[PATCH] v9fs: new multiplexer implementation
[powerpc.git] / fs / 9p / mux.c
index 8835b57..62b6ad0 100644 (file)
@@ -4,7 +4,7 @@
  * Protocol Multiplexer
  *
  *  Copyright (C) 2004 by Eric Van Hensbergen <ericvh@gmail.com>
- *  Copyright (C) 2004 by Latchesar Ionkov <lucho@ionkov.net>
+ *  Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -28,6 +28,7 @@
 #include <linux/module.h>
 #include <linux/errno.h>
 #include <linux/fs.h>
+#include <linux/poll.h>
 #include <linux/kthread.h>
 #include <linux/idr.h>
 
 #include "conv.h"
 #include "mux.h"
 
-/**
- * dprintcond - print condition of session info
- * @v9ses: session info structure
- * @req: RPC request structure
- *
- */
+#define ERREQFLUSH     1
+#define SCHED_TIMEOUT  10
+#define MAXPOLLWADDR   2
+
+enum {
+       Rworksched = 1,         /* read work scheduled or running */
+       Rpending = 2,           /* can read */
+       Wworksched = 4,         /* write work scheduled or running */
+       Wpending = 8,           /* can write */
+};
+
+struct v9fs_mux_poll_task;
+
+struct v9fs_req {
+       int tag;
+       struct v9fs_fcall *tcall;
+       struct v9fs_fcall *rcall;
+       int err;
+       v9fs_mux_req_callback cb;
+       void *cba;
+       struct list_head req_list;
+};
+
+struct v9fs_mux_data {
+       spinlock_t lock;
+       struct list_head mux_list;
+       struct v9fs_mux_poll_task *poll_task;
+       int msize;
+       unsigned char *extended;
+       struct v9fs_transport *trans;
+       struct v9fs_idpool tidpool;
+       int err;
+       wait_queue_head_t equeue;
+       struct list_head req_list;
+       struct list_head unsent_req_list;
+       int rpos;
+       char *rbuf;
+       int wpos;
+       int wsize;
+       char *wbuf;
+       wait_queue_t poll_wait[MAXPOLLWADDR];
+       wait_queue_head_t *poll_waddr[MAXPOLLWADDR];
+       poll_table pt;
+       struct work_struct rq;
+       struct work_struct wq;
+       unsigned long wsched;
+};
+
+struct v9fs_mux_poll_task {
+       struct task_struct *task;
+       struct list_head mux_list;
+       int muxnum;
+};
+
+struct v9fs_mux_rpc {
+       struct v9fs_mux_data *m;
+       struct v9fs_req *req;
+       int err;
+       struct v9fs_fcall *rcall;
+       wait_queue_head_t wqueue;
+};
+
+static int v9fs_poll_proc(void *);
+static void v9fs_read_work(void *);
+static void v9fs_write_work(void *);
+static void v9fs_pollwait(struct file *filp, wait_queue_head_t * wait_address,
+                         poll_table * p);
+
+static DECLARE_MUTEX(v9fs_mux_task_lock);
+static struct workqueue_struct *v9fs_mux_wq;
+
+static int v9fs_mux_num;
+static int v9fs_mux_poll_task_num;
+static struct v9fs_mux_poll_task v9fs_mux_poll_tasks[100];
+
+void v9fs_mux_global_init(void)
+{
+       int i;
+
+       for (i = 0; i < ARRAY_SIZE(v9fs_mux_poll_tasks); i++)
+               v9fs_mux_poll_tasks[i].task = NULL;
+
+       v9fs_mux_wq = create_workqueue("v9fs");
+}
 
-static inline int
-dprintcond(struct v9fs_session_info *v9ses, struct v9fs_rpcreq *req)
+void v9fs_mux_global_exit(void)
 {
-       dprintk(DEBUG_MUX, "condition: %d, %p\n", v9ses->transport->status,
-               req->rcall);
-       return 0;
+       destroy_workqueue(v9fs_mux_wq);
 }
 
 /**
- * xread - force read of a certain number of bytes
- * @v9ses: session info structure
- * @ptr: pointer to buffer
- * @sz: number of bytes to read
+ * v9fs_mux_calc_poll_procs - calculates the number of polling procs
+ * based on the number of mounted v9fs filesystems.
  *
- * Chuck Cranor CS-533 project1
+ * The current implementation returns sqrt of the number of mounts.
  */
+inline int v9fs_mux_calc_poll_procs(int muxnum)
+{
+       int n;
+
+       if (v9fs_mux_poll_task_num)
+               n = muxnum / v9fs_mux_poll_task_num +
+                   (muxnum % v9fs_mux_poll_task_num ? 1 : 0);
+       else
+               n = 1;
+
+       if (n > ARRAY_SIZE(v9fs_mux_poll_tasks))
+               n = ARRAY_SIZE(v9fs_mux_poll_tasks);
 
-static int xread(struct v9fs_session_info *v9ses, void *ptr, unsigned long sz)
+       return n;
+}
+
+static void v9fs_mux_poll_start(struct v9fs_mux_data *m)
 {
-       int rd = 0;
-       int ret = 0;
-       while (rd < sz) {
-               ret = v9ses->transport->read(v9ses->transport, ptr, sz - rd);
-               if (ret <= 0) {
-                       dprintk(DEBUG_ERROR, "xread errno %d\n", ret);
-                       return ret;
+       int i, n;
+       struct v9fs_mux_poll_task *vpt, *vptlast;
+
+       dprintk(DEBUG_MUX, "mux %p muxnum %d procnum %d\n", m, v9fs_mux_num,
+               v9fs_mux_poll_task_num);
+       up(&v9fs_mux_task_lock);
+
+       n = v9fs_mux_calc_poll_procs(v9fs_mux_num + 1);
+       if (n > v9fs_mux_poll_task_num) {
+               for (i = 0; i < ARRAY_SIZE(v9fs_mux_poll_tasks); i++) {
+                       if (v9fs_mux_poll_tasks[i].task == NULL) {
+                               vpt = &v9fs_mux_poll_tasks[i];
+                               dprintk(DEBUG_MUX, "create proc %p\n", vpt);
+                               vpt->task = kthread_create(v9fs_poll_proc,
+                                       vpt, "v9fs-poll");
+                               INIT_LIST_HEAD(&vpt->mux_list);
+                               vpt->muxnum = 0;
+                               v9fs_mux_poll_task_num++;
+                               wake_up_process(vpt->task);
+                               break;
+                       }
                }
-               rd += ret;
-               ptr += ret;
-       }
-       return (rd);
-}
 
-/**
- * read_message - read a full 9P2000 fcall packet
- * @v9ses: session info structure
- * @rcall: fcall structure to read into
- * @rcalllen: size of fcall buffer
- *
- */
+               if (i >= ARRAY_SIZE(v9fs_mux_poll_tasks))
+                       dprintk(DEBUG_ERROR, "warning: no free poll slots\n");
+       }
 
-static int
-read_message(struct v9fs_session_info *v9ses,
-            struct v9fs_fcall *rcall, int rcalllen)
-{
-       unsigned char buf[4];
-       void *data;
-       int size = 0;
-       int res = 0;
-
-       res = xread(v9ses, buf, sizeof(buf));
-       if (res < 0) {
-               dprintk(DEBUG_ERROR,
-                       "Reading of count field failed returned: %d\n", res);
-               return res;
+       n = (v9fs_mux_num + 1) / v9fs_mux_poll_task_num +
+           ((v9fs_mux_num + 1) % v9fs_mux_poll_task_num ? 1 : 0);
+
+       vptlast = NULL;
+       for (i = 0; i < ARRAY_SIZE(v9fs_mux_poll_tasks); i++) {
+               vpt = &v9fs_mux_poll_tasks[i];
+               if (vpt->task != NULL) {
+                       vptlast = vpt;
+                       if (vpt->muxnum < n) {
+                               dprintk(DEBUG_MUX, "put in proc %d\n", i);
+                               list_add(&m->mux_list, &vpt->mux_list);
+                               vpt->muxnum++;
+                               m->poll_task = vpt;
+                               memset(&m->poll_waddr, 0, sizeof(m->poll_waddr));
+                               init_poll_funcptr(&m->pt, v9fs_pollwait);
+                               break;
+                       }
+               }
        }
 
-       if (res < 4) {
-               dprintk(DEBUG_ERROR,
-                       "Reading of count field failed returned: %d\n", res);
-               return -EIO;
+       if (i >= ARRAY_SIZE(v9fs_mux_poll_tasks)) {
+               dprintk(DEBUG_MUX, "put in proc %d\n", i);
+               list_add(&m->mux_list, &vptlast->mux_list);
+               vptlast->muxnum++;
+               m->poll_task = vpt;
+               memset(&m->poll_waddr, 0, sizeof(m->poll_waddr));
+               init_poll_funcptr(&m->pt, v9fs_pollwait);
        }
 
-       size = buf[0] | (buf[1] << 8) | (buf[2] << 16) | (buf[3] << 24);
-       dprintk(DEBUG_MUX, "got a packet count: %d\n", size);
+       v9fs_mux_num++;
+       down(&v9fs_mux_task_lock);
+}
 
-       /* adjust for the four bytes of size */
-       size -= 4;
+static void v9fs_mux_poll_stop(struct v9fs_mux_data *m)
+{
+       int i;
+       struct v9fs_mux_poll_task *vpt;
+
+       up(&v9fs_mux_task_lock);
+       vpt = m->poll_task;
+       list_del(&m->mux_list);
+       for(i = 0; i < ARRAY_SIZE(m->poll_waddr); i++) {
+               if (m->poll_waddr[i] != NULL) {
+                       remove_wait_queue(m->poll_waddr[i], &m->poll_wait[i]);
+                       m->poll_waddr[i] = NULL;
+               }
+       }
+       vpt->muxnum--;
+       if (!vpt->muxnum) {
+               dprintk(DEBUG_MUX, "destroy proc %p\n", vpt);
+               send_sig(SIGKILL, vpt->task, 1);
+               vpt->task = NULL;
+               v9fs_mux_poll_task_num--;
+       }
+       v9fs_mux_num--;
+       down(&v9fs_mux_task_lock);
+}
 
-       if (size > v9ses->maxdata) {
-               dprintk(DEBUG_ERROR, "packet too big: %d\n", size);
-               return -E2BIG;
+/**
+ * v9fs_mux_init - allocate and initialize the per-session mux data
+ * Creates the polling task if this is the first session.
+ *
+ * @trans - transport structure
+ * @msize - maximum message size
+ * @extended - pointer to the extended flag
+ */
+struct v9fs_mux_data *v9fs_mux_init(struct v9fs_transport *trans, int msize,
+                                   unsigned char *extended)
+{
+       int i, n;
+       struct v9fs_mux_data *m, *mtmp;
+
+       dprintk(DEBUG_MUX, "transport %p msize %d\n", trans, msize);
+       m = kmalloc(sizeof(struct v9fs_mux_data) + 2 * msize, GFP_KERNEL);
+       if (!m)
+               return ERR_PTR(-ENOMEM);
+
+       spin_lock_init(&m->lock);
+       INIT_LIST_HEAD(&m->mux_list);
+       m->msize = msize;
+       m->extended = extended;
+       m->trans = trans;
+       idr_init(&m->tidpool.pool);
+       init_MUTEX(&m->tidpool.lock);
+       m->err = 0;
+       init_waitqueue_head(&m->equeue);
+       INIT_LIST_HEAD(&m->req_list);
+       INIT_LIST_HEAD(&m->unsent_req_list);
+       m->rpos = 0;
+       m->rbuf = (char *)m + sizeof(struct v9fs_mux_data);
+       m->wpos = m->wsize = 0;
+       m->wbuf = m->rbuf + msize;
+       INIT_WORK(&m->rq, v9fs_read_work, m);
+       INIT_WORK(&m->wq, v9fs_write_work, m);
+       m->wsched = 0;
+       memset(&m->poll_waddr, 0, sizeof(m->poll_waddr));
+       v9fs_mux_poll_start(m);
+
+       n = trans->poll(trans, &m->pt);
+       if (n & POLLIN) {
+               dprintk(DEBUG_MUX, "mux %p can read\n", m);
+               set_bit(Rpending, &m->wsched);
        }
 
