fuse: don't use inode in fuse_file_poll
[safe/jmp/linux-2.6] / fs / fuse / file.c
index f0f0f27..c5de60e 100644 (file)
@@ -1,6 +1,6 @@
 /*
   FUSE: Filesystem in Userspace
-  Copyright (C) 2001-2006  Miklos Szeredi <miklos@szeredi.hu>
+  Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
 
   This program can be distributed under the terms of the GNU GPL.
   See the file COPYING.
 
 static const struct file_operations fuse_direct_io_file_operations;
 
-static int fuse_send_open(struct inode *inode, struct file *file, int isdir,
-                         struct fuse_open_out *outargp)
+static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
+                         int opcode, struct fuse_open_out *outargp)
 {
-       struct fuse_conn *fc = get_fuse_conn(inode);
        struct fuse_open_in inarg;
        struct fuse_req *req;
        int err;
@@ -31,35 +30,45 @@ static int fuse_send_open(struct inode *inode, struct file *file, int isdir,
        inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
        if (!fc->atomic_o_trunc)
                inarg.flags &= ~O_TRUNC;
-       req->in.h.opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
-       req->in.h.nodeid = get_node_id(inode);
+       req->in.h.opcode = opcode;
+       req->in.h.nodeid = nodeid;
        req->in.numargs = 1;
        req->in.args[0].size = sizeof(inarg);
        req->in.args[0].value = &inarg;
        req->out.numargs = 1;
        req->out.args[0].size = sizeof(*outargp);
        req->out.args[0].value = outargp;
-       request_send(fc, req);
+       fuse_request_send(fc, req);
        err = req->out.h.error;
        fuse_put_request(fc, req);
 
        return err;
 }
 
-struct fuse_file *fuse_file_alloc(void)
+struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
 {
        struct fuse_file *ff;
+
        ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
-       if (ff) {
-               ff->reserved_req = fuse_request_alloc();
-               if (!ff->reserved_req) {
-                       kfree(ff);
-                       ff = NULL;
-               } else {
-                       INIT_LIST_HEAD(&ff->write_entry);
-                       atomic_set(&ff->count, 0);
-               }
+       if (unlikely(!ff))
+               return NULL;
+
+       ff->fc = fc;
+       ff->reserved_req = fuse_request_alloc();
+       if (unlikely(!ff->reserved_req)) {
+               kfree(ff);
+               return NULL;
        }
+
+       INIT_LIST_HEAD(&ff->write_entry);
+       atomic_set(&ff->count, 0);
+       RB_CLEAR_NODE(&ff->polled_node);
+       init_waitqueue_head(&ff->poll_wait);
+
+       spin_lock(&fc->lock);
+       ff->kh = ++fc->khctr;
+       spin_unlock(&fc->lock);
+
        return ff;
 }
 
@@ -69,7 +78,7 @@ void fuse_file_free(struct fuse_file *ff)
        kfree(ff);
 }
 
-static struct fuse_file *fuse_file_get(struct fuse_file *ff)
+struct fuse_file *fuse_file_get(struct fuse_file *ff)
 {
        atomic_inc(&ff->count);
        return ff;
@@ -77,38 +86,64 @@ static struct fuse_file *fuse_file_get(struct fuse_file *ff)
 
 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 {
-       dput(req->misc.release.dentry);
-       mntput(req->misc.release.vfsmount);
-       fuse_put_request(fc, req);
+       path_put(&req->misc.release.path);
 }
 
 static void fuse_file_put(struct fuse_file *ff)
 {
        if (atomic_dec_and_test(&ff->count)) {
                struct fuse_req *req = ff->reserved_req;
-               struct inode *inode = req->misc.release.dentry->d_inode;
-               struct fuse_conn *fc = get_fuse_conn(inode);
+
                req->end = fuse_release_end;
-               request_send_background(fc, req);
+               fuse_request_send_background(ff->fc, req);
                kfree(ff);
        }
 }
 
-void fuse_finish_open(struct inode *inode, struct file *file,
-                     struct fuse_file *ff, struct fuse_open_out *outarg)
+static int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
+                       bool isdir)
 {
-       if (outarg->open_flags & FOPEN_DIRECT_IO)
+       struct fuse_open_out outarg;
+       struct fuse_file *ff;
+       int err;
+       int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
+
+       ff = fuse_file_alloc(fc);
+       if (!ff)
+               return -ENOMEM;
+
+       err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
+       if (err) {
+               fuse_file_free(ff);
+               return err;
+       }
+
+       if (isdir)
+               outarg.open_flags &= ~FOPEN_DIRECT_IO;
+
+       ff->fh = outarg.fh;
+       ff->nodeid = nodeid;
+       ff->open_flags = outarg.open_flags;
+       file->private_data = fuse_file_get(ff);
+
+       return 0;
+}
+
+void fuse_finish_open(struct inode *inode, struct file *file)
+{
+       struct fuse_file *ff = file->private_data;
+
+       if (ff->open_flags & FOPEN_DIRECT_IO)
                file->f_op = &fuse_direct_io_file_operations;
-       if (!(outarg->open_flags & FOPEN_KEEP_CACHE))
+       if (!(ff->open_flags & FOPEN_KEEP_CACHE))
                invalidate_inode_pages2(inode->i_mapping);
-       ff->fh = outarg->fh;
-       file->private_data = fuse_file_get(ff);
+       if (ff->open_flags & FOPEN_NONSEEKABLE)
+               nonseekable_open(inode, file);
 }
 
-int fuse_open_common(struct inode *inode, struct file *file, int isdir)
+int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 {
-       struct fuse_open_out outarg;
-       struct fuse_file *ff;
+       struct fuse_conn *fc = get_fuse_conn(inode);
        int err;
 
        /* VFS checks this, but only _after_ ->open() */
@@ -119,73 +154,83 @@ int fuse_open_common(struct inode *inode, struct file *file, int isdir)
        if (err)
                return err;
 
-       ff = fuse_file_alloc();
-       if (!ff)
-               return -ENOMEM;
-
-       err = fuse_send_open(inode, file, isdir, &outarg);
+       err = fuse_do_open(fc, get_node_id(inode), file, isdir);
        if (err)
-               fuse_file_free(ff);
-       else {
-               if (isdir)
-                       outarg.open_flags &= ~FOPEN_DIRECT_IO;
-               fuse_finish_open(inode, file, ff, &outarg);
-       }
+               return err;
 
