[PATCHv8,23/34] time: Allocate per-timens vvar page

Submitted by Jann Horn via Containers on Nov. 12, 2019, 1:27 a.m.

Details

Message ID 20191112012724.250792-24-dima@arista.com
State New
Series "kernel: Introduce Time Namespace"
Headers show

Commit Message

Jann Horn via Containers Nov. 12, 2019, 1:27 a.m.
VDSO support for Time namespace needs to set up a page with the same
layout as VVAR. That timens page will be placed on position of VVAR page
inside namespace. That page contains time namespace clock offsets and it
has vdso_data->seq set to 1 to enforce the slow path and
vdso_data->clock_mode set to VCLOCK_TIMENS to enforce the time namespace
handling path.

Allocate the timens page during namespace creation. Setup the offsets
when the first task enters the ns and freeze them to guarantee the pace
of monotonic/boottime clocks and to avoid breakage of applications.

The design decision is to have a global offset_lock which is used
during namespace offsets set up and to freeze offsets when first task
joins the new ns. That is better in terms of memory usage comparing
to having per-ns mutex that's used only during the set up period.

Suggested-by: Andy Lutomirski <luto@kernel.org>
Based-on-work-by: Thomas Gleixner <tglx@linutronix.de>
Co-developed-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
---
 include/linux/time_namespace.h |   3 +
 kernel/time/namespace.c        | 103 ++++++++++++++++++++++++++++++++-
 2 files changed, 105 insertions(+), 1 deletion(-)

Patch hide | download patch | download mbox

diff --git a/include/linux/time_namespace.h b/include/linux/time_namespace.h
index 90e82654b4ce..cdb438ba0037 100644
--- a/include/linux/time_namespace.h
+++ b/include/linux/time_namespace.h
@@ -23,6 +23,9 @@  struct time_namespace {
 	struct ucounts *ucounts;
 	struct ns_common ns;
 	struct timens_offsets offsets;
+	struct page *vvar_page;
+	/* Disallow changing offsets after any task joined namespace. */
+	bool frozen_offsets;
 } __randomize_layout;
 extern struct time_namespace init_time_ns;
 
diff --git a/kernel/time/namespace.c b/kernel/time/namespace.c
index 1a0fbaa5d2d4..e14cd1ca387d 100644
--- a/kernel/time/namespace.c
+++ b/kernel/time/namespace.c
@@ -15,6 +15,7 @@ 
 #include <linux/cred.h>
 #include <linux/err.h>
 #include <linux/mm.h>
+#include <vdso/datapage.h>
 
 ktime_t do_timens_ktime_to_host(clockid_t clockid, ktime_t tim,
 				struct timens_offsets *ns_offsets)
@@ -90,16 +91,23 @@  static struct time_namespace *clone_time_ns(struct user_namespace *user_ns,
 
 	kref_init(&ns->kref);
 
+	ns->vvar_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
+	if (!ns->vvar_page)
+		goto fail_free;
+
 	err = ns_alloc_inum(&ns->ns);
 	if (err)
-		goto fail_free;
+		goto fail_free_page;
 
 	ns->ucounts = ucounts;
 	ns->ns.ops = &timens_operations;
 	ns->user_ns = get_user_ns(user_ns);
 	ns->offsets = old_ns->offsets;
+	ns->frozen_offsets = false;
 	return ns;
 
+fail_free_page:
+	__free_page(ns->vvar_page);
 fail_free:
 	kfree(ns);
 fail_dec:
@@ -128,6 +136,93 @@  struct time_namespace *copy_time_ns(unsigned long flags,
 	return clone_time_ns(user_ns, old_ns);
 }
 
