fuse: fix thinko in max I/O size calucation
[safe/jmp/linux-2.6] / fs / fuse / inode.c
index d359d8d..3141690 100644 (file)
@@ -1,6 +1,6 @@
 /*
   FUSE: Filesystem in Userspace
-  Copyright (C) 2001-2005  Miklos Szeredi <miklos@szeredi.hu>
+  Copyright (C) 2001-2006  Miklos Szeredi <miklos@szeredi.hu>
 
   This program can be distributed under the terms of the GNU GPL.
   See the file COPYING.
 #include <linux/pagemap.h>
 #include <linux/slab.h>
 #include <linux/file.h>
-#include <linux/mount.h>
 #include <linux/seq_file.h>
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/parser.h>
 #include <linux/statfs.h>
+#include <linux/random.h>
+#include <linux/sched.h>
 
 MODULE_AUTHOR("Miklos Szeredi <miklos@szeredi.hu>");
 MODULE_DESCRIPTION("Filesystem in Userspace");
 MODULE_LICENSE("GPL");
 
-spinlock_t fuse_lock;
-static kmem_cache_t *fuse_inode_cachep;
-static struct subsystem connections_subsys;
-
-struct fuse_conn_attr {
-       struct attribute attr;
-       ssize_t (*show)(struct fuse_conn *, char *);
-       ssize_t (*store)(struct fuse_conn *, const char *, size_t);
-};
+static struct kmem_cache *fuse_inode_cachep;
+struct list_head fuse_conn_list;
+DEFINE_MUTEX(fuse_mutex);
 
 #define FUSE_SUPER_MAGIC 0x65735546
 
+#define FUSE_DEFAULT_BLKSIZE 512
+
 struct fuse_mount_data {
        int fd;
        unsigned rootmode;
@@ -45,6 +42,7 @@ struct fuse_mount_data {
        unsigned group_id_present : 1;
        unsigned flags;
        unsigned max_read;
+       unsigned blksize;
 };
 
 static struct inode *fuse_alloc_inode(struct super_block *sb)
@@ -52,14 +50,20 @@ static struct inode *fuse_alloc_inode(struct super_block *sb)
        struct inode *inode;
        struct fuse_inode *fi;
 
-       inode = kmem_cache_alloc(fuse_inode_cachep, SLAB_KERNEL);
+       inode = kmem_cache_alloc(fuse_inode_cachep, GFP_KERNEL);
        if (!inode)
                return NULL;
 
        fi = get_fuse_inode(inode);
-       fi->i_time = jiffies - 1;
+       fi->i_time = 0;
        fi->nodeid = 0;
        fi->nlookup = 0;
+       fi->attr_version = 0;
+       fi->writectr = 0;
+       INIT_LIST_HEAD(&fi->write_files);
+       INIT_LIST_HEAD(&fi->queued_writes);
+       INIT_LIST_HEAD(&fi->writepages);
+       init_waitqueue_head(&fi->page_waitq);
        fi->forget_req = fuse_request_alloc();
        if (!fi->forget_req) {
                kmem_cache_free(fuse_inode_cachep, inode);
@@ -72,18 +76,15 @@ static struct inode *fuse_alloc_inode(struct super_block *sb)
 static void fuse_destroy_inode(struct inode *inode)
 {
        struct fuse_inode *fi = get_fuse_inode(inode);
+       BUG_ON(!list_empty(&fi->write_files));
+       BUG_ON(!list_empty(&fi->queued_writes));
        if (fi->forget_req)
                fuse_request_free(fi->forget_req);
        kmem_cache_free(fuse_inode_cachep, inode);
 }
 
-static void fuse_read_inode(struct inode *inode)
-{
-       /* No op */
-}
-
 void fuse_send_forget(struct fuse_conn *fc, struct fuse_req *req,
-                     unsigned long nodeid, u64 nlookup)
+                     u64 nodeid, u64 nlookup)
 {
        struct fuse_forget_in *inarg = &req->misc.forget_in;
        inarg->nlookup = nlookup;
@@ -105,18 +106,36 @@ static void fuse_clear_inode(struct inode *inode)
        }
 }
 