-       return err;
+       fuse_finish_open(inode, file);
+
+       return 0;
 }
 
-void fuse_release_fill(struct fuse_file *ff, u64 nodeid, int flags, int opcode)
+static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 {
+       struct fuse_conn *fc = ff->fc;
        struct fuse_req *req = ff->reserved_req;
        struct fuse_release_in *inarg = &req->misc.release.in;
 
+       spin_lock(&fc->lock);
+       list_del(&ff->write_entry);
+       if (!RB_EMPTY_NODE(&ff->polled_node))
+               rb_erase(&ff->polled_node, &fc->polled_files);
+       spin_unlock(&fc->lock);
+
+       wake_up_interruptible_sync(&ff->poll_wait);
+
        inarg->fh = ff->fh;
        inarg->flags = flags;
        req->in.h.opcode = opcode;
-       req->in.h.nodeid = nodeid;
+       req->in.h.nodeid = ff->nodeid;
        req->in.numargs = 1;
        req->in.args[0].size = sizeof(struct fuse_release_in);
        req->in.args[0].value = inarg;
 }
 
-int fuse_release_common(struct inode *inode, struct file *file, int isdir)
+void fuse_release_common(struct file *file, int opcode)
 {
-       struct fuse_file *ff = file->private_data;
-       if (ff) {
-               struct fuse_conn *fc = get_fuse_conn(inode);
-               struct fuse_req *req = ff->reserved_req;
+       struct fuse_file *ff;
+       struct fuse_req *req;
 
-               fuse_release_fill(ff, get_node_id(inode), file->f_flags,
-                                 isdir ? FUSE_RELEASEDIR : FUSE_RELEASE);
+       ff = file->private_data;
+       if (unlikely(!ff))
+               return;
 
-               /* Hold vfsmount and dentry until release is finished */
-               req->misc.release.vfsmount = mntget(file->f_path.mnt);
-               req->misc.release.dentry = dget(file->f_path.dentry);
+       req = ff->reserved_req;
+       fuse_prepare_release(ff, file->f_flags, opcode);
 
-               spin_lock(&fc->lock);
-               list_del(&ff->write_entry);
-               spin_unlock(&fc->lock);
-               /*
-                * Normally this will send the RELEASE request,
-                * however if some asynchronous READ or WRITE requests
-                * are outstanding, the sending will be delayed
-                */
-               fuse_file_put(ff);
-       }
+       /* Hold vfsmount and dentry until release is finished */
+       path_get(&file->f_path);
+       req->misc.release.path = file->f_path;
 
-       /* Return value is ignored by VFS */
-       return 0;
+       /*
+        * Normally this will send the RELEASE request, however if
+        * some asynchronous READ or WRITE requests are outstanding,
+        * the sending will be delayed.
+        */
+       fuse_file_put(ff);
 }
 
 static int fuse_open(struct inode *inode, struct file *file)
 {
-       return fuse_open_common(inode, file, 0);
+       return fuse_open_common(inode, file, false);
 }
 
 static int fuse_release(struct inode *inode, struct file *file)
 {
-       return fuse_release_common(inode, file, 0);
+       fuse_release_common(file, FUSE_RELEASE);
+
+       /* return value is ignored by VFS */
+       return 0;
+}
+
+void fuse_sync_release(struct fuse_file *ff, int flags)
+{
+       WARN_ON(atomic_read(&ff->count) > 1);
+       fuse_prepare_release(ff, flags, FUSE_RELEASE);
+       ff->reserved_req->force = 1;
+       fuse_request_send(ff->fc, ff->reserved_req);
+       fuse_put_request(ff->fc, ff->reserved_req);
+       kfree(ff);
 }
 
 /*
@@ -278,7 +323,7 @@ static int fuse_flush(struct file *file, fl_owner_t id)
        req->in.args[0].size = sizeof(inarg);
        req->in.args[0].value = &inarg;
        req->force = 1;
-       request_send(fc, req);
+       fuse_request_send(fc, req);
        err = req->out.h.error;
        fuse_put_request(fc, req);
        if (err == -ENOSYS) {
@@ -342,7 +387,7 @@ int fuse_fsync_common(struct file *file, struct dentry *de, int datasync,
        req->in.numargs = 1;
        req->in.args[0].size = sizeof(inarg);
        req->in.args[0].value = &inarg;
-       request_send(fc, req);
+       fuse_request_send(fc, req);
        err = req->out.h.error;
        fuse_put_request(fc, req);
        if (err == -ENOSYS) {
@@ -360,10 +405,10 @@ static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
        return fuse_fsync_common(file, de, datasync, 0);
 }
 
-void fuse_read_fill(struct fuse_req *req, struct file *file,
-                   struct inode *inode, loff_t pos, size_t count, int opcode)
+void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
+                   size_t count, int opcode)
 {
-       struct fuse_read_in *inarg = &req->misc.read_in;
+       struct fuse_read_in *inarg = &req->misc.read.in;
        struct fuse_file *ff = file->private_data;
 
        inarg->fh = ff->fh;
@@ -371,38 +416,55 @@ void fuse_read_fill(struct fuse_req *req, struct file *file,
        inarg->size = count;
        inarg->flags = file->f_flags;
        req->in.h.opcode = opcode;
-       req->in.h.nodeid = get_node_id(inode);
+       req->in.h.nodeid = ff->nodeid;
        req->in.numargs = 1;
        req->in.args[0].size = sizeof(struct fuse_read_in);
        req->in.args[0].value = inarg;
-       req->out.argpages = 1;
        req->out.argvar = 1;
        req->out.numargs = 1;
        req->out.args[0].size = count;
 }
 
 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
-                            struct inode *inode, loff_t pos, size_t count,
-                            fl_owner_t owner)
+                            loff_t pos, size_t count, fl_owner_t owner)
 {
-       struct fuse_conn *fc = get_fuse_conn(inode);
+       struct fuse_file *ff = file->private_data;
+       struct fuse_conn *fc = ff->fc;
 
-       fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
+       fuse_read_fill(req, file, pos, count, FUSE_READ);
        if (owner != NULL) {
-               struct fuse_read_in *inarg = &req->misc.read_in;
+               struct fuse_read_in *inarg = &req->misc.read.in;
 
                inarg->read_flags |= FUSE_READ_LOCKOWNER;
                inarg->lock_owner = fuse_lock_owner_id(fc, owner);
        }
-       request_send(fc, req);
+       fuse_request_send(fc, req);
        return req->out.args[0].size;
 }
 
