[v2,1/3] kernel/sysctl: support setting sysctl parameters from kernel command line

Commit Message

Vlastimil Babka April 14, 2020, 11:32 a.m. UTC

A recently proposed patch to add vm_swappiness command line parameter in
addition to existing sysctl [1] made me wonder why we don't have a general
support for passing sysctl parameters via command line. Googling found only
somebody else wondering the same [2], but I haven't found any prior discussion
with reasons why not to do this.

Settings the vm_swappiness issue aside (the underlying issue might be solved in
a different way), quick search of kernel-parameters.txt shows there are already
some that exist as both sysctl and kernel parameter - hung_task_panic,
nmi_watchdog, numa_zonelist_order, traceoff_on_warning. A general mechanism
would remove the need to add more of those one-offs and might be handy in
situations where configuration by e.g. /etc/sysctl.d/ is impractical.

Hence, this patch adds a new parse_args() pass that looks for parameters
prefixed by 'sysctl.' and tries to interpret them as writes to the
corresponding sys/ files using an temporary in-kernel procfs mount. This
mechanism was suggested by Eric W. Biederman [3], as it handles all dynamically
registered sysctl tables, even though we don't handle modular sysctls. Errors
due to e.g. invalid parameter name or value are reported in the kernel log.

The processing is hooked right before the init process is loaded, as some
handlers might be more complicated than simple setters and might need some
subsystems to be initialized. At the moment the init process can be started and
eventually execute a process writing to /proc/sys/ then it should be also fine
to do that from the kernel.

Sysctls registered later on module load time are not set by this mechanism -
it's expected that in such scenarios, setting sysctl values from userspace is
practical enough.

[1] https://lore.kernel.org/r/[e-mail]/
[2] https://unix.stackexchange.com/questions/558802/how-to-set-sysctl-using-kernel-command-line-parameter
[3] https://lore.kernel.org/r/[e-mail]/

Signed-off-by: Vlastimil Babka <[e-mail]>
Reviewed-by: Luis Chamberlain <[e-mail]>
---
 .../admin-guide/kernel-parameters.txt         |   9 ++
 fs/proc/proc_sysctl.c                         | 107 ++++++++++++++++++
 include/linux/sysctl.h                        |   4 +
 init/main.c                                   |   2 +
 4 files changed, 122 insertions(+)

Comments

Kees Cook April 15, 2020, 3:08 a.m. UTC | #1

On Tue, Apr 14, 2020 at 01:32:20PM +0200, Vlastimil Babka wrote:
> A recently proposed patch to add vm_swappiness command line parameter in
> addition to existing sysctl [1] made me wonder why we don't have a general
> support for passing sysctl parameters via command line. Googling found only
> somebody else wondering the same [2], but I haven't found any prior discussion
> with reasons why not to do this.
> 
> Settings the vm_swappiness issue aside (the underlying issue might be solved in
> a different way), quick search of kernel-parameters.txt shows there are already
> some that exist as both sysctl and kernel parameter - hung_task_panic,
> nmi_watchdog, numa_zonelist_order, traceoff_on_warning. A general mechanism
> would remove the need to add more of those one-offs and might be handy in
> situations where configuration by e.g. /etc/sysctl.d/ is impractical.
> 
> Hence, this patch adds a new parse_args() pass that looks for parameters
> prefixed by 'sysctl.' and tries to interpret them as writes to the
> corresponding sys/ files using an temporary in-kernel procfs mount. This
> mechanism was suggested by Eric W. Biederman [3], as it handles all dynamically
> registered sysctl tables, even though we don't handle modular sysctls. Errors
> due to e.g. invalid parameter name or value are reported in the kernel log.
> 
> The processing is hooked right before the init process is loaded, as some
> handlers might be more complicated than simple setters and might need some
> subsystems to be initialized. At the moment the init process can be started and
> eventually execute a process writing to /proc/sys/ then it should be also fine
> to do that from the kernel.
> 
> Sysctls registered later on module load time are not set by this mechanism -
> it's expected that in such scenarios, setting sysctl values from userspace is
> practical enough.
> 
> [1] https://lore.kernel.org/r/[e-mail]/
> [2] https://unix.stackexchange.com/questions/558802/how-to-set-sysctl-using-kernel-command-line-parameter
> [3] https://lore.kernel.org/r/[e-mail]/
> 
> Signed-off-by: Vlastimil Babka <[e-mail]>

Looks great!