-void fuse_change_attributes(struct inode *inode, struct fuse_attr *attr)
+static int fuse_remount_fs(struct super_block *sb, int *flags, char *data)
+{
+       if (*flags & MS_MANDLOCK)
+               return -EINVAL;
+
+       return 0;
+}
+
+void fuse_truncate(struct address_space *mapping, loff_t offset)
 {
-       if (S_ISREG(inode->i_mode) && i_size_read(inode) != attr->size)
-               invalidate_inode_pages(inode->i_mapping);
+       /* See vmtruncate() */
+       unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
+       truncate_inode_pages(mapping, offset);
+       unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
+}
+
+void fuse_change_attributes_common(struct inode *inode, struct fuse_attr *attr,
+                                  u64 attr_valid)
+{
+       struct fuse_conn *fc = get_fuse_conn(inode);
+       struct fuse_inode *fi = get_fuse_inode(inode);
+
+       fi->attr_version = ++fc->attr_version;
+       fi->i_time = attr_valid;
 
        inode->i_ino     = attr->ino;
-       inode->i_mode    = (inode->i_mode & S_IFMT) + (attr->mode & 07777);
+       inode->i_mode    = (inode->i_mode & S_IFMT) | (attr->mode & 07777);
        inode->i_nlink   = attr->nlink;
        inode->i_uid     = attr->uid;
        inode->i_gid     = attr->gid;
-       i_size_write(inode, attr->size);
-       inode->i_blksize = PAGE_CACHE_SIZE;
        inode->i_blocks  = attr->blocks;
        inode->i_atime.tv_sec   = attr->atime;
        inode->i_atime.tv_nsec  = attr->atimensec;
@@ -124,12 +143,52 @@ void fuse_change_attributes(struct inode *inode, struct fuse_attr *attr)
        inode->i_mtime.tv_nsec  = attr->mtimensec;
        inode->i_ctime.tv_sec   = attr->ctime;
        inode->i_ctime.tv_nsec  = attr->ctimensec;
+
+       if (attr->blksize != 0)
+               inode->i_blkbits = ilog2(attr->blksize);
+       else
+               inode->i_blkbits = inode->i_sb->s_blocksize_bits;
+
+       /*
+        * Don't set the sticky bit in i_mode, unless we want the VFS
+        * to check permissions.  This prevents failures due to the
+        * check in may_delete().
+        */
+       fi->orig_i_mode = inode->i_mode;
+       if (!(fc->flags & FUSE_DEFAULT_PERMISSIONS))
+               inode->i_mode &= ~S_ISVTX;
+}
+
+void fuse_change_attributes(struct inode *inode, struct fuse_attr *attr,
+                           u64 attr_valid, u64 attr_version)
+{
+       struct fuse_conn *fc = get_fuse_conn(inode);
+       struct fuse_inode *fi = get_fuse_inode(inode);
+       loff_t oldsize;
+
+       spin_lock(&fc->lock);
+       if (attr_version != 0 && fi->attr_version > attr_version) {
+               spin_unlock(&fc->lock);
+               return;
+       }
+
+       fuse_change_attributes_common(inode, attr, attr_valid);
+
+       oldsize = inode->i_size;
+       i_size_write(inode, attr->size);
+       spin_unlock(&fc->lock);
+
+       if (S_ISREG(inode->i_mode) && oldsize != attr->size) {
+               if (attr->size < oldsize)
+                       fuse_truncate(inode->i_mapping, attr->size);
+               invalidate_inode_pages2(inode->i_mapping);
+       }
 }
 
 static void fuse_init_inode(struct inode *inode, struct fuse_attr *attr)
 {
        inode->i_mode = attr->mode & S_IFMT;
-       i_size_write(inode, attr->size);
+       inode->i_size = attr->size;
        if (S_ISREG(inode->i_mode)) {
                fuse_init_common(inode);
                fuse_init_file_inode(inode);
@@ -148,7 +207,7 @@ static void fuse_init_inode(struct inode *inode, struct fuse_attr *attr)
 
 static int fuse_inode_eq(struct inode *inode, void *_nodeidp)
 {
-       unsigned long nodeid = *(unsigned long *) _nodeidp;
+       u64 nodeid = *(u64 *) _nodeidp;
        if (get_node_id(inode) == nodeid)
                return 1;
        else
@@ -157,18 +216,18 @@ static int fuse_inode_eq(struct inode *inode, void *_nodeidp)
 
 static int fuse_inode_set(struct inode *inode, void *_nodeidp)
 {
-       unsigned long nodeid = *(unsigned long *) _nodeidp;
+       u64 nodeid = *(u64 *) _nodeidp;
        get_fuse_inode(inode)->nodeid = nodeid;
        return 0;
 }
 
-struct inode *fuse_iget(struct super_block *sb, unsigned long nodeid,
-                       int generation, struct fuse_attr *attr)
+struct inode *fuse_iget(struct super_block *sb, u64 nodeid,
+                       int generation, struct fuse_attr *attr,
+                       u64 attr_valid, u64 attr_version)
 {
        struct inode *inode;
        struct fuse_inode *fi;
        struct fuse_conn *fc = get_fuse_conn_super(sb);
-       int retried = 0;
 
  retry:
        inode = iget5_locked(sb, nodeid, fuse_inode_eq, fuse_inode_set, &nodeid);
@@ -182,17 +241,18 @@ struct inode *fuse_iget(struct super_block *sb, unsigned long nodeid,
                fuse_init_inode(inode, attr);
                unlock_new_inode(inode);
        } else if ((inode->i_mode ^ attr->mode) & S_IFMT) {
-               BUG_ON(retried);
                /* Inode has changed type, any I/O on the old should fail */
                make_bad_inode(inode);
                iput(inode);
-               retried = 1;
                goto retry;
        }
 
        fi = get_fuse_inode(inode);
+       spin_lock(&fc->lock);
        fi->nlookup ++;
-       fuse_change_attributes(inode, attr);
+       spin_unlock(&fc->lock);
+       fuse_change_attributes(inode, attr, attr_valid, attr_version);
+
        return inode;
 }
 
@@ -201,24 +261,37 @@ static void fuse_umount_begin(struct super_block *sb)
        fuse_abort_conn(get_fuse_conn_super(sb));
 }
 
+static void fuse_send_destroy(struct fuse_conn *fc)
+{
+       struct fuse_req *req = fc->destroy_req;
+       if (req && fc->conn_init) {
+               fc->destroy_req = NULL;
+               req->in.h.opcode = FUSE_DESTROY;
+               req->force = 1;
+               request_send(fc, req);
+               fuse_put_request(fc, req);
+       }
+}
+
 static void fuse_put_super(struct super_block *sb)
 {
        struct fuse_conn *fc = get_fuse_conn_super(sb);
 
-       down_write(&fc->sbput_sem);
-       while (!list_empty(&fc->background))
-               fuse_release_background(list_entry(fc->background.next,
-                                                  struct fuse_req, bg_entry));
-
-       spin_lock(&fuse_lock);
-       fc->mounted = 0;
+       fuse_send_destroy(fc);
+       spin_lock(&fc->lock);
        fc->connected = 0;
-       spin_unlock(&fuse_lock);
-       up_write(&fc->sbput_sem);
+       fc->blocked = 0;
+       spin_unlock(&fc->lock);
        /* Flush all readers on this fs */
+       kill_fasync(&fc->fasync, SIGIO, POLL_IN);
        wake_up_all(&fc->waitq);
-       kobject_del(&fc->kobj);
-       kobject_put(&fc->kobj);
+       wake_up_all(&fc->blocked_waitq);
+       wake_up_all(&fc->reserved_req_waitq);
+       mutex_lock(&fuse_mutex);
+       list_del(&fc->entry);
+       fuse_ctl_remove_conn(fc);
+       mutex_unlock(&fuse_mutex);
+       fuse_conn_put(fc);
 }
 
 static void convert_fuse_statfs(struct kstatfs *stbuf, struct fuse_kstatfs *attr)
@@ -235,20 +308,27 @@ static void convert_fuse_statfs(struct kstatfs *stbuf, struct fuse_kstatfs *attr
        /* fsid is left zero */
 }
 
-static int fuse_statfs(struct super_block *sb, struct kstatfs *buf)
+static int fuse_statfs(struct dentry *dentry, struct kstatfs *buf)
 {
+       struct super_block *sb = dentry->d_sb;
        struct fuse_conn *fc = get_fuse_conn_super(sb);
        struct fuse_req *req;
        struct fuse_statfs_out outarg;
        int err;
 
-        req = fuse_get_request(fc);
-       if (!req)
-               return -EINTR;
+       if (!fuse_allow_task(fc, current)) {
+               buf->f_type = FUSE_SUPER_MAGIC;
+               return 0;
+       }
+
+       req = fuse_get_req(fc);
+       if (IS_ERR(req))
+               return PTR_ERR(req);
 
        memset(&outarg, 0, sizeof(outarg));
        req->in.numargs = 0;
        req->in.h.opcode = FUSE_STATFS;
+       req->in.h.nodeid = get_node_id(dentry->d_inode);
        req->out.numargs = 1;
        req->out.args[0].size =
                fc->minor < 4 ? FUSE_COMPAT_STATFS_SIZE : sizeof(outarg);
@@ -269,6 +349,7 @@ enum {
        OPT_DEFAULT_PERMISSIONS,
        OPT_ALLOW_OTHER,
        OPT_MAX_READ,
+       OPT_BLKSIZE,
        OPT_ERR
 };
 
@@ -280,14 +361,16 @@ static match_table_t tokens = {
        {OPT_DEFAULT_PERMISSIONS,       "default_permissions"},
        {OPT_ALLOW_OTHER,               "allow_other"},
        {OPT_MAX_READ,                  "max_read=%u"},
+       {OPT_BLKSIZE,                   "blksize=%u"},
        {OPT_ERR,                       NULL}
 };
 
