the socket if it is blocked */
+               rpc_delay(task, HZ>>4);
+       }
+       return status;
+}
+
+/*
+ * Close down a transport socket
+ */
+static void
+xprt_close(struct rpc_xprt *xprt)
+{
+       struct socket   *sock = xprt->sock;
+       struct sock     *sk = xprt->inet;
+
+       if (!sk)
+               return;
+
+       write_lock_bh(&sk->sk_callback_lock);
+       xprt->inet = NULL;
+       xprt->sock = NULL;
+
+       sk->sk_user_data    = NULL;
+       sk->sk_data_ready   = xprt->old_data_ready;
+       sk->sk_state_change = xprt->old_state_change;
+       sk->sk_write_space  = xprt->old_write_space;
+       write_unlock_bh(&sk->sk_callback_lock);
+
+       sk->sk_no_check  = 0;
+
+       sock_release(sock);
+}
+
+static void xprt_socket_destroy(struct rpc_xprt *xprt)
+{
+       cancel_delayed_work(&xprt->sock_connect);
+       flush_scheduled_work();
+
+       xprt_disconnect(xprt);
+       xprt_close(xprt);
+       kfree(xprt->slot);
+}
+
+/*
+ * Input handler for RPC replies. Called from a bottom half and hence
+ * atomic.
+ */
+static void
+udp_data_ready(struct sock *sk, int len)
+{
+       struct rpc_task *task;
+       struct rpc_xprt *xprt;
+       struct rpc_rqst *rovr;
+       struct sk_buff  *skb;
+       int err, repsize, copied;
+       u32 _xid, *xp;
+
+       read_lock(&sk->sk_callback_lock);
+       dprintk("RPC:      udp_data_ready...\n");
+       if (!(xprt = xprt_from_sock(sk))) {
+               printk("RPC:      udp_data_ready request not found!\n");
+               goto out;
+       }
+
+       dprintk("RPC:      udp_data_ready client %p\n", xprt);
+
+       if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
+               goto out;
+
+       if (xprt->shutdown)
+               goto dropit;
+
+       repsize = skb->len - sizeof(struct udphdr);
+       if (repsize < 4) {
+               printk("RPC: impossible RPC reply size %d!\n", repsize);
+               goto dropit;
+       }
+
+       /* Copy the XID from the skb... */
+       xp = skb_header_pointer(skb, sizeof(struct udphdr),
+                               sizeof(_xid), &_xid);
+       if (xp == NULL)
+               goto dropit;
+
+       /* Look up and lock the request corresponding to the given XID */
+       spin_lock(&xprt->sock_lock);
+       rovr = xprt_lookup_rqst(xprt, *xp);
+       if (!rovr)
+               goto out_unlock;
+       task = rovr->rq_task;
+
+       dprintk("RPC: %4d received reply\n", task->tk_pid);
+
+       if ((copied = rovr->rq_private_buf.buflen) > repsize)
+               copied = repsize;
+
+       /* Suck it into the iovec, verify checksum if not done by hw. */
+       if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
+               goto out_unlock;
+
+       /* Something worked... */
+       dst_confirm(skb->dst);
+
+       xprt_complete_rqst(xprt, rovr, copied);
+
+ out_unlock:
+       spin_unlock(&xprt->sock_lock);
+ dropit:
+       skb_free_datagram(sk, skb);
+ out:
+       read_unlock(&sk->sk_callback_lock);
+}
+
+/*
+ * Copy from an skb into memory and shrink the skb.
