Message ID | 20230209080936.148489-3-mikko.rapeli@linaro.org |
---|---|
State | Accepted, archived |
Commit | 9c63970fce3a3d6029745252a6ec2bf9b9da862d |
Headers | show |
Series | fix oeqa runtime test framework when qemu hangs | expand |
On Thu, 2023-02-09 at 10:09 +0200, Mikko Rapeli wrote: > When qemu machine hangs, the ssh commands done by tests > are not timing out. do_testimage() task has last logs like this: > > DEBUG: time: 1673531086.3155053, endtime: 1673531686.315502 > > The test process is stuck for hours, or for ever if the > executing command or test case did not set a timeout correctly. > The default 300 second timeout is not working when target hangs. > Note that timeout is really a "inactive timeout" since data returned > by the process will reset the timeout. > > Make the process stdout non-blocking so read() will always return > right away using os.set_blocking() available in python 3.5 and later. > > Then change from python codec reader to plain read() and make > the ssh subprocess stdout non-blocking. Even with select() > making sure the file had input to be read, the codec reader was > trying to find more stuff and blocking for ever when process hangs. > > While at it, add a small timeout to read data in larger chunks if > possible. This avoids reading data one or few characters at a time > and makes the debug logs more readable. > > close() the stdout file in all cases after read loop is complete. > > Then make sure to wait or kill the ssh subprocess in all cases. > Just reading the output stream and receiving EOF there does not mean > that the process exited, and wait() needs a timeout if the process > is hanging. In the end kill the process and return the return value > and captured output utf-8 encoded, just like before these changes. > > This fixes ssh run() related deadlocks when a qemu target hangs > completely. > > Signed-off-by: Mikko Rapeli <mikko.rapeli@linaro.org> > --- > meta/lib/oeqa/core/target/ssh.py | 33 +++++++++++++++++++++++++------- > 1 file changed, 26 insertions(+), 7 deletions(-) > > diff --git a/meta/lib/oeqa/core/target/ssh.py b/meta/lib/oeqa/core/target/ssh.py > index 13fd5b2a49..466a795eb4 100644 > --- a/meta/lib/oeqa/core/target/ssh.py > +++ b/meta/lib/oeqa/core/target/ssh.py > @@ -229,22 +229,28 @@ def SSHCall(command, logger, timeout=None, **opts): > if timeout: > endtime = starttime + timeout > eof = False > + os.set_blocking(process.stdout.fileno(), False) > while time.time() < endtime and not eof: > - logger.debug('time: %s, endtime: %s' % (time.time(), endtime)) > try: > + logger.debug('Waiting for process output: time: %s, endtime: %s' % (time.time(), endtime)) > if select.select([process.stdout], [], [], 5)[0] != []: > - reader = codecs.getreader('utf-8')(process.stdout, 'ignore') > - data = reader.read(1024, 4096) > + # wait a bit for more data, tries to avoid reading single characters > + time.sleep(0.2) > + data = process.stdout.read() > if not data: > - process.stdout.close() > eof = True I'm not sure I understand the use of eof now. If stdout is non- blocking, it would be perfectly normal for it to return with no data but that doesn't mean we're at eof, it just means there is no data ready? I suspect the code might be ok but we should stop calling it eof since it no longer is? I also tend to dislike sleep() codepaths as it usually means the code is sub optimal. I was going to make further suggestions but I stepped back and looked at the bigger picture of the code. The code is basically doing: if timeout: <complex code> else: output = process.communicate() Python 3.3 added a timeout parameter to subprocess.communicate(). Can we just use that instead? Or is there something special about our needs which means we can't? Cheers, Richard
Hi, On Thu, Feb 09, 2023 at 09:15:10AM +0000, Richard Purdie wrote: > On Thu, 2023-02-09 at 10:09 +0200, Mikko Rapeli wrote: > > When qemu machine hangs, the ssh commands done by tests > > are not timing out. do_testimage() task has last logs like this: > > > > DEBUG: time: 1673531086.3155053, endtime: 1673531686.315502 > > > > The test process is stuck for hours, or for ever if the > > executing command or test case did not set a timeout correctly. > > The default 300 second timeout is not working when target hangs. > > Note that timeout is really a "inactive timeout" since data returned > > by the process will reset the timeout. > > > > Make the process stdout non-blocking so read() will always return > > right away using os.set_blocking() available in python 3.5 and later. > > > > Then change from python codec reader to plain read() and make > > the ssh subprocess stdout non-blocking. Even with select() > > making sure the file had input to be read, the codec reader was > > trying to find more stuff and blocking for ever when process hangs. > > > > While at it, add a small