Acked-by: Kees Cook <[e-mail]>

-Kees

> Reviewed-by: Luis Chamberlain <[e-mail]>
> ---
>  .../admin-guide/kernel-parameters.txt         |   9 ++
>  fs/proc/proc_sysctl.c                         | 107 ++++++++++++++++++
>  include/linux/sysctl.h                        |   4 +
>  init/main.c                                   |   2 +
>  4 files changed, 122 insertions(+)
> 
> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> index f2a93c8679e8..fd38f1e8a2bf 100644
> --- a/Documentation/admin-guide/kernel-parameters.txt
> +++ b/Documentation/admin-guide/kernel-parameters.txt
> @@ -4891,6 +4891,15 @@
>  
>  	switches=	[HW,M68k]
>  
> +	sysctl.*=	[KNL]
> +			Set a sysctl parameter, right before loading the init
> +			process, as if the value was written to the respective
> +			/proc/sys/... file. Both '.' and '/' are recognized as
> +			separators. Unrecognized parameters and invalid values
> +			are reported in the kernel log. Sysctls registered
> +			later by a loaded module cannot be set this way.
> +			Example: sysctl.vm.swappiness=40
> +
>  	sysfs.deprecated=0|1 [KNL]
>  			Enable/disable old style sysfs layout for old udev
>  			on older distributions. When this option is enabled
> diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c
> index b6f5d459b087..7804da5f5be0 100644
> --- a/fs/proc/proc_sysctl.c
> +++ b/fs/proc/proc_sysctl.c
> @@ -14,6 +14,7 @@
>  #include <linux/mm.h>
>  #include <linux/module.h>
>  #include <linux/bpf-cgroup.h>
> +#include <linux/mount.h>
>  #include "internal.h"
>  
>  static const struct dentry_operations proc_sys_dentry_operations;
> @@ -1692,3 +1693,109 @@ int __init proc_sys_init(void)
>  
>  	return sysctl_init();
>  }
> +
> +/* Set sysctl value passed on kernel command line. */
> +static int process_sysctl_arg(char *param, char *val,
> +			       const char *unused, void *arg)
> +{
> +	char *path;
> +	struct vfsmount **proc_mnt = arg;
> +	struct file_system_type *proc_fs_type;
> +	struct file *file;
> +	int len;
> +	int err;
> +	loff_t pos = 0;
> +	ssize_t wret;
> +
> +	if (strncmp(param, "sysctl", sizeof("sysctl") - 1))
> +		return 0;
> +
> +	param += sizeof("sysctl") - 1;
> +
> +	if (param[0] != '/' && param[0] != '.')
> +		return 0;
> +
> +	param++;
> +
> +	/*
> +	 * To set sysctl options, we use a temporary mount of proc, look up the
> +	 * respective sys/ file and write to it. To avoid mounting it when no
> +	 * options were given, we mount it only when the first sysctl option is
> +	 * found. Why not a persistent mount? There are problems with a
> +	 * persistent mount of proc in that it forces userspace not to use any
> +	 * proc mount options.
> +	 */
> +	if (!*proc_mnt) {
> +		proc_fs_type = get_fs_type("proc");
> +		if (!proc_fs_type) {
> +			pr_err("Failed to find procfs to set sysctl from command line\n");
> +			return 0;
> +		}
> +		*proc_mnt = kern_mount(proc_fs_type);
> +		put_filesystem(proc_fs_type);
> +		if (IS_ERR(*proc_mnt)) {
> +			pr_err("Failed to mount procfs to set sysctl from command line\n");
> +			return 0;
> +		}
> +	}
> +
> +	path = kasprintf(GFP_KERNEL, "sys/%s", param);
> +	if (!path)
> +		panic("%s: Failed to allocate path for %s\n", __func__, param);
> +	strreplace(path, '.', '/');
> +
> +	file = file_open_root((*proc_mnt)->mnt_root, *proc_mnt, path, O_WRONLY, 0);
> +	if (IS_ERR(file)) {
> +		err = PTR_ERR(file);
> +		if (err == -ENOENT)
> +			pr_err("Failed to set sysctl parameter '%s=%s': parameter not found\n",
> +				param, val);
> +		else if (err == -EACCES)
> +			pr_err("Failed to set sysctl parameter '%s=%s': permission denied (read-only?)\n",
> +				param, val);
> +		else
> +			pr_err("Error %pe opening proc file to set sysctl parameter '%s=%s'\n",
> +				file, param, val);
> +		goto out;
> +	}
> +	len = strlen(val);
> +	wret = kernel_write(file, val, len, &pos);
> +	if (wret < 0) {
> +		err = wret;
> +		if (err == -EINVAL)
> +			pr_err("Failed to set sysctl parameter '%s=%s': invalid value\n",
> +				param, val);
> +		else
> +			pr_err("Error %pe writing to proc file to set sysctl parameter '%s=%s'\n",
> +				ERR_PTR(err), param, val);
> +	} else if (wret != len) {
> +		pr_err("Wrote only %zd bytes of %d writing to proc file %s to set sysctl parameter '%s=%s\n",
> +			wret, len, path, param, val);
> +	}
> +
> +	err = filp_close(file, NULL);
> +	if (err)
> +		pr_err("Error %pe closing proc file to set sysctl parameter '%s=%s\n",
> +			ERR_PTR(err), param, val);
> +out:
> +	kfree(path);
> +	return 0;
> +}
> +
> +void do_sysctl_args(void)
> +{
> +	char *command_line;
> +	struct vfsmount *proc_mnt = NULL;
> +
> +	command_line = kstrdup(saved_command_line, GFP_KERNEL);
> +	if (!command_line)
> +		panic("%s: Failed to allocate copy of command line\n", __func__);
> +
> +	parse_args("Setting sysctl args", command_line,
> +		   NULL, 0, -1, -1, &proc_mnt, process_sysctl_arg);
> +
> +	if (proc_mnt)
> +		kern_unmount(proc_mnt);
> +
> +	kfree(command_line);
> +}
> diff --git a/include/linux/sysctl.h b/include/linux/sysctl.h
> index 02fa84493f23..bee62d30ebdb 100644
> --- a/include/linux/sysctl.h
> +++ b/include/linux/sysctl.h
> @@ -206,6 +206,7 @@ struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
>  void unregister_sysctl_table(struct ctl_table_header * table);
>  
>  extern int sysctl_init(void);
> +void do_sysctl_args(void);
>  
>  extern struct ctl_table sysctl_mount_point[];
>  
> @@ -236,6 +237,9 @@ static inline void setup_sysctl_set(struct ctl_table_set *p,
>  {
>  }
>  
> +static inline void do_sysctl_args(void)
> +{
> +}
>  #endif /* CONFIG_SYSCTL */
>  
>  int sysctl_max_threads(struct ctl_table *table, int write,
> diff --git a/init/main.c b/init/main.c
> index a48617f2e5e5..7b43118215d6 100644
> --- a/init/main.c
> +++ b/init/main.c
> @@ -1372,6 +1372,8 @@ static int __ref kernel_init(void *unused)
>  
>  	rcu_end_inkernel_boot();
>  
> +	do_sysctl_args();
> +
>  	if (ramdisk_execute_command) {
>  		ret = run_init_process(ramdisk_execute_command);
>  		if (!ret)
> -- 
> 2.26.0
>