+static void fuse_read_update_size(struct inode *inode, loff_t size,
+                                 u64 attr_ver)
+{
+       struct fuse_conn *fc = get_fuse_conn(inode);
+       struct fuse_inode *fi = get_fuse_inode(inode);
+
+       spin_lock(&fc->lock);
+       if (attr_ver == fi->attr_version && size < inode->i_size) {
+               fi->attr_version = ++fc->attr_version;
+               i_size_write(inode, size);
+       }
+       spin_unlock(&fc->lock);
+}
+
 static int fuse_readpage(struct file *file, struct page *page)
 {
        struct inode *inode = page->mapping->host;
        struct fuse_conn *fc = get_fuse_conn(inode);
        struct fuse_req *req;
+       size_t num_read;
+       loff_t pos = page_offset(page);
+       size_t count = PAGE_CACHE_SIZE;
+       u64 attr_ver;
        int err;
 
        err = -EIO;
@@ -421,15 +483,26 @@ static int fuse_readpage(struct file *file, struct page *page)
        if (IS_ERR(req))
                goto out;
 
+       attr_ver = fuse_get_attr_version(fc);
+
        req->out.page_zeroing = 1;
+       req->out.argpages = 1;
        req->num_pages = 1;
        req->pages[0] = page;
-       fuse_send_read(req, file, inode, page_offset(page), PAGE_CACHE_SIZE,
-                      NULL);
+       num_read = fuse_send_read(req, file, pos, count, NULL);
        err = req->out.h.error;
        fuse_put_request(fc, req);
-       if (!err)
+
+       if (!err) {
+               /*
+                * Short read means EOF.  If file size is larger, truncate it
+                */
+               if (num_read < count)
+                       fuse_read_update_size(inode, pos + num_read, attr_ver);
+
                SetPageUptodate(page);
+       }
+
        fuse_invalidate_attr(inode); /* atime changed */
  out:
        unlock_page(page);
@@ -439,8 +512,19 @@ static int fuse_readpage(struct file *file, struct page *page)
 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 {
        int i;
+       size_t count = req->misc.read.in.size;
+       size_t num_read = req->out.args[0].size;
+       struct inode *inode = req->pages[0]->mapping->host;
 
-       fuse_invalidate_attr(req->pages[0]->mapping->host); /* atime changed */
+       /*
+        * Short read means EOF.  If file size is larger, truncate it
+        */
+       if (!req->out.h.error && num_read < count) {
+               loff_t pos = page_offset(req->pages[0]) + num_read;
+               fuse_read_update_size(inode, pos, req->misc.read.attr_ver);
+       }
+
+       fuse_invalidate_attr(inode); /* atime changed */
 
        for (i = 0; i < req->num_pages; i++) {
                struct page *page = req->pages[i];
@@ -452,25 +536,27 @@ static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
        }
        if (req->ff)
                fuse_file_put(req->ff);
-       fuse_put_request(fc, req);
 }
 
-static void fuse_send_readpages(struct fuse_req *req, struct file *file,
-                               struct inode *inode)
+static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 {
-       struct fuse_conn *fc = get_fuse_conn(inode);
+       struct fuse_file *ff = file->private_data;
+       struct fuse_conn *fc = ff->fc;
        loff_t pos = page_offset(req->pages[0]);
        size_t count = req->num_pages << PAGE_CACHE_SHIFT;
+
+       req->out.argpages = 1;
        req->out.page_zeroing = 1;
-       fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
+       fuse_read_fill(req, file, pos, count, FUSE_READ);
+       req->misc.read.attr_ver = fuse_get_attr_version(fc);
        if (fc->async_read) {
-               struct fuse_file *ff = file->private_data;
                req->ff = fuse_file_get(ff);
                req->end = fuse_readpages_end;
-               request_send_background(fc, req);
+               fuse_request_send_background(fc, req);
        } else {
-               request_send(fc, req);
+               fuse_request_send(fc, req);
                fuse_readpages_end(fc, req);
+               fuse_put_request(fc, req);
        }
 }
 
@@ -493,7 +579,7 @@ static int fuse_readpages_fill(void *_data, struct page *page)
            (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
             (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
             req->pages[req->num_pages - 1]->index + 1 != page->index)) {
-               fuse_send_readpages(req, data->file, inode);
+               fuse_send_readpages(req, data->file);
                data->req = req = fuse_get_req(fc);
                if (IS_ERR(req)) {
                        unlock_page(page);
@@ -501,7 +587,7 @@ static int fuse_readpages_fill(void *_data, struct page *page)
                }
        }
        req->pages[req->num_pages] = page;
-       req->num_pages ++;
+       req->num_pages++;
        return 0;
 }
 
@@ -527,7 +613,7 @@ static int fuse_readpages(struct file *file, struct address_space *mapping,
        err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
        if (!err) {
                if (data.req->num_pages)
-                       fuse_send_readpages(data.req, file, inode);
+                       fuse_send_readpages(data.req, file);
                else
                        fuse_put_request(fc, data.req);
        }
@@ -554,25 +640,19 @@ static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
        return generic_file_aio_read(iocb, iov, nr_segs, pos);
 }
 
-static void fuse_write_fill(struct fuse_req *req, struct file *file,
-                           struct fuse_file *ff, struct inode *inode,
-                           loff_t pos, size_t count, int writepage)
+static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
+                           loff_t pos, size_t count)
 {
-       struct fuse_conn *fc = get_fuse_conn(inode);
        struct fuse_write_in *inarg = &req->misc.write.in;
        struct fuse_write_out *outarg = &req->misc.write.out;
 
-       memset(inarg, 0, sizeof(struct fuse_write_in));
        inarg->fh = ff->fh;
        inarg->offset = pos;
        inarg->size = count;
-       inarg->write_flags = writepage ? FUSE_WRITE_CACHE : 0;
-       inarg->flags = file ? file->f_flags : 0;
        req->in.h.opcode = FUSE_WRITE;
-       req->in.h.nodeid = get_node_id(inode);
-       req->in.argpages = 1;
+       req->in.h.nodeid = ff->nodeid;
        req->in.numargs = 2;
-       if (fc->minor < 9)
+       if (ff->fc->minor < 9)
                req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
        else
                req->in.args[0].size = sizeof(struct fuse_write_in);
@@ -584,17 +664,19 @@ static void fuse_write_fill(struct fuse_req *req, struct file *file,
 }
 
