Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752956AbdHWALL (ORCPT ); Tue, 22 Aug 2017 20:11:11 -0400 Received: from smtp-fw-6001.amazon.com ([52.95.48.154]:28539 "EHLO smtp-fw-6001.amazon.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752635AbdHWAKj (ORCPT ); Tue, 22 Aug 2017 20:10:39 -0400 X-IronPort-AV: E=Sophos;i="5.41,414,1498521600"; d="scan'208";a="306480112" From: Vallish Vaidyeshwara To: , , , , , CC: , , Subject: [PATCH v2 1/2] net: enable high resolution timer mode to timeout datagram sockets Date: Wed, 23 Aug 2017 00:10:26 +0000 Message-ID: <1503447027-44399-2-git-send-email-vallish@amazon.com> X-Mailer: git-send-email 2.7.3.AMZN In-Reply-To: <1503447027-44399-1-git-send-email-vallish@amazon.com> References: <1503447027-44399-1-git-send-email-vallish@amazon.com> MIME-Version: 1.0 Content-Type: text/plain Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 4935 Lines: 118 Enable high resolution timer mode to time SO_RCVTIMEO value used with setsockopt(2) on AF_UNIX and AF_INET datagram sockets. By default, SO_RCVTIMEO uses low resolution timer which is good for most of socket use cases. Background: Kernel timer wheel was refactored in 4.8 to avoid drawbacks with previous implementation: https://lwn.net/Articles/691064/ Unlike the previous "kernel timer wheel" implementation in 4.4 which aimed for accuracy by paying cost for cascading tracked timers at the boundary of 256 jiffies, the new timer wheel implementation gets rid of cascading latency by paying a price for being less accurate for far off timers. Use Case: New implementation is good for most of socket use cases. However we have a use case where our application is sensitive to socket timeout including long timeouts. Please refer to test code as part of this patch series. One of the test runs with a timeout value of 180 seconds timed out at 190 seconds. [root@]# ./datagram_sock_timeout 180000 datagram_sock_timeout failed: took 190.00 seconds [root@]# The same program when run on a 4.4 kernel would timeout more acurately and the kernel added slack was not noticeable to user application. Interesting text: a) Standards for setsockopt: http://pubs.opengroup.org/onlinepubs/009695399/functions/setsockopt.html SO_RCVTIMEO Sets the timeout value that specifies the maximum amount of time an input function waits until it completes. It accepts a timeval structure with the number of seconds and microseconds specifying the limit on how long to wait for an input operation to complete. If a receive operation has blocked for this much time without receiving additional data, it shall return with a partial count or errno set to [EAGAIN] or [EWOULDBLOCK] if no data is received. The default for this option is zero, which indicates that a receive operation shall not time out. This option takes a timeval structure. Note that not all implementations allow this option to be set. This only talks about the maximum time and the current behavior indeed follows this standard. System call does not return before the time specified and it does return EAGAIN. b) Man page for SETSOCKOPT(3P): The option_name argument specifies a single option to set. It can be one of the socket-level options defined in and described in Section 2.10.16, Use of Options. If option_name is equal to SO_RCVTIMEO or SO_SNDTIMEO and the implementation supports setting the option, it is unspecified whether the struct timeval pointed to by option_value is stored as provided by this function or is rounded up to align with the resolution of the clock being used. If setsockopt() is called with option_name equal to SO_ACCEPTCONN, SO_ERROR, or SO_TYPE, the behavior is unspecified. Behavior is unspecified. 3) Man page for SELECT(2): Note that the timeout interval will be rounded up to the system clock granularity, and kernel scheduling delays mean that the blocking interval may overrun by a small amount. If both fields of the timeval structure are zero, then select() returns immediately. (This is useful for polling.) If timeout is NULL (no timeout), select() can block indefinitely. Select system call guarantees timeout interval and inturn uses highres timer. Reported-by: Manjula Peiris Reviewed-by: Eduardo Valentin Reviewed-by: Anchal Agarwal Signed-off-by: Vallish Vaidyeshwara --- net/core/datagram.c | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/net/core/datagram.c b/net/core/datagram.c index ee5647b..832c147 100644 --- a/net/core/datagram.c +++ b/net/core/datagram.c @@ -80,6 +80,7 @@ static int receiver_wake_function(wait_queue_entry_t *wait, unsigned int mode, i return 0; return autoremove_wake_function(wait, mode, sync, key); } + /* * Wait for the last received packet to be different from skb */ @@ -87,6 +88,9 @@ int __skb_wait_for_more_packets(struct sock *sk, int *err, long *timeo_p, const struct sk_buff *skb) { int error; + ktime_t expires; + struct timespec64 time; + unsigned long pre_sched_time; DEFINE_WAIT_FUNC(wait, receiver_wake_function); prepare_to_wait_exclusive(sk_sleep(sk), &wait, TASK_INTERRUPTIBLE); @@ -116,7 +120,15 @@ int __skb_wait_for_more_packets(struct sock *sk, int *err, long *timeo_p, goto interrupted; error = 0; - *timeo_p = schedule_timeout(*timeo_p); + /* Wait using highres timer */ + time.tv_sec = *timeo_p / HZ; + time.tv_nsec = jiffies_to_nsecs(*timeo_p % HZ); + expires = ktime_add_safe(ktime_get(), timespec64_to_ktime(time)); + pre_sched_time = jiffies; + if (schedule_hrtimeout(&expires, HRTIMER_MODE_ABS)) + *timeo_p = jiffies - pre_sched_time; + else + *timeo_p = 0; out: finish_wait(sk_sleep(sk), &wait); return error; -- 2.7.3.AMZN