Masami Hiramatsu (Google) April 15, 2020, 9:03 a.m. UTC | #2

On Tue, 14 Apr 2020 13:32:20 +0200
Vlastimil Babka <[e-mail]> wrote:

> A recently proposed patch to add vm_swappiness command line parameter in
> addition to existing sysctl [1] made me wonder why we don't have a general
> support for passing sysctl parameters via command line. Googling found only
> somebody else wondering the same [2], but I haven't found any prior discussion
> with reasons why not to do this.
> 
> Settings the vm_swappiness issue aside (the underlying issue might be solved in
> a different way), quick search of kernel-parameters.txt shows there are already
> some that exist as both sysctl and kernel parameter - hung_task_panic,
> nmi_watchdog, numa_zonelist_order, traceoff_on_warning. A general mechanism
> would remove the need to add more of those one-offs and might be handy in
> situations where configuration by e.g. /etc/sysctl.d/ is impractical.
> 
> Hence, this patch adds a new parse_args() pass that looks for parameters
> prefixed by 'sysctl.' and tries to interpret them as writes to the
> corresponding sys/ files using an temporary in-kernel procfs mount. This
> mechanism was suggested by Eric W. Biederman [3], as it handles all dynamically
> registered sysctl tables, even though we don't handle modular sysctls. Errors
> due to e.g. invalid parameter name or value are reported in the kernel log.
> 
> The processing is hooked right before the init process is loaded, as some
> handlers might be more complicated than simple setters and might need some
> subsystems to be initialized. At the moment the init process can be started and
> eventually execute a process writing to /proc/sys/ then it should be also fine
> to do that from the kernel.
> 
> Sysctls registered later on module load time are not set by this mechanism -
> it's expected that in such scenarios, setting sysctl values from userspace is
> practical enough.
> 
> [1] https://lore.kernel.org/r/[e-mail]/
> [2] https://unix.stackexchange.com/questions/558802/how-to-set-sysctl-using-kernel-command-line-parameter
> [3] https://lore.kernel.org/r/[e-mail]/
> 
> Signed-off-by: Vlastimil Babka <[e-mail]>
> Reviewed-by: Luis Chamberlain <[e-mail]>

[...]

> diff --git a/init/main.c b/init/main.c
> index a48617f2e5e5..7b43118215d6 100644
> --- a/init/main.c
> +++ b/init/main.c
> @@ -1372,6 +1372,8 @@ static int __ref kernel_init(void *unused)
>  
>  	rcu_end_inkernel_boot();
>  
> +	do_sysctl_args();
> +

Ah, I see. Since the sysctl is designed to be called after all __init calls were
done, it shouldn't use bootconfig directly because bootconfig is full of __init
call. OK, anyway we can use "kernel.sysctl" prefixed bootconfig for these.

Reviewed-by: Masami Hiramatsu <[e-mail]>

Thank you,

Luis Chamberlain April 16, 2020, 1:29 a.m. UTC | #3

On Wed, Apr 15, 2020 at 06:03:55PM +0900, Masami Hiramatsu wrote:
> On Tue, 14 Apr 2020 13:32:20 +0200
> Vlastimil Babka <[e-mail]> wrote:
> > diff --git a/init/main.c b/init/main.c
> > index a48617f2e5e5..7b43118215d6 100644
> > --- a/init/main.c
> > +++ b/init/main.c
> > @@ -1372,6 +1372,8 @@ static int __ref kernel_init(void *unused)
> >  
> >  	rcu_end_inkernel_boot();
> >  
> > +	do_sysctl_args();
> > +
> 
> Ah, I see. Since the sysctl is designed to be called after all __init calls were
> done, it shouldn't use bootconfig directly because bootconfig is full of __init
> call.

The idea is bootconfig would be useful in the sense of a library set of
helpers which could be modified to remove __init, and then used to
instrument the cmdline depending on certain debugging kconfig entries.

We currently have no way to purposely extend / break the cmdline for
debugging purposes, so, bootconfig's parsers, since it already has a
way to extend the cmdlineline, might make it much easier to do this
later.

Without bootconfig, if we wanted to add new kconfig to, for example,
add new funny cmdline arguments to test they worked or not, we'd have
to devise our own set of helpers now. ie, new functionality. bootconfig
however already has existing functionality to tweak the cmdline, and so
some code could be leveraged there for this purpose.

  Luis

Masami Hiramatsu (Google) April 16, 2020, 10:49 a.m. UTC | #4

Hi Luis,

On Thu, 16 Apr 2020 01:29:31 +0000
Luis Chamberlain <[e-mail]> wrote:

> On Wed, Apr 15, 2020 at 06:03:55PM +0900, Masami Hiramatsu wrote:
> > On Tue, 14 Apr 2020 13:32:20 +0200
> > Vlastimil Babka <[e-mail]> wrote:
> > > diff --git a/init/main.c b/init/main.c
> > > index a48617f2e5e5..7b43118215d6 100644
> > > --- a/init/main.c
> > > +++ b/init/main.c
> > > @@ -1372,6 +1372,8 @@ static int __ref kernel_init(void *unused)
> > >  
> > >  	rcu_end_inkernel_boot();
> > >  
> > > +	do_sysctl_args();
> > > +
> > 
> > Ah, I see. Since the sysctl is designed to be called after all __init calls were
> > done, it shouldn't use bootconfig directly because bootconfig is full of __init
> > call.
> 
> The idea is bootconfig would be useful in the sense of a library set of
> helpers which could be modified to remove __init, and then used to
> instrument the cmdline depending on certain debugging kconfig entries.

Would you mean making bootconfig (parser and APIs) be more generic so that
other subsystem can reuse it with their data?
Or just make it available after boot? (I think this latter one will be
useful for module initialization)

> We currently have no way to purposely extend / break the cmdline for
> debugging purposes, so, bootconfig's parsers, since it already has a
> way to extend the cmdlineline, might make it much easier to do this
> later.
> 
> Without bootconfig, if we wanted to add new kconfig to, for example,
> add new funny cmdline arguments to test they worked or not, we'd have
> to devise our own set of helpers now. ie, new functionality. bootconfig
> however already has existing functionality to tweak the cmdline, and so
> some code could be leveraged there for this purpose.

Hmm, you can use the bootconfig as a "supplemental" kernel command line,
but not tweak (like modify/replace) it. Would you like to change the
kernel command line parameter on-line?

Thank you,

Luis Chamberlain April 16, 2020, 3:53 p.m. UTC | #5

On Thu, Apr 16, 2020 at 07:49:55PM +0900, Masami Hiramatsu wrote:
> Hi Luis,
> 
> On Thu, 16 Apr 2020 01:29:31 +0000
> Luis Chamberlain <[e-mail]> wrote:
> 
> > On Wed, Apr 15, 2020 at 06:03:55PM +0900, Masami Hiramatsu wrote:
> > > On Tue, 14 Apr 2020 13:32:20 +0200
> > > Vlastimil Babka <[e-mail]> wrote:
> > > > diff --git a/init/main.c b/init/main.c
> > > > index a48617f2e5e5..7b43118215d6 100644
> > > > --- a/init/main.c
> > > > +++ b/init/main.c
> > > > @@ -1372,6 +1372,8 @@ static int __ref kernel_init(void *unused)
> > > >  
> > > >  	rcu_end_inkernel_boot();
> > > >  
> > > > +	do_sysctl_args();
> > > > +
> > > 
> > > Ah, I see. Since the sysctl is designed to be called after all __init calls were
> > > done, it shouldn't use bootconfig directly because bootconfig is full of __init
> > > call.
> > 
> > The idea is bootconfig would be useful in the sense of a library set of
> > helpers which could be modified to remove __init, and then used to
> > instrument the cmdline depending on certain debugging kconfig entries.
> 
> Would you mean making bootconfig (parser and APIs) be more generic so that
> other subsystem can reuse it with their data?
> Or just make it available after boot? (I think this latter one will be
> useful for module initialization)

The later. First use case that comes to mind is debugging cmdline, so
to see if what one adds is what ends up happening at run time after
boot.

> > We currently have no way to purposely extend / break the cmdline for
> > debugging purposes, so, bootconfig's parsers, since it already has a
> > way to extend the cmdlineline, might make it much easier to do this
> > later.
> > 
> > Without bootconfig, if we wanted to add new kconfig to, for example,
> > add new funny cmdline arguments to test they worked or not, we'd have
> > to devise our own set of helpers now. ie, new functionality. bootconfig
> > however already has existing functionality to tweak the cmdline, and so
> > some code could be leveraged there for this purpose.
> 
> Hmm, you can use the bootconfig as a "supplemental" kernel command line,
> but not tweak (like modify/replace) it. Would you like to change the
> kernel command line parameter on-line?

It would be during boot. To augment it as if the user had used certain
parameters on boot. But if only a new path is tested, and we can't
reproduce as if the user had *not* used bootconfig, this idea would
only be useful to test bootconfig parsing, nothing else. The hope was
to do both.

  Luis

Masami Hiramatsu (Google) April 17, 2020, 10:34 a.m. UTC | #6

On Thu, 16 Apr 2020 15:53:27 +0000
Luis Chamberlain <[e-mail]> wrote:

> On Thu, Apr 16, 2020 at 07:49:55PM +0900, Masami Hiramatsu wrote:
> > Hi Luis,
> > 
> > On Thu, 16 Apr 2020 01:29:31 +0000
> > Luis Chamberlain <[e-mail]> wrote:
> > 
> > > On Wed, Apr 15, 2020 at 06:03:55PM +0900, Masami Hiramatsu wrote:
> > > > On Tue, 14 Apr 2020 13:32:20 +0200
> > > > Vlastimil Babka <[e-mail]> wrote:
> > > > > diff --git a/init/main.c b/init/main.c
> > > > > index a48617f2e5e5..7b43118215d6 100644
> > > > > --- a/init/main.c
> > > > > +++ b/init/main.c
> > > > > @@ -1372,6 +1372,8 @@ static int __ref kernel_init(void *unused)
> > > > >  
> > > > >  	rcu_end_inkernel_boot();
> > > > >  
> > > > > +	do_sysctl_args();
> > > > > +
> > > > 
> > > > Ah, I see. Since the sysctl is designed to be called after all __init calls were
> > > > done, it shouldn't use bootconfig directly because bootconfig is full of __init
> > > > call.
> > > 
> > > The idea is bootconfig would be useful in the sense of a library set of
> > > helpers which could be modified to remove __init, and then used to
> > > instrument the cmdline depending on certain debugging kconfig entries.
> > 
> > Would you mean making bootconfig (parser and APIs) be more generic so that
> > other subsystem can reuse it with their data?
> > Or just make it available after boot? (I think this latter one will be
> > useful for module initialization)
> 
> The later. First use case that comes to mind is debugging cmdline, so
> to see if what one adds is what ends up happening at run time after
> boot.

Hmm, I think that's not so easy to debug command line after boot, because
the kernel command line is parsed (and handlers are executed) already in
boot time. We can not repeat it after boot.

> > > We currently have no way to purposely extend / break the cmdline for
> > > debugging purposes, so, bootconfig's parsers, since it already has a
> > > way to extend the cmdlineline, might make it much easier to do this
> > > later.
> > > 
> > > Without bootconfig, if we wanted to add new kconfig to, for example,
> > > add new funny cmdline arguments to test they worked or not, we'd have
> > > to devise our own set of helpers now. ie, new functionality. bootconfig
> > > however already has existing functionality to tweak the cmdline, and so
> > > some code could be leveraged there for this purpose.
> > 
> > Hmm, you can use the bootconfig as a "supplemental" kernel command line,
> > but not tweak (like modify/replace) it. Would you like to change the
> > kernel command line parameter on-line?
> 
> It would be during boot. To augment it as if the user had used certain
> parameters on boot. But if only a new path is tested, and we can't
> reproduce as if the user had *not* used bootconfig, this idea would
> only be useful to test bootconfig parsing, nothing else. The hope was
> to do both.

As you may know, the bootconfig already supports "additional" kernel
command line. All keys which starts "kernel" is copied into kernel
command line at early boot timing. So if you want to write a test
parameter in the bootconfig, you can do it.

However, it is not a good idea to execute command line handlers
twice because it can be destructive or can append all parameters
(e.g. "console=" .)