 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
-                             struct inode *inode, loff_t pos, size_t count,
-                             fl_owner_t owner)
+                             loff_t pos, size_t count, fl_owner_t owner)
 {
-       struct fuse_conn *fc = get_fuse_conn(inode);
-       fuse_write_fill(req, file, file->private_data, inode, pos, count, 0);
+       struct fuse_file *ff = file->private_data;
+       struct fuse_conn *fc = ff->fc;
+       struct fuse_write_in *inarg = &req->misc.write.in;
+
+       fuse_write_fill(req, ff, pos, count);
+       inarg->flags = file->f_flags;
        if (owner != NULL) {
-               struct fuse_write_in *inarg = &req->misc.write.in;
                inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
                inarg->lock_owner = fuse_lock_owner_id(fc, owner);
        }
-       request_send(fc, req);
+       fuse_request_send(fc, req);
        return req->misc.write.out.size;
 }
 
@@ -604,7 +686,7 @@ static int fuse_write_begin(struct file *file, struct address_space *mapping,
 {
        pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 
-       *pagep = __grab_cache_page(mapping, index);
+       *pagep = grab_cache_page_write_begin(mapping, index, flags);
        if (!*pagep)
                return -ENOMEM;
        return 0;
@@ -644,10 +726,11 @@ static int fuse_buffered_write(struct file *file, struct inode *inode,
        if (IS_ERR(req))
                return PTR_ERR(req);
 
+       req->in.argpages = 1;
        req->num_pages = 1;
        req->pages[0] = page;
        req->page_offset = offset;
-       nres = fuse_send_write(req, file, inode, pos, count, NULL);
+       nres = fuse_send_write(req, file, pos, count, NULL);
        err = req->out.h.error;
        fuse_put_request(fc, req);
        if (!err && !nres)
@@ -677,6 +760,201 @@ static int fuse_write_end(struct file *file, struct address_space *mapping,
        return res;
 }
 
+static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
+                                   struct inode *inode, loff_t pos,
+                                   size_t count)
+{
+       size_t res;
+       unsigned offset;
+       unsigned i;
+
+       for (i = 0; i < req->num_pages; i++)
+               fuse_wait_on_page_writeback(inode, req->pages[i]->index);
+
+       res = fuse_send_write(req, file, pos, count, NULL);
+
+       offset = req->page_offset;
+       count = res;
+       for (i = 0; i < req->num_pages; i++) {
+               struct page *page = req->pages[i];
+
+               if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
+                       SetPageUptodate(page);
+
+               if (count > PAGE_CACHE_SIZE - offset)
+                       count -= PAGE_CACHE_SIZE - offset;
+               else
+                       count = 0;
+               offset = 0;
+
+               unlock_page(page);
+               page_cache_release(page);
+       }
+
+       return res;
+}
+
+static ssize_t fuse_fill_write_pages(struct fuse_req *req,
+                              struct address_space *mapping,
+                              struct iov_iter *ii, loff_t pos)
+{
+       struct fuse_conn *fc = get_fuse_conn(mapping->host);
+       unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
+       size_t count = 0;
+       int err;
+
+       req->in.argpages = 1;
+       req->page_offset = offset;
+
+       do {
+               size_t tmp;
+               struct page *page;
+               pgoff_t index = pos >> PAGE_CACHE_SHIFT;
+               size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
+                                    iov_iter_count(ii));
+
+               bytes = min_t(size_t, bytes, fc->max_write - count);
+
+ again:
+               err = -EFAULT;
+               if (iov_iter_fault_in_readable(ii, bytes))
+                       break;
+
+               err = -ENOMEM;
+               page = grab_cache_page_write_begin(mapping, index, 0);
+               if (!page)
+                       break;
+
+               pagefault_disable();
+               tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
+               pagefault_enable();
+               flush_dcache_page(page);
+
+               if (!tmp) {
+                       unlock_page(page);
+                       page_cache_release(page);
+                       bytes = min(bytes, iov_iter_single_seg_count(ii));
+                       goto again;
+               }
+
+               err = 0;
+               req->pages[req->num_pages] = page;
+               req->num_pages++;
+
+               iov_iter_advance(ii, tmp);
+               count += tmp;
+               pos += tmp;
+               offset += tmp;
+               if (offset == PAGE_CACHE_SIZE)
+                       offset = 0;
+
+               if (!fc->big_writes)
+                       break;
+       } while (iov_iter_count(ii) && count < fc->max_write &&
+                req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
+
+       return count > 0 ? count : err;
+}
+
+static ssize_t fuse_perform_write(struct file *file,
+                                 struct address_space *mapping,
+                                 struct iov_iter *ii, loff_t pos)
+{
+       struct inode *inode = mapping->host;
+       struct fuse_conn *fc = get_fuse_conn(inode);
+       int err = 0;
+       ssize_t res = 0;
+
+       if (is_bad_inode(inode))
+               return -EIO;
+
+       do {
+               struct fuse_req *req;
+               ssize_t count;
+
+               req = fuse_get_req(fc);
+               if (IS_ERR(req)) {
+                       err = PTR_ERR(req);
+                       break;
+               }
+
+               count = fuse_fill_write_pages(req, mapping, ii, pos);
+               if (count <= 0) {
+                       err = count;
+               } else {
+                       size_t num_written;
+
+                       num_written = fuse_send_write_pages(req, file, inode,
+                                                           pos, count);
+                       err = req->out.h.error;
+                       if (!err) {
+                               res += num_written;
+                               pos += num_written;
+
+                               /* break out of the loop on short write */
+                               if (num_written != count)
+                                       err = -EIO;
+                       }
+               }
+               fuse_put_request(fc, req);
+       } while (!err && iov_iter_count(ii));
+
+       if (res > 0)
+               fuse_write_update_size(inode, pos);
+
+       fuse_invalidate_attr(inode);
+
+       return res > 0 ? res : err;
+}
+
+static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
+                                  unsigned long nr_segs, loff_t pos)
+{
+       struct file *file = iocb->ki_filp;
+       struct address_space *mapping = file->f_mapping;
+       size_t count = 0;
+       ssize_t written = 0;
+       struct inode *inode = mapping->host;
+       ssize_t err;
+       struct iov_iter i;
+
+       WARN_ON(iocb->ki_pos != pos);
+
+       err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
+       if (err)
+               return err;
+
+       mutex_lock(&inode->i_mutex);
+       vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
+
+       /* We can write back this queue in page reclaim */
+       current->backing_dev_info = mapping->backing_dev_info;
+
+       err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
+       if (err)
+               goto out;
+
+       if (count == 0)
+               goto out;
+
+       err = file_remove_suid(file);
+       if (err)
+               goto out;
+
+       file_update_time(file);
+
+       iov_iter_init(&i, iov, nr_segs, count, 0);
+       written = fuse_perform_write(file, mapping, &i, pos);
+       if (written >= 0)
+               iocb->ki_pos = pos + written;
+
+out:
+       current->backing_dev_info = NULL;
+       mutex_unlock(&inode->i_mutex);
+
+       return written ? written : err;
+}
+
 static void fuse_release_user_pages(struct fuse_req *req, int write)
 {
        unsigned i;
@@ -690,21 +968,28 @@ static void fuse_release_user_pages(struct fuse_req *req, int write)
 }
 
