support booting of arbitrary non-blockdevice file systems

support booting of arbitrary non-blockdevice file systems

[Christoph Hellwig]

Hi all,

this series adds support to boot off arbitrary non-blockdevice root file
systems, based off an earlier patch from Vivek.

[PATCH 1/2] init: split get_fs_names

[Christoph Hellwig]

Split get_fs_names into one function that splits up the command line
argument, and one that gets the list of all registered file systems.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 init/do_mounts.c | 48 ++++++++++++++++++++++++++----------------------
 1 file changed, 26 insertions(+), 22 deletions(-)

diff --git a/init/do_mounts.c b/init/do_mounts.c
index 74aede860de7..ec32de3ad52b 100644
--- a/init/do_mounts.c
+++ b/init/do_mounts.c
@@ -338,30 +338,31 @@ __setup("rootflags=", root_data_setup);
 __setup("rootfstype=", fs_names_setup);
 __setup("rootdelay=", root_delay_setup);
 
-static void __init get_fs_names(char *page)
+static void __init split_fs_names(char *page, char *names)
 {
-	char *s = page;
-
-	if (root_fs_names) {
-		strcpy(page, root_fs_names);
-		while (*s++) {
-			if (s[-1] == ',')
-				s[-1] = '\0';
-		}
-	} else {
-		int len = get_filesystem_list(page);
-		char *p, *next;
+	strcpy(page, root_fs_names);
+	while (*page++) {
+		if (page[-1] == ',')
+			page[-1] = '\0';
+	}
+	*page = '\0';
+}
 
-		page[len] = '\0';
-		for (p = page-1; p; p = next) {
-			next = strchr(++p, '\n');
-			if (*p++ != '\t')
-				continue;
-			while ((*s++ = *p++) != '\n')
-				;
-			s[-1] = '\0';
-		}
+static void __init get_all_fs_names(char *page)
+{
+	int len = get_filesystem_list(page);
+	char *s = page, *p, *next;
+
+	page[len] = '\0';
+	for (p = page - 1; p; p = next) {
+		next = strchr(++p, '\n');
+		if (*p++ != '\t')
+			continue;
+		while ((*s++ = *p++) != '\n')
+			;
+		s[-1] = '\0';
 	}
+
 	*s = '\0';
 }
 
@@ -411,7 +412,10 @@ void __init mount_block_root(char *name, int flags)
 
 	scnprintf(b, BDEVNAME_SIZE, "unknown-block(%u,%u)",
 		  MAJOR(ROOT_DEV), MINOR(ROOT_DEV));
-	get_fs_names(fs_names);
+	if (root_fs_names)
+		split_fs_names(fs_names, root_fs_names);
+	else
+		get_all_fs_names(fs_names);
 retry:
 	for (p = fs_names; *p; p += strlen(p)+1) {
 		int err = do_mount_root(name, p, flags, root_mount_data);
-- 
2.30.2

[PATCH 2/2] init: allow mounting arbitrary non-blockdevice filesystems as root

[Christoph Hellwig]

Currently the only non-blockdevice filesystems that can be used as the
initial root filesystem are NFS and CIFS, which use the magic
"root=/dev/nfs" and "root=/dev/cifs" syntax that requires the root
device file system details to come from filesystem specific kernel
command line options.

Add a little bit of new code that allows to just pass arbitrary
string mount options to any non-blockdevice filesystems so that it can
be mounted as the root file system.

For example a virtiofs root file system can be mounted using the
following syntax:

"root=myfs rootfstype=virtiofs rw"

Based on an earlier patch from Vivek Goyal <vgoyal@redhat.com>.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 init/do_mounts.c | 40 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 40 insertions(+)

diff --git a/init/do_mounts.c b/init/do_mounts.c
index ec32de3ad52b..64c60cb72ecb 100644
--- a/init/do_mounts.c
+++ b/init/do_mounts.c
@@ -534,6 +534,44 @@ static int __init mount_cifs_root(void)
 }
 #endif
 