timeout to read data in larger chunks if > > possible. This avoids reading data one or few characters at a time > > and makes the debug logs more readable. > > > > close() the stdout file in all cases after read loop is complete. > > > > Then make sure to wait or kill the ssh subprocess in all cases. > > Just reading the output stream and receiving EOF there does not mean > > that the process exited, and wait() needs a timeout if the process > > is hanging. In the end kill the process and return the return value > > and captured output utf-8 encoded, just like before these changes. > > > > This fixes ssh run() related deadlocks when a qemu target hangs > > completely. > > > > Signed-off-by: Mikko Rapeli <mikko.rapeli@linaro.org> > > --- > > meta/lib/oeqa/core/target/ssh.py | 33 +++++++++++++++++++++++++------- > > 1 file changed, 26 insertions(+), 7 deletions(-) > > > > diff --git a/meta/lib/oeqa/core/target/ssh.py b/meta/lib/oeqa/core/target/ssh.py > > index 13fd5b2a49..466a795eb4 100644 > > --- a/meta/lib/oeqa/core/target/ssh.py > > +++ b/meta/lib/oeqa/core/target/ssh.py > > @@ -229,22 +229,28 @@ def SSHCall(command, logger, timeout=None, **opts): > > if timeout: > > endtime = starttime + timeout > > eof = False > > + os.set_blocking(process.stdout.fileno(), False) > > while time.time() < endtime and not eof: > > - logger.debug('time: %s, endtime: %s' % (time.time(), endtime)) > > try: > > + logger.debug('Waiting for process output: time: %s, endtime: %s' % (time.time(), endtime)) > > if select.select([process.stdout], [], [], 5)[0] != []: > > - reader = codecs.getreader('utf-8')(process.stdout, 'ignore') > > - data = reader.read(1024, 4096) > > + # wait a bit for more data, tries to avoid reading single characters > > + time.sleep(0.2) > > + data = process.stdout.read() > > if not data: > > - process.stdout.close() > > eof = True > > I'm not sure I understand the use of eof now. If stdout is non- > blocking, it would be perfectly normal for it to return with no data > but that doesn't mean we're at eof, it just means there is no data > ready? The select() makes sure that there must be data available for read(), or eof when the file descriptor is closed. But closed file descriptor does not mean that the subprocess is dead. For that we must wait() with a timeout and possibly after that kill(). > I suspect the code might be ok but we should stop calling it eof since > it no longer is? > > I also tend to dislike sleep() codepaths as it usually means the code > is sub optimal. I was going to make further suggestions but I stepped > back and looked at the bigger picture of the code. I think sleep() is ok in this case to reduce busy looping over serial console and read() more data than what select() initially saw. Reading single character at a time is a bit silly and just wastes CPU time when we don't need ms level accuracy. > The code is basically doing: > > if timeout: > <complex code> > else: > output = process.communicate() > > > Python 3.3 added a timeout parameter to subprocess.communicate(). Can > we just use that instead? Or is there something special about our needs > which means we can't? The current code uses timeout as "inactivity timeout" and thus code reads output using select() and read() until the end or if there is no output until timeout. Code could use communicate() with a timeout but would still need to use select() and read() loop to check if there was any output. I did not see a simple way to rewrite this in a more pythonic way. The reader iterators did the wrong thing and ended up dead locking when subprocess hangs. Cheers, -Mikko
On Thu, 2023-02-09 at 10:09 +0200, Mikko Rapeli wrote: > When qemu machine hangs, the ssh commands done by tests > are not timing out. do_testimage() task has last logs like this: > > DEBUG: time: 1673531086.3155053, endtime: 1673531686.315502 > > The test process is stuck for hours, or for ever if the > executing command or test case did not set a timeout correctly. > The default 300 second timeout is not working when target hangs. > Note that timeout is really a "inactive timeout" since data returned > by the process will reset the timeout. > > Make the process stdout non-blocking so read() will always return > right away using os.set_blocking() available in python 3.5 and later. > > Then change from python codec reader to plain read() and make > the ssh subprocess stdout non-blocking. Even with select() > making sure the file had input to be read, the codec reader was > trying to find more stuff and blocking for ever when process hangs. > > While at it, add a small timeout to read data in larger chunks if > possible. This avoids reading data one or few characters at a time > and makes the debug logs more readable. > > close() the stdout file in all cases after read loop is complete. > > Then make sure to wait or kill the ssh