+static struct timens_offset offset_from_ts(struct timespec64 off)
+{
+	struct timens_offset ret;
+
+	ret.sec = off.tv_sec;
+	ret.nsec = off.tv_nsec;
+
+	return ret;
+}
+
+/*
+ * A time namespace VVAR page has the same layout as the VVAR page which
+ * contains the system wide VDSO data.
+ *
+ * For a normal task the VVAR pages are installed in the normal ordering:
+ *     VVAR
+ *     PVCLOCK
+ *     HVCLOCK
+ *     TIMENS   <- Not really required
+ *
+ * Now for a timens task the pages are installed in the following order:
+ *     TIMENS
+ *     PVCLOCK
+ *     HVCLOCK
+ *     VVAR
+ *
+ * The check for vdso_data->clock_mode is in the unlikely path of
+ * the seq begin magic. So for the non-timens case most of the time
+ * 'seq' is even, so the branch is not taken.
+ *
+ * If 'seq' is odd, i.e. a concurrent update is in progress, the extra check
+ * for vdso_data->clock_mode is a non-issue. The task is spin waiting for the
+ * update to finish and for 'seq' to become even anyway.
+ *
+ * Timens page has vdso_data->clock_mode set to VCLOCK_TIMENS which enforces
+ * the time namespace handling path.
+ */
+static void timens_setup_vdso_data(struct vdso_data *vdata,
+				   struct time_namespace *ns)
+{
+	struct timens_offset *offset = vdata->offset;
+	struct timens_offset monotonic = offset_from_ts(ns->offsets.monotonic);
+	struct timens_offset boottime = offset_from_ts(ns->offsets.boottime);
+
+	vdata->seq			= 1;
+	vdata->clock_mode		= VCLOCK_TIMENS;
+	offset[CLOCK_MONOTONIC]		= monotonic;
+	offset[CLOCK_MONOTONIC_RAW]	= monotonic;
+	offset[CLOCK_MONOTONIC_COARSE]	= monotonic;
+	offset[CLOCK_BOOTTIME]		= boottime;
+	offset[CLOCK_BOOTTIME_ALARM]	= boottime;
+}
+
+/*
+ * Protects possibly multiple offsets writers racing each other
+ * and tasks entering the namespace.
+ */
+static DEFINE_MUTEX(offset_lock);
+
+static void timens_set_vvar_page(struct task_struct *task,
+				struct time_namespace *ns)
+{
+	struct vdso_data *vdata;
+	unsigned int i;
+
+	if (ns == &init_time_ns)
+		return;
+
+	/* Fast-path, taken by every task in namespace except the first. */
+	if (likely(ns->frozen_offsets))
+		return;
+
+	mutex_lock(&offset_lock);
+	/* Nothing to-do: vvar_page has been already initialized. */
+	if (ns->frozen_offsets)
+		goto out;
+
+	ns->frozen_offsets = true;
+	vdata = arch_get_vdso_data(page_address(ns->vvar_page));
+
+	for (i = 0; i < CS_BASES; i++)
+		timens_setup_vdso_data(&vdata[i], ns);
+
+out:
+	mutex_unlock(&offset_lock);
+}
+
 void free_time_ns(struct kref *kref)
 {
 	struct time_namespace *ns;
@@ -136,6 +231,7 @@  void free_time_ns(struct kref *kref)
 	dec_time_namespaces(ns->ucounts);
 	put_user_ns(ns->user_ns);
 	ns_free_inum(&ns->ns);
+	__free_page(ns->vvar_page);
 	kfree(ns);
 }
 
@@ -192,6 +288,8 @@  static int timens_install(struct nsproxy *nsproxy, struct ns_common *new)
 	    !ns_capable(current_user_ns(), CAP_SYS_ADMIN))
 		return -EPERM;
 
+	timens_set_vvar_page(current, ns);
+
 	get_time_ns(ns);
 	put_time_ns(nsproxy->time_ns);
 	nsproxy->time_ns = ns;
@@ -211,6 +309,8 @@  int timens_on_fork(struct nsproxy *nsproxy, struct task_struct *tsk)
 	if (nsproxy->time_ns == nsproxy->time_ns_for_children)
 		return 0;
 
+	timens_set_vvar_page(tsk, ns);
+
 	get_time_ns(ns);
 	put_time_ns(nsproxy->time_ns);
 	nsproxy->time_ns = ns;
@@ -246,6 +346,7 @@  struct time_namespace init_time_ns = {
 	.user_ns	= &init_user_ns,
 	.ns.inum	= PROC_TIME_INIT_INO,
 	.ns.ops		= &timens_operations,
+	.frozen_offsets	= true,
 };
 
 static int __init time_ns_init(void)