-       data = kmalloc(size, GFP_KERNEL);
-       if (!data) {
-               eprintk(KERN_WARNING, "out of memory\n");
-               return -ENOMEM;
+       if (n & POLLOUT) {
+               dprintk(DEBUG_MUX, "mux %p can write\n", m);
+               set_bit(Wpending, &m->wsched);
        }
 
-       res = xread(v9ses, data, size);
-       if (res < size) {
-               dprintk(DEBUG_ERROR, "Reading of fcall failed returned: %d\n",
-                       res);
-               kfree(data);
-               return res;
+       for(i = 0; i < ARRAY_SIZE(m->poll_waddr); i++) {
+               if (IS_ERR(m->poll_waddr[i])) {
+                       v9fs_mux_poll_stop(m);
+                       mtmp = (void *)m->poll_waddr;   /* the error code */
+                       kfree(m);
+                       m = mtmp;
+                       break;
+               }
        }
 
-       /* we now have an in-memory string that is the reply.
-        * deserialize it. There is very little to go wrong at this point
-        * save for v9fs_alloc errors.
-        */
-       res = v9fs_deserialize_fcall(v9ses, size, data, v9ses->maxdata,
-                                    rcall, rcalllen);
+       return m;
+}
 
-       kfree(data);
+/**
+ * v9fs_mux_destroy - cancels all pending requests and frees mux resources
+ */
+void v9fs_mux_destroy(struct v9fs_mux_data *m)
+{
+       dprintk(DEBUG_MUX, "mux %p prev %p next %p\n", m,
+               m->mux_list.prev, m->mux_list.next);
+       v9fs_mux_cancel(m, -ECONNRESET);
+
+       if (!list_empty(&m->req_list)) {
+               /* wait until all processes waiting on this session exit */
+               dprintk(DEBUG_MUX, "mux %p waiting for empty request queue\n",
+                       m);
+               wait_event_timeout(m->equeue, (list_empty(&m->req_list)), 5000);
+               dprintk(DEBUG_MUX, "mux %p request queue empty: %d\n", m,
+                       list_empty(&m->req_list));
+       }
 
-       if (res < 0)
-               return res;
+       v9fs_mux_poll_stop(m);
+       m->trans = NULL;
 
-       return 0;
+       kfree(m);
 }
 
 /**
- * v9fs_recv - receive an RPC response for a particular tag
- * @v9ses: session info structure
- * @req: RPC request structure
- *
+ * v9fs_pollwait - called by files poll operation to add v9fs-poll task
+ *     to files wait queue
  */
-
-static int v9fs_recv(struct v9fs_session_info *v9ses, struct v9fs_rpcreq *req)
+static void
+v9fs_pollwait(struct file *filp, wait_queue_head_t * wait_address,
+             poll_table * p)
 {
-       int ret = 0;
-
-       dprintk(DEBUG_MUX, "waiting for response: %d\n", req->tcall->tag);
-       ret = wait_event_interruptible(v9ses->read_wait,
-                      ((v9ses->transport->status != Connected) ||
-                       (req->rcall != 0) || (req->err < 0) ||
-                       dprintcond(v9ses, req)));
+       int i;
+       struct v9fs_mux_data *m;
 
-       dprintk(DEBUG_MUX, "got it: rcall %p\n", req->rcall);
+       m = container_of(p, struct v9fs_mux_data, pt);
+       for(i = 0; i < ARRAY_SIZE(m->poll_waddr); i++)
+               if (m->poll_waddr[i] == NULL)
+                       break;
 
-       spin_lock(&v9ses->muxlock);
-       list_del(&req->next);
-       spin_unlock(&v9ses->muxlock);
+       if (i >= ARRAY_SIZE(m->poll_waddr)) {
+               dprintk(DEBUG_ERROR, "not enough wait_address slots\n");
+               return;
+       }
 
-       if (req->err < 0)
-               return req->err;
+       m->poll_waddr[i] = wait_address;
 
-       if (v9ses->transport->status == Disconnected)
-               return -ECONNRESET;
+       if (!wait_address) {
+               dprintk(DEBUG_ERROR, "no wait_address\n");
+               m->poll_waddr[i] = ERR_PTR(-EIO);
+               return;
+       }
 
-       return ret;
+       init_waitqueue_entry(&m->poll_wait[i], m->poll_task->task);
+       add_wait_queue(wait_address, &m->poll_wait[i]);
 }
 
