Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Tue, 16 Feb 2021 11:39:11 -0800
To:     netdev@vger.kernel.org, linux-kselftest@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        viro@zeniv.linux.org.uk, fruggeri@arista.com
Subject: epoll: different edge-triggered behavior bewteen pipe and
 socketpair
User-Agent: Heirloom mailx 12.5 7/5/10
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-Id: <20210216193911.88C9795C05D0@us180.sjc.aristanetworks.com>
From:   fruggeri@arista.com (Francesco Ruggeri)
Precedence: bulk

pipe() and socketpair() have different behavior wrt edge-triggered
read epoll, in that no event is generated when data is written into
a non-empty pipe, but an event is generated if socketpair() is used
instead.
This simple modification of the epoll2 testlet from 
tools/testing/selftests/filesystems/epoll/epoll_wakeup_test.c
(it just adds a second write) shows the different behavior.
The testlet passes with pipe() but fails with socketpair() with 5.10.
They both fail with 4.19.
Is it fair to assume that 5.10 pipe's behavior is the correct one?

Thanks,
Francesco Ruggeri

/*
 *          t0
 *           | (ew)
 *          e0
 *           | (et)
 *          s0
 */
TEST(epoll2)
{
	int efd;
	int sfd[2];
	struct epoll_event e;

	ASSERT_EQ(socketpair(AF_UNIX, SOCK_STREAM, 0, sfd), 0);
	//ASSERT_EQ(pipe(sfd), 0);

	efd = epoll_create(1);
	ASSERT_GE(efd, 0);

	e.events = EPOLLIN | EPOLLET;
	ASSERT_EQ(epoll_ctl(efd, EPOLL_CTL_ADD, sfd[0], &e), 0);

	ASSERT_EQ(write(sfd[1], "w", 1), 1);
	EXPECT_EQ(epoll_wait(efd, &e, 1, 0), 1);

	ASSERT_EQ(write(sfd[1], "w", 1), 1);
	EXPECT_EQ(epoll_wait(efd, &e, 1, 0), 0);

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