 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
-                              unsigned nbytes, int write)
+                              size_t *nbytesp, int write)
 {
+       size_t nbytes = *nbytesp;
        unsigned long user_addr = (unsigned long) buf;
        unsigned offset = user_addr & ~PAGE_MASK;
        int npages;
 
-       /* This doesn't work with nfsd */
-       if (!current->mm)
-               return -EPERM;
+       /* Special case for kernel I/O: can copy directly into the buffer */
+       if (segment_eq(get_fs(), KERNEL_DS)) {
+               if (write)
+                       req->in.args[1].value = (void *) user_addr;
+               else
+                       req->out.args[0].value = (void *) user_addr;
+
+               return 0;
+       }
 
-       nbytes = min(nbytes, (unsigned) FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
+       nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
        npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
-       npages = min(max(npages, 1), FUSE_MAX_PAGES_PER_REQ);
+       npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
        down_read(&current->mm->mmap_sem);
-       npages = get_user_pages(current, current->mm, user_addr, npages, write,
+       npages = get_user_pages(current, current->mm, user_addr, npages, !write,
                                0, req->pages, NULL);
        up_read(&current->mm->mmap_sem);
        if (npages < 0)
@@ -712,42 +997,47 @@ static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
 
        req->num_pages = npages;
        req->page_offset = offset;
+
+       if (write)
+               req->in.argpages = 1;
+       else
+               req->out.argpages = 1;
+
+       nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
+       *nbytesp = min(*nbytesp, nbytes);
+
        return 0;
 }
 
 static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
                              size_t count, loff_t *ppos, int write)
 {
-       struct inode *inode = file->f_path.dentry->d_inode;
-       struct fuse_conn *fc = get_fuse_conn(inode);
+       struct fuse_file *ff = file->private_data;
+       struct fuse_conn *fc = ff->fc;
        size_t nmax = write ? fc->max_write : fc->max_read;
        loff_t pos = *ppos;
        ssize_t res = 0;
        struct fuse_req *req;
 
-       if (is_bad_inode(inode))
-               return -EIO;
-
        req = fuse_get_req(fc);
        if (IS_ERR(req))
                return PTR_ERR(req);
 
        while (count) {
                size_t nres;
+               fl_owner_t owner = current->files;
                size_t nbytes = min(count, nmax);
-               int err = fuse_get_user_pages(req, buf, nbytes, !write);
+               int err = fuse_get_user_pages(req, buf, &nbytes, write);
                if (err) {
                        res = err;
                        break;
                }
-               nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
-               nbytes = min(count, nbytes);
+
                if (write)
-                       nres = fuse_send_write(req, file, inode, pos, nbytes,
-                                              current->files);
+                       nres = fuse_send_write(req, file, pos, nbytes, owner);
                else
-                       nres = fuse_send_read(req, file, inode, pos, nbytes,
-                                             current->files);
+                       nres = fuse_send_read(req, file, pos, nbytes, owner);
+
                fuse_release_user_pages(req, !write);
                if (req->out.h.error) {
                        if (!res)
@@ -771,12 +1061,8 @@ static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
                }
        }
        fuse_put_request(fc, req);
-       if (res > 0) {
-               if (write)
-                       fuse_write_update_size(inode, pos);
+       if (res > 0)
                *ppos = pos;
-       }
-       fuse_invalidate_attr(inode);
 
        return res;
 }
@@ -784,7 +1070,17 @@ static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
                                     size_t count, loff_t *ppos)
 {
-       return fuse_direct_io(file, buf, count, ppos, 0);
+       ssize_t res;
+       struct inode *inode = file->f_path.dentry->d_inode;
+
+       if (is_bad_inode(inode))
+               return -EIO;
+
+       res = fuse_direct_io(file, buf, count, ppos, 0);
+
+       fuse_invalidate_attr(inode);
+
+       return res;
 }
 
 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
@@ -792,12 +1088,22 @@ static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
 {
        struct inode *inode = file->f_path.dentry->d_inode;
        ssize_t res;
+
+       if (is_bad_inode(inode))
+               return -EIO;
+
        /* Don't allow parallel writes to the same file */
        mutex_lock(&inode->i_mutex);
        res = generic_write_checks(file, ppos, &count, 0);
-       if (!res)
+       if (!res) {
                res = fuse_direct_io(file, buf, count, ppos, 1);
+               if (res > 0)
+                       fuse_write_update_size(inode, *ppos);
+       }
        mutex_unlock(&inode->i_mutex);
+
+       fuse_invalidate_attr(inode);
+
        return res;
 }
 
@@ -805,7 +1111,6 @@ static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
 {
        __free_page(req->pages[0]);
        fuse_file_put(req->ff);
-       fuse_put_request(fc, req);
 }
 
 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
@@ -823,6 +1128,8 @@ static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
 
 /* Called under fc->lock, may release and reacquire it */
 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
+__releases(&fc->lock)
+__acquires(&fc->lock)
 {
        struct fuse_inode *fi = get_fuse_inode(req->inode);
        loff_t size = i_size_read(req->inode);
@@ -842,13 +1149,14 @@ static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
 
        req->in.args[1].size = inarg->size;
        fi->writectr++;
-       request_send_background_locked(fc, req);
+       fuse_request_send_background_locked(fc, req);
        return;
 
  out_free:
        fuse_writepage_finish(fc, req);
        spin_unlock(&fc->lock);
        fuse_writepage_free(fc, req);
+       fuse_put_request(fc, req);
        spin_lock(&fc->lock);
 }
 
@@ -859,6 +1167,8 @@ static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
  * Called with fc->lock
  */
 void fuse_flush_writepages(struct inode *inode)
+__releases(&fc->lock)
+__acquires(&fc->lock)
 {
        struct fuse_conn *fc = get_fuse_conn(inode);
        struct fuse_inode *fi = get_fuse_inode(inode);
@@ -910,9 +1220,11 @@ static int fuse_writepage_locked(struct page *page)
        req->ff = fuse_file_get(ff);
        spin_unlock(&fc->lock);
 
-       fuse_write_fill(req, NULL, ff, inode, page_offset(page), 0, 1);
+       fuse_write_fill(req, ff, page_offset(page), 0);
 
        copy_highpage(tmp_page, page);
+       req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
+       req->in.argpages = 1;
        req->num_pages = 1;
        req->pages[0] = tmp_page;
        req->page_offset = 0;
@@ -984,8 +1296,9 @@ static void fuse_vma_close(struct vm_area_struct *vma)
  * - sync(2)
  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
  */
-static int fuse_page_mkwrite(struct vm_area_struct *vma, struct page *page)
+static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 {
+       struct page *page = vmf->page;
        /*
         * Don't use page->mapping as it may become NULL from a
         * concurrent truncate.
@@ -1023,6 +1336,17 @@ static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
        return 0;
 }
 
+static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
+{
+       /* Can't provide the coherency needed for MAP_SHARED */
+       if (vma->vm_flags & VM_MAYSHARE)
+               return -ENODEV;
+
+       invalidate_inode_pages2(file->f_mapping);
+
+       return generic_file_mmap(file, vma);
+}
+
 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
                                  struct file_lock *fl)
 {
@@ -1088,7 +1412,7 @@ static int fuse_getlk(struct file *file, struct file_lock *fl)
        req->out.numargs = 1;
        req->out.args[0].size = sizeof(outarg);
        req->out.args[0].value = &outarg;
-       request_send(fc, req);
+       fuse_request_send(fc, req);
        err = req->out.h.error;
        fuse_put_request(fc, req);
        if (!err)
@@ -1106,6 +1430,11 @@ static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
        pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
        int err;
 
+       if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
+               /* NLM needs asynchronous locks, which we don't support yet */
+               return -ENOLCK;
+       }
+
        /* Unlock on close is handled by the flush method */
        if (fl->fl_flags & FL_CLOSE)
                return 0;
@@ -1115,7 +1444,7 @@ static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
                return PTR_ERR(req);
 
        fuse_lk_fill(req, file, fl, opcode, pid, flock);
-       request_send(fc, req);
+       fuse_request_send(fc, req);
        err = req->out.h.error;
        /* locking is restartable */
        if (err == -EINTR)
@@ -1130,7 +1459,9 @@ static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
        struct fuse_conn *fc = get_fuse_conn(inode);
        int err;
 
-       if (cmd == F_GETLK) {
+       if (cmd == F_CANCELLK) {
+               err = 0;
+       } else if (cmd == F_GETLK) {
                if (fc->no_lock) {
                        posix_test_lock(file, fl);
                        err = 0;
@@ -1138,7 +1469,7 @@ static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
                        err = fuse_getlk(file, fl);
        } else {
                if (fc->no_lock)
-                       err = posix_lock_file_wait(file, fl);
+                       err = posix_lock_file(file, fl, NULL);
                else
                        err = fuse_setlk(file, fl, 0);
        }
@@ -1189,7 +1520,7 @@ static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
        req->out.numargs = 1;
        req->out.args[0].size = sizeof(outarg);
        req->out.args[0].value = &outarg;
-       request_send(fc, req);
+       fuse_request_send(fc, req);
        err = req->out.h.error;
        fuse_put_request(fc, req);
        if (err == -ENOSYS)
@@ -1198,12 +1529,447 @@ static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
        return err ? 0 : outarg.block;
 }
 