subprocess in all cases. > Just reading the output stream and receiving EOF there does not mean > that the process exited, and wait() needs a timeout if the process > is hanging. In the end kill the process and return the return value > and captured output utf-8 encoded, just like before these changes. > > This fixes ssh run() related deadlocks when a qemu target hangs > completely. > > Signed-off-by: Mikko Rapeli <mikko.rapeli@linaro.org> > --- > meta/lib/oeqa/core/target/ssh.py | 33 +++++++++++++++++++++++++------- > 1 file changed, 26 insertions(+), 7 deletions(-) > > diff --git a/meta/lib/oeqa/core/target/ssh.py b/meta/lib/oeqa/core/target/ssh.py > index 13fd5b2a49..466a795eb4 100644 > --- a/meta/lib/oeqa/core/target/ssh.py > +++ b/meta/lib/oeqa/core/target/ssh.py > @@ -229,22 +229,28 @@ def SSHCall(command, logger, timeout=None, **opts): > if timeout: > endtime = starttime + timeout > eof = False > + os.set_blocking(process.stdout.fileno(), False) > while time.time() < endtime and not eof: > - logger.debug('time: %s, endtime: %s' % (time.time(), endtime)) > try: > + logger.debug('Waiting for process output: time: %s, endtime: %s' % (time.time(), endtime)) > if select.select([process.stdout], [], [], 5)[0] != []: > - reader = codecs.getreader('utf-8')(process.stdout, 'ignore') > - data = reader.read(1024, 4096) > + # wait a bit for more data, tries to avoid reading single characters > + time.sleep(0.2) > + data = process.stdout.read() > if not data: > - process.stdout.close() > eof = True > else: > - output += data > - logger.debug('Partial data from SSH call:\n%s' % data) > + # ignore errors to capture as much as possible > + string = data.decode('utf-8', errors='ignore') I've been trying to remember why we have the reader here and I think the issue was you can't call decode on partial data :(. If it happens that there is a multi byte character on the stream crossing the boundary then you'll get errors. "ignoring" them just means you'll corrupt data. The case is rare but with long running commands with utf8 data like some of the ptests, it can be an issue. That is why the code was written the way it was... Cheers, Richard
Hi, On Thu, Feb 09, 2023 at 10:20:36AM +0000, Richard Purdie wrote: > On Thu, 2023-02-09 at 10:09 +0200, Mikko Rapeli wrote: > > When qemu machine hangs, the ssh commands done by tests > > are not timing out. do_testimage() task has last logs like this: > > > > DEBUG: time: 1673531086.3155053, endtime: 1673531686.315502 > > > > The test process is stuck for hours, or for ever if the > > executing command or test case did not set a timeout correctly. > > The default 300 second timeout is not working when target hangs. > > Note that timeout is really a "inactive timeout" since data returned > > by the process will reset the timeout. > > > > Make the process stdout non-blocking so read() will always return > > right away using os.set_blocking() available in python 3.5 and later. > > > > Then change from python codec reader to plain read() and make > > the ssh subprocess stdout non-blocking. Even with select() > > making sure the file had input to be read, the codec reader was > > trying to find more stuff and blocking for ever when process hangs. > > > > While at it, add a small timeout to read data in larger chunks if > > possible. This avoids reading data one or few characters at a time > > and makes the debug logs more readable. > > > > close() the stdout file in all cases after read loop is complete. > > > > Then make sure to wait or kill the ssh subprocess in all cases. > > Just reading the output stream and receiving EOF there does not mean > > that the process exited, and wait() needs a timeout if the process > > is hanging. In the end kill the process and return the return value > > and captured output utf-8 encoded, just like before these changes. > > > > This fixes ssh run() related deadlocks when a qemu target hangs > > completely. > > > > Signed-off-by: Mikko Rapeli <mikko.rapeli@linaro.org> > > --- > > meta/lib/oeqa/core/target/ssh.py | 33 +++++++++++++++++++++++++------- > > 1 file changed, 26 insertions(+), 7 deletions(-) > > > > diff --git a/meta/lib/oeqa/core/target/ssh.py b/meta/lib/oeqa/core/target/ssh.py > > index 13fd5b2a49..466a795eb4 100644 > > --- a/meta/lib/oeqa/core/target/ssh.py > > +++ b/meta/lib/oeqa/core/target/ssh.py > > @@ -229,22 +229,28 @@ def SSHCall(command, logger, timeout=None, **opts): > > if timeout: > > endtime = starttime + timeout > > eof = False > > + os.set_blocking(process.stdout.fileno(), False) > > while time.time() < endtime and not eof: > > - logger.debug('time: %s, endtime: %s' % (time.time(), endtime)) > > try: > > + logger.debug('Waiting for process output: time: %s, endtime: %s' % (time.time(), endtime)) > > if select.select([process.stdout], [], [], 5)[0] != []: > > - reader = codecs.getreader('utf-8')(process.stdout, 'ignore') > > - data = reader.read(1024, 4096) > > + # wait a bit for more data, tries to avoid reading single characters > > + time.sleep(0.2) > > + data = process.stdout.read() > > if not data: > > - process.stdout.close() > > eof = True > > else: > > - output += data > > - logger.debug('Partial data from SSH call:\n%s' % data) > > + # ignore errors to capture as much as possible > > + string = data.decode('utf-8', errors='ignore') > > I've been trying to remember why we have the reader here and I think > the issue was you can't call decode on partial data :(. I get this. Sadly readers read more than file has available and they don't have timeouts... > If it happens that there is a multi byte character on the stream > crossing the boundary then you'll get errors. "ignoring" them just > means you'll corrupt data. The case is rare but with long running > commands with utf8 data like some of the ptests, it can be an issue. > > That is why the code was written the way it was... For logging purposes errors are fine. Should the raw data be captured to output then? Or another conversion to utf-8 after all data has been read? Reading more than one character at a time for sure helps here but granted there may be windows where we hit the in character bytes.. Cheers, -Mikko
On Thu, 2023-02-09 at 12:35 +0200, Mikko Rapeli wrote: > Hi, > > On Thu, Feb 09, 2023 at 10:20:36AM +0000, Richard Purdie wrote: > > On Thu, 2023-02-09 at 10:09 +0200, Mikko Rapeli wrote: > > > When qemu machine hangs, the ssh commands done by tests > > > are not timing out. do_testimage() task has last logs like this: > > > > > > DEBUG: time: 1673531086.3155053, endtime: 1673531686.315502 > > > > > > The test process is stuck for hours, or for ever if the > > > executing command or test case did not set a timeout correctly. > > > The default 300 second timeout is not working when target hangs. > > > Note that timeout is really a "inactive timeout" since data returned > > > by the process will reset the timeout. > > > > > > Make the process stdout non-blocking so read() will always return > > > right away using os.set_blocking() available in python 3.5 and later. > > > > > > Then change from python codec reader to plain read() and make > > > the ssh subprocess stdout non-blocking. Even with select() > > > making sure the file had input to be read, the codec reader was > > > trying to find more stuff and blocking for ever when process hangs. > > > > > > While at it, add a small timeout to read data in larger chunks if > > > possible. This avoids reading data one or few characters at a time > > > and makes the debug logs more readable. > > > > > > close() the stdout file in all cases after read loop is complete. > > > > > > Then make sure to wait or kill the ssh subprocess in all cases. > > > Just reading the output stream and receiving EOF there does not mean > > > that the process exited, and wait() needs a timeout if the process > > > is hanging. In the end kill the process and return the return value > > > and captured output utf-8 encoded, just like before these changes. > > > > > > This fixes ssh run() related deadlocks when a qemu target hangs > > > completely. > > > > > > Signed-off-by: Mikko Rapeli <mikko.rapeli@linaro.org> > > > --- > > > meta/lib/oeqa/core/target/ssh.py | 33 +++++++++++++++++++++++++------- > > > 1 file changed, 26 insertions(+), 7 deletions(-) > > > > > > diff --git a/meta/lib/oeqa/core/target/ssh.py b/meta/lib/oeqa/core/target/ssh.py > > > index 13fd5b2a49..466a795eb4 100644 > > > --- a/meta/lib/oeqa/core/target/ssh.py > > > +++ b/meta/lib/oeqa/core/target/ssh.py > > > @@ -229,22 +229,28 @@ def SSHCall(command, logger, timeout=None, **opts): > > > if timeout: > > > endtime = starttime + timeout > > > eof = False > > > + os.set_blocking(process.stdout.fileno(), False) > > > while time.time() < endtime and not eof: > > > - logger.debug('time: %s, endtime: %s' % (time.time(), endtime)) > > > try: > > > + logger.debug('Waiting for process output: time: %s, endtime: %s' % (time.time(), endtime)) > > > if select.select([process.stdout], [], [], 5)[0] != []: > > > - reader = codecs.getreader('utf-8')(process.stdout, 'ignore') > > > - data = reader.read(1024, 4096) > > > + # wait a bit for more data, tries to avoid reading single characters > > > + time.sleep(0.2) > > > + data = process.stdout.read() > > > if not data: > > > - process.stdout.close() > > > eof = True > > > else: > > > - output += data > > > - logger.debug('Partial data from SSH call:\n%s' % data) > > > + # ignore errors to capture as much as possible > > > + string = data.decode('utf-8', errors='ignore') > > > > I've been trying to remember why we have the reader here and I think > > the issue was you can't call decode on partial data :(. > > I get this. Sadly readers read more than file has available