-static int parse_fuse_opt(char *opt, struct fuse_mount_data *d)
+static int parse_fuse_opt(char *opt, struct fuse_mount_data *d, int is_bdev)
 {
        char *p;
        memset(d, 0, sizeof(struct fuse_mount_data));
        d->max_read = ~0;
+       d->blksize = FUSE_DEFAULT_BLKSIZE;
 
        while ((p = strsep(&opt, ",")) != NULL) {
                int token;
@@ -308,6 +391,8 @@ static int parse_fuse_opt(char *opt, struct fuse_mount_data *d)
                case OPT_ROOTMODE:
                        if (match_octal(&args[0], &value))
                                return 0;
+                       if (!fuse_valid_type(value))
+                               return 0;
                        d->rootmode = value;
                        d->rootmode_present = 1;
                        break;
@@ -340,6 +425,12 @@ static int parse_fuse_opt(char *opt, struct fuse_mount_data *d)
                        d->max_read = value;
                        break;
 
+               case OPT_BLKSIZE:
+                       if (!is_bdev || match_int(&args[0], &value))
+                               return 0;
+                       d->blksize = value;
+                       break;
+
                default:
                        return 0;
                }
@@ -364,83 +455,90 @@ static int fuse_show_options(struct seq_file *m, struct vfsmount *mnt)
                seq_puts(m, ",allow_other");
        if (fc->max_read != ~0)
                seq_printf(m, ",max_read=%u", fc->max_read);
+       if (mnt->mnt_sb->s_bdev &&
+           mnt->mnt_sb->s_blocksize != FUSE_DEFAULT_BLKSIZE)
+               seq_printf(m, ",blksize=%lu", mnt->mnt_sb->s_blocksize);
        return 0;
 }
 
-static void fuse_conn_release(struct kobject *kobj)
-{
-       struct fuse_conn *fc = get_fuse_conn_kobj(kobj);
-
-       while (!list_empty(&fc->unused_list)) {
-               struct fuse_req *req;
-               req = list_entry(fc->unused_list.next, struct fuse_req, list);
-               list_del(&req->list);
-               fuse_request_free(req);
-       }
-       kfree(fc);
-}
-
-static struct fuse_conn *new_conn(void)
+static struct fuse_conn *new_conn(struct super_block *sb)
 {
        struct fuse_conn *fc;
+       int err;
 
        fc = kzalloc(sizeof(*fc), GFP_KERNEL);
        if (fc) {
-               int i;
+               spin_lock_init(&fc->lock);
+               mutex_init(&fc->inst_mutex);
+               atomic_set(&fc->count, 1);
                init_waitqueue_head(&fc->waitq);
+               init_waitqueue_head(&fc->blocked_waitq);
+               init_waitqueue_head(&fc->reserved_req_waitq);
                INIT_LIST_HEAD(&fc->pending);
                INIT_LIST_HEAD(&fc->processing);
                INIT_LIST_HEAD(&fc->io);
-               INIT_LIST_HEAD(&fc->unused_list);
-               INIT_LIST_HEAD(&fc->background);
-               sema_init(&fc->outstanding_sem, 1); /* One for INIT */
-               init_rwsem(&fc->sbput_sem);
-               kobj_set_kset_s(fc, connections_subsys);
-               kobject_init(&fc->kobj);
-               for (i = 0; i < FUSE_MAX_OUTSTANDING; i++) {
-                       struct fuse_req *req = fuse_request_alloc();
-                       if (!req) {
-                               kobject_put(&fc->kobj);
-                               return NULL;
-                       }
-                       list_add(&req->list, &fc->unused_list);
-               }
+               INIT_LIST_HEAD(&fc->interrupts);
+               INIT_LIST_HEAD(&fc->bg_queue);
+               atomic_set(&fc->num_waiting, 0);
                fc->bdi.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
                fc->bdi.unplug_io_fn = default_unplug_io_fn;
+               /* fuse does it's own writeback accounting */
+               fc->bdi.capabilities = BDI_CAP_NO_ACCT_WB;
+               fc->dev = sb->s_dev;
+               err = bdi_init(&fc->bdi);
+               if (err)
+                       goto error_kfree;
+               if (sb->s_bdev) {
+                       err = bdi_register(&fc->bdi, NULL, "%u:%u-fuseblk",
+                                          MAJOR(fc->dev), MINOR(fc->dev));
+               } else {
+                       err = bdi_register_dev(&fc->bdi, fc->dev);
+               }
+               if (err)
+                       goto error_bdi_destroy;
+               /*
+                * For a single fuse filesystem use max 1% of dirty +
+                * writeback threshold.
+                *
+                * This gives about 1M of write buffer for memory maps on a
+                * machine with 1G and 10% dirty_ratio, which should be more
+                * than enough.
+                *
+                * Privileged users can raise it by writing to
+                *
+                *    /sys/class/bdi/<bdi>/max_ratio
+                */
+               bdi_set_max_ratio(&fc->bdi, 1);
                fc->reqctr = 0;
+               fc->blocked = 1;
+               fc->attr_version = 1;
+               get_random_bytes(&fc->scramble_key, sizeof(fc->scramble_key));
        }
        return fc;
+
+error_bdi_destroy:
+       bdi_destroy(&fc->bdi);
+error_kfree:
+       mutex_destroy(&fc->inst_mutex);
+       kfree(fc);
+       return NULL;
 }
 