+static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
+{
+       loff_t retval;
+       struct inode *inode = file->f_path.dentry->d_inode;
+
+       mutex_lock(&inode->i_mutex);
+       switch (origin) {
+       case SEEK_END:
+               retval = fuse_update_attributes(inode, NULL, file, NULL);
+               if (retval)
+                       goto exit;
+               offset += i_size_read(inode);
+               break;
+       case SEEK_CUR:
+               offset += file->f_pos;
+       }
+       retval = -EINVAL;
+       if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
+               if (offset != file->f_pos) {
+                       file->f_pos = offset;
+                       file->f_version = 0;
+               }
+               retval = offset;
+       }
+exit:
+       mutex_unlock(&inode->i_mutex);
+       return retval;
+}
+
+static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
+                       unsigned int nr_segs, size_t bytes, bool to_user)
+{
+       struct iov_iter ii;
+       int page_idx = 0;
+
+       if (!bytes)
+               return 0;
+
+       iov_iter_init(&ii, iov, nr_segs, bytes, 0);
+
+       while (iov_iter_count(&ii)) {
+               struct page *page = pages[page_idx++];
+               size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
+               void *kaddr, *map;
+
+               kaddr = map = kmap(page);
+
+               while (todo) {
+                       char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
+                       size_t iov_len = ii.iov->iov_len - ii.iov_offset;
+                       size_t copy = min(todo, iov_len);
+                       size_t left;
+
+                       if (!to_user)
+                               left = copy_from_user(kaddr, uaddr, copy);
+                       else
+                               left = copy_to_user(uaddr, kaddr, copy);
+
+                       if (unlikely(left))
+                               return -EFAULT;
+
+                       iov_iter_advance(&ii, copy);
+                       todo -= copy;
+                       kaddr += copy;
+               }
+
+               kunmap(map);
+       }
+
+       return 0;
+}
+
+/*
+ * For ioctls, there is no generic way to determine how much memory
+ * needs to be read and/or written.  Furthermore, ioctls are allowed
+ * to dereference the passed pointer, so the parameter requires deep
+ * copying but FUSE has no idea whatsoever about what to copy in or
+ * out.
+ *
+ * This is solved by allowing FUSE server to retry ioctl with
+ * necessary in/out iovecs.  Let's assume the ioctl implementation
+ * needs to read in the following structure.
+ *
+ * struct a {
+ *     char    *buf;
+ *     size_t  buflen;
+ * }
+ *
+ * On the first callout to FUSE server, inarg->in_size and
+ * inarg->out_size will be NULL; then, the server completes the ioctl
+ * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
+ * the actual iov array to
+ *
+ * { { .iov_base = inarg.arg,  .iov_len = sizeof(struct a) } }
+ *
+ * which tells FUSE to copy in the requested area and retry the ioctl.
+ * On the second round, the server has access to the structure and
+ * from that it can tell what to look for next, so on the invocation,
+ * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
+ *
+ * { { .iov_base = inarg.arg,  .iov_len = sizeof(struct a)     },
+ *   { .iov_base = a.buf,      .iov_len = a.buflen             } }
+ *
+ * FUSE will copy both struct a and the pointed buffer from the
+ * process doing the ioctl and retry ioctl with both struct a and the
+ * buffer.
+ *
+ * This time, FUSE server has everything it needs and completes ioctl
+ * without FUSE_IOCTL_RETRY which finishes the ioctl call.
+ *
+ * Copying data out works the same way.
+ *
+ * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
+ * automatically initializes in and out iovs by decoding @cmd with
+ * _IOC_* macros and the server is not allowed to request RETRY.  This
+ * limits ioctl data transfers to well-formed ioctls and is the forced
+ * behavior for all FUSE servers.
+ */
+static long fuse_do_ioctl(struct file *file, unsigned int cmd,
+                         unsigned long arg, unsigned int flags)
+{
+       struct fuse_file *ff = file->private_data;
+       struct fuse_conn *fc = ff->fc;
+       struct fuse_ioctl_in inarg = {
+               .fh = ff->fh,
+               .cmd = cmd,
+               .arg = arg,
+               .flags = flags
+       };
+       struct fuse_ioctl_out outarg;
+       struct fuse_req *req = NULL;
+       struct page **pages = NULL;
+       struct page *iov_page = NULL;
+       struct iovec *in_iov = NULL, *out_iov = NULL;
+       unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
+       size_t in_size, out_size, transferred;
+       int err;
+
+       /* assume all the iovs returned by client always fits in a page */
+       BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
+
+       err = -ENOMEM;
+       pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
+       iov_page = alloc_page(GFP_KERNEL);
+       if (!pages || !iov_page)
+               goto out;
+
+       /*
+        * If restricted, initialize IO parameters as encoded in @cmd.
+        * RETRY from server is not allowed.
+        */
+       if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
+               struct iovec *iov = page_address(iov_page);
+
+               iov->iov_base = (void __user *)arg;
+               iov->iov_len = _IOC_SIZE(cmd);
+
+               if (_IOC_DIR(cmd) & _IOC_WRITE) {
+                       in_iov = iov;
+                       in_iovs = 1;
+               }
+
+               if (_IOC_DIR(cmd) & _IOC_READ) {
+                       out_iov = iov;
+                       out_iovs = 1;
+               }
+       }
+
+ retry:
+       inarg.in_size = in_size = iov_length(in_iov, in_iovs);
+       inarg.out_size = out_size = iov_length(out_iov, out_iovs);
+
+       /*
+        * Out data can be used either for actual out data or iovs,
+        * make sure there always is at least one page.
+        */
+       out_size = max_t(size_t, out_size, PAGE_SIZE);
+       max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
+
+       /* make sure there are enough buffer pages and init request with them */
+       err = -ENOMEM;
+       if (max_pages > FUSE_MAX_PAGES_PER_REQ)
+               goto out;
+       while (num_pages < max_pages) {
+               pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
+               if (!pages[num_pages])
+                       goto out;
+               num_pages++;
+       }
+
+       req = fuse_get_req(fc);
+       if (IS_ERR(req)) {
+               err = PTR_ERR(req);
+               req = NULL;
+               goto out;
+       }
+       memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
+       req->num_pages = num_pages;
+
+       /* okay, let's send it to the client */
+       req->in.h.opcode = FUSE_IOCTL;
+       req->in.h.nodeid = ff->nodeid;
+       req->in.numargs = 1;
+       req->in.args[0].size = sizeof(inarg);
+       req->in.args[0].value = &inarg;
+       if (in_size) {
+               req->in.numargs++;
+               req->in.args[1].size = in_size;
+               req->in.argpages = 1;
+
+               err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
+                                          false);
+               if (err)
+                       goto out;
+       }
+
+       req->out.numargs = 2;
+       req->out.args[0].size = sizeof(outarg);
+       req->out.args[0].value = &outarg;
+       req->out.args[1].size = out_size;
+       req->out.argpages = 1;
+       req->out.argvar = 1;
+
+       fuse_request_send(fc, req);
+       err = req->out.h.error;
+       transferred = req->out.args[1].size;
+       fuse_put_request(fc, req);
+       req = NULL;
+       if (err)
+               goto out;
+
+       /* did it ask for retry? */
