[PATCH] devfs: Remove the devfs_fs_kernel.h file from the tree
[safe/jmp/linux-2.6] / drivers / block / nbd.c
index 485345c..b164ee0 100644 (file)
@@ -7,39 +7,9 @@
  * Copyright 1997-2000 Pavel Machek <pavel@ucw.cz>
  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  *
- * (part of code stolen from loop.c)
+ * This file is released under GPLv2 or later.
  *
- * 97-3-25 compiled 0-th version, not yet tested it 
- *   (it did not work, BTW) (later that day) HEY! it works!
- *   (bit later) hmm, not that much... 2:00am next day:
- *   yes, it works, but it gives something like 50kB/sec
- * 97-4-01 complete rewrite to make it possible for many requests at 
- *   once to be processed
- * 97-4-11 Making protocol independent of endianity etc.
- * 97-9-13 Cosmetic changes
- * 98-5-13 Attempt to make 64-bit-clean on 64-bit machines
- * 99-1-11 Attempt to make 64-bit-clean on 32-bit machines <ankry@mif.pg.gda.pl>
- * 01-2-27 Fix to store proper blockcount for kernel (calculated using
- *   BLOCK_SIZE_BITS, not device blocksize) <aga@permonline.ru>
- * 01-3-11 Make nbd work with new Linux block layer code. It now supports
- *   plugging like all the other block devices. Also added in MSG_MORE to
- *   reduce number of partial TCP segments sent. <steve@chygwyn.com>
- * 01-12-6 Fix deadlock condition by making queue locks independent of
- *   the transmit lock. <steve@chygwyn.com>
- * 02-10-11 Allow hung xmit to be aborted via SIGKILL & various fixes.
- *   <Paul.Clements@SteelEye.com> <James.Bottomley@SteelEye.com>
- * 03-06-22 Make nbd work with new linux 2.5 block layer design. This fixes
- *   memory corruption from module removal and possible memory corruption
- *   from sending/receiving disk data. <ldl@aros.net>
- * 03-06-23 Cosmetic changes. <ldl@aros.net>
- * 03-06-23 Enhance diagnostics support. <ldl@aros.net>
- * 03-06-24 Remove unneeded blksize_bits field from nbd_device struct.
- *   <ldl@aros.net>
- * 03-06-24 Cleanup PARANOIA usage & code. <ldl@aros.net>
- * 04-02-19 Remove PARANOIA, plus various cleanups (Paul Clements)
- * possible FIXME: make set_sock / set_blksize / set_size / do_it one syscall
- * why not: would need access_ok and friends, would share yet another
- *          structure with userland
+ * (part of code stolen from loop.c)
  */
 
 #include <linux/major.h>
 #include <linux/errno.h>
 #include <linux/file.h>
 #include <linux/ioctl.h>
+#include <linux/compiler.h>
+#include <linux/err.h>
+#include <linux/kernel.h>
 #include <net/sock.h>
 
-#include <linux/devfs_fs_kernel.h>
-
 #include <asm/uaccess.h>
+#include <asm/system.h>
 #include <asm/types.h>
 
 #include <linux/nbd.h>
@@ -170,7 +142,6 @@ static int sock_xmit(struct socket *sock, int send, void *buf, int size,
                msg.msg_namelen = 0;
                msg.msg_control = NULL;
                msg.msg_controllen = 0;
-               msg.msg_namelen = 0;
                msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 
                if (send)
@@ -230,14 +201,6 @@ static int nbd_send_req(struct nbd_device *lo, struct request *req)
        request.len = htonl(size);
        memcpy(request.handle, &req, sizeof(req));
 
-       down(&lo->tx_lock);
-
-       if (!sock || !lo->sock) {
-               printk(KERN_ERR "%s: Attempted send on closed socket\n",
-                               lo->disk->disk_name);
-               goto error_out;
-       }
-
        dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
                        lo->disk->disk_name, req,
                        nbdcmd_to_ascii(nbd_cmd(req)),
@@ -276,11 +239,9 @@ static int nbd_send_req(struct nbd_device *lo, struct request *req)
                        }
                }
        }
-       up(&lo->tx_lock);
        return 0;
 
 error_out:
-       up(&lo->tx_lock);
        return 1;
 }
 
@@ -289,9 +250,14 @@ static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
        struct request *req;
        struct list_head *tmp;
        struct request *xreq;
+       int err;
 
        memcpy(&xreq, handle, sizeof(xreq));
 