For the new feature can natively use the bootconfig, for example
boot-time tracing (kernel/trace/trace_boot.c) is something like
this sysctl on boot, and natively uses the bootconfig because
the tracing parameter is too complex for kernel command line :)

Thank you,

Luis Chamberlain April 17, 2020, 4:59 p.m. UTC | #7

On Fri, Apr 17, 2020 at 07:34:42PM +0900, Masami Hiramatsu wrote:
> On Thu, 16 Apr 2020 15:53:27 +0000
> Luis Chamberlain <[e-mail]> wrote:
> 
> > On Thu, Apr 16, 2020 at 07:49:55PM +0900, Masami Hiramatsu wrote:
> > > Hi Luis,
> > > 
> > > On Thu, 16 Apr 2020 01:29:31 +0000
> > > Luis Chamberlain <[e-mail]> wrote:
> > > 
> > > > On Wed, Apr 15, 2020 at 06:03:55PM +0900, Masami Hiramatsu wrote:
> > > > > On Tue, 14 Apr 2020 13:32:20 +0200
> > > > > Vlastimil Babka <[e-mail]> wrote:
> > > > > > diff --git a/init/main.c b/init/main.c
> > > > > > index a48617f2e5e5..7b43118215d6 100644
> > > > > > --- a/init/main.c
> > > > > > +++ b/init/main.c
> > > > > > @@ -1372,6 +1372,8 @@ static int __ref kernel_init(void *unused)
> > > > > >  
> > > > > >  	rcu_end_inkernel_boot();
> > > > > >  
> > > > > > +	do_sysctl_args();
> > > > > > +
> > > > > 
> > > > > Ah, I see. Since the sysctl is designed to be called after all __init calls were
> > > > > done, it shouldn't use bootconfig directly because bootconfig is full of __init
> > > > > call.
> > > > 
> > > > The idea is bootconfig would be useful in the sense of a library set of
> > > > helpers which could be modified to remove __init, and then used to
> > > > instrument the cmdline depending on certain debugging kconfig entries.
> > > 
> > > Would you mean making bootconfig (parser and APIs) be more generic so that
> > > other subsystem can reuse it with their data?
> > > Or just make it available after boot? (I think this latter one will be
> > > useful for module initialization)
> > 
> > The later. First use case that comes to mind is debugging cmdline, so
> > to see if what one adds is what ends up happening at run time after
> > boot.
> 
> Hmm, I think that's not so easy to debug command line after boot, because
> the kernel command line is parsed (and handlers are executed) already in
> boot time. We can not repeat it after boot.

Unless you put into the command line everything you need to test on one
boot.

> > > > We currently have no way to purposely extend / break the cmdline for
> > > > debugging purposes, so, bootconfig's parsers, since it already has a
> > > > way to extend the cmdlineline, might make it much easier to do this
> > > > later.
> > > > 
> > > > Without bootconfig, if we wanted to add new kconfig to, for example,
> > > > add new funny cmdline arguments to test they worked or not, we'd have
> > > > to devise our own set of helpers now. ie, new functionality. bootconfig
> > > > however already has existing functionality to tweak the cmdline, and so
> > > > some code could be leveraged there for this purpose.
> > > 
> > > Hmm, you can use the bootconfig as a "supplemental" kernel command line,
> > > but not tweak (like modify/replace) it. Would you like to change the
> > > kernel command line parameter on-line?
> > 
> > It would be during boot. To augment it as if the user had used certain
> > parameters on boot. But if only a new path is tested, and we can't
> > reproduce as if the user had *not* used bootconfig, this idea would
> > only be useful to test bootconfig parsing, nothing else. The hope was
> > to do both.
> 
> As you may know, the bootconfig already supports "additional" kernel
> command line. All keys which starts "kernel" is copied into kernel
> command line at early boot timing. So if you want to write a test
> parameter in the bootconfig, you can do it.

There are two bootparams paths now, the old way, and the new bootconfig
path. Extending test coverage to test bootconfig seems rather easier to
consider. However the hope was that there may be some existing code
within bootconfig which would also allow one to test the old cmdline
path, as if the cmdline had certain params present. It doesn't seem
to be the case.

> However, it is not a good idea to execute command line handlers
> twice because it can be destructive or can append all parameters
> (e.g. "console=" .)

I see so bootconfig is mutually exclusive with the old cmdline?

> For the new feature can natively use the bootconfig, for example
> boot-time tracing (kernel/trace/trace_boot.c) is something like
> this sysctl on boot, and natively uses the bootconfig because
> the tracing parameter is too complex for kernel command line :)

Neat.

  Luis

Masami Hiramatsu (Google) April 18, 2020, 10:12 a.m. UTC | #8

On Fri, 17 Apr 2020 16:59:39 +0000
Luis Chamberlain <[e-mail]> wrote:

> > > > > > Ah, I see. Since the sysctl is designed to be called after all __init calls were
> > > > > > done, it shouldn't use bootconfig directly because bootconfig is full of __init
> > > > > > call.
> > > > > 
> > > > > The idea is bootconfig would be useful in the sense of a library set of
> > > > > helpers which could be modified to remove __init, and then used to
> > > > > instrument the cmdline depending on certain debugging kconfig entries.
> > > > 
> > > > Would you mean making bootconfig (parser and APIs) be more generic so that
> > > > other subsystem can reuse it with their data?
> > > > Or just make it available after boot? (I think this latter one will be
> > > > useful for module initialization)
> > > 
> > > The later. First use case that comes to mind is debugging cmdline, so
> > > to see if what one adds is what ends up happening at run time after
> > > boot.
> > 
> > Hmm, I think that's not so easy to debug command line after boot, because
> > the kernel command line is parsed (and handlers are executed) already in
> > boot time. We can not repeat it after boot.
> 
> Unless you put into the command line everything you need to test on one
> boot.