 /**
- * v9fs_send - send a 9P request
- * @v9ses: session info structure
- * @req: RPC request to send
- *
+ * v9fs_poll_mux - polls a mux and schedules read or write works if necessary
  */
-
-static int v9fs_send(struct v9fs_session_info *v9ses, struct v9fs_rpcreq *req)
+static inline void v9fs_poll_mux(struct v9fs_mux_data *m)
 {
-       int ret = -1;
-       void *data = NULL;
-       struct v9fs_fcall *tcall = req->tcall;
-
-       data = kmalloc(v9ses->maxdata + V9FS_IOHDRSZ, GFP_KERNEL);
-       if (!data)
-               return -ENOMEM;
-
-       tcall->size = 0;        /* enforce size recalculation */
-       ret =
-           v9fs_serialize_fcall(v9ses, tcall, data,
-                                v9ses->maxdata + V9FS_IOHDRSZ);
-       if (ret < 0)
-               goto free_data;
-
-       spin_lock(&v9ses->muxlock);
-       list_add(&req->next, &v9ses->mux_fcalls);
-       spin_unlock(&v9ses->muxlock);
-
-       dprintk(DEBUG_MUX, "sending message: tag %d size %d\n", tcall->tag,
-               tcall->size);
-       ret = v9ses->transport->write(v9ses->transport, data, tcall->size);
-
-       if (ret != tcall->size) {
-               spin_lock(&v9ses->muxlock);
-               list_del(&req->next);
-               kfree(req->rcall);
+       int n;
 
-               spin_unlock(&v9ses->muxlock);
-               if (ret >= 0)
-                       ret = -EREMOTEIO;
-       } else
-               ret = 0;
+       if (m->err < 0)
+               return;
+
+       n = m->trans->poll(m->trans, NULL);
+       if (n < 0 || n & (POLLERR | POLLHUP | POLLNVAL)) {
+               dprintk(DEBUG_MUX, "error mux %p err %d\n", m, n);
+               if (n >= 0)
+                       n = -ECONNRESET;
+               v9fs_mux_cancel(m, n);
+       }
+
+       if (n & POLLIN) {
+               set_bit(Rpending, &m->wsched);
+               dprintk(DEBUG_MUX, "mux %p can read\n", m);
+               if (!test_and_set_bit(Rworksched, &m->wsched)) {
+                       dprintk(DEBUG_MUX, "schedule read work mux %p\n", m);
+                       queue_work(v9fs_mux_wq, &m->rq);
+               }
+       }
 
-      free_data:
-       kfree(data);
-       return ret;
+       if (n & POLLOUT) {
+               set_bit(Wpending, &m->wsched);
+               dprintk(DEBUG_MUX, "mux %p can write\n", m);
+               if ((m->wsize || !list_empty(&m->unsent_req_list))
+                   && !test_and_set_bit(Wworksched, &m->wsched)) {
+                       dprintk(DEBUG_MUX, "schedule write work mux %p\n", m);
+                       queue_work(v9fs_mux_wq, &m->wq);
+               }
+       }
 }
 
 /**
- * v9fs_mux_rpc - send a request, receive a response
- * @v9ses: session info structure
- * @tcall: fcall to send
- * @rcall: buffer to place response into
- *
+ * v9fs_poll_proc - polls all v9fs transports for new events and queues
+ *     the appropriate work to the work queue
  */
-
-long
-v9fs_mux_rpc(struct v9fs_session_info *v9ses, struct v9fs_fcall *tcall,
-            struct v9fs_fcall **rcall)
+static int v9fs_poll_proc(void *a)
 {
-       int tid = -1;
-       struct v9fs_fcall *fcall = NULL;
-       struct v9fs_rpcreq req;
-       int ret = -1;
+       struct v9fs_mux_data *m, *mtmp;
+       struct v9fs_mux_poll_task *vpt;
 
-       if (!v9ses)
-               return -EINVAL;
+       vpt = a;
+       dprintk(DEBUG_MUX, "start %p %p\n", current, vpt);
+       allow_signal(SIGKILL);
+       while (!kthread_should_stop()) {
+               set_current_state(TASK_INTERRUPTIBLE);
+               if (signal_pending(current))
+                       break;
 
-       if (!v9ses->transport || v9ses->transport->status != Connected)
-               return -EIO;
+               list_for_each_entry_safe(m, mtmp, &vpt->mux_list, mux_list) {
+                       v9fs_poll_mux(m);
+               }
+
+               dprintk(DEBUG_MUX, "sleeping...\n");
+               schedule_timeout(SCHED_TIMEOUT * HZ);
+       }
 
-       if (rcall)
-               *rcall = NULL;
+       __set_current_state(TASK_RUNNING);
+       dprintk(DEBUG_MUX, "finish\n");
+       return 0;
+}
 
-       if (tcall->id != TVERSION) {
-               tid = v9fs_get_idpool(&v9ses->tidpool);
-               if (tid < 0)
-                       return -ENOMEM;
+static inline int v9fs_write_req(struct v9fs_mux_data *m, struct v9fs_req *req)
+{
+       int n;
+
+       list_move_tail(&req->req_list, &m->req_list);
+       n = v9fs_serialize_fcall(req->tcall, m->wbuf, m->msize, *m->extended);
+       if (n < 0) {
+               req->err = n;
+               list_del(&req->req_list);
+               if (req->cb) {
+                       spin_unlock(&m->lock);
+                       (*req->cb) (req->cba, req->tcall, req->rcall, req->err);
+                       req->cb = NULL;
+                       spin_lock(&m->lock);
+               } else
+                       kfree(req->rcall);
+
+               kfree(req);
        }
 
-       tcall->tag = tid;
+       return n;
+}
 
-       req.tcall = tcall;
-       req.err = 0;
-       req.rcall = NULL;
+/**
+ * v9fs_write_work - called when a transport can send some data
+ */
+static void v9fs_write_work(void *a)
+{
+       int n, err;
+       struct v9fs_mux_data *m;
+       struct v9fs_req *req, *rtmp;
 
-       ret = v9fs_send(v9ses, &req);
+       m = a;
 
