summaryrefslogtreecommitdiffstats
path: root/meta/recipes-devtools/unfs3/unfs3
diff options
context:
space:
mode:
authorHaiqing Bai <Haiqing.Bai@windriver.com>2019-10-24 10:33:04 +0800
committerRichard Purdie <richard.purdie@linuxfoundation.org>2019-10-24 12:03:38 +0100
commit1f6784d2e839f81749d21ad1b615a9f7bb0e64d6 (patch)
tree8e336a35bf9a8176162ba833b7e0f978facc7ef3 /meta/recipes-devtools/unfs3/unfs3
parent3c39d4158677b97253df63f23b74c3a9dd5539f6 (diff)
downloadopenembedded-core-1f6784d2e839f81749d21ad1b615a9f7bb0e64d6.tar.gz
unfs3: fixed the issue that unfsd consumes 100% CPU
The 'accept' function on the socket of unfsd daemon is always in below error state: accept(4, 0x7ffd5e6dddc0, [128]) = -1 EINVAL (Invalid argument) accept(6, 0x7ffd5e6dddc0, [128]) = -1 EINVAL (Invalid argument) And 'strace -c -p <the pid of unfsd>' shows: % time seconds usecs/call calls errors syscall ------ ----------- ----------- --------- --------- ---------------- 70.87 0.005392 0 513886 513886 accept 29.13 0.002216 0 256943 poll This error state is in the 'for' loop of the daemon, so it consumes 100% CPU. The reason is that 'listen' is not called for the TCP socket before 'accept'. Actually the called 'svc_tli_create' from libtirpc will not call 'listen' on a bound socket. Signed-off-by: Haiqing Bai <Haiqing.Bai@windriver.com> Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Diffstat (limited to 'meta/recipes-devtools/unfs3/unfs3')
-rw-r--r--meta/recipes-devtools/unfs3/unfs3/0001-Add-listen-action-for-a-tcp-socket.patch54
1 files changed, 54 insertions, 0 deletions
diff --git a/meta/recipes-devtools/unfs3/unfs3/0001-Add-listen-action-for-a-tcp-socket.patch b/meta/recipes-devtools/unfs3/unfs3/0001-Add-listen-action-for-a-tcp-socket.patch
new file mode 100644
index 0000000000..e9b9d3df46
--- /dev/null
+++ b/meta/recipes-devtools/unfs3/unfs3/0001-Add-listen-action-for-a-tcp-socket.patch
@@ -0,0 +1,54 @@
+From b42ab8e1aca951dd06c113159491b3fd5cf06f2e Mon Sep 17 00:00:00 2001
+From: Haiqing Bai <Haiqing.Bai@windriver.com>
+Date: Thu, 24 Oct 2019 09:39:04 +0800
+Subject: [PATCH] Add "listen" action for a tcp socket which does not call
+ 'listen' after 'bind'
+
+It is found that /usr/bin/unfsd customus 100% cpu after starting qemu with 'nfs'
+option, and below lots of error messages shows when strace the process:
+
+poll([{fd=3, events=POLLIN|POLLPRI|POLLRDNORM|POLLRDBAND},{fd=4, events=POLLIN|POLLPRI|POLLRDNORM|POLLRDBAND},
+{fd=5, events=POLLIN|POLLPRI|POLLRDNORM|POLLRDBAND},{fd=6, events =POLLIN|POLLPRI|POLLRDNORM|POLLRDBAND}],
+4, 2000) = 2 ([{fd=4, revents=POLLHUP},{fd=6, revents=POLLHUP}])
+accept(4, 0x7ffd5e6dddc0, [128]) = -1 EINVAL (Invalid argument)
+accept(6, 0x7ffd5e6dddc0, [128]) = -1 EINVAL (Invalid argument)
+
+% time seconds usecs/call calls errors syscall
+------ ----------- ----------- --------- --------- ----------------
+ 70.87 0.005392 0 513886 513886 accept
+ 29.13 0.002216 0 256943 poll
+ 0.00 0.000000 0 4 read
+
+The root cause is that 'listen' is not called for the binded
+socket. The depended libtipc does not call 'listen' if found
+the incomming socket is binded, so 'accept' reports the error
+in the 'for' loop and cpu consumed.
+
+Upstream-Status: Pending
+
+Signed-off-by: Haiqing Bai <Haiqing.Bai@windriver.com>
+---
+ daemon.c | 7 +++++++
+ 1 file changed, 7 insertions(+)
+
+diff --git a/daemon.c b/daemon.c
+index 028a181..4c85903 100644
+--- a/daemon.c
++++ b/daemon.c
+@@ -814,6 +814,13 @@ static SVCXPRT *create_tcp_transport(unsigned int port)
+ fprintf(stderr, "Couldn't bind to tcp port %d\n", port);
+ exit(1);
+ }
++
++ if (listen(sock, SOMAXCONN) < 0) {
++ perror("listen");
++ fprintf(stderr, "Couldn't listen on the address \n");
++ close(sock);
++ exit(1);
++ }
+ }
+
+ transp = svctcp_create(sock, 0, 0);
+--
+1.9.1
+