I got it. But it requires completely another thing, because the command line
handlers are __init function, which is also released after boot, and bootconfig
is not designed for update (read only) so that we can remove locks.

> 
> > > > > We currently have no way to purposely extend / break the cmdline for
> > > > > debugging purposes, so, bootconfig's parsers, since it already has a
> > > > > way to extend the cmdlineline, might make it much easier to do this
> > > > > later.
> > > > > 
> > > > > Without bootconfig, if we wanted to add new kconfig to, for example,
> > > > > add new funny cmdline arguments to test they worked or not, we'd have
> > > > > to devise our own set of helpers now. ie, new functionality. bootconfig
> > > > > however already has existing functionality to tweak the cmdline, and so
> > > > > some code could be leveraged there for this purpose.
> > > > 
> > > > Hmm, you can use the bootconfig as a "supplemental" kernel command line,
> > > > but not tweak (like modify/replace) it. Would you like to change the
> > > > kernel command line parameter on-line?
> > > 
> > > It would be during boot. To augment it as if the user had used certain
> > > parameters on boot. But if only a new path is tested, and we can't
> > > reproduce as if the user had *not* used bootconfig, this idea would
> > > only be useful to test bootconfig parsing, nothing else. The hope was
> > > to do both.
> > 
> > As you may know, the bootconfig already supports "additional" kernel
> > command line. All keys which starts "kernel" is copied into kernel
> > command line at early boot timing. So if you want to write a test
> > parameter in the bootconfig, you can do it.
> 
> There are two bootparams paths now, the old way, and the new bootconfig
> path. Extending test coverage to test bootconfig seems rather easier to
> consider. However the hope was that there may be some existing code
> within bootconfig which would also allow one to test the old cmdline
> path, as if the cmdline had certain params present. It doesn't seem
> to be the case.

Hmm, maybe you misunderstand how the bootconfig passes the options
to command line. At the very early timing, it copies the kernel
boot options into the existing legacy kernel command line. See
setup_command_line()@init/main.c. The extra_command_line is what the
bootconfig generated.

This extra_command_line is mixed with the old cmdline parameters
and it is passed to the handlers via parse_args() on each init-level.
(some handlers just set the option to its own variables and
 use those variables in their __init functions)

On the other hand, new code which uses the bootconfig, will use the
bootconfig APIs (xbc_*) to find their related options in its __init
functions.

> > However, it is not a good idea to execute command line handlers
> > twice because it can be destructive or can append all parameters
> > (e.g. "console=" .)
> 
> I see so bootconfig is mutually exclusive with the old cmdline?

If you put options under "kernel" or "init", those are inclusive to the
old cmdline, but other keys are not visible from the old cmdline.

Thank you,

> 
> > For the new feature can natively use the bootconfig, for example
> > boot-time tracing (kernel/trace/trace_boot.c) is something like
> > this sysctl on boot, and natively uses the bootconfig because
> > the tracing parameter is too complex for kernel command line :)
> 
> Neat.
> 
>   Luis

Patch

diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index f2a93c8679e8..fd38f1e8a2bf 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -4891,6 +4891,15 @@ 
 
 	switches=	[HW,M68k]
 
+	sysctl.*=	[KNL]
+			Set a sysctl parameter, right before loading the init
+			process, as if the value was written to the respective
+			/proc/sys/... file. Both '.' and '/' are recognized as
+			separators. Unrecognized parameters and invalid values
+			are reported in the kernel log. Sysctls registered
+			later by a loaded module cannot be set this way.
+			Example: sysctl.vm.swappiness=40
+
 	sysfs.deprecated=0|1 [KNL]
 			Enable/disable old style sysfs layout for old udev
 			on older distributions. When this option is enabled
diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c
index b6f5d459b087..7804da5f5be0 100644
--- a/fs/proc/proc_sysctl.c
+++ b/fs/proc/proc_sysctl.c
@@ -14,6 +14,7 @@ 
 #include <linux/mm.h>
 #include <linux/module.h>
 #include <linux/bpf-cgroup.h>
+#include <linux/mount.h>
 #include "internal.h"
 
 static const struct dentry_operations proc_sys_dentry_operations;
@@ -1692,3 +1693,109 @@  int __init proc_sys_init(void)
 