-static struct fuse_conn *get_conn(struct file *file, struct super_block *sb)
+void fuse_conn_put(struct fuse_conn *fc)
 {
-       struct fuse_conn *fc;
-       int err;
-
-       err = -EINVAL;
-       if (file->f_op != &fuse_dev_operations)
-               goto out_err;
-
-       err = -ENOMEM;
-       fc = new_conn();
-       if (!fc)
-               goto out_err;
-
-       spin_lock(&fuse_lock);
-       err = -EINVAL;
-       if (file->private_data)
-               goto out_unlock;
+       if (atomic_dec_and_test(&fc->count)) {
+               if (fc->destroy_req)
+                       fuse_request_free(fc->destroy_req);
+               mutex_destroy(&fc->inst_mutex);
+               bdi_destroy(&fc->bdi);
+               kfree(fc);
+       }
+}
 
-       kobject_get(&fc->kobj);
-       file->private_data = fc;
-       spin_unlock(&fuse_lock);
+struct fuse_conn *fuse_conn_get(struct fuse_conn *fc)
+{
+       atomic_inc(&fc->count);
        return fc;
-
- out_unlock:
-       spin_unlock(&fuse_lock);
-       kobject_put(&fc->kobj);
- out_err:
-       return ERR_PTR(err);
 }
 
 static struct inode *get_root_inode(struct super_block *sb, unsigned mode)
@@ -450,28 +548,79 @@ static struct inode *get_root_inode(struct super_block *sb, unsigned mode)
 
        attr.mode = mode;
        attr.ino = FUSE_ROOT_ID;
-       return fuse_iget(sb, 1, 0, &attr);
+       attr.nlink = 1;
+       return fuse_iget(sb, 1, 0, &attr, 0, 0);
 }
 
-static struct super_operations fuse_super_operations = {
+static const struct super_operations fuse_super_operations = {
        .alloc_inode    = fuse_alloc_inode,
        .destroy_inode  = fuse_destroy_inode,
-       .read_inode     = fuse_read_inode,
        .clear_inode    = fuse_clear_inode,
+       .drop_inode     = generic_delete_inode,
+       .remount_fs     = fuse_remount_fs,
        .put_super      = fuse_put_super,
        .umount_begin   = fuse_umount_begin,
        .statfs         = fuse_statfs,
        .show_options   = fuse_show_options,
 };
 
