[PATCH v3 37/46] perf/x86/intel/cmt: add cont_monitoring to perf cgroup

[David Carrillo-Cisneros <davidcc@google.com>]

Expose the attribute intel_cmt.cont_monitoring to perf cgroups using
the newly introduced hook PERF_CGROUP_ARCH_CGRP_SUBSYS_ATTS.

The format of new attribute is a semi-colon separated list of per-package
hexadecimal flags with missing an empty string assigned to zero.

  echo "1;2" > g1/intel_cmt.cont_monitoring

Implies 0x1 for pkg 0 and 0x2 for pkg 1, 0x0 for all other pkgs.

This patch introduces the basic ideas of per-package uflags through
a cgroup attribute. The format can be changed to match Intel CAT's
schemata's file format once that is settled.

Signed-off-by: David Carrillo-Cisneros <davidcc@google.com>
---
 arch/x86/events/intel/cmt.c       | 173 ++++++++++++++++++++++++++++++++++++++
 arch/x86/include/asm/perf_event.h |  10 +++
 2 files changed, 183 insertions(+)

diff --git a/arch/x86/events/intel/cmt.c b/arch/x86/events/intel/cmt.c
index 194038b..3ade923 100644
--- a/arch/x86/events/intel/cmt.c
+++ b/arch/x86/events/intel/cmt.c
@@ -2323,4 +2323,177 @@ inline void __intel_cmt_no_event_sched_in(void)
 #endif
 }
 
+#ifdef CONFIG_CGROUP_PERF
+
+static int cmt_monitoring_seq_show(struct seq_file *m, void *v)
+{
+	struct perf_cgroup *cgrp = perf_cgroup_from_css(seq_css(m));
+	struct monr *monr = NULL;
+	int p, nr_pkgs = topology_max_packages();
+
+	mutex_lock(&cmt_mutex);
+
+	if (perf_cgroup_mon_started(cgrp))
+		monr = monr_from_perf_cgroup(cgrp);
+
+	for (p = 0; p < nr_pkgs; p++) {
+		seq_printf(m, "%x", monr ? monr->pkg_uflags[p] : 0);
+		if (p + 1 < nr_pkgs)
+			seq_puts(m, ";");
+	}
+
+	seq_puts(m, "\n");
+
+	mutex_unlock(&cmt_mutex);
+	return 0;
+}
+
+/**
+ * Parses uflags string of the form: m1;m2;...;mP where m* are hexadecimal
+ * masks and P is less or equal than topology_max_packages(). Values not
+ * provided in the mask are assumed to be zero.
+ * On success, it allocates and returns an array.
+ */
+static enum cmt_user_flags *parse_pkg_uflags(char *buf, size_t nbytes)
+{
+	enum cmt_user_flags *uflags;
+	char *local_buf, *b, *m;
+	int err = 0, pmask, nr_pkgs = topology_max_packages();
+	u16 p = 0;
+
+	uflags = kcalloc(nr_pkgs, sizeof(*uflags), GFP_KERNEL);
+	if (!uflags)
+		return ERR_PTR(-ENOMEM);
+
+	local_buf = kcalloc(nbytes, sizeof(char), GFP_KERNEL);
+	if (!local_buf) {
+		err = -ENOMEM;
+		goto error;
+	}
+	memcpy(local_buf, buf, nbytes);
+	b = local_buf;
+	while ((m = strsep(&b, ";")) != NULL) {
+		if (p >= nr_pkgs) {
+			err = -EINVAL;
+			goto error;
+		}
+		if (!*m) {
+			uflags[p] = 0;
+			continue;
+		}
+		err = kstrtoint(m, 16, &pmask);
+		if (err)
+			goto error;
+		uflags[p] = pmask;
+		if (uflags[p] > CMT_UF_MAX) {
+			err = -EINVAL;
+			goto error;
+		}
+		/*
+		 * Non-zero flags must have CMT_UF_HAS_USER set. Otherwise
+		 * there could be that monrs are not used.
+		 */
+		if (uflags[p] && (!(uflags[p] & CMT_UF_HAS_USER))) {
+			err = -EINVAL;
+			goto error;
+		}
+		p++;
+	}
+
+	kfree(local_buf);
+
+	return uflags;
+
+error:
+	kfree(local_buf);
+	kfree(uflags);
+
+	return ERR_PTR(err);
+}
+
+static ssize_t cmt_monitoring_write(struct kernfs_open_file *of,
+		char *buf, size_t nbytes, loff_t off)
+{
+	struct cgroup_subsys_state *css = of_css(of);
+	struct monr *monr;
+	int err = 0;
+	bool is_mon;
+	enum cmt_user_flags *uflags;
+
+	/* root is read-only */
+	if (css == get_root_perf_css())
+		return -EINVAL;
+
+	buf = strstrip(buf);
+
+	mutex_lock(&cmt_mutex);
+	monr_hrchy_acquire_mutexes();
+
+	/* Monitoring active, use new flags. */
+	uflags = parse_pkg_uflags(buf, nbytes);
+	if (IS_ERR(uflags)) {
+		err = PTR_ERR(uflags);
+		goto exit_unlock;
+	}
+
+	is_mon = perf_cgroup_mon_started(perf_cgroup_from_css(css));
+	if (!is_mon) {
+		if (pkg_uflags_has_user(uflags)) {
+			err = __css_start_monitoring(css);
+			if (err)
+				goto exit_free;
+		} else {
+			/*
+			 * uflags must be all zero or parse_pkg_uflags
+			 * would have failed.
+			 */
+			goto exit_free;
+		}
+	}
+
+	/* At this point the monr is guaranteed to be this css's monr. */
+	monr = monr_from_css(css);
+
+	/* Disregard if flags have not changed. */
+	if (!memcmp(uflags, monr->pkg_uflags, pkg_uflags_size))
+		goto exit_free;
+
+	/*
+	 * will update monr->pkg_flags. Do not exit on error, continue to
+	 * clean up monr if unused.
+	 */
+	err = monr_apply_uflags(monr, uflags);
+
+	if (!monr_has_user(monr))
+		monr_destroy(monr);
+
+exit_free:
+	kfree(uflags);
+exit_unlock:
+	monr_hrchy_release_mutexes();
+	mutex_unlock(&cmt_mutex);
+	return err ?: nbytes;
+}
+
+struct cftype perf_event_cgrp_arch_subsys_cftypes[] = {
+	{
+		/*
+		 * allows per-package specification of uflags. It takes a
+		 * semi-colon separated list of hex uflags values. The hex
+		 * value in the i-th position corresponds to the uflags of
+		 * the package with logical id == i + 1 . Empty and missing hex
+		 * values receive 0.
+		 * e.g. "1;2" -> uflag 0x1 for pkg 0, uflag 0x2 for pkg 2,
+		 * and 0x0 for all others.
+		 */
+		.name = "cmt_monitoring",
+		.seq_show = cmt_monitoring_seq_show,
+		.write = cmt_monitoring_write,
+	},
+
+	{}	/* terminate */
+};
+
+#endif
+
 device_initcall(intel_cmt_init);
diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
index 783bdbb..babee97 100644
--- a/arch/x86/include/asm/perf_event.h
+++ b/arch/x86/include/asm/perf_event.h
@@ -315,6 +315,16 @@ int perf_cgroup_arch_css_online(struct cgroup_subsys_state *css);
 	perf_cgroup_arch_css_offline
 void perf_cgroup_arch_css_offline(struct cgroup_subsys_state *css);
 
+extern struct cftype perf_event_cgrp_arch_subsys_cftypes[];
+
+#define PERF_CGROUP_ARCH_CGRP_SUBSYS_ATTS \
+	.dfl_cftypes = perf_event_cgrp_arch_subsys_cftypes, \
+	.legacy_cftypes = perf_event_cgrp_arch_subsys_cftypes,
+
+#else
+
+#define PERF_CGROUP_ARCH_CGRP_SUBSYS_ATTS
+
 #endif
 #endif
 
-- 
2.8.0.rc3.226.g39d4020