 	return sysctl_init();
 }
+
+/* Set sysctl value passed on kernel command line. */
+static int process_sysctl_arg(char *param, char *val,
+			       const char *unused, void *arg)
+{
+	char *path;
+	struct vfsmount **proc_mnt = arg;
+	struct file_system_type *proc_fs_type;
+	struct file *file;
+	int len;
+	int err;
+	loff_t pos = 0;
+	ssize_t wret;
+
+	if (strncmp(param, "sysctl", sizeof("sysctl") - 1))
+		return 0;
+
+	param += sizeof("sysctl") - 1;
+
+	if (param[0] != '/' && param[0] != '.')
+		return 0;
+
+	param++;
+
+	/*
+	 * To set sysctl options, we use a temporary mount of proc, look up the
+	 * respective sys/ file and write to it. To avoid mounting it when no
+	 * options were given, we mount it only when the first sysctl option is
+	 * found. Why not a persistent mount? There are problems with a
+	 * persistent mount of proc in that it forces userspace not to use any
+	 * proc mount options.
+	 */
+	if (!*proc_mnt) {
+		proc_fs_type = get_fs_type("proc");
+		if (!proc_fs_type) {
+			pr_err("Failed to find procfs to set sysctl from command line\n");
+			return 0;
+		}
+		*proc_mnt = kern_mount(proc_fs_type);
+		put_filesystem(proc_fs_type);
+		if (IS_ERR(*proc_mnt)) {
+			pr_err("Failed to mount procfs to set sysctl from command line\n");
+			return 0;
+		}
+	}
+
+	path = kasprintf(GFP_KERNEL, "sys/%s", param);
+	if (!path)
+		panic("%s: Failed to allocate path for %s\n", __func__, param);
+	strreplace(path, '.', '/');
+
+	file = file_open_root((*proc_mnt)->mnt_root, *proc_mnt, path, O_WRONLY, 0);
+	if (IS_ERR(file)) {
+		err = PTR_ERR(file);
+		if (err == -ENOENT)
+			pr_err("Failed to set sysctl parameter '%s=%s': parameter not found\n",
+				param, val);
+		else if (err == -EACCES)
+			pr_err("Failed to set sysctl parameter '%s=%s': permission denied (read-only?)\n",
+				param, val);
+		else
+			pr_err("Error %pe opening proc file to set sysctl parameter '%s=%s'\n",
+				file, param, val);
+		goto out;
+	}
+	len = strlen(val);
+	wret = kernel_write(file, val, len, &pos);
+	if (wret < 0) {
+		err = wret;
+		if (err == -EINVAL)
+			pr_err("Failed to set sysctl parameter '%s=%s': invalid value\n",
+				param, val);
+		else
+			pr_err("Error %pe writing to proc file to set sysctl parameter '%s=%s'\n",
+				ERR_PTR(err), param, val);
+	} else if (wret != len) {
+		pr_err("Wrote only %zd bytes of %d writing to proc file %s to set sysctl parameter '%s=%s\n",
+			wret, len, path, param, val);
+	}
+
+	err = filp_close(file, NULL);
+	if (err)
+		pr_err("Error %pe closing proc file to set sysctl parameter '%s=%s\n",
+			ERR_PTR(err), param, val);
+out:
+	kfree(path);
+	return 0;
+}
+
+void do_sysctl_args(void)
+{
+	char *command_line;
+	struct vfsmount *proc_mnt = NULL;
+
+	command_line = kstrdup(saved_command_line, GFP_KERNEL);
+	if (!command_line)
+		panic("%s: Failed to allocate copy of command line\n", __func__);
+
+	parse_args("Setting sysctl args", command_line,
+		   NULL, 0, -1, -1, &proc_mnt, process_sysctl_arg);
+
+	if (proc_mnt)
+		kern_unmount(proc_mnt);
+
+	kfree(command_line);
+}
diff --git a/include/linux/sysctl.h b/include/linux/sysctl.h
index 02fa84493f23..bee62d30ebdb 100644
--- a/include/linux/sysctl.h
+++ b/include/linux/sysctl.h
@@ -206,6 +206,7 @@  struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
 void unregister_sysctl_table(struct ctl_table_header * table);
 
 extern int sysctl_init(void);
+void do_sysctl_args(void);
 
 extern struct ctl_table sysctl_mount_point[];
 
@@ -236,6 +237,9 @@  static inline void setup_sysctl_set(struct ctl_table_set *p,
 {
 }
 
+static inline void do_sysctl_args(void)
+{
+}
 #endif /* CONFIG_SYSCTL */
 
 int sysctl_max_threads(struct ctl_table *table, int write,
diff --git a/init/main.c b/init/main.c
index a48617f2e5e5..7b43118215d6 100644
--- a/init/main.c
+++ b/init/main.c
@@ -1372,6 +1372,8 @@  static int __ref kernel_init(void *unused)
 
 	rcu_end_inkernel_boot();
 
+	do_sysctl_args();
+
 	if (ramdisk_execute_command) {
 		ret = run_init_process(ramdisk_execute_command);
 		if (!ret)