-static unsigned long long conn_id(void)
+static void process_init_reply(struct fuse_conn *fc, struct fuse_req *req)
+{
+       struct fuse_init_out *arg = &req->misc.init_out;
+
+       if (req->out.h.error || arg->major != FUSE_KERNEL_VERSION)
+               fc->conn_error = 1;
+       else {
+               unsigned long ra_pages;
+
+               if (arg->minor >= 6) {
+                       ra_pages = arg->max_readahead / PAGE_CACHE_SIZE;
+                       if (arg->flags & FUSE_ASYNC_READ)
+                               fc->async_read = 1;
+                       if (!(arg->flags & FUSE_POSIX_LOCKS))
+                               fc->no_lock = 1;
+                       if (arg->flags & FUSE_ATOMIC_O_TRUNC)
+                               fc->atomic_o_trunc = 1;
+                       if (arg->flags & FUSE_BIG_WRITES)
+                               fc->big_writes = 1;
+               } else {
+                       ra_pages = fc->max_read / PAGE_CACHE_SIZE;
+                       fc->no_lock = 1;
+               }
+
+               fc->bdi.ra_pages = min(fc->bdi.ra_pages, ra_pages);
+               fc->minor = arg->minor;
+               fc->max_write = arg->minor < 5 ? 4096 : arg->max_write;
+               fc->max_write = max_t(unsigned, 4096, fc->max_write);
+               fc->conn_init = 1;
+       }
+       fuse_put_request(fc, req);
+       fc->blocked = 0;
+       wake_up_all(&fc->blocked_waitq);
+}
+
+static void fuse_send_init(struct fuse_conn *fc, struct fuse_req *req)
 {
-       static unsigned long long ctr = 1;
-       unsigned long long val;
-       spin_lock(&fuse_lock);
-       val = ctr++;
-       spin_unlock(&fuse_lock);
-       return val;
+       struct fuse_init_in *arg = &req->misc.init_in;
+
+       arg->major = FUSE_KERNEL_VERSION;
+       arg->minor = FUSE_KERNEL_MINOR_VERSION;
+       arg->max_readahead = fc->bdi.ra_pages * PAGE_CACHE_SIZE;
+       arg->flags |= FUSE_ASYNC_READ | FUSE_POSIX_LOCKS | FUSE_ATOMIC_O_TRUNC |
+               FUSE_BIG_WRITES;
+       req->in.h.opcode = FUSE_INIT;
+       req->in.numargs = 1;
+       req->in.args[0].size = sizeof(*arg);
+       req->in.args[0].value = arg;
+       req->out.numargs = 1;
+       /* Variable length arguement used for backward compatibility
+          with interface version < 7.5.  Rest of init_out is zeroed
+          by do_get_request(), so a short reply is not a problem */
+       req->out.argvar = 1;
+       req->out.args[0].size = sizeof(struct fuse_init_out);
+       req->out.args[0].value = &req->misc.init_out;
+       req->end = process_init_reply;
+       request_send_background(fc, req);
 }
 
 static int fuse_fill_super(struct super_block *sb, void *data, int silent)
@@ -481,13 +630,25 @@ static int fuse_fill_super(struct super_block *sb, void *data, int silent)
        struct fuse_mount_data d;
        struct file *file;
        struct dentry *root_dentry;
+       struct fuse_req *init_req;
        int err;
+       int is_bdev = sb->s_bdev != NULL;
 
-       if (!parse_fuse_opt((char *) data, &d))
+       if (sb->s_flags & MS_MANDLOCK)
                return -EINVAL;
 
-       sb->s_blocksize = PAGE_CACHE_SIZE;
-       sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
+       if (!parse_fuse_opt((char *) data, &d, is_bdev))
+               return -EINVAL;
+
+       if (is_bdev) {
+#ifdef CONFIG_BLOCK
+               if (!sb_set_blocksize(sb, d.blksize))
+                       return -EINVAL;
+#endif
+       } else {
+               sb->s_blocksize = PAGE_CACHE_SIZE;
+               sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
+       }
        sb->s_magic = FUSE_SUPER_MAGIC;
        sb->s_op = &fuse_super_operations;
        sb->s_maxbytes = MAX_LFS_FILESIZE;
@@ -496,17 +657,17 @@ static int fuse_fill_super(struct super_block *sb, void *data, int silent)
        if (!file)
                return -EINVAL;
 
-       fc = get_conn(file, sb);
-       fput(file);
-       if (IS_ERR(fc))
-               return PTR_ERR(fc);
+       if (file->f_op != &fuse_dev_operations)
+               return -EINVAL;
+
+       fc = new_conn(sb);
+       if (!fc)
+               return -ENOMEM;
 
        fc->flags = d.flags;
        fc->user_id = d.user_id;
        fc->group_id = d.group_id;
-       fc->max_read = d.max_read;
-       if (fc->max_read / PAGE_CACHE_SIZE < fc->bdi.ra_pages)
-               fc->bdi.ra_pages = fc->max_read / PAGE_CACHE_SIZE;
+       fc->max_read = max_t(unsigned, 4096, d.max_read);
 
        /* Used by get_root_inode() */
        sb->s_fs_info = fc;
@@ -522,116 +683,109 @@ static int fuse_fill_super(struct super_block *sb, void *data, int silent)
                goto err;
        }
 
-       err = kobject_set_name(&fc->kobj, "%llu", conn_id());
-       if (err)
+       init_req = fuse_request_alloc();
+       if (!init_req)
                goto err_put_root;
 
-       err = kobject_add(&fc->kobj);
+       if (is_bdev) {
+               fc->destroy_req = fuse_request_alloc();
+               if (!fc->destroy_req)
+                       goto err_put_root;
+       }
+
+       mutex_lock(&fuse_mutex);
+       err = -EINVAL;
+       if (file->private_data)
+               goto err_unlock;
+
+       err = fuse_ctl_add_conn(fc);
        if (err)
-               goto err_put_root;
+               goto err_unlock;
 