-       if (ret < 0) {
-               if (tcall->id != TVERSION)
-                       v9fs_put_idpool(tid, &v9ses->tidpool);
-               dprintk(DEBUG_MUX, "error %d\n", ret);
-               return ret;
+       if (m->err < 0) {
+               clear_bit(Wworksched, &m->wsched);
+               return;
        }
 
-       ret = v9fs_recv(v9ses, &req);
-
-       fcall = req.rcall;
-
-       dprintk(DEBUG_MUX, "received: tag=%x, ret=%d\n", tcall->tag, ret);
-       if (ret == -ERESTARTSYS) {
-               if (v9ses->transport->status != Disconnected
-                   && tcall->id != TFLUSH) {
-                       unsigned long flags;
-
-                       dprintk(DEBUG_MUX, "flushing the tag: %d\n",
-                               tcall->tag);
-                       clear_thread_flag(TIF_SIGPENDING);
-                       v9fs_t_flush(v9ses, tcall->tag);
-                       spin_lock_irqsave(&current->sighand->siglock, flags);
-                       recalc_sigpending();
-                       spin_unlock_irqrestore(&current->sighand->siglock,
-                                              flags);
-                       dprintk(DEBUG_MUX, "flushing done\n");
+       if (!m->wsize) {
+               if (list_empty(&m->unsent_req_list)) {
+                       clear_bit(Wworksched, &m->wsched);
+                       return;
                }
 
-               goto release_req;
-       } else if (ret < 0)
-               goto release_req;
-
-       if (!fcall)
-               ret = -EIO;
-       else {
-               if (fcall->id == RERROR) {
-                       ret = v9fs_errstr2errno(fcall->params.rerror.error);
-                       if (ret == 0) { /* string match failed */
-                               if (fcall->params.rerror.errno)
-                                       ret = -(fcall->params.rerror.errno);
-                               else
-                                       ret = -ESERVERFAULT;
-                       }
-               } else if (fcall->id != tcall->id + 1) {
-                       dprintk(DEBUG_ERROR,
-                               "fcall mismatch: expected %d, got %d\n",
-                               tcall->id + 1, fcall->id);
-                       ret = -EIO;
+               err = 0;
+               spin_lock(&m->lock);
+               list_for_each_entry_safe(req, rtmp, &m->unsent_req_list,
+                                        req_list) {
+                       err = v9fs_write_req(m, req);
+                       if (err > 0)
+                               break;
                }
+
+               m->wsize = err;
+               m->wpos = 0;
+               spin_unlock(&m->lock);
        }
 
-      release_req:
-       if (tcall->id != TVERSION)
-               v9fs_put_idpool(tid, &v9ses->tidpool);
-       if (rcall)
-               *rcall = fcall;
-       else
-               kfree(fcall);
+       dprintk(DEBUG_MUX, "mux %p pos %d size %d\n", m, m->wpos, m->wsize);
+       clear_bit(Wpending, &m->wsched);
+       err = m->trans->write(m->trans, m->wbuf + m->wpos, m->wsize - m->wpos);
+       dprintk(DEBUG_MUX, "mux %p sent %d bytes\n", m, err);
+       if (err == -EAGAIN) {
+               clear_bit(Wworksched, &m->wsched);
+               return;
+       }
+
+       if (err <= 0)
+               goto error;
+
+       m->wpos += err;
+       if (m->wpos == m->wsize)
+               m->wpos = m->wsize = 0;
+
+       if (m->wsize == 0 && !list_empty(&m->unsent_req_list)) {
+               if (test_and_clear_bit(Wpending, &m->wsched))
+                       n = POLLOUT;
+               else
+                       n = m->trans->poll(m->trans, NULL);
+
+               if (n & POLLOUT) {
+                       dprintk(DEBUG_MUX, "schedule write work mux %p\n", m);
+                       queue_work(v9fs_mux_wq, &m->wq);
+               } else
+                       clear_bit(Wworksched, &m->wsched);
+       } else
+               clear_bit(Wworksched, &m->wsched);
 
-       return ret;
+       return;
+
+      error:
+       v9fs_mux_cancel(m, err);
+       clear_bit(Wworksched, &m->wsched);
 }
 
-/**
- * v9fs_mux_cancel_requests - cancels all pending requests
- *
- * @v9ses: session info structure
- * @err: error code to return to the requests
- */
-void v9fs_mux_cancel_requests(struct v9fs_session_info *v9ses, int err)
+static void process_request(struct v9fs_mux_data *m, struct v9fs_req *req)
 {
-       struct v9fs_rpcreq *rptr;
-       struct v9fs_rpcreq *rreq;
+       int ecode, tag;
+       char *ename;
+
+       tag = req->tag;
+       if (req->rcall->id == RERROR && !req->err) {
+               ecode = req->rcall->params.rerror.errno;
+               ename = req->rcall->params.rerror.error;
 
-       dprintk(DEBUG_MUX, " %d\n", err);
-       spin_lock(&v9ses->muxlock);
-       list_for_each_entry_safe(rreq, rptr, &v9ses->mux_fcalls, next) {
-               rreq->err = err;
+               dprintk(DEBUG_MUX, "Rerror %s\n", ename);
+
+               if (*m->extended)
+                       req->err = -ecode;
+
+               if (!req->err) {
+                       req->err = v9fs_errstr2errno(ename);
+
+                       if (!req->err) {        /* string match failed */
+                               dprintk(DEBUG_ERROR, "unknown error: %s\n",
+                                       ename);
+                       }
+
+                       if (!req->err)
+                               req->err = -ESERVERFAULT;
+               }
+       } else if (req->tcall && req->rcall->id != req->tcall->id + 1) {
+               dprintk(DEBUG_ERROR, "fcall mismatch: expected %d, got %d\n",
+                       req->tcall->id + 1, req->rcall->id);
+               if (!req->err)
+                       req->err = -EIO;
        }
-       spin_unlock(&v9ses->muxlock);
-       wake_up_all(&v9ses->read_wait);
+
+       if (req->cb && req->err != ERREQFLUSH) {
+               dprintk(DEBUG_MUX, "calling callback tcall %p rcall %p\n",
+                       req->tcall, req->rcall);
+
+               (*req->cb) (req->cba, req->tcall, req->rcall, req->err);
+               req->cb = NULL;
+       } else
+               kfree(req->rcall);
+
+       if (tag != V9FS_NOTAG)
+               v9fs_put_idpool(tag, &m->tidpool);
+
+       wake_up(&m->equeue);
+       kfree(req);
 }
 
