Re: [PATCH net-next v4] rds: Expose feature parameters via sysfs (and ELF)

[Allison Henderson <allison.henderson@oracle.com>]

On Fri, 2025-06-20 at 20:26 -0400, Konrad Rzeszutek Wilk wrote:
> We would like to have a programmatic way for applications
> to query which of the features defined in include/uapi/linux/rds.h
> are actually implemented by the kernel.
> 
> The problem is that applications can be built against newer
> kernel (or older) and they may have the feature implemented or not.
> 
> The lack of a certain feature would signify that the kernel
> does not support it. The presence of it signifies the existence
> of it.
> 
> This would provide the application to query the sysfs and figure
> out what is supported on a running system.
> 
> This patch would expose this extra sysfs file:
> 
> /sys/kernel/rds/features/ioctl_get_tos
> /sys/kernel/rds/features/ioctl_set_tos
> /sys/kernel/rds/features/socket_cancel_sent_to
> /sys/kernel/rds/features/socket_cong_monitor
> /sys/kernel/rds/features/socket_free_mr
> /sys/kernel/rds/features/socket_get_mr
> /sys/kernel/rds/features/socket_get_mr_for_dest
> /sys/kernel/rds/features/socket_recverr
> /sys/kernel/rds/features/socket_so_rxpath_latency
> /sys/kernel/rds/features/socket_so_transport
> 
> With the value of 'supported' in them. In the future this value
> could change to say 'deprecated' or have other values (for example
> different versions) or can be runtime changed.
> 
> The choice to use sysfs and this particular way is modeled on the
> filesystems usage exposing their features.
> 
> Alternative solution such as exposing one file ('features') with
> each feature enumerated (which cgroup does) is a bit limited in
> that it does not provide means to provide extra content in the future
> for each feature. For example if one of the features had three
> modes and one wanted to set a particular one at runtime - that
> does not exist in cgroup (albeit it can be implemented but it would
> be quite hectic to have just one single attribute).
> 
> Another solution of using an ioctl to expose a bitmask has the
> disadvantage of being less flexible in the future and while it can
> have a bit of supported/unsupported, it is not clear how one would
> change modes or expose versions. It is most certainly feasible
> but it can get seriously complex fast.
> 
> As such this mechanism offers the basic support we require
> now and offers the flexibility for the future.
> 
> Lastly, we also utilize the ELF note macro to expose these via
> so that applications that have not yet initialized RDS transport
> can inspect the kernel module to see if they have the appropiate
> support and choose an alternative protocol if they wish so.
> 
> Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> ---
> v3: Add the missing documentation
>     Remove the CONFIG_SYSFS #ifdef machinations
>     Redo it with each feature as a seperate file in 'features'
>     directory.
> v4: Follow the syntax in Documentation/ABI/stable/sysfs-module
>     Tack on the ret value if we fail kobject_create_and_add
> ---
>  Documentation/ABI/stable/sysfs-transport-rds | 43 +++++++++++++
>  net/rds/af_rds.c                             | 63 +++++++++++++++++++-
>  2 files changed, 105 insertions(+), 1 deletion(-)
>  create mode 100644 Documentation/ABI/stable/sysfs-transport-rds
> 
> diff --git a/Documentation/ABI/stable/sysfs-transport-rds b/Documentation/ABI/stable/sysfs-transport-rds
> new file mode 100644
> index 000000000000..749a043d4e88
> --- /dev/null
> +++ b/Documentation/ABI/stable/sysfs-transport-rds
> @@ -0,0 +1,43 @@
> +What:		/sys/kernel/rds/features/*
> +Date:		June 2025
> +KernelVersion:	6.17
> +Contact:	rds-devel@oss.oracle.com
> +Description:	This directory contains the features that this kernel
> +		has been built with and supports. They correspond
> +		to the include/uapi/linux/rds.h features.
> +
> +		The intent is for applications compiled against rds.h
> +		to be able to query and find out what features the
> +		driver supports. The current expected value is 'supported'.
> +
> +What:		/sys/kernel/rds/features/ioctl_[get,set]_tos
> +Description:	Allows the user to set on the socket a type of
> +		service(tos) value associated forever.
> +
> +What:		/sys/kernel/rds/features/socket_cancel_sent_to
> +Description:	Allows to cancel all pending messages to a given destination.
> +
> +What:		/sys/kernel/rds/features/socket_cong_monitor
> +Description:	RDS provides explicit congestion monitoring for a socket using
> +		a 64-bit mask. Each bit in the mask corresponds to a group of ports.
> +
> +		When a congestion update arrives, RDS checks the set of ports
> +		that became uncongested against the bit mask.
> +
> +		If they overlap, a control messages is enqueued on the socket,
> +		and the application is woken up.
> +
> +What:		/sys/kernel/rds/features/socket_[get_mr.get_mr_for_dest,free_mr]

One last nit here: comma between get_mr and get_mr_for_des

Other than that, this version looks good to me!  Thank you for the update!  What that fix, you can add my rvb:
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>