+       list_add_tail(&fc->entry, &fuse_conn_list);
        sb->s_root = root_dentry;
-       spin_lock(&fuse_lock);
-       fc->mounted = 1;
        fc->connected = 1;
-       spin_unlock(&fuse_lock);
+       file->private_data = fuse_conn_get(fc);
+       mutex_unlock(&fuse_mutex);
+       /*
+        * atomic_dec_and_test() in fput() provides the necessary
+        * memory barrier for file->private_data to be visible on all
+        * CPUs after this
+        */
+       fput(file);
 
-       fuse_send_init(fc);
+       fuse_send_init(fc, init_req);
 
        return 0;
 
+ err_unlock:
+       mutex_unlock(&fuse_mutex);
+       fuse_request_free(init_req);
  err_put_root:
        dput(root_dentry);
  err:
-       kobject_put(&fc->kobj);
+       fput(file);
+       fuse_conn_put(fc);
        return err;
 }
 
-static struct super_block *fuse_get_sb(struct file_system_type *fs_type,
-                                      int flags, const char *dev_name,
-                                      void *raw_data)
+static int fuse_get_sb(struct file_system_type *fs_type,
+                      int flags, const char *dev_name,
+                      void *raw_data, struct vfsmount *mnt)
 {
-       return get_sb_nodev(fs_type, flags, raw_data, fuse_fill_super);
+       return get_sb_nodev(fs_type, flags, raw_data, fuse_fill_super, mnt);
 }
 
 static struct file_system_type fuse_fs_type = {
        .owner          = THIS_MODULE,
        .name           = "fuse",
+       .fs_flags       = FS_HAS_SUBTYPE,
        .get_sb         = fuse_get_sb,
        .kill_sb        = kill_anon_super,
 };
 
-static ssize_t fuse_conn_waiting_show(struct fuse_conn *fc, char *page)
-{
-       return sprintf(page, "%i\n", atomic_read(&fc->num_waiting));
-}
-
-static ssize_t fuse_conn_abort_store(struct fuse_conn *fc, const char *page,
-                                    size_t count)
+#ifdef CONFIG_BLOCK
+static int fuse_get_sb_blk(struct file_system_type *fs_type,
+                          int flags, const char *dev_name,
+                          void *raw_data, struct vfsmount *mnt)
 {
-       fuse_abort_conn(fc);
-       return count;
+       return get_sb_bdev(fs_type, flags, dev_name, raw_data, fuse_fill_super,
+                          mnt);
 }
 
-static struct fuse_conn_attr fuse_conn_waiting =
-       __ATTR(waiting, 0400, fuse_conn_waiting_show, NULL);
-static struct fuse_conn_attr fuse_conn_abort =
-       __ATTR(abort, 0600, NULL, fuse_conn_abort_store);
-
-static struct attribute *fuse_conn_attrs[] = {
-       &fuse_conn_waiting.attr,
-       &fuse_conn_abort.attr,
-       NULL,
+static struct file_system_type fuseblk_fs_type = {
+       .owner          = THIS_MODULE,
+       .name           = "fuseblk",
+       .get_sb         = fuse_get_sb_blk,
+       .kill_sb        = kill_block_super,
+       .fs_flags       = FS_REQUIRES_DEV | FS_HAS_SUBTYPE,
 };
 
-static ssize_t fuse_conn_attr_show(struct kobject *kobj,
-                                  struct attribute *attr,
-                                  char *page)
+static inline int register_fuseblk(void)
 {
-       struct fuse_conn_attr *fca =
-               container_of(attr, struct fuse_conn_attr, attr);
-
-       if (fca->show)
-               return fca->show(get_fuse_conn_kobj(kobj), page);
-       else
-               return -EACCES;
+       return register_filesystem(&fuseblk_fs_type);
 }
 
-static ssize_t fuse_conn_attr_store(struct kobject *kobj,
-                                   struct attribute *attr,
-                                   const char *page, size_t count)
+static inline void unregister_fuseblk(void)
 {
-       struct fuse_conn_attr *fca =
-               container_of(attr, struct fuse_conn_attr, attr);
-
-       if (fca->store)
-               return fca->store(get_fuse_conn_kobj(kobj), page, count);
-       else
-               return -EACCES;
+       unregister_filesystem(&fuseblk_fs_type);
+}
+#else
+static inline int register_fuseblk(void)
+{
+       return 0;
 }
 
-static struct sysfs_ops fuse_conn_sysfs_ops = {
-       .show   = &fuse_conn_attr_show,
-       .store  = &fuse_conn_attr_store,
-};
-
-static struct kobj_type ktype_fuse_conn = {
-       .release        = fuse_conn_release,
-       .sysfs_ops      = &fuse_conn_sysfs_ops,
-       .default_attrs  = fuse_conn_attrs,
-};
-
-static decl_subsys(fuse, NULL, NULL);
-static decl_subsys(connections, &ktype_fuse_conn, NULL);
+static inline void unregister_fuseblk(void)
+{
+}
+#endif
 
-static void fuse_inode_init_once(void *foo, kmem_cache_t *cachep,
-                                unsigned long flags)
+static void fuse_inode_init_once(struct kmem_cache *cachep, void *foo)
 {
        struct inode * inode = foo;
 
-       if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
-           SLAB_CTOR_CONSTRUCTOR)
-               inode_init_once(inode);
+       inode_init_once(inode);
 }
 
 static int __init fuse_fs_init(void)
@@ -640,53 +794,68 @@ static int __init fuse_fs_init(void)
 
        err = register_filesystem(&fuse_fs_type);
        if (err)
-               printk("fuse: failed to register filesystem\n");
-       else {
-               fuse_inode_cachep = kmem_cache_create("fuse_inode",
-                                                     sizeof(struct fuse_inode),
-                                                     0, SLAB_HWCACHE_ALIGN,
-                                                     fuse_inode_init_once, NULL);
-               if (!fuse_inode_cachep) {
-                       unregister_filesystem(&fuse_fs_type);
-                       err = -ENOMEM;
-               }
-       }
+               goto out;
 
+       err = register_fuseblk();
+       if (err)
+               goto out_unreg;
+
+       fuse_inode_cachep = kmem_cache_create("fuse_inode",
+                                             sizeof(struct fuse_inode),
+                                             0, SLAB_HWCACHE_ALIGN,
+                                             fuse_inode_init_once);
+       err = -ENOMEM;
+       if (!fuse_inode_cachep)
+               goto out_unreg2;
+
+       return 0;
+
+ out_unreg2:
+       unregister_fuseblk();
+ out_unreg:
+       unregister_filesystem(&fuse_fs_type);
+ out:
        return err;
 }
 
 static void fuse_fs_cleanup(void)
 {
        unregister_filesystem(&fuse_fs_type);
+       unregister_fuseblk();
        kmem_cache_destroy(fuse_inode_cachep);
 }
 
+static struct kobject *fuse_kobj;
+static struct kobject *connections_kobj;
+
 static int fuse_sysfs_init(void)
 {
        int err;
 
-       kset_set_kset_s(&fuse_subsys, fs_subsys);
-       err = subsystem_register(&fuse_subsys);
-       if (err)
+       fuse_kobj = kobject_create_and_add("fuse", fs_kobj);
+       if (!fuse_kobj) {
+               err = -ENOMEM;
                goto out_err;
+       }
 
-       kset_set_kset_s(&connections_subsys, fuse_subsys);
-       err = subsystem_register(&connections_subsys);
-       if (err)
+       connections_kobj = kobject_create_and_add("connections", fuse_kobj);
+       if (!connections_kobj) {
+               err = -ENOMEM;
                goto out_fuse_unregister;
+       }
 
        return 0;
 
  out_fuse_unregister:
-       subsystem_unregister(&fuse_subsys);
+       kobject_put(fuse_kobj);
  out_err:
        return err;
 }
 
 static void fuse_sysfs_cleanup(void)
 {
-       subsystem_unregister(&connections_subsys);
-       subsystem_unregister(&fuse_subsys);
+       kobject_put(connections_kobj);
+       kobject_put(fuse_kobj);
 }
 
 static int __init fuse_init(void)
@@ -696,7 +865,7 @@ static int __init fuse_init(void)
        printk("fuse init (API version %i.%i)\n",
               FUSE_KERNEL_VERSION, FUSE_KERNEL_MINOR_VERSION);
 
-       spin_lock_init(&fuse_lock);
+       INIT_LIST_HEAD(&fuse_conn_list);
        res = fuse_fs_init();
        if (res)
                goto err;
@@ -709,8 +878,14 @@ static int __init fuse_init(void)
        if (res)
                goto err_dev_cleanup;
 
+       res = fuse_ctl_init();
+       if (res)
+               goto err_sysfs_cleanup;
+
        return 0;
 
+ err_sysfs_cleanup:
+       fuse_sysfs_cleanup();
  err_dev_cleanup:
        fuse_dev_cleanup();
  err_fs_cleanup:
@@ -723,6 +898,7 @@ static void __exit fuse_exit(void)
 {
        printk(KERN_DEBUG "fuse exit\n");
 
+       fuse_ctl_cleanup();
        fuse_sysfs_cleanup();
        fuse_fs_cleanup();
        fuse_dev_cleanup();