rpms/gpsd
7
0
Fork 0
Code Issues Pull Requests Releases Activity

fix busy wait when reading from gpsd socket

Browse Source
This commit is contained in:
Miroslav Lichvar 2024-02-20 15:52:05 +01:00
parent fec40c08af
commit 75e324c9e3
2 changed files with 88 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

84
gpsd-busywait.patch Normal file
View File

@ -0,0 +1,84 @@
commit e5ba7aa2af74fd22ebbd5c4a6624edcf983863de
Author: Michal Schmidt <mschmidt@redhat.com>
Date: Fri Aug 4 16:53:01 2023 +0200
gps/gps.py.in: no busy-waiting when reading from gpsd socket
ubxtool keeps one CPU 100% busy while it waits for data to read from the
gpsd socket. Running it under strace showed that it calls select() with
zero timeout in a loop:
...
11:02:34.049629 pselect6(4, [3], [], [], {tv_sec=0, tv_nsec=0}, NULL) = 0 (Timeout)
11:02:34.049649 pselect6(4, [3], [], [], {tv_sec=0, tv_nsec=0}, NULL) = 0 (Timeout)
11:02:34.049670 pselect6(4, [3], [], [], {tv_sec=0, tv_nsec=0}, NULL) = 0 (Timeout)
...
The busy waiting can be eliminated by passing the actual timeout value
to select(). In the reading loop in gps.py, the remaining time can be
easily calculated and passed as the argument to the self.ser.waiting()
function (which is basically a select() wrapper).
Fixing this problem exposed a bug in how the received bytes are decoded.
decode_func may not consume all input at once. Consumable input may be
left in self.out until decode_func returns zero, indicating that it
could not process any more input. So decode_func must be called in a
loop each time a buffer is received from the socket. The busy waiting
was hiding this issue, because decode_func was being called all the
time.
The "elif self.input_is_device:" branch probably needs similar
treatment, but I am testing only the gpsd usecase.
diff --git a/gps/gps.py.in b/gps/gps.py.in
index 623a750a0..14d7707ab 100644
--- a/gps/gps.py.in
+++ b/gps/gps.py.in
@@ -384,10 +384,11 @@ class gps_io(object):
if self.gpsd_host is not None:
# gpsd input
start = monotonic()
- while (monotonic() - start) < input_wait:
+ remaining_time = input_wait
+ while remaining_time > 0:
# First priority is to be sure the input buffer is read.
# This is to prevent input buffer overuns
- if 0 < self.ser.waiting():
+ if 0 < self.ser.waiting(remaining_time):
# We have serial input waiting, get it
# No timeout possible
# RTCM3 JSON can be over 4.4k long, so go big
@@ -397,17 +398,22 @@ class gps_io(object):
raw_fd.write(polybytes(new_out))
self.out += new_out
- consumed = decode_func(self.out)
- # TODO: the decoder shall return a some current
- # statement_identifier # to fill last_statement_identifier
- last_statement_identifier = None
- #
- self.out = self.out[consumed:]
- if ((expect_statement_identifier and
- (expect_statement_identifier ==
- last_statement_identifier))):
- # Got what we were waiting for. Done?
- ret_code = 0
+ while True:
+ consumed = decode_func(self.out)
+ if consumed == 0:
+ break
+ # TODO: the decoder shall return a some current
+ # statement_identifier # to fill last_statement_identifier
+ last_statement_identifier = None
+ #
+ self.out = self.out[consumed:]
+ if ((expect_statement_identifier and
+ (expect_statement_identifier ==
+ last_statement_identifier))):
+ # Got what we were waiting for. Done?
+ ret_code = 0
+
+ remaining_time = start + input_wait - monotonic()
elif self.input_is_device:
# input is a serial device

4
gpsd.spec
View File

@ -22,6 +22,9 @@ Source11: gpsd.sysconfig
Patch1: gpsd-ipv6.patch
# fix some issues reported by coverity and shellcheck
Patch2: gpsd-scanfixes.patch
# fix busy wait when reading from gpsd socket
Patch3: gpsd-busywait.patch
BuildRequires: gcc
BuildRequires: dbus-devel
BuildRequires: ncurses-devel
@ -141,6 +144,7 @@ do not reference it or depend on it in any way.
%setup -q
%patch -P 1 -p1 -b .ipv6
%patch -P 2 -p1 -b .scanfixes
%patch -P 3 -p1 -b .busywait
# don't try reloading systemd when installing in the build root
sed -i 's|systemctl daemon-reload|true|' SConscript