+       err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
+       if (unlikely(err))
+               goto out;
+
        spin_lock(&lo->queue_lock);
        list_for_each(tmp, &lo->queue_head) {
                req = list_entry(tmp, struct request, queuelist);
@@ -302,7 +268,11 @@ static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
                return req;
        }
        spin_unlock(&lo->queue_lock);
-       return NULL;
+
+       err = -ENOENT;
+
+out:
+       return ERR_PTR(err);
 }
 
 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
@@ -331,7 +301,11 @@ static struct request *nbd_read_stat(struct nbd_device *lo)
                goto harderror;
        }
        req = nbd_find_request(lo, reply.handle);
-       if (req == NULL) {
+       if (unlikely(IS_ERR(req))) {
+               result = PTR_ERR(req);
+               if (result != -ENOENT)
+                       goto harderror;
+
                printk(KERN_ERR "%s: Unexpected reply (%p)\n",
                                lo->disk->disk_name, reply.handle);
                result = -EBADR;
@@ -395,19 +369,24 @@ static void nbd_clear_que(struct nbd_device *lo)
 
        BUG_ON(lo->magic != LO_MAGIC);
 
-       do {
-               req = NULL;
-               spin_lock(&lo->queue_lock);
-               if (!list_empty(&lo->queue_head)) {
-                       req = list_entry(lo->queue_head.next, struct request, queuelist);
-                       list_del_init(&req->queuelist);
-               }
-               spin_unlock(&lo->queue_lock);
-               if (req) {
-                       req->errors++;
-                       nbd_end_request(req);
-               }
-       } while (req);
+       /*
+        * Because we have set lo->sock to NULL under the tx_lock, all
+        * modifications to the list must have completed by now.  For
+        * the same reason, the active_req must be NULL.
+        *
+        * As a consequence, we don't need to take the spin lock while
+        * purging the list here.
+        */
+       BUG_ON(lo->sock);
+       BUG_ON(lo->active_req);
+
+       while (!list_empty(&lo->queue_head)) {
+               req = list_entry(lo->queue_head.next, struct request,
+                                queuelist);
+               list_del_init(&req->queuelist);
+               req->errors++;
+               nbd_end_request(req);
+       }
 }
 
 /*
@@ -435,11 +414,6 @@ static void do_nbd_request(request_queue_t * q)
 
                BUG_ON(lo->magic != LO_MAGIC);
 
-               if (!lo->file) {
-                       printk(KERN_ERR "%s: Request when not-ready\n",
-                                       lo->disk->disk_name);
-                       goto error_out;
-               }
                nbd_cmd(req) = NBD_CMD_READ;
                if (rq_data_dir(req) == WRITE) {
                        nbd_cmd(req) = NBD_CMD_WRITE;
@@ -453,32 +427,34 @@ static void do_nbd_request(request_queue_t * q)
                req->errors = 0;
                spin_unlock_irq(q->queue_lock);
 
-               spin_lock(&lo->queue_lock);
-
-               if (!lo->file) {
-                       spin_unlock(&lo->queue_lock);
-                       printk(KERN_ERR "%s: failed between accept and semaphore, file lost\n",
-                                       lo->disk->disk_name);
+               mutex_lock(&lo->tx_lock);
+               if (unlikely(!lo->sock)) {
+                       mutex_unlock(&lo->tx_lock);
+                       printk(KERN_ERR "%s: Attempted send on closed socket\n",
+                              lo->disk->disk_name);
                        req->errors++;
                        nbd_end_request(req);
                        spin_lock_irq(q->queue_lock);
                        continue;
                }
 
-               list_add(&req->queuelist, &lo->queue_head);
-               spin_unlock(&lo->queue_lock);
+               lo->active_req = req;
 
                if (nbd_send_req(lo, req) != 0) {
                        printk(KERN_ERR "%s: Request send failed\n",
                                        lo->disk->disk_name);
-                       if (nbd_find_request(lo, (char *)&req) != NULL) {
-                               /* we still own req */
-                               req->errors++;
-                               nbd_end_request(req);
-                       } else /* we're racing with nbd_clear_que */
-                               printk(KERN_DEBUG "nbd: can't find req\n");
+                       req->errors++;
+                       nbd_end_request(req);
+               } else {
+                       spin_lock(&lo->queue_lock);
+                       list_add(&req->queuelist, &lo->queue_head);
+                       spin_unlock(&lo->queue_lock);
                }
 
+               lo->active_req = NULL;
+               mutex_unlock(&lo->tx_lock);
+               wake_up_all(&lo->active_wq);
+
                spin_lock_irq(q->queue_lock);
                continue;
 
@@ -526,20 +502,13 @@ static int nbd_ioctl(struct inode *inode, struct file *file,
  
        case NBD_CLEAR_SOCK:
                error = 0;
-               down(&lo->tx_lock);
+               mutex_lock(&lo->tx_lock);
                lo->sock = NULL;
-               up(&lo->tx_lock);
-               spin_lock(&lo->queue_lock);
+               mutex_unlock(&lo->tx_lock);
                file = lo->file;
                lo->file = NULL;
-               spin_unlock(&lo->queue_lock);
                nbd_clear_que(lo);
-               spin_lock(&lo->queue_lock);
-               if (!list_empty(&lo->queue_head)) {
-                       printk(KERN_ERR "nbd: disconnect: some requests are in progress -> please try again.\n");
-                       error = -EBUSY;
-               }
-               spin_unlock(&lo->queue_lock);
+               BUG_ON(!list_empty(&lo->queue_head));
                if (file)
                        fput(file);
                return error;
@@ -589,7 +558,7 @@ static int nbd_ioctl(struct inode *inode, struct file *file,
                 * FIXME: This code is duplicated from sys_shutdown, but
                 * there should be a more generic interface rather than
                 * calling socket ops directly here */
-               down(&lo->tx_lock);
+               mutex_lock(&lo->tx_lock);
                if (lo->sock) {
                        printk(KERN_WARNING "%s: shutting down socket\n",
                                lo->disk->disk_name);
@@ -597,25 +566,20 @@ static int nbd_ioctl(struct inode *inode, struct file *file,
                                SEND_SHUTDOWN|RCV_SHUTDOWN);
                        lo->sock = NULL;
                }
-               up(&lo->tx_lock);
-               spin_lock(&lo->queue_lock);
+               mutex_unlock(&lo->tx_lock);
                file = lo->file;
                lo->file = NULL;
-               spin_unlock(&lo->queue_lock);
                nbd_clear_que(lo);
                printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
                if (file)
                        fput(file);
                return lo->harderror;
        case NBD_CLEAR_QUE:
-               down(&lo->tx_lock);
-               if (lo->sock) {
-                       up(&lo->tx_lock);
-                       return 0; /* probably should be error, but that would
-                                  * break "nbd-client -d", so just return 0 */
-               }
-               up(&lo->tx_lock);
-               nbd_clear_que(lo);
+               /*
+                * This is for compatibility only.  The queue is always cleared
+                * by NBD_DO_IT or NBD_CLEAR_SOCK.
+                */
+               BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
                return 0;
        case NBD_PRINT_DEBUG:
                printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
@@ -643,10 +607,7 @@ static int __init nbd_init(void)
        int err = -ENOMEM;
        int i;
 
-       if (sizeof(struct nbd_request) != 28) {
-               printk(KERN_CRIT "nbd: sizeof nbd_request needs to be 28 in order to work!\n" );
-               return -EIO;
-       }
+       BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 
        if (nbds_max > MAX_NBD) {
                printk(KERN_CRIT "nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD,
@@ -679,7 +640,6 @@ static int __init nbd_init(void)
        printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
        dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 
-       devfs_mk_dir("nbd");
        for (i = 0; i < nbds_max; i++) {
                struct gendisk *disk = nbd_dev[i].disk;
                nbd_dev[i].file = NULL;
@@ -687,7 +647,8 @@ static int __init nbd_init(void)
                nbd_dev[i].flags = 0;
                spin_lock_init(&nbd_dev[i].queue_lock);
                INIT_LIST_HEAD(&nbd_dev[i].queue_head);
-               init_MUTEX(&nbd_dev[i].tx_lock);
+               mutex_init(&nbd_dev[i].tx_lock);
+               init_waitqueue_head(&nbd_dev[i].active_wq);
                nbd_dev[i].blksize = 1024;
                nbd_dev[i].bytesize = 0x7ffffc00ULL << 10; /* 2TB */
                disk->major = NBD_MAJOR;
@@ -722,7 +683,6 @@ static void __exit nbd_cleanup(void)
                        put_disk(disk);
                }
        }
-       devfs_remove("nbd");
        unregister_blkdev(NBD_MAJOR, "nbd");
        printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 }