 /**
- * v9fs_recvproc - kproc to handle demultiplexing responses
- * @data: session info structure
- *
+ * v9fs_read_work - called when there is some data to be read from a transport
  */
-
-static int v9fs_recvproc(void *data)
+static void v9fs_read_work(void *a)
 {
-       struct v9fs_session_info *v9ses = (struct v9fs_session_info *)data;
-       struct v9fs_fcall *rcall = NULL;
-       struct v9fs_rpcreq *rptr;
-       struct v9fs_rpcreq *req;
-       struct v9fs_rpcreq *rreq;
-       int err = 0;
+       int n, err, rcallen;
+       struct v9fs_mux_data *m;
+       struct v9fs_req *req, *rptr, *rreq;
+       struct v9fs_fcall *rcall;
+
+       m = a;
+
+       if (m->err < 0)
+               return;
+
+       rcall = NULL;
+       dprintk(DEBUG_MUX, "start mux %p pos %d\n", m, m->rpos);
+       clear_bit(Rpending, &m->wsched);
+       err = m->trans->read(m->trans, m->rbuf + m->rpos, m->msize - m->rpos);
+       dprintk(DEBUG_MUX, "mux %p got %d bytes\n", m, err);
+       if (err == -EAGAIN) {
+               clear_bit(Rworksched, &m->wsched);
+               return;
+       }
 
-       allow_signal(SIGKILL);
-       set_current_state(TASK_INTERRUPTIBLE);
-       complete(&v9ses->proccmpl);
-       while (!kthread_should_stop() && err >= 0) {
-               req = rptr = rreq = NULL;
+       if (err <= 0)
+               goto error;
 
-               rcall = kmalloc(v9ses->maxdata + V9FS_IOHDRSZ, GFP_KERNEL);
-               if (!rcall) {
-                       eprintk(KERN_ERR, "no memory for buffers\n");
+       m->rpos += err;
+       while (m->rpos > 4) {
+               n = le32_to_cpu(*(__le32 *) m->rbuf);
+               if (n >= m->msize) {
+                       dprintk(DEBUG_ERROR,
+                               "requested packet size too big: %d\n", n);
+                       err = -EIO;
+                       goto error;
+               }
+
+               if (m->rpos < n)
                        break;
+
+               rcallen = n + V9FS_FCALLHDRSZ;
+               rcall = kmalloc(rcallen, GFP_KERNEL);
+               if (!rcall) {
+                       err = -ENOMEM;
+                       goto error;
                }
 
-               err = read_message(v9ses, rcall, v9ses->maxdata + V9FS_IOHDRSZ);
-               spin_lock(&v9ses->muxlock);
+               dump_data(m->rbuf, n);
+               err = v9fs_deserialize_fcall(m->rbuf, n, rcall, rcallen,
+                                            *m->extended);
                if (err < 0) {
-                       list_for_each_entry_safe(rreq, rptr, &v9ses->mux_fcalls, next) {
-                               rreq->err = err;
-                       }
-                       if(err != -ERESTARTSYS)
-                               eprintk(KERN_ERR,
-                                       "Transport error while reading message %d\n", err);
-               } else {
-                       list_for_each_entry_safe(rreq, rptr, &v9ses->mux_fcalls, next) {
-                               if (rreq->tcall->tag == rcall->tag) {
-                                       req = rreq;
-                                       req->rcall = rcall;
-                                       break;
-                               }
-                       }
+                       kfree(rcall);
+                       goto error;
                }
 
-               if (req && (req->tcall->id == TFLUSH)) {
-                       struct v9fs_rpcreq *treq = NULL;
-                       list_for_each_entry_safe(treq, rptr, &v9ses->mux_fcalls, next) {
-                               if (treq->tcall->tag ==
-                                   req->tcall->params.tflush.oldtag) {
-                                       list_del(&rptr->next);
-                                       kfree(treq->rcall);
-                                       break;
-                               }
+               dprintk(DEBUG_MUX, "mux %p fcall id %d tag %d\n", m, rcall->id,
+                       rcall->tag);
+
+               req = NULL;
+               spin_lock(&m->lock);
+               list_for_each_entry_safe(rreq, rptr, &m->req_list, req_list) {
+                       if (rreq->tag == rcall->tag) {
+                               req = rreq;
+                               req->rcall = rcall;
+                               list_del(&req->req_list);
+                               spin_unlock(&m->lock);
+                               process_request(m, req);
+                               break;
                        }
                }
 
-               spin_unlock(&v9ses->muxlock);
-
                if (!req) {
-                       if (err >= 0)
+                       spin_unlock(&m->lock);
+                       if (err >= 0 && rcall->id != RFLUSH)
                                dprintk(DEBUG_ERROR,
-                                       "unexpected response: id %d tag %d\n",
-                                       rcall->id, rcall->tag);
-
+                                       "unexpected response mux %p id %d tag %d\n",
+                                       m, rcall->id, rcall->tag);
                        kfree(rcall);
                }
 
-               wake_up_all(&v9ses->read_wait);
-               set_current_state(TASK_INTERRUPTIBLE);
+               if (m->rpos > n)
+                       memmove(m->rbuf, m->rbuf + n, m->rpos - n);
+               m->rpos -= n;
        }
 
-       v9ses->transport->close(v9ses->transport);
-
-       /* Inform all pending processes about the failure */
-       wake_up_all(&v9ses->read_wait);
-
-       if (signal_pending(current))
-               complete(&v9ses->proccmpl);
+       if (!list_empty(&m->req_list)) {
+               if (test_and_clear_bit(Rpending, &m->wsched))
+                       n = POLLIN;
+               else
+                       n = m->trans->poll(m->trans, NULL);
+
+               if (n & POLLIN) {
+                       dprintk(DEBUG_MUX, "schedule read work mux %p\n", m);
+                       queue_work(v9fs_mux_wq, &m->rq);
+               } else
+                       clear_bit(Rworksched, &m->wsched);
+       } else
+               clear_bit(Rworksched, &m->wsched);
 
-       dprintk(DEBUG_MUX, "recvproc: end\n");
-       v9ses->recvproc = NULL;
+       return;
 
-       return err >= 0;
+      error:
+       v9fs_mux_cancel(m, err);
+       clear_bit(Rworksched, &m->wsched);
 }
 
 /**
- * v9fs_mux_init - initialize multiplexer (spawn kproc)
- * @v9ses: session info structure
- * @dev_name: mount device information (to create unique kproc)
+ * v9fs_send_request - send 9P request
+ * The function can sleep until the request is scheduled for sending.
+ * The function can be interrupted. Return from the function is not
+ * a guarantee that the request is sent succesfully. Can return errors
+ * that can be retrieved by PTR_ERR macros.
  *
+ * @m: mux data
+ * @tc: request to be sent
+ * @cb: callback function to call when response is received
+ * @cba: parameter to pass to the callback function
  */
+static struct v9fs_req *v9fs_send_request(struct v9fs_mux_data *m,
+                                         struct v9fs_fcall *tc,
+                                         v9fs_mux_req_callback cb, void *cba)
+{
+       int n;
+       struct v9fs_req *req;
+
+       dprintk(DEBUG_MUX, "mux %p task %p tcall %p id %d\n", m, current,
+               tc, tc->id);
+       if (m->err < 0)
+               return ERR_PTR(m->err);
+
+       req = kmalloc(sizeof(struct v9fs_req), GFP_KERNEL);
+       if (!req)
+               return ERR_PTR(-ENOMEM);
 
-int v9fs_mux_init(struct v9fs_session_info *v9ses, const char *dev_name)
+       if (tc->id == TVERSION)
+               n = V9FS_NOTAG;
+       else
+               n = v9fs_get_idpool(&m->tidpool);
+
+       if (n < 0)
+               return ERR_PTR(-ENOMEM);
+
+       tc->tag = n;
+       req->tag = n;
+       req->tcall = tc;
+       req->rcall = NULL;
+       req->err = 0;
+       req->cb = cb;
+       req->cba = cba;
+
+       spin_lock(&m->lock);
+       list_add_tail(&req->req_list, &m->unsent_req_list);
+       spin_unlock(&m->lock);
+
+       if (test_and_clear_bit(Wpending, &m->wsched))
+               n = POLLOUT;
+       else
+               n = m->trans->poll(m->trans, NULL);
+
+       if (n & POLLOUT && !test_and_set_bit(Wworksched, &m->wsched))
+               queue_work(v9fs_mux_wq, &m->wq);
+
+       return req;
+}
+
+static inline void
+v9fs_mux_flush_cb(void *a, struct v9fs_fcall *tc, struct v9fs_fcall *rc,
+                 int err)
 {
-       char procname[60];
-
-       strncpy(procname, dev_name, sizeof(procname));
-       procname[sizeof(procname) - 1] = 0;
-
-       init_waitqueue_head(&v9ses->read_wait);
-       init_completion(&v9ses->fcread);
-       init_completion(&v9ses->proccmpl);
-       spin_lock_init(&v9ses->muxlock);
-       INIT_LIST_HEAD(&v9ses->mux_fcalls);
-       v9ses->recvproc = NULL;
-       v9ses->curfcall = NULL;
-
-       v9ses->recvproc = kthread_create(v9fs_recvproc, v9ses,
-                                        "v9fs_recvproc %s", procname);
-
-       if (IS_ERR(v9ses->recvproc)) {
-               eprintk(KERN_ERR, "cannot create receiving thread\n");
-               v9fs_session_close(v9ses);
-               return -ECONNABORTED;
+       v9fs_mux_req_callback cb;
+       int tag;
+       struct v9fs_mux_data *m;
+       struct v9fs_req *req, *rptr;
+
+       m = a;
+       dprintk(DEBUG_MUX, "mux %p tc %p rc %p err %d oldtag %d\n", m, tc,
+               rc, err, tc->params.tflush.oldtag);
+
+       spin_lock(&m->lock);
+       cb = NULL;
+       tag = tc->params.tflush.oldtag;
+       list_for_each_entry_safe(req, rptr, &m->req_list, req_list) {
+               if (req->tag == tag) {
+                       list_del(&req->req_list);
+                       if (req->cb) {
+                               cb = req->cb;
+                               req->cb = NULL;
+                               spin_unlock(&m->lock);
+                               (*cb) (req->cba, req->tcall, req->rcall,
+                                      req->err);
+                       }
+                       kfree(req);
+                       wake_up(&m->equeue);
+                       break;
+               }
+       }
+
+       if (!cb)
+               spin_unlock(&m->lock);
+
+       if (v9fs_check_idpool(tag, &m->tidpool))
+               v9fs_put_idpool(tag, &m->tidpool);
+
+       kfree(tc);
+       kfree(rc);
+}
+
+static void
+v9fs_mux_flush_request(struct v9fs_mux_data *m, struct v9fs_req *req)
+{
+       struct v9fs_fcall *fc;
+
+       dprintk(DEBUG_MUX, "mux %p req %p tag %d\n", m, req, req->tag);
+
+       fc = kmalloc(sizeof(struct v9fs_fcall), GFP_KERNEL);
+       fc->id = TFLUSH;
+       fc->params.tflush.oldtag = req->tag;
+
+       v9fs_send_request(m, fc, v9fs_mux_flush_cb, m);
+}
+
+static void
+v9fs_mux_rpc_cb(void *a, struct v9fs_fcall *tc, struct v9fs_fcall *rc, int err)
+{
+       struct v9fs_mux_rpc *r;
+
+       if (err == ERREQFLUSH) {
+               dprintk(DEBUG_MUX, "err req flush\n");
+               return;
+       }
+
+       r = a;
+       dprintk(DEBUG_MUX, "mux %p req %p tc %p rc %p err %d\n", r->m, r->req,
+               tc, rc, err);
+       r->rcall = rc;
+       r->err = err;
+       wake_up(&r->wqueue);
+}
+
+/**
+ * v9fs_mux_rpc - sends 9P request and waits until a response is available.
+ *     The function can be interrupted.
+ * @m: mux data
+ * @tc: request to be sent
+ * @rc: pointer where a pointer to the response is stored
+ */
+int
+v9fs_mux_rpc(struct v9fs_mux_data *m, struct v9fs_fcall *tc,
+            struct v9fs_fcall **rc)
+{
+       int err;
+       unsigned long flags;
+       struct v9fs_req *req;
+       struct v9fs_mux_rpc r;
+
+       r.err = 0;
+       r.rcall = NULL;
+       r.m = m;
+       init_waitqueue_head(&r.wqueue);
+
+       if (rc)
+               *rc = NULL;
+
+       req = v9fs_send_request(m, tc, v9fs_mux_rpc_cb, &r);
+       if (IS_ERR(req)) {
+               err = PTR_ERR(req);
+               dprintk(DEBUG_MUX, "error %d\n", err);
+               return PTR_ERR(req);
+       }
+
+       r.req = req;
+       dprintk(DEBUG_MUX, "mux %p tc %p tag %d rpc %p req %p\n", m, tc,
+               req->tag, &r, req);
+       err = wait_event_interruptible(r.wqueue, r.rcall != NULL || r.err < 0);
+       if (r.err < 0)
+               err = r.err;
+
+       if (err == -ERESTARTSYS && m->trans->status == Connected && m->err == 0) {
+               spin_lock(&m->lock);
+               req->tcall = NULL;
+               req->err = ERREQFLUSH;
+               spin_unlock(&m->lock);
+
+               clear_thread_flag(TIF_SIGPENDING);
+               v9fs_mux_flush_request(m, req);
+               spin_lock_irqsave(&current->sighand->siglock, flags);
+               recalc_sigpending();
+               spin_unlock_irqrestore(&current->sighand->siglock, flags);
        }
 
-       wake_up_process(v9ses->recvproc);
-       wait_for_completion(&v9ses->proccmpl);
+       if (!err) {
+               if (r.rcall)
+                       dprintk(DEBUG_MUX, "got response id %d tag %d\n",
+                               r.rcall->id, r.rcall->tag);
+
+               if (rc)
+                       *rc = r.rcall;
+               else
+                       kfree(r.rcall);
+       } else {
+               kfree(r.rcall);
+               dprintk(DEBUG_MUX, "got error %d\n", err);
+               if (err > 0)
+                       err = -EIO;
+       }
+
+       return err;
+}
+
+/**
+ * v9fs_mux_rpcnb - sends 9P request without waiting for response.
+ * @m: mux data
+ * @tc: request to be sent
+ * @cb: callback function to be called when response arrives
+ * @cba: value to pass to the callback function
+ */
+int v9fs_mux_rpcnb(struct v9fs_mux_data *m, struct v9fs_fcall *tc,
+                  v9fs_mux_req_callback cb, void *a)
+{
+       int err;
+       struct v9fs_req *req;
+
+       req = v9fs_send_request(m, tc, cb, a);
+       if (IS_ERR(req)) {
+               err = PTR_ERR(req);
+               dprintk(DEBUG_MUX, "error %d\n", err);
+               return PTR_ERR(req);
+       }
 
+       dprintk(DEBUG_MUX, "mux %p tc %p tag %d\n", m, tc, req->tag);
        return 0;
 }
+
+/**
+ * v9fs_mux_cancel - cancel all pending requests with error
+ * @m: mux data
+ * @err: error code
+ */
+void v9fs_mux_cancel(struct v9fs_mux_data *m, int err)
+{
+       struct v9fs_req *req, *rtmp;
+       LIST_HEAD(cancel_list);
+
+       dprintk(DEBUG_MUX, "mux %p err %d\n", m, err);
+       m->err = err;
+       spin_lock(&m->lock);
+       list_for_each_entry_safe(req, rtmp, &m->req_list, req_list) {
+               list_move(&req->req_list, &cancel_list);
+       }
+       spin_unlock(&m->lock);
+
+       list_for_each_entry_safe(req, rtmp, &cancel_list, req_list) {
+               list_del(&req->req_list);
+               if (!req->err)
+                       req->err = err;
+
+               if (req->cb)
+                       (*req->cb) (req->cba, req->tcall, req->rcall, req->err);
+               else
+                       kfree(req->rcall);
+
+               kfree(req);
+       }
+
+       wake_up(&m->equeue);
+}