and they don't have > timeouts... > > > If it happens that there is a multi byte character on the stream > > crossing the boundary then you'll get errors. "ignoring" them just > > means you'll corrupt data. The case is rare but with long running > > commands with utf8 data like some of the ptests, it can be an issue. > > > > That is why the code was written the way it was... > > For logging purposes errors are fine. Should the raw data be captured to > output then? Or another conversion to utf-8 after all data has been > read? Converting at the end is one option. The other option would be to feed the reader our own stream object where it's read method would only return the data we have. I think you can create dummy "file like" objects in python, it might be easiest with io.BytesIO(). The codec reader could then interact with that as a dummy stream. Cheers, Richard
On Thu, 9 Feb 2023 at 10:36, Mikko Rapeli <mikko.rapeli@linaro.org> wrote: > I think sleep() is ok in this case to reduce busy looping over serial > console and read() more data than what select() initially saw. Reading > single character at a time is a bit silly and just wastes CPU time when > we don't need ms level accuracy. Something doesn't compute here. The writer should be sending data in chunks, and there are buffers all the way to the reader, so where does the splitting into single characters happen? Alex
On Thu, 2023-02-09 at 11:53 +0100, Alexander Kanavin wrote: > On Thu, 9 Feb 2023 at 10:36, Mikko Rapeli <mikko.rapeli@linaro.org> wrote: > > I think sleep() is ok in this case to reduce busy looping over serial > > console and read() more data than what select() initially saw. Reading > > single character at a time is a bit silly and just wastes CPU time when > > we don't need ms level accuracy. > > Something doesn't compute here. The writer should be sending data in > chunks, and there are buffers all the way to the reader, so where does > the splitting into single characters happen? Just guessing but in the case of the serial emulation in qemu, 115200bps is probably "slow" compared to the system speed so they arrive as bytes? Cheers, Richard
Hi, On Thu, Feb 09, 2023 at 11:53:00AM +0100, Alexander Kanavin wrote: > On Thu, 9 Feb 2023 at 10:36, Mikko Rapeli <mikko.rapeli@linaro.org> wrote: > > I think sleep() is ok in this case to reduce busy looping over serial > > console and read() more data than what select() initially saw. Reading > > single character at a time is a bit silly and just wastes CPU time when > > we don't need ms level accuracy. > > Something doesn't compute here. The writer should be sending data in > chunks, and there are buffers all the way to the reader, so where does > the splitting into single characters happen? Sorry I mixed with the serial console patch, but the output was similar over ssh. select() returns with any data, even single character. Fancy readers will read larger chunks but will read more data than is available and thus block for ever. Plain read() will read the data which is available from select() and thus really small amounts like single characters if that's available over the socket. This is visible in the partial read output of do_testimage() task output log which becomes quite unreadable without change. Cheers, -Mikko
On Thu, 9 Feb 2023 at 12:12, Mikko Rapeli <mikko.rapeli@linaro.org> wrote: > select() returns with any data, even single character. Fancy readers > will read larger chunks but will read more data than is available and thus > block for ever. Plain read() will read the data which is available from > select() and thus really small amounts like single characters if that's > available over the socket. This is visible in the partial read output of > do_testimage() task output log which becomes quite unreadable without change. Yes, but this does not explain why single characters happen to begin with - RP offered a hypothesis, but the problem is that data arrives that way, not the tight-loop manner in which we're reading it. Alex
Hi, On Thu, Feb 09, 2023 at 12:42:09PM +0100, Alexander Kanavin wrote: > On Thu, 9 Feb 2023 at 12:12, Mikko Rapeli <mikko.rapeli@linaro.org> wrote: > > select() returns with any data, even single character. Fancy readers > > will read larger chunks but will read more data than is available and thus > > block for ever. Plain read() will read the data which is available from > > select() and thus really small amounts like single characters if that's > > available over the socket. This is visible in the partial read output of > > do_testimage() task output log which becomes quite unreadable without change. > > Yes, but this does not explain why single characters happen to begin > with - RP offered a hypothesis, but the problem is that data arrives > that way, not the tight-loop manner in which we're reading it. Without the sleