Fedora kernel-2.6.17-1.2142_FC4 patched with stable patch-2.6.17.4-vs2.0.2-rc26.diff
[linux-2.6.git] / drivers / block / nbd.c
index b98b70b..8bca490 100644 (file)
@@ -38,7 +38,7 @@
  * 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 verify_area and friends, would share yet another 
+ * why not: would need access_ok and friends, would share yet another
  *          structure with userland
  */
 
 #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>
@@ -78,6 +82,7 @@
 #define DBG_RX          0x0200
 #define DBG_TX          0x0400
 static unsigned int debugflags;
+static unsigned int nbds_max = 16;
 #endif /* NDEBUG */
 
 static struct nbd_device nbd_dev[MAX_NBD];
@@ -92,7 +97,7 @@ static struct nbd_device nbd_dev[MAX_NBD];
  *    a single lock.
  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  */
-static spinlock_t nbd_lock = SPIN_LOCK_UNLOCKED;
+static DEFINE_SPINLOCK(nbd_lock);
 
 #ifndef NDEBUG
 static const char *ioctl_cmd_to_ascii(int cmd)
@@ -128,26 +133,14 @@ static void nbd_end_request(struct request *req)
 {
        int uptodate = (req->errors == 0) ? 1 : 0;
        request_queue_t *q = req->q;
-       struct nbd_device *lo = req->rq_disk->private_data;
        unsigned long flags;
 
        dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
                        req, uptodate? "done": "failed");
 
-       spin_lock(&lo->queue_lock);
-       while (req->ref_count > 1) { /* still in send */
-               spin_unlock(&lo->queue_lock);
-               printk(KERN_DEBUG "%s: request %p still in use (%d), waiting\n",
-                   lo->disk->disk_name, req, req->ref_count);
-               set_current_state(TASK_UNINTERRUPTIBLE);
-               schedule_timeout(HZ); /* wait a second */
-               spin_lock(&lo->queue_lock);
-       }
-       spin_unlock(&lo->queue_lock);
-
        spin_lock_irqsave(q->queue_lock, flags);
        if (!end_that_request_first(req, uptodate, req->nr_sectors)) {
-               end_that_request_last(req);
+               end_that_request_last(req, uptodate);
        }
        spin_unlock_irqrestore(q->queue_lock, flags);
 }
@@ -181,7 +174,6 @@ 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)
@@ -228,7 +220,7 @@ static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec,
        return result;
 }
 
-void nbd_send_req(struct nbd_device *lo, struct request *req)
+static int nbd_send_req(struct nbd_device *lo, struct request *req)
 {
        int result, i, flags;
        struct nbd_request request;
@@ -241,14 +233,6 @@ void 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)),
@@ -287,12 +271,10 @@ void nbd_send_req(struct nbd_device *lo, struct request *req)
                        }
                }
        }
-       up(&lo->tx_lock);
-       return;
+       return 0;
 
 error_out:
-       up(&lo->tx_lock);
-       req->errors++;
+       return 1;
 }
 
 static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
@@ -300,9 +282,14 @@ 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);
@@ -313,7 +300,11 @@ 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)
@@ -327,7 +318,7 @@ static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
 }
 
 /* NULL returned = something went wrong, inform userspace */
-struct request *nbd_read_stat(struct nbd_device *lo)
+static struct request *nbd_read_stat(struct nbd_device *lo)
 {
        int result;
        struct nbd_reply reply;
@@ -342,7 +333,11 @@ 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;
@@ -389,7 +384,7 @@ harderror:
        return NULL;
 }
 
-void nbd_do_it(struct nbd_device *lo)
+static void nbd_do_it(struct nbd_device *lo)
 {
        struct request *req;
 
@@ -400,25 +395,30 @@ void nbd_do_it(struct nbd_device *lo)
        return;
 }
 
-void nbd_clear_que(struct nbd_device *lo)
+static void nbd_clear_que(struct nbd_device *lo)
 {
        struct request *req;
 
        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);
+       }
 }
 
 /*
@@ -446,11 +446,6 @@ 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;
@@ -464,39 +459,34 @@ 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);
-               req->ref_count++; /* make sure req does not get freed */
-               spin_unlock(&lo->queue_lock);
-
-               nbd_send_req(lo, req);
+               lo->active_req = req;
 
-               if (req->errors) {
+               if (nbd_send_req(lo, req) != 0) {
                        printk(KERN_ERR "%s: Request send failed\n",
                                        lo->disk->disk_name);
+                       req->errors++;
+                       nbd_end_request(req);
+               } else {
                        spin_lock(&lo->queue_lock);
-                       list_del_init(&req->queuelist);
-                       req->ref_count--;
+                       list_add(&req->queuelist, &lo->queue_head);
                        spin_unlock(&lo->queue_lock);
-                       nbd_end_request(req);
-                       spin_lock_irq(q->queue_lock);
-                       continue;
                }
 
-               spin_lock(&lo->queue_lock);
-               req->ref_count--;
-               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;
 
@@ -544,20 +534,13 @@ 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;
@@ -568,7 +551,7 @@ static int nbd_ioctl(struct inode *inode, struct file *file,
                file = fget(arg);
                if (file) {
                        inode = file->f_dentry->d_inode;
-                       if (inode->i_sock) {
+                       if (S_ISSOCK(inode->i_mode)) {
                                lo->file = file;
                                lo->sock = SOCKET_I(inode);
                                error = 0;
@@ -607,7 +590,7 @@ 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);
@@ -615,25 +598,20 @@ 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",
@@ -661,12 +639,15 @@ 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,
+                               nbds_max);
+               return -EINVAL;
        }
 
-       for (i = 0; i < MAX_NBD; i++) {
+       for (i = 0; i < nbds_max; i++) {
                struct gendisk *disk = alloc_disk(1);
                if (!disk)
                        goto out;
@@ -692,14 +673,15 @@ static int __init nbd_init(void)
        dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 
        devfs_mk_dir("nbd");
-       for (i = 0; i < MAX_NBD; i++) {
+       for (i = 0; i < nbds_max; i++) {
                struct gendisk *disk = nbd_dev[i].disk;
                nbd_dev[i].file = NULL;
                nbd_dev[i].magic = LO_MAGIC;
                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;
@@ -725,8 +707,9 @@ out:
 static void __exit nbd_cleanup(void)
 {
        int i;
-       for (i = 0; i < MAX_NBD; i++) {
+       for (i = 0; i < nbds_max; i++) {
                struct gendisk *disk = nbd_dev[i].disk;
+               nbd_dev[i].magic = 0;
                if (disk) {
                        del_gendisk(disk);
                        blk_cleanup_queue(disk->queue);
@@ -744,7 +727,9 @@ module_exit(nbd_cleanup);
 MODULE_DESCRIPTION("Network Block Device");
 MODULE_LICENSE("GPL");
 
+module_param(nbds_max, int, 0444);
+MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize.");
 #ifndef NDEBUG
-MODULE_PARM(debugflags, "i");
+module_param(debugflags, int, 0644);
 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 #endif