Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   hev <r@hev.cc>
To:     linux-fsdevel@vger.kernel.org
Cc:     Heiher <r@hev.cc>, Al Viro <viro@ZenIV.linux.org.uk>,
        Andrew Morton <akpm@linux-foundation.org>,
        Davide Libenzi <davidel@xmailserver.org>,
        Davidlohr Bueso <dave@stgolabs.net>,
        Dominik Brodowski <linux@dominikbrodowski.net>,
        Eric Wong <e@80x24.org>, Jason Baron <jbaron@akamai.com>,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Roman Penyaev <rpenyaev@suse.de>,
        Sridhar Samudrala <sridhar.samudrala@intel.com>,
        linux-kernel@vger.kernel.org
Subject: [PATCH RESEND v3] fs/epoll: Remove unnecessary wakeups of nested epoll that in ET mode
Date:   Tue, 24 Sep 2019 22:26:54 +0800
Message-Id: <20190924142654.5742-1-r@hev.cc>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

From: Heiher <r@hev.cc>

Take the case where we have:

        t0
         | (ew)
        e0
         | (et)
        e1
         | (lt)
        s0

t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew: epoll_wait
et: edge-trigger
lt: level-trigger

When s0 fires an event, e1 catches the event, and then e0 catches an event from
e1. After this, There is a thread t0 do epoll_wait() many times on e0, it should
only get one event in total, because e1 is a dded to e0 in edge-triggered mode.

This patch only allows wakeup(&ep->poll_wait) in ep_scan_ready_list if one of
the conditions is met:

 1. depth == 0.
 2. There have event is added to ep->ovflist during processing.

Test code:
 #include <unistd.h>
 #include <sys/epoll.h>
 #include <sys/socket.h>

 int main(int argc, char *argv[])
 {
 	int sfd[2];
 	int efd[2];
 	struct epoll_event e;

 	if (socketpair(AF_UNIX, SOCK_STREAM, 0, sfd) < 0)
 		goto out;

 	efd[0] = epoll_create(1);
 	if (efd[0] < 0)
 		goto out;

 	efd[1] = epoll_create(1);
 	if (efd[1] < 0)
 		goto out;

 	e.events = EPOLLIN;
 	if (epoll_ctl(efd[1], EPOLL_CTL_ADD, sfd[0], &e) < 0)
 		goto out;

 	e.events = EPOLLIN | EPOLLET;
 	if (epoll_ctl(efd[0], EPOLL_CTL_ADD, efd[1], &e) < 0)
 		goto out;

 	if (write(sfd[1], "w", 1) != 1)
 		goto out;

 	if (epoll_wait(efd[0], &e, 1, 0) != 1)
 		goto out;

 	if (epoll_wait(efd[0], &e, 1, 0) != 0)
 		goto out;

 	close(efd[0]);
 	close(efd[1]);
 	close(sfd[0]);
 	close(sfd[1]);

 	return 0;

 out:
 	return -1;
 }

More tests:
 https://github.com/heiher/epoll-wakeup

Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Cc: Eric Wong <e@80x24.org>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Roman Penyaev <rpenyaev@suse.de>
Cc: Sridhar Samudrala <sridhar.samudrala@intel.com>
Cc: linux-kernel@vger.kernel.org
Cc: linux-fsdevel@vger.kernel.org
Signed-off-by: hev <r@hev.cc>
---
 fs/eventpoll.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)

diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index c4159bcc05d9..a05249400901 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -685,6 +685,9 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep,
 	if (!ep_locked)
 		mutex_lock_nested(&ep->mtx, depth);
 
+	if (!depth || list_empty_careful(&ep->rdllist))
+		pwake++;
+
 	/*
 	 * Steal the ready list, and re-init the original one to the
 	 * empty list. Also, set ep->ovflist to NULL so that events
@@ -704,12 +707,21 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep,
 	res = (*sproc)(ep, &txlist, priv);
 
 	write_lock_irq(&ep->lock);
+	nepi = READ_ONCE(ep->ovflist);
+	/*
+	 * We only need to wakeup nested epoll fds if something has been queued
+	 * to the overflow list, since the ep_poll() traverses the rdllist
+	 * during recursive poll and thus events on the overflow list may not be
+	 * visible yet.
+	 */
+	if (nepi != NULL)
+		pwake++;
 	/*
 	 * During the time we spent inside the "sproc" callback, some
 	 * other events might have been queued by the poll callback.
 	 * We re-insert them inside the main ready-list here.
 	 */
-	for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
+	for (; (epi = nepi) != NULL;
 	     nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
 		/*
 		 * We need to check if the item is already in the list.
@@ -755,7 +767,7 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep,
 		mutex_unlock(&ep->mtx);
 
 	/* We have to call this outside the lock */
-	if (pwake)
+	if (pwake == 2)
 		ep_poll_safewake(&ep->poll_wait);
 
 	return res;
-- 
2.23.0