+ */
+static inline size_t
+tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
+{
+       if (len > desc->count)
+               len = desc->count;
+       if (skb_copy_bits(desc->skb, desc->offset, p, len)) {
+               dprintk("RPC:      failed to copy %zu bytes from skb. %zu bytes remain\n",
+                               len, desc->count);
+               return 0;
+       }
+       desc->offset += len;
+       desc->count -= len;
+       dprintk("RPC:      copied %zu bytes from skb. %zu bytes remain\n",
+                       len, desc->count);
+       return len;
+}
+
+/*
+ * TCP read fragment marker
+ */
+static inline void
+tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
+{
+       size_t len, used;
+       char *p;
+
+       p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
+       len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
+       used = tcp_copy_data(desc, p, len);
+       xprt->tcp_offset += used;
+       if (used != len)
+               return;
+       xprt->tcp_reclen = ntohl(xprt->tcp_recm);
+       if (xprt->tcp_reclen & 0x80000000)
+               xprt->tcp_flags |= XPRT_LAST_FRAG;
+       else
+               xprt->tcp_flags &= ~XPRT_LAST_FRAG;
+       xprt->tcp_reclen &= 0x7fffffff;
+       xprt->tcp_flags &= ~XPRT_COPY_RECM;
+       xprt->tcp_offset = 0;
+       /* Sanity check of the record length */
+       if (xprt->tcp_reclen < 4) {
+               printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
+               xprt_disconnect(xprt);
+       }
+       dprintk("RPC:      reading TCP record fragment of length %d\n",
+                       xprt->tcp_reclen);
+}
+
+static void
+tcp_check_recm(struct rpc_xprt *xprt)
+{
+       dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u, tcp_flags = %lx\n",
+                       xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen, xprt->tcp_flags);
+       if (xprt->tcp_offset == xprt->tcp_reclen) {
+               xprt->tcp_flags |= XPRT_COPY_RECM;
+               xprt->tcp_offset = 0;
+               if (xprt->tcp_flags & XPRT_LAST_FRAG) {
+                       xprt->tcp_flags &= ~XPRT_COPY_DATA;
+                       xprt->tcp_flags |= XPRT_COPY_XID;
+                       xprt->tcp_copied = 0;
+               }
+       }
+}
+
+/*
+ * TCP read xid
+ */
+static inline void
+tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
+{
+       size_t len, used;
+       char *p;
+
+       len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
+       dprintk("RPC:      reading XID (%Zu bytes)\n", len);
+       p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
+       used = tcp_copy_data(desc, p, len);
+       xprt->tcp_offset += used;
+       if (used != len)
+               return;
+       xprt->tcp_flags &= ~XPRT_COPY_XID;
+       xprt->tcp_flags |= XPRT_COPY_DATA;
+       xprt->tcp_copied = 4;
+       dprintk("RPC:      reading reply for XID %08x\n",
+                                               ntohl(xprt->tcp_xid));
+       tcp_check_recm(xprt);
+}
+
+/*
+ * TCP read and complete request
+ */
+static inline void
+tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
+{
+       struct rpc_rqst *req;
+       struct xdr_buf *rcvbuf;
+       size_t len;
+       ssize_t r;
+
+       /* Find and lock the request corresponding to this xid */
+       spin_lock(&xprt->sock_lock);
+       req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
+       if (!req) {
+               xprt->tcp_flags &= ~XPRT_COPY_DATA;
+               dprintk("RPC:      XID %08x request not found!\n",
+                               ntohl(xprt->tcp_xid));
+               spin_unlock(&xprt->sock_lock);
+               return;
+       }
+
+       rcvbuf = &req->rq_private_buf;
+       len = desc->count;
+       if (len > xprt->tcp_reclen - xprt->tcp_offset) {
+               skb_reader_t my_desc;
+
+               len = xprt->tcp_reclen - xprt->tcp_offset;
+               memcpy(&my_desc, desc, sizeof(my_desc));
+               my_desc.count = len;
+               r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
+                                         &my_desc, tcp_copy_data);
+               desc->count -= r;
+               desc->offset += r;
+       } else
+               r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
+                                         desc, tcp_copy_data);
+
+       if (r > 0) {
+               xprt->tcp_copied += r;
+               xprt->tcp_offset += r;
+       }
+       if (r != len) {
+               /* Error when copying to the receive buffer,
+                * usually because we weren't able to allocate
+                * additional buffer pages. All we can do now
+                * is turn off XPRT_COPY_DATA, so the request
+                * will not receive any additional updates,
+                * and time out.
+                * Any remaining data from this record will
+                * be discarded.
+                */
+               xprt->tcp_flags &= ~XPRT_COPY_DATA;
+               dprintk("RPC:      XID %08x truncated request\n",
+                               ntohl(xprt->tcp_xid));
+               dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
+                               xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
+               goto out;
+       }
+
+       dprintk("RPC:      XID %08x read %Zd bytes\n",
+                       ntohl(xprt->tcp_xid), r);
+       dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
+                       xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
+
+       if (xprt->tcp_copied == req->rq_private_buf.buflen)
+               xprt->tcp_flags &= ~XPRT_COPY_DATA;
+       else if (xprt->tcp_offset == xprt->tcp_reclen) {
+               if (xprt->tcp_flags & XPRT_LAST_FRAG)
+                       xprt->tcp_flags &= ~XPRT_COPY_DATA;
+       }
+
+out:
+       if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
+               dprintk("RPC: %4d received reply complete\n",
+                               req->rq_task->tk_pid);
+               xprt_complete_rqst(xprt, req, xprt->tcp_copied);
+       }
+       spin_unlock(&xprt->sock_lock);
+       tcp_check_recm(xprt);
+}
+
+/*
+ * TCP discard extra bytes from a short read
+ */
+static inline void
+tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
+{
+       size_t len;
+
+       len = xprt->tcp_reclen - xprt->tcp_offset;
+       if (len > desc->count)
+               len = desc->count;
+       desc->count -= len;
+       desc->offset += len;
+       xprt->tcp_offset += len;
+       dprintk("RPC:      discarded %Zu bytes\n", len);
+       tcp_check_recm(xprt);
+}
+
+/*
+ * TCP record receive routine
+ * We first have to grab the record marker, then the XID, then the data.
+ */
+static int
+tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
+               unsigned int offset, size_t len)
+{
+       struct rpc_xprt *xprt = rd_desc->arg.data;
+       skb_reader_t desc = {
+               .skb    = skb,
+               .offset = offset,
+               .count  = len,
+               .csum   = 0
+               };
+
+       dprintk("RPC:      tcp_data_recv\n");
+       do {
+               /* Read in a new fragment marker if necessary */
+               /* Can we ever really expect to get completely empty fragments? */
+               if (xprt->tcp_flags & XPRT_COPY_RECM) {
+                       tcp_read_fraghdr(xprt, &desc);
+                       continue;
+               }
+               /* Read in the xid if necessary */
+               if (xprt->tcp_flags & XPRT_COPY_XID) {
+                       tcp_read_xid(xprt, &desc);
+                       continue;
+               }
+               /* Read in the request data */
+               if (xprt->tcp_flags & XPRT_COPY_DATA) {
+                       tcp_read_request(xprt, &desc);
+                       continue;
+               }
+               /* Skip over any trailing bytes on short reads */
+               tcp_read_discard(xprt, &desc);
+       } while (desc.count);
+       dprintk("RPC:      tcp_data_recv done\n");
+       return len - desc.count;
+}
+
+static void tcp_data_ready(struct sock *sk, int bytes)
+{
+       struct rpc_xprt *xprt;
+       read_descriptor_t rd_desc;
+
+       read_lock(&sk->sk_callback_lock);
+       dprintk("RPC:      tcp_data_ready...\n");
+       if (!(xprt = xprt_from_sock(sk))) {
+               printk("RPC:      tcp_data_ready socket info not found!\n");
+               goto out;
+       }
+       if (xprt->shutdown)
+               goto out;
+
+       /* We use rd_desc to pass struct xprt to tcp_data_recv */
+       rd_desc.arg.data = xprt;
+       rd_desc.count = 65536;
+       tcp_read_sock(sk, &rd_desc, tcp_data_recv);
+out:
+       read_unlock(&sk->sk_callback_lock);
+}
+
+static void
+tcp_state_change(struct sock *sk)
+{
+       struct rpc_xprt *xprt;
+
+       read_lock(&sk->sk_callback_lock);
+       if (!(xprt = xprt_from_sock(sk)))
+               goto out;
+       dprintk("RPC:      tcp_state_change client %p...\n", xprt);
+       dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
+                               sk->sk_state, xprt_connected(xprt),
+                               sock_flag(sk, SOCK_DEAD),
+                               sock_flag(sk, SOCK_ZAPPED));
+
+       switch (sk->sk_state) {
+       case TCP_ESTABLISHED:
+               spin_lock_bh(&xprt->sock_lock);
+               if (!xprt_test_and_set_connected(xprt)) {
+                       /* Reset TCP record info */
+                       xprt->tcp_offset = 0;
+                       xprt->tcp_reclen = 0;
+                       xprt->tcp_copied = 0;
+                       xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
+                       rpc_wake_up(&xprt->pending);
+               }
+               spin_unlock_bh(&xprt->sock_lock);
+               break;
+       case TCP_SYN_SENT:
+       case TCP_SYN_RECV:
+               break;
+       default:
+               xprt_disconnect(xprt);
+               break;
+       }
+ out:
+       read_unlock(&sk->sk_callback_lock);
+}
+
+/*
+ * Called when more output buffer space is available for this socket.
+ * We try not to wake our writers until they can make "significant"
+ * progress, otherwise we'll waste resources thrashing sock_sendmsg
+ * with a bunch of small requests.
