Message-Id: <20230315021850.2788946-5-anjali.k.kulkarni@oracle.com>
Date:   Tue, 14 Mar 2023 19:18:49 -0700
From:   Anjali Kulkarni <anjali.k.kulkarni@...cle.com>
To:     davem@...emloft.net
Cc:     edumazet@...gle.com, kuba@...nel.org, pabeni@...hat.com,
        zbr@...emap.net, brauner@...nel.org, johannes@...solutions.net,
        ecree.xilinx@...il.com, leon@...nel.org, keescook@...omium.org,
        socketcan@...tkopp.net, petrm@...dia.com,
        linux-kernel@...r.kernel.org, netdev@...r.kernel.org,
        anjali.k.kulkarni@...cle.com
Subject: [PATCH v2 4/5] connector/cn_proc: Performance improvements

This patch adds the capability to filter messages sent by the proc
connector on the event type supplied in the message from the client
to the connector. Thus, a client can register to listen for only exit
or fork or a mix or all of the events. The patch takes care that
existing clients using old mechanism of not sending the event type work
without any changes. This will greatly enhance performance - currently, we
need to listen to all events, which increases the times per event. Handling
8K exits takes about 70ms, whereas 8K-forks + 8K-exits takes about 150ms
& handling 8K-forks + 8K-exits + 8K-execs takes 200ms. There are currently
9 different types of events, all being sent to proc event listeners. Also,
measuring the time using pidfds for monitoring 8K process exits took 4
times longer - 200ms, as compared to 70ms using only exit notifications of
proc connector.

We also want to add a new event PROC_EVENT_NONZERO_EXIT, which is
only sent by kernel to a listening application when any process exiting,
has a non-zero exit status. Reason for this is as follows:
Oracle DB runs on a large scale with 100000s of short lived processes,
starting up and exiting quickly. A process monitoring DB daemon which
tracks and cleans after processes that have died without a proper exit
hence not done their own cleanup, needs notifications only when a process
exited ungracefully(which should be rare).

This change will give Oracle DB huge performance savings - it takes 50ms
to scan about 8K PIDs in /proc, about 500ms for 100K PIDs. DB does this
check every 3 secs, so over an hour we save 10secs for 100K PIDs.
This kind of a new event could also be useful to other applications like
Google's lmkd daemon, which needs a killed process's exit notification.

The proc_filter.c test file is updated to reflect the new filtering.

Signed-off-by: Anjali Kulkarni <anjali.k.kulkarni@...cle.com>
---
 drivers/connector/cn_proc.c     | 59 +++++++++++++++++++++++++++++----
 include/uapi/linux/cn_proc.h    | 19 +++++++++++
 samples/connector/proc_filter.c | 47 +++++++++++++++++++++++---
 3 files changed, 115 insertions(+), 10 deletions(-)

diff --git a/drivers/connector/cn_proc.c b/drivers/connector/cn_proc.c
index ef3820b43b5c..465c9c7a8f8b 100644
--- a/drivers/connector/cn_proc.c
+++ b/drivers/connector/cn_proc.c
@@ -50,22 +50,45 @@ static DEFINE_PER_CPU(struct local_event, local_event) = {
 
 int cn_filter(struct sock *dsk, struct sk_buff *skb, void *data)
 {
+	uintptr_t val;
+	__u32 what, exit_code, *ptr;
 	enum proc_cn_mcast_op mc_op;
 
-	if (!dsk)
+	if (!dsk || !data)
 		return 0;
 
+	ptr = (__u32 *)data;
+	what = *ptr++;
+	exit_code = *ptr;
+	val = ((struct proc_input *)(dsk->sk_user_data))->event_type;
 	mc_op = ((struct proc_input *)(dsk->sk_user_data))->mcast_op;
 
 	if (mc_op == PROC_CN_MCAST_IGNORE)
 		return 1;
 
-	return 0;
+	if ((__u32)val == PROC_EVENT_ALL)
+		return 0;
+	/*
+	 * Drop packet if we have to report only non-zero exit status
+	 * (PROC_EVENT_NONZERO_EXIT) and exit status is 0
+	 */
+	if (((__u32)val & PROC_EVENT_NONZERO_EXIT) &&
+	    (what == PROC_EVENT_EXIT)) {
+		if (exit_code)
+			return 0;
+		else
+			return 1;
+	}
+	if ((__u32)val & what)
+		return 0;
+	return 1;
 }
 EXPORT_SYMBOL_GPL(cn_filter);
 
 static inline void send_msg(struct cn_msg *msg)
 {
+	__u32 filter_data[2];
+
 	local_lock(&local_event.lock);
 
 	msg->seq = __this_cpu_inc_return(local_event.count) - 1;
@@ -77,8 +100,15 @@ static inline void send_msg(struct cn_msg *msg)
 	 *
 	 * If cn_netlink_send() fails, the data is not sent.
 	 */
+	filter_data[0] = ((struct proc_event *)msg->data)->what;
+	if (filter_data[0] == PROC_EVENT_EXIT) {
+		filter_data[1] =
+		((struct proc_event *)msg->data)->event_data.exit.exit_code;
+	} else {
+		filter_data[1] = 0;
+	}
 	cn_netlink_send_mult(msg, msg->len, 0, CN_IDX_PROC, GFP_NOWAIT,
-			     cn_filter, NULL);
+			     cn_filter, (void *)filter_data);
 
 	local_unlock(&local_event.lock);
 }
@@ -358,12 +388,15 @@ static void cn_proc_ack(int err, int rcvd_seq, int rcvd_ack)
 
 /**
  * cn_proc_mcast_ctl
- * @data: message sent from userspace via the connector
+ * @msg: message sent from userspace via the connector
+ * @nsp: NETLINK_CB of the client's socket buffer
  */
 static void cn_proc_mcast_ctl(struct cn_msg *msg,
 			      struct netlink_skb_parms *nsp)
 {
 	enum proc_cn_mcast_op mc_op = 0, prev_mc_op = 0;
+	struct proc_input *pinput = NULL;
+	enum proc_cn_event ev_type = 0;
 	int err = 0, initial = 0;
 	struct sock *sk = NULL;
 
@@ -382,11 +415,21 @@ static void cn_proc_mcast_ctl(struct cn_msg *msg,
 		goto out;
 	}
 
-	if (msg->len == sizeof(mc_op))
+	if (msg->len == sizeof(*pinput)) {
+		pinput = (struct proc_input *)msg->data;
+		mc_op = pinput->mcast_op;
+		ev_type = pinput->event_type;
+	} else if (msg->len == sizeof(mc_op)) {
 		mc_op = *((enum proc_cn_mcast_op *)msg->data);
-	else
+		ev_type = PROC_EVENT_ALL;
+	} else
 		return;
 
+	ev_type = valid_event((enum proc_cn_event)ev_type);
+
+	if (ev_type == PROC_EVENT_NONE)
+		ev_type = PROC_EVENT_ALL;
+
 	if (nsp->sk) {
 		sk = nsp->sk;
 		if (sk->sk_user_data == NULL) {
@@ -397,6 +440,8 @@ static void cn_proc_mcast_ctl(struct cn_msg *msg,
 			prev_mc_op =
 			((struct proc_input *)(sk->sk_user_data))->mcast_op;
 		}
+		((struct proc_input *)(sk->sk_user_data))->event_type =
+			ev_type;
 		((struct proc_input *)(sk->sk_user_data))->mcast_op = mc_op;
 	}
 
@@ -408,6 +453,8 @@ static void cn_proc_mcast_ctl(struct cn_msg *msg,
 	case PROC_CN_MCAST_IGNORE:
 		if (!initial && (prev_mc_op != PROC_CN_MCAST_IGNORE))
 			atomic_dec(&proc_event_num_listeners);
+		((struct proc_input *)(sk->sk_user_data))->event_type =
+			PROC_EVENT_NONE;
 		break;
 	default:
 		err = EINVAL;
diff --git a/include/uapi/linux/cn_proc.h b/include/uapi/linux/cn_proc.h
index 6a06fb424313..f2afb7cc4926 100644
--- a/include/uapi/linux/cn_proc.h
+++ b/include/uapi/linux/cn_proc.h
@@ -30,6 +30,15 @@ enum proc_cn_mcast_op {
 	PROC_CN_MCAST_IGNORE = 2
 };
 
+#define PROC_EVENT_ALL (PROC_EVENT_FORK | PROC_EVENT_EXEC | PROC_EVENT_UID |  \
+			PROC_EVENT_GID | PROC_EVENT_SID | PROC_EVENT_PTRACE | \
+			PROC_EVENT_COMM | PROC_EVENT_NONZERO_EXIT |           \
+			PROC_EVENT_COREDUMP | PROC_EVENT_EXIT)
+
+/*
+ * If you add an entry in proc_cn_event, make sure you add it in
+ * PROC_EVENT_ALL above as well.
+ */
 enum proc_cn_event {
 	/* Use successive bits so the enums can be used to record
 	 * sets of events as well
@@ -45,15 +54,25 @@ enum proc_cn_event {
 	/* "next" should be 0x00000400 */
 	/* "last" is the last process event: exit,
 	 * while "next to last" is coredumping event
+	 * before that is report only if process dies
+	 * with non-zero exit status
 	 */
+	PROC_EVENT_NONZERO_EXIT = 0x20000000,
 	PROC_EVENT_COREDUMP = 0x40000000,
 	PROC_EVENT_EXIT = 0x80000000
 };
 
 struct proc_input {
 	enum proc_cn_mcast_op mcast_op;
+	enum proc_cn_event event_type;
 };
 
+static inline enum proc_cn_event valid_event(enum proc_cn_event ev_type)
+{
+	ev_type &= PROC_EVENT_ALL;
+	return ev_type;
+}
+
 /*
  * From the user's point of view, the process
  * ID is the thread group ID and thread ID is the internal
diff --git a/samples/connector/proc_filter.c b/samples/connector/proc_filter.c
index 25202f5bc126..63504fc5f002 100644
--- a/samples/connector/proc_filter.c
+++ b/samples/connector/proc_filter.c
@@ -15,22 +15,33 @@
 #include <errno.h>
 #include <signal.h>
 
+#define FILTER
+
+#ifdef FILTER
+#define NL_MESSAGE_SIZE (sizeof(struct nlmsghdr) + sizeof(struct cn_msg) + \
+			 sizeof(struct proc_input))
+#else
 #define NL_MESSAGE_SIZE (sizeof(struct nlmsghdr) + sizeof(struct cn_msg) + \
 			 sizeof(int))
+#endif
 
 #define MAX_EVENTS 1
 
+volatile static int interrupted;
+static int nl_sock, ret_errno, tcount;
+static struct epoll_event evn;
+
 #ifdef ENABLE_PRINTS
 #define Printf printf
 #else
 #define Printf
 #endif
 
-volatile static int interrupted;
-static int nl_sock, ret_errno, tcount;
-static struct epoll_event evn;
-
+#ifdef FILTER
+int send_message(struct proc_input *pinp)
+#else
 int send_message(enum proc_cn_mcast_op mcast_op)
+#endif
 {
 	char buff[NL_MESSAGE_SIZE];
 	struct nlmsghdr *hdr;
@@ -50,8 +61,14 @@ int send_message(enum proc_cn_mcast_op mcast_op)
 	msg->ack = 0;
 	msg->flags = 0;
 
+#ifdef FILTER
+	msg->len = sizeof(struct proc_input);
+	((struct proc_input *)msg->data)->mcast_op = pinp->mcast_op;
+	((struct proc_input *)msg->data)->event_type = pinp->event_type;
+#else
 	msg->len = sizeof(int);
 	*(int *)msg->data = mcast_op;
+#endif
 
 	if (send(nl_sock, hdr, hdr->nlmsg_len, 0) == -1) {
 		ret_errno = errno;
@@ -61,7 +78,11 @@ int send_message(enum proc_cn_mcast_op mcast_op)
 	return 0;
 }
 
+#ifdef FILTER
+int register_proc_netlink(int *efd, struct proc_input *input)
+#else
 int register_proc_netlink(int *efd, enum proc_cn_mcast_op mcast_op)
+#endif
 {
 	struct sockaddr_nl sa_nl;
 	int err = 0, epoll_fd;
@@ -92,7 +113,11 @@ int register_proc_netlink(int *efd, enum proc_cn_mcast_op mcast_op)
 		return -2;
 	}
 
+#ifdef FILTER
+	err = send_message(input);
+#else
 	err = send_message(mcast_op);
+#endif
 	if (err < 0)
 		return err;
 
@@ -223,10 +248,19 @@ int main(int argc, char *argv[])
 {
 	int epoll_fd, err;
 	struct proc_event proc_ev;
+#ifdef FILTER
+	struct proc_input input;
+#endif
 
 	signal(SIGINT, sigint);
 
+#ifdef FILTER
+	input.event_type = PROC_EVENT_NONZERO_EXIT;
+	input.mcast_op = PROC_CN_MCAST_LISTEN;
+	err = register_proc_netlink(&epoll_fd, &input);
+#else
 	err = register_proc_netlink(&epoll_fd, PROC_CN_MCAST_LISTEN);
+#endif
 	if (err < 0) {
 		if (err == -2)
 			close(nl_sock);
@@ -252,7 +286,12 @@ int main(int argc, char *argv[])
 		}
 	}
 
+#ifdef FILTER
+	input.mcast_op = PROC_CN_MCAST_IGNORE;
+	send_message(&input);
+#else
 	send_message(PROC_CN_MCAST_IGNORE);
+#endif
 
 	close(epoll_fd);
 	close(nl_sock);
-- 
2.39.2

Powered by blists - more mailing lists