+       if (outarg.flags & FUSE_IOCTL_RETRY) {
+               char *vaddr;
+
+               /* no retry if in restricted mode */
+               err = -EIO;
+               if (!(flags & FUSE_IOCTL_UNRESTRICTED))
+                       goto out;
+
+               in_iovs = outarg.in_iovs;
+               out_iovs = outarg.out_iovs;
+
+               /*
+                * Make sure things are in boundary, separate checks
+                * are to protect against overflow.
+                */
+               err = -ENOMEM;
+               if (in_iovs > FUSE_IOCTL_MAX_IOV ||
+                   out_iovs > FUSE_IOCTL_MAX_IOV ||
+                   in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
+                       goto out;
+
+               err = -EIO;
+               if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
+                       goto out;
+
+               /* okay, copy in iovs and retry */
+               vaddr = kmap_atomic(pages[0], KM_USER0);
+               memcpy(page_address(iov_page), vaddr, transferred);
+               kunmap_atomic(vaddr, KM_USER0);
+
+               in_iov = page_address(iov_page);
+               out_iov = in_iov + in_iovs;
+
+               goto retry;
+       }
+
+       err = -EIO;
+       if (transferred > inarg.out_size)
+               goto out;
+
+       err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
+ out:
+       if (req)
+               fuse_put_request(fc, req);
+       if (iov_page)
+               __free_page(iov_page);
+       while (num_pages)
+               __free_page(pages[--num_pages]);
+       kfree(pages);
+
+       return err ? err : outarg.result;
+}
+
+static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
+                                  unsigned long arg, unsigned int flags)
+{
+       struct inode *inode = file->f_dentry->d_inode;
+       struct fuse_conn *fc = get_fuse_conn(inode);
+
+       if (!fuse_allow_task(fc, current))
+               return -EACCES;
+
+       if (is_bad_inode(inode))
+               return -EIO;
+
+       return fuse_do_ioctl(file, cmd, arg, flags);
+}
+
+static long fuse_file_ioctl(struct file *file, unsigned int cmd,
+                           unsigned long arg)
+{
+       return fuse_file_ioctl_common(file, cmd, arg, 0);
+}
+
+static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
+                                  unsigned long arg)
+{
+       return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
+}
+
+/*
+ * All files which have been polled are linked to RB tree
+ * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
+ * find the matching one.
+ */
+static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
+                                             struct rb_node **parent_out)
+{
+       struct rb_node **link = &fc->polled_files.rb_node;
+       struct rb_node *last = NULL;
+
+       while (*link) {
+               struct fuse_file *ff;
+
+               last = *link;
+               ff = rb_entry(last, struct fuse_file, polled_node);
+
+               if (kh < ff->kh)
+                       link = &last->rb_left;
+               else if (kh > ff->kh)
+                       link = &last->rb_right;
+               else
+                       return link;
+       }
+
+       if (parent_out)
+               *parent_out = last;
+       return link;
+}
+
+/*
+ * The file is about to be polled.  Make sure it's on the polled_files
+ * RB tree.  Note that files once added to the polled_files tree are
+ * not removed before the file is released.  This is because a file
+ * polled once is likely to be polled again.
+ */
+static void fuse_register_polled_file(struct fuse_conn *fc,
+                                     struct fuse_file *ff)
+{
+       spin_lock(&fc->lock);
+       if (RB_EMPTY_NODE(&ff->polled_node)) {
+               struct rb_node **link, *parent;
+
+               link = fuse_find_polled_node(fc, ff->kh, &parent);
+               BUG_ON(*link);
+               rb_link_node(&ff->polled_node, parent, link);
+               rb_insert_color(&ff->polled_node, &fc->polled_files);
+       }
+       spin_unlock(&fc->lock);
+}
+
+static unsigned fuse_file_poll(struct file *file, poll_table *wait)
+{
+       struct fuse_file *ff = file->private_data;
+       struct fuse_conn *fc = ff->fc;
+       struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
+       struct fuse_poll_out outarg;
+       struct fuse_req *req;
+       int err;
+
+       if (fc->no_poll)
+               return DEFAULT_POLLMASK;
+
+       poll_wait(file, &ff->poll_wait, wait);
+
+       /*
+        * Ask for notification iff there's someone waiting for it.
+        * The client may ignore the flag and always notify.
+        */
+       if (waitqueue_active(&ff->poll_wait)) {
+               inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
+               fuse_register_polled_file(fc, ff);
+       }
+
+       req = fuse_get_req(fc);
+       if (IS_ERR(req))
+               return PTR_ERR(req);
+
+       req->in.h.opcode = FUSE_POLL;
+       req->in.h.nodeid = ff->nodeid;
+       req->in.numargs = 1;
+       req->in.args[0].size = sizeof(inarg);
+       req->in.args[0].value = &inarg;
+       req->out.numargs = 1;
+       req->out.args[0].size = sizeof(outarg);
+       req->out.args[0].value = &outarg;
+       fuse_request_send(fc, req);
+       err = req->out.h.error;
+       fuse_put_request(fc, req);
+
+       if (!err)
+               return outarg.revents;
+       if (err == -ENOSYS) {
+               fc->no_poll = 1;
+               return DEFAULT_POLLMASK;
+       }
+       return POLLERR;
+}
+
+/*
+ * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
+ * wakes up the poll waiters.
+ */
+int fuse_notify_poll_wakeup(struct fuse_conn *fc,
+                           struct fuse_notify_poll_wakeup_out *outarg)
+{
+       u64 kh = outarg->kh;
+       struct rb_node **link;
+
+       spin_lock(&fc->lock);
+
+       link = fuse_find_polled_node(fc, kh, NULL);
+       if (*link) {
+               struct fuse_file *ff;
+
+               ff = rb_entry(*link, struct fuse_file, polled_node);
+               wake_up_interruptible_sync(&ff->poll_wait);
+       }
+
+       spin_unlock(&fc->lock);
+       return 0;
+}
+
 static const struct file_operations fuse_file_operations = {
-       .llseek         = generic_file_llseek,
+       .llseek         = fuse_file_llseek,
        .read           = do_sync_read,
        .aio_read       = fuse_file_aio_read,
        .write          = do_sync_write,
-       .aio_write      = generic_file_aio_write,
+       .aio_write      = fuse_file_aio_write,
        .mmap           = fuse_file_mmap,
        .open           = fuse_open,
        .flush          = fuse_flush,
@@ -1212,19 +1978,26 @@ static const struct file_operations fuse_file_operations = {
        .lock           = fuse_file_lock,
        .flock          = fuse_file_flock,
        .splice_read    = generic_file_splice_read,
+       .unlocked_ioctl = fuse_file_ioctl,
+       .compat_ioctl   = fuse_file_compat_ioctl,
+       .poll           = fuse_file_poll,
 };
 
 static const struct file_operations fuse_direct_io_file_operations = {
-       .llseek         = generic_file_llseek,
+       .llseek         = fuse_file_llseek,
        .read           = fuse_direct_read,
        .write          = fuse_direct_write,
+       .mmap           = fuse_direct_mmap,
        .open           = fuse_open,
        .flush          = fuse_flush,
        .release        = fuse_release,
        .fsync          = fuse_fsync,
        .lock           = fuse_file_lock,
        .flock          = fuse_file_flock,
-       /* no mmap and splice_read */
+       .unlocked_ioctl = fuse_file_ioctl,
+       .compat_ioctl   = fuse_file_compat_ioctl,
+       .poll           = fuse_file_poll,
+       /* no splice_read */
 };
 
 static const struct address_space_operations fuse_file_aops  = {