+ */
+static void
+xprt_write_space(struct sock *sk)
+{
+       struct rpc_xprt *xprt;
+       struct socket   *sock;
+
+       read_lock(&sk->sk_callback_lock);
+       if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
+               goto out;
+       if (xprt->shutdown)
+               goto out;
+
+       /* Wait until we have enough socket memory */
+       if (xprt->stream) {
+               /* from net/core/stream.c:sk_stream_write_space */
+               if (sk_stream_wspace(sk) < sk_stream_min_wspace(sk))
+                       goto out;
+       } else {
+               /* from net/core/sock.c:sock_def_write_space */
+               if (!sock_writeable(sk))
+                       goto out;
+       }
+
+       if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
+               goto out;
+
+       spin_lock_bh(&xprt->sock_lock);
+       if (xprt->snd_task)
+               rpc_wake_up_task(xprt->snd_task);
+       spin_unlock_bh(&xprt->sock_lock);
+out:
+       read_unlock(&sk->sk_callback_lock);
+}
+
+/*
+ * Set socket buffer length
+ */
+static void
+xprt_sock_setbufsize(struct rpc_xprt *xprt)
+{
+       struct sock *sk = xprt->inet;
+
+       if (xprt->stream)
+               return;
+       if (xprt->rcvsize) {
+               sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
+               sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
+       }
+       if (xprt->sndsize) {
+               sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
+               sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
+               sk->sk_write_space(sk);
+       }
+}
+
+/*
+ * Bind to a reserved port
+ */
+static inline int xprt_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
+{
+       struct sockaddr_in myaddr = {
+               .sin_family = AF_INET,
+       };
+       int             err, port;
+
+       /* Were we already bound to a given port? Try to reuse it */
+       port = xprt->port;
+       do {
+               myaddr.sin_port = htons(port);
+               err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
+                                               sizeof(myaddr));
+               if (err == 0) {
+                       xprt->port = port;
+                       return 0;
+               }
+               if (--port == 0)
+                       port = XPRT_MAX_RESVPORT;
+       } while (err == -EADDRINUSE && port != xprt->port);
+
+       printk("RPC: Can't bind to reserved port (%d).\n", -err);
+       return err;
+}
+
+static void
+xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
+{
+       struct sock     *sk = sock->sk;
+
+       if (xprt->inet)
+               return;
+
+       write_lock_bh(&sk->sk_callback_lock);
+       sk->sk_user_data = xprt;
+       xprt->old_data_ready = sk->sk_data_ready;
+       xprt->old_state_change = sk->sk_state_change;
+       xprt->old_write_space = sk->sk_write_space;
+       if (xprt->prot == IPPROTO_UDP) {
+               sk->sk_data_ready = udp_data_ready;
+               sk->sk_no_check = UDP_CSUM_NORCV;
+               xprt_set_connected(xprt);
+       } else {
+               tcp_sk(sk)->nonagle = 1;        /* disable Nagle's algorithm */
+               sk->sk_data_ready = tcp_data_ready;
+               sk->sk_state_change = tcp_state_change;
+               xprt_clear_connected(xprt);
+       }
+       sk->sk_write_space = xprt_write_space;
+
+       /* Reset to new socket */
+       xprt->sock = sock;
+       xprt->inet = sk;
+       write_unlock_bh(&sk->sk_callback_lock);
+
+       return;
+}
+
+/*
+ * Datastream sockets are created here, but xprt_connect will create
+ * and connect stream sockets.
+ */
+static struct socket * xprt_create_socket(struct rpc_xprt *xprt, int proto, int resvport)
+{
+       struct socket   *sock;
+       int             type, err;
+
+       dprintk("RPC:      xprt_create_socket(%s %d)\n",
+                          (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
+
+       type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
+
+       if ((err = sock_create_kern(PF_INET, type, proto, &sock)) < 0) {
+               printk("RPC: can't create socket (%d).\n", -err);
+               return NULL;
+       }
+
+       /* If the caller has the capability, bind to a reserved port */
+       if (resvport && xprt_bindresvport(xprt, sock) < 0) {
+               printk("RPC: can't bind to reserved port.\n");
+               goto failed;
+       }
+
+       return sock;
+
+failed:
+       sock_release(sock);
+       return NULL;
+}
+
+static void xprt_socket_connect(void *args)
+{
+       struct rpc_xprt *xprt = (struct rpc_xprt *)args;
+       struct socket *sock = xprt->sock;
+       int status = -EIO;
+
+       if (xprt->shutdown || xprt->addr.sin_port == 0)
+               goto out;
+
+       /*
+        * Start by resetting any existing state
+        */
+       xprt_close(xprt);
+       sock = xprt_create_socket(xprt, xprt->prot, xprt->resvport);
+       if (sock == NULL) {
+               /* couldn't create socket or bind to reserved port;
+                * this is likely a permanent error, so cause an abort */
+               goto out;
+       }
+       xprt_bind_socket(xprt, sock);
+       xprt_sock_setbufsize(xprt);
+
+       status = 0;
+       if (!xprt->stream)
+               goto out;
+
+       /*
+        * Tell the socket layer to start connecting...
+        */
+       status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
+                       sizeof(xprt->addr), O_NONBLOCK);
+       dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
+                       xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
+       if (status < 0) {
+               switch (status) {
+                       case -EINPROGRESS:
+                       case -EALREADY:
+                               goto out_clear;
+               }
+       }
+out:
+       if (status < 0)
+               rpc_wake_up_status(&xprt->pending, status);
+       else
+               rpc_wake_up(&xprt->pending);
+out_clear:
+       smp_mb__before_clear_bit();
+       clear_bit(XPRT_CONNECTING, &xprt->sockstate);
+       smp_mb__after_clear_bit();
+}
+
+static void
+xprt_connect_sock(struct rpc_task *task)
+{
+       struct rpc_xprt *xprt = task->tk_xprt;
+
+       if (!test_and_set_bit(XPRT_CONNECTING, &xprt->sockstate)) {
+               /* Note: if we are here due to a dropped connection
+                *       we delay reconnecting by RPC_REESTABLISH_TIMEOUT/HZ
+                *       seconds
+                */
+               if (xprt->sock != NULL)
+                       schedule_delayed_work(&xprt->sock_connect,
+                                       RPC_REESTABLISH_TIMEOUT);
+               else {
+                       schedule_work(&xprt->sock_connect);
+                       /* flush_scheduled_work can sleep... */
+                       if (!RPC_IS_ASYNC(task))
+                               flush_scheduled_work();
+               }
+       }
+}
+
+/*
+ * Set default timeout parameters
+ */
+static void
+xprt_default_timeout(struct rpc_timeout *to, int proto)
+{
+       if (proto == IPPROTO_UDP)
+               xprt_set_timeout(to, 5,  5 * HZ);
+       else
+               xprt_set_timeout(to, 2, 60 * HZ);
+}
+
+static struct rpc_xprt_ops xprt_socket_ops = {
+       .set_buffer_size        = xprt_sock_setbufsize,
+       .connect                = xprt_connect_sock,
+       .send_request           = xprt_send_request,
+       .close                  = xprt_close,
+       .destroy                = xprt_socket_destroy,
+};
+
+extern unsigned int xprt_udp_slot_table_entries;
+extern unsigned int xprt_tcp_slot_table_entries;
+
+int xs_setup_udp(struct rpc_xprt *xprt, struct rpc_timeout *to)
+{
+       size_t slot_table_size;
+
+       dprintk("RPC:      setting up udp-ipv4 transport...\n");
+
+       xprt->max_reqs = xprt_udp_slot_table_entries;
+       slot_table_size = xprt->max_reqs * sizeof(xprt->slot[0]);
+       xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
+       if (xprt->slot == NULL)
+               return -ENOMEM;
+       memset(xprt->slot, 0, slot_table_size);
+
+       xprt->prot = IPPROTO_UDP;
+       xprt->port = XPRT_MAX_RESVPORT;
+       xprt->stream = 0;
+       xprt->nocong = 0;
+       xprt->cwnd = RPC_INITCWND;
+       xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
+       /* XXX: header size can vary due to auth type, IPv6, etc. */
+       xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
+
+       INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
+
+       xprt->ops = &xprt_socket_ops;
+
+       if (to)
+               xprt->timeout = *to;
+       else
+               xprt_default_timeout(to, xprt->prot);
+
+       return 0;
+}
+
+int xs_setup_tcp(struct rpc_xprt *xprt, struct rpc_timeout *to)
+{
+       size_t slot_table_size;
+
+       dprintk("RPC:      setting up tcp-ipv4 transport...\n");
+
+       xprt->max_reqs = xprt_tcp_slot_table_entries;
+       slot_table_size = xprt->max_reqs * sizeof(xprt->slot[0]);
+       xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
+       if (xprt->slot == NULL)
+               return -ENOMEM;
+       memset(xprt->slot, 0, slot_table_size);
+
+       xprt->prot = IPPROTO_TCP;
+       xprt->port = XPRT_MAX_RESVPORT;
+       xprt->stream = 1;
+       xprt->nocong = 1;
+       xprt->cwnd = RPC_MAXCWND(xprt);
+       xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
+       xprt->max_payload = (1U << 31) - 1;
+
+       INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
+
+       xprt->ops = &xprt_socket_ops;
+
+       if (to)
+               xprt->timeout = *to;
+       else
+               xprt_default_timeout(to, xprt->prot);
+
+       return 0;
+}