+static int __init try_mount_nodev(char *fstype)
+{
+	struct file_system_type *fs = get_fs_type(fstype);
+	int err = -EINVAL;
+
+	if (!fs)
+		return -EINVAL;
+	if (!(fs->fs_flags & (FS_REQUIRES_DEV | FS_BINARY_MOUNTDATA)))
+		err = do_mount_root(root_device_name, fstype, root_mountflags,
+					root_mount_data);
+	put_filesystem(fs);
+
+	if (err != -EACCES && err != -EINVAL)
+		panic("VFS: Unable to mount root \"%s\" (%s), err=%d\n",
+			      root_device_name, fstype, err);
+	return err;
+}
+
+static int __init mount_nodev_root(void)
+{
+	char *fs_names, *p;
+	int err = -EINVAL;
+
+	fs_names = (void *)__get_free_page(GFP_KERNEL);
+	if (!fs_names)
+		return -EINVAL;
+	split_fs_names(fs_names, root_fs_names);
+
+	for (p = fs_names; *p; p += strlen(p) + 1) {
+		err = try_mount_nodev(p);
+		if (!err)
+			break;
+	}
+
+	free_page((unsigned long)fs_names);
+	return err;
+}
+
 void __init mount_root(void)
 {
 #ifdef CONFIG_ROOT_NFS
@@ -550,6 +588,8 @@ void __init mount_root(void)
 		return;
 	}
 #endif
+	if (ROOT_DEV == 0 && mount_nodev_root() == 0)
+		return;
 #ifdef CONFIG_BLOCK
 	{
 		int err = create_dev("/dev/root", ROOT_DEV);
-- 
2.30.2

Re: [PATCH 2/2] init: allow mounting arbitrary non-blockdevice filesystems as root

[Vivek Goyal]

On Thu, Jun 17, 2021 at 05:36:49PM +0200, Christoph Hellwig wrote:

[..]
> +static int __init try_mount_nodev(char *fstype)
> +{
> +	struct file_system_type *fs = get_fs_type(fstype);
> +	int err = -EINVAL;
> +
> +	if (!fs)
> +		return -EINVAL;
> +	if (!(fs->fs_flags & (FS_REQUIRES_DEV | FS_BINARY_MOUNTDATA)))

Not sure what FS_BINARY_MOUNTDATA is why fs should not have that set. nfs
seems to set it too. So that means they can't use try_mount_nodev().

> +		err = do_mount_root(root_device_name, fstype, root_mountflags,
> +					root_mount_data);
> +	put_filesystem(fs);
> +
> +	if (err != -EACCES && err != -EINVAL)

In case of success err == 0, but we still panic(). We will need to
check for success as well.
	if (err && err != -EACCES && err != -EINVAL)

> +		panic("VFS: Unable to mount root \"%s\" (%s), err=%d\n",
> +			      root_device_name, fstype, err);
> +	return err;
> +}
> +
> +static int __init mount_nodev_root(void)
> +{
> +	char *fs_names, *p;
> +	int err = -EINVAL;
> +
> +	fs_names = (void *)__get_free_page(GFP_KERNEL);
> +	if (!fs_names)
> +		return -EINVAL;
> +	split_fs_names(fs_names, root_fs_names);

root_fs_names can be NULL and it crashes with NULL pointer dereference.

Vivek

> +
> +	for (p = fs_names; *p; p += strlen(p) + 1) {
> +		err = try_mount_nodev(p);
> +		if (!err)
> +			break;
> +	}
> +
> +	free_page((unsigned long)fs_names);
> +	return err;
> +}
> +
>  void __init mount_root(void)
>  {
>  #ifdef CONFIG_ROOT_NFS
> @@ -550,6 +588,8 @@ void __init mount_root(void)
>  		return;
>  	}
>  #endif
> +	if (ROOT_DEV == 0 && mount_nodev_root() == 0)
> +		return;
>  #ifdef CONFIG_BLOCK
>  	{
>  		int err = create_dev("/dev/root", ROOT_DEV);
> -- 
> 2.30.2
> 

Re: [PATCH 2/2] init: allow mounting arbitrary non-blockdevice filesystems as root

[Christoph Hellwig]

On Thu, Jun 17, 2021 at 12:26:10PM -0400, Vivek Goyal wrote:
> Not sure what FS_BINARY_MOUNTDATA is why fs should not have that set. nfs
> seems to set it too. So that means they can't use try_mount_nodev().

We can't really pass actual binary mountdata using the string separation
scheme used by the rootfstype= option.  But given that NFS only uses
binary mountdata for legacy reasons and people get what they ask for
using the option I think we can drop the check.

> In case of success err == 0, but we still panic(). We will need to
> check for success as well.

Indeed.

> root_fs_names can be NULL and it crashes with NULL pointer dereference.

True.

What do you think of this version?

---
From 141caa79a619b5f5d100eeb8e94ecf8b3b1c9af7 Mon Sep 17 00:00:00 2001
From: Christoph Hellwig <hch@lst.de>
Date: Fri, 18 Jun 2021 15:10:39 +0200
Subject: init: allow mounting arbitrary non-blockdevice filesystems as root

Currently the only non-blockdevice filesystems that can be used as the
initial root filesystem are NFS and CIFS, which use the magic
"root=/dev/nfs" and "root=/dev/cifs" syntax that requires the root
device file system details to come from filesystem specific kernel
command line options.

Add a little bit of new code that allows to just pass arbitrary
string mount options to any non-blockdevice filesystems so that it can
be mounted as the root file system.

For example a virtiofs root file system can be mounted using the
following syntax:

"root=myfs rootfstype=virtiofs rw"

Based on an earlier patch from Vivek Goyal <vgoyal@redhat.com>.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 init/do_mounts.c | 43 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 43 insertions(+)

diff --git a/init/do_mounts.c b/init/do_mounts.c
index ec32de3ad52b..66c47193e9ee 100644
--- a/init/do_mounts.c
+++ b/init/do_mounts.c
@@ -534,6 +534,45 @@ static int __init mount_cifs_root(void)
 }
 #endif
 
+static bool __init fs_is_nodev(char *fstype)
+{
+	struct file_system_type *fs = get_fs_type(fstype);
+	bool ret = false;
+
+	if (fs) {
+		ret = !(fs->fs_flags & FS_REQUIRES_DEV);
+		put_filesystem(fs);
+	}
+
+	return ret;
+}
+
+static int __init mount_nodev_root(void)
+{
+	char *fs_names, *fstype;
+	int err = -EINVAL;
+
+	fs_names = (void *)__get_free_page(GFP_KERNEL);
+	if (!fs_names)
+		return -EINVAL;
+	split_fs_names(fs_names, root_fs_names);
+
+	for (fstype = fs_names; *fstype; fstype += strlen(fstype) + 1) {
+		if (!fs_is_nodev(fstype))
+			continue;
+		err = do_mount_root(root_device_name, fstype, root_mountflags,
+				    root_mount_data);
+		if (!err)
+			break;
+		if (err != -EACCES && err != -EINVAL)
+			panic("VFS: Unable to mount root \"%s\" (%s), err=%d\n",
+			      root_device_name, fstype, err);
+	}
+
+	free_page((unsigned long)fs_names);
+	return err;
+}
+
 void __init mount_root(void)
 {
 #ifdef CONFIG_ROOT_NFS
@@ -550,6 +589,10 @@ void __init mount_root(void)
 		return;
 	}
 #endif
+	if (ROOT_DEV == 0 && root_fs_names) {
+		if (mount_nodev_root() == 0)
+			return;
+	}
 #ifdef CONFIG_BLOCK
 	{
 		int err = create_dev("/dev/root", ROOT_DEV);
-- 
2.30.2

Re: [PATCH 2/2] init: allow mounting arbitrary non-blockdevice filesystems as root

[Vivek Goyal]

On Fri, Jun 18, 2021 at 03:20:38PM +0200, Christoph Hellwig wrote:
> On Thu, Jun 17, 2021 at 12:26:10PM -0400, Vivek Goyal wrote:
> > Not sure what FS_BINARY_MOUNTDATA is why fs should not have that set. nfs
> > seems to set it too. So that means they can't use try_mount_nodev().
> 
> We can't really pass actual binary mountdata using the string separation
> scheme used by the rootfstype= option.  But given that NFS only uses
> binary mountdata for legacy reasons and people get what they ask for
> using the option I think we can drop the check.
> 
> > In case of success err == 0, but we still panic(). We will need to
> > check for success as well.
> 
> Indeed.
> 
> > root_fs_names can be NULL and it crashes with NULL pointer dereference.
> 
> True.
> 
> What do you think of this version?

[ Cc Dominique Martinet ]

Hi Christoph,

This one works well for me. Just one minor nit. root_device_name, can
be NULL as well if user passes following.

"rootfstype=virtiofs rw".

And currently mount_nodev_root() is assuming root_device_name is not NULL
and we crash with null pointer dereference.

May be something like this.

        if (ROOT_DEV == 0 && root_device_name && root_fs_names) {
                if (mount_nodev_root() == 0)
                        return;
        }

Strangely people are using "rootfstype=virtiofs rw" to mount virtiofs
as rootfs. They give their filesystem a tag named "/dev/root". And
currently it works and they can mount virtiofs as rootfs.

With above change, current hackish way will also continue to work and
not break existing setups.

Thanks
Vivek

> 
> ---
> From 141caa79a619b5f5d100eeb8e94ecf8b3b1c9af7 Mon Sep 17 00:00:00 2001
> From: Christoph Hellwig <hch@lst.de>
> Date: Fri, 18 Jun 2021 15:10:39 +0200
> Subject: init: allow mounting arbitrary non-blockdevice filesystems as root
> 
> Currently the only non-blockdevice filesystems that can be used as the
> initial root filesystem are NFS and CIFS, which use the magic
> "root=/dev/nfs" and "root=/dev/cifs" syntax that requires the root
> device file system details to come from filesystem specific kernel
> command line options.
> 
> Add a little bit of new code that allows to just pass arbitrary
> string mount options to any non-blockdevice filesystems so that it can
> be mounted as the root file system.
> 
> For example a virtiofs root file system can be mounted using the
> following syntax:
> 
> "root=myfs rootfstype=virtiofs rw"
> 
> Based on an earlier patch from Vivek Goyal <vgoyal@redhat.com>.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
>  init/do_mounts.c | 43 +++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 43 insertions(+)
> 
> diff --git a/init/do_mounts.c b/init/do_mounts.c
> index ec32de3ad52b..66c47193e9ee 100644
> --- a/init/do_mounts.c
> +++ b/init/do_mounts.c
> @@ -534,6 +534,45 @@ static int __init mount_cifs_root(void)
>  }
>  #endif
>  
> +static bool __init fs_is_nodev(char *fstype)
> +{
> +	struct file_system_type *fs = get_fs_type(fstype);
> +	bool ret = false;
> +
> +	if (fs) {
> +		ret = !(fs->fs_flags & FS_REQUIRES_DEV);
> +		put_filesystem(fs);
> +	}
> +
> +	return ret;
> +}
> +
> +static int __init mount_nodev_root(void)
> +{
> +	char *fs_names, *fstype;
> +	int err = -EINVAL;
> +
> +	fs_names = (void *)__get_free_page(GFP_KERNEL);
> +	if (!fs_names)
> +		return -EINVAL;
> +	split_fs_names(fs_names, root_fs_names);
> +
> +	for (fstype = fs_names; *fstype; fstype += strlen(fstype) + 1) {
> +		if (!fs_is_nodev(fstype))
> +			continue;
> +		err = do_mount_root(root_device_name, fstype, root_mountflags,
> +				    root_mount_data);
> +		if (!err)
> +			break;
> +		if (err != -EACCES && err != -EINVAL)
> +			panic("VFS: Unable to mount root \"%s\" (%s), err=%d\n",
> +			      root_device_name, fstype, err);
> +	}
> +
> +	free_page((unsigned long)fs_names);
> +	return err;
> +}
> +
>  void __init mount_root(void)
>  {
>  #ifdef CONFIG_ROOT_NFS
> @@ -550,6 +589,10 @@ void __init mount_root(void)
>  		return;
>  	}
>  #endif
> +	if (ROOT_DEV == 0 && root_fs_names) {
> +		if (mount_nodev_root() == 0)
> +			return;
> +	}
>  #ifdef CONFIG_BLOCK
>  	{
>  		int err = create_dev("/dev/root", ROOT_DEV);
> -- 
> 2.30.2
> 

Re: [Virtio-fs] support booting of arbitrary non-blockdevice file systems

[Stefan Hajnoczi]

[-- Attachment #1: Type: text/plain, Size: 329 bytes --]

On Thu, Jun 17, 2021 at 05:36:47PM +0200, Christoph Hellwig wrote:
> this series adds support to boot off arbitrary non-blockdevice root file
> systems, based off an earlier patch from Vivek.

Cool, thanks for working on generic syntax for mounting non-blockdevice
file systems. Looks good modulo the comments Vivek had.

Stefan

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]