Thank you!
Allison


> +Description:	RDS allows a process to register or release memory ranges for
> +		RDMA.
> +
> +What:		/sys/kernel/rds/features/socket_recverr
> +Description:	RDS will send RDMA notification messages to the application for
> +		any RDMA operation that fails. By default this is off.
> +
> +What:		/sys/kernel/rds/features/socket_so_rxpath_latency
> +Description:	Receive path latency in various stages of receive path.
> +
> +What:		/sys/kernel/rds/features/socket_so_transport
> +Description:	Attach the socket to the underlaying transport (TCP, RDMA
> +		or loop) before invoking bind on the socket.
> diff --git a/net/rds/af_rds.c b/net/rds/af_rds.c
> index 8435a20968ef..f920cc6f0b2b 100644
> --- a/net/rds/af_rds.c
> +++ b/net/rds/af_rds.c
> @@ -32,7 +32,9 @@
>   */
>  #include <linux/module.h>
>  #include <linux/errno.h>
> +#include <linux/elfnote.h>
>  #include <linux/kernel.h>
> +#include <linux/kobject.h>
>  #include <linux/gfp.h>
>  #include <linux/in.h>
>  #include <linux/ipv6.h>
> @@ -871,6 +873,53 @@ static void rds6_sock_info(struct socket *sock, unsigned int len,
>  }
>  #endif
>  
> +static ssize_t supported_show(struct kobject *kobj, struct kobj_attribute *attr,
> +			     char *buf)
> +{
> +	return sysfs_emit(buf, "supported\n");
> +}
> +
> +#define SYSFS_ATTR(_name)					\
> +ELFNOTE64("rds." #_name, 0, 1);				\
> +static struct kobj_attribute rds_attr_##_name = {		\
> +	.attr = {.name = __stringify(_name), .mode = 0444 },	\
> +	.show = supported_show,					\
> +}
> +
> +SYSFS_ATTR(ioctl_set_tos);
> +SYSFS_ATTR(ioctl_get_tos);
> +SYSFS_ATTR(socket_cancel_sent_to);
> +SYSFS_ATTR(socket_cong_monitor);
> +SYSFS_ATTR(socket_get_mr);
> +SYSFS_ATTR(socket_get_mr_for_dest);
> +SYSFS_ATTR(socket_free_mr);
> +SYSFS_ATTR(socket_recverr);
> +SYSFS_ATTR(socket_so_rxpath_latency);
> +SYSFS_ATTR(socket_so_transport);
> +
> +#define ATTR_LIST(_name) &rds_attr_##_name.attr
> +
> +static struct attribute *rds_feat_attrs[] = {
> +	ATTR_LIST(ioctl_set_tos),
> +	ATTR_LIST(ioctl_get_tos),
> +	ATTR_LIST(socket_cancel_sent_to),
> +	ATTR_LIST(socket_cong_monitor),
> +	ATTR_LIST(socket_get_mr),
> +	ATTR_LIST(socket_get_mr_for_dest),
> +	ATTR_LIST(socket_free_mr),
> +	ATTR_LIST(socket_recverr),
> +	ATTR_LIST(socket_so_rxpath_latency),
> +	ATTR_LIST(socket_so_transport),
> +	NULL,
> +};
> +
> +static const struct attribute_group rds_feat_group = {
> +	.attrs = rds_feat_attrs,
> +	.name = "features",
> +};
> +
> +static struct kobject *rds_sysfs_kobj;
> +
>  static void rds_exit(void)
>  {
>  	sock_unregister(rds_family_ops.family);
> @@ -882,6 +931,8 @@ static void rds_exit(void)
>  	rds_stats_exit();
>  	rds_page_exit();
>  	rds_bind_lock_destroy();
> +	sysfs_remove_group(rds_sysfs_kobj, &rds_feat_group);
> +	kobject_put(rds_sysfs_kobj);
>  	rds_info_deregister_func(RDS_INFO_SOCKETS, rds_sock_info);
>  	rds_info_deregister_func(RDS_INFO_RECV_MESSAGES, rds_sock_inc_info);
>  #if IS_ENABLED(CONFIG_IPV6)
> @@ -923,6 +974,15 @@ static int __init rds_init(void)
>  	if (ret)
>  		goto out_proto;
>  
> +	rds_sysfs_kobj = kobject_create_and_add("rds", kernel_kobj);
> +	if (!rds_sysfs_kobj) {
> +		ret = -ENOMEM;
> +		goto out_proto;
> +	}
> +	ret = sysfs_create_group(rds_sysfs_kobj, &rds_feat_group);
> +	if (ret)
> +		goto out_kobject;
> +
>  	rds_info_register_func(RDS_INFO_SOCKETS, rds_sock_info);
>  	rds_info_register_func(RDS_INFO_RECV_MESSAGES, rds_sock_inc_info);
>  #if IS_ENABLED(CONFIG_IPV6)
> @@ -931,7 +991,8 @@ static int __init rds_init(void)
>  #endif
>  
>  	goto out;
> -
> +out_kobject:
> +	kobject_put(rds_sysfs_kobj);
>  out_proto:
>  	proto_unregister(&rds_proto);
>  out_stats: