Skip to contentSkip to conversation

drivers/xen/hypervisor: Expose Xen SIF flags to userspace

PATCH Superseded Per Bilse 2 participants 3 messages Dec 2, 2022 – Dec 12, 2022 Xen hypervisor interface

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Per Bilse

author
#0 lore raw

PATCH 1/1 drivers/xen/hypervisor: Expose Xen SIF flags to userspace

/proc/xen is a legacy pseudo filesystem which predates Xen support getting merged into Linux. It has largely been replaced with more normal locations for data (/sys/hypervisor/ for info, /dev/xen/ for user devices). We want to compile xenfs support out of the dom0 kernel.

There is one item which only exists in /proc/xen, namely /proc/xen/capabilities with "control_d" being the signal of "you're in the control domain". This ultimately comes from the SIF flags provided at VM start.

This patch exposes all SIF flags in /sys/hypervisor/start_flags/ as boolean files, one for each bit, returning '1' if set, '0' otherwise. Two known flags, 'privileged' and 'initdomain', are explicitly named, and all remaining flags can be accessed via generically named files, as suggested by Andrew Cooper.

This patch replaces previous suggestion for SIF flags exposure in its entirety.

---
 Documentation/ABI/stable/sysfs-hypervisor-xen |  8 +++
 drivers/xen/sys-hypervisor.c                  | 70 +++++++++++++++++--
 2 files changed, 74 insertions(+), 4 deletions(-)
Documentation/ABI/stable/sysfs-hypervisor-xen +8 -0
index 748593c64568..f52f98548184 100644
@@ -120,3 +120,11 @@ Contact:	[email protected]
 Description:	If running under Xen:
 		The Xen version is in the format <major>.<minor><extra>
 		This is the <minor> part of it.
+
+What:		/sys/hypervisor/start_flags/*
+Date:		December 2022
+KernelVersion:	6.1.0
+Contact:	[email protected]
+Description:	If running under Xen:
+		All bits in Xen's start-flags are represented as
+		boolean files, returning '1' if set, '0' otherwise.
drivers/xen/sys-hypervisor.c +66 -4

97 lines not on this page · raw message

Signature · 1 line
2.30.2
maintainer
#1 lore raw
Quoted 31 lines · Per Bilse

/proc/xen is a legacy pseudo filesystem which predates Xen support getting merged into Linux. It has largely been replaced with more normal locations for data (/sys/hypervisor/ for info, /dev/xen/ for user devices). We want to compile xenfs support out of the dom0 kernel.

There is one item which only exists in /proc/xen, namely /proc/xen/capabilities with "control_d" being the signal of "you're in the control domain". This ultimately comes from the SIF flags provided at VM start.

This patch exposes all SIF flags in /sys/hypervisor/start_flags/ as boolean files, one for each bit, returning '1' if set, '0' otherwise. Two known flags, 'privileged' and 'initdomain', are explicitly named, and all remaining flags can be accessed via generically named files, as suggested by Andrew Cooper.

This patch replaces previous suggestion for SIF flags exposure in its entirety.

---
  Documentation/ABI/stable/sysfs-hypervisor-xen |  8 +++
  drivers/xen/sys-hypervisor.c                  | 70 +++++++++++++++++--
  2 files changed, 74 insertions(+), 4 deletions(-)
Documentation/ABI/stable/sysfs-hypervisor-xen +0 -0
index 748593c64568..f52f98548184 100644
@@ -120,3 +120,11 @@ Contact:	[email protected]
  Description:	If running under Xen:

Documentation/ABI/stable/sysfs-hypervisor-xen Contact: [email protected] · line 130

Quoted 10 lines
Documentation/ABI/stable/sysfs-hypervisor-xen +8 -0
  		The Xen version is in the format <major>.<minor><extra>
  		This is the <minor> part of it.
+
+What:		/sys/hypervisor/start_flags/*
+Date:		December 2022
+KernelVersion:	6.1.0
+Contact:	[email protected]
+Description:	If running under Xen:
+		All bits in Xen's start-flags are represented as
+		boolean files, returning '1' if set, '0' otherwise.

I think at least the files which want to be used by e.g. systemd ("initdomain" as replacement for the "control_d" string in capabilities, but I think "privileged" as well) should be explicitly added to this description, as those are meant to be used as a stable ABI.

Quoted 38 lines
drivers/xen/sys-hypervisor.c +22 -1
index fcb0792f090e..e15d3ff2c56f 100644
@@ -31,7 +31,10 @@ struct hyp_sysfs_attr {
  	struct attribute attr;
  	ssize_t (*show)(struct hyp_sysfs_attr *, char *);
  	ssize_t (*store)(struct hyp_sysfs_attr *, const char *, size_t);
-	void *hyp_attr_data;
+	union {
+		void *hyp_attr_data;
+		unsigned long hyp_attr_value;
+	};
  };
 
  static ssize_t type_show(struct hyp_sysfs_attr *attr, char *buffer)
@@ -399,6 +402,61 @@ static int __init xen_sysfs_properties_init(void)
  	return sysfs_create_group(hypervisor_kobj, &xen_properties_group);
  }
 
+#define FLAG_UNAME "unknown"
+#define FLAG_UNAME_FMT FLAG_UNAME "%02u"
+#define FLAG_UNAME_MAX sizeof(FLAG_UNAME "XX")
+#define FLAG_COUNT (sizeof(xen_start_flags) * BITS_PER_BYTE)
+static_assert(sizeof(xen_start_flags)
+		<= sizeof_field(struct hyp_sysfs_attr, hyp_attr_value));
+
+static ssize_t flag_show(struct hyp_sysfs_attr *attr, char *buffer)
+{
+	char *p = buffer;
+
+	*p++ = '0' + ((xen_start_flags & attr->hyp_attr_value) != 0);
+	*p++ = '\n';
+	return p - buffer;
+}
+
+/*
+* Add new, known flags here.  No other changes are required, but

drivers/xen/sys-hypervisor.c static int __init xen_sysfs_properties_init(void) · line 432

Quoted 10 lines
drivers/xen/sys-hypervisor.c +10 -0
+* note that each known flag wastes one entry in flag_unames[].
+* The code/complexity machinations to avoid this isn't worth it
+* for a few entries, but keep it in mind.
+*/
+static struct hyp_sysfs_attr flag_attrs[FLAG_COUNT] = {
+	[ilog2(SIF_PRIVILEGED)] = {
+		.attr = { .name = "privileged", .mode = 0444 },
+		.show = flag_show,
+		.hyp_attr_value = SIF_PRIVILEGED
+	},

What about:

#define FLAG_NODE(bit, node)					\
	[ilog2(bit)] = {					\
		.attr = { .name = #node, .mode = 0444 },	\
		.show = flag_show,				\
		.hyp_attr_value = bit				\
	}

FLAG_NODE(SIF_PRIVILEGED, privileged),
FLAG_NODE(SIF_INITDOMAIN, initdomain)

drivers/xen/sys-hypervisor.c static int __init xen_sysfs_properties_init(void) · line 437

Quoted 5 lines
drivers/xen/sys-hypervisor.c +5 -0
+	[ilog2(SIF_INITDOMAIN)] = {
+		.attr = { .name = "initdomain", .mode = 0444 },
+		.show = flag_show,
+		.hyp_attr_value = SIF_INITDOMAIN
+	}

In order to avoid a consumer having to look into the sources for any other set flag, maybe add names for currently defined flags, too? Or just skip the other flags and add a single additional node ("flags"?) with the whole value of xen_start_flags either in hex or binary form?

Please note that this is a suggestion only, I'm not insisting on it.

Juergen

author
#2 lore raw
Quoted 9 lines · Juergen Gross
Quoted 3 lines · Per Bilse

+Description:    If running under Xen:
+        All bits in Xen's start-flags are represented as +        boolean files, returning '1' if set, '0' otherwise.

I think at least the files which want to be used by e.g. systemd ("initdomain" as replacement for the "control_d" string in capabilities, but I think "privileged" as well) should be explicitly added to this description, as those are meant to be used as a stable ABI.

OK, yes, it's a bit vague as is.

Quoted 7 lines
Quoted 6 lines

+static struct hyp_sysfs_attr flag_attrs[FLAG_COUNT] = { +    [ilog2(SIF_PRIVILEGED)] = {
+        .attr = { .name = "privileged", .mode = 0444 }, +        .show = flag_show,
+        .hyp_attr_value = SIF_PRIVILEGED
+    },

Quoted 8 lines

What about:

#define FLAG_NODE(bit, node)                    \
    [ilog2(bit)] = {                    \
        .attr = { .name = #node, .mode = 0444 },    \
        .show = flag_show,                \
        .hyp_attr_value = bit                \
    }

Yes ... I was thinking that maybe flags wouldn't normally share most characteristics, but they probably do.

Quoted 2 lines

In order to avoid a consumer having to look into the sources for any other
set flag, maybe add names for currently defined flags, too? Or just skip > the other flags and add a single additional node ("flags"?) with the

whole

Quoted 1 line

value of xen_start_flags either in hex or binary form?

The current format is as suggested by Andrew Cooper in email about a previous incarnation
(https://lore.kernel.org/lkml/[email protected]/);
specifically, that of the known flags only initdomain and privileged are relevant outside a kernel context, and that all other flags be exposed individually.

Quoted 1 line

Please note that this is a suggestion only, I'm not insisting on it.

Please, many thanks for the feedback, it's always good to hear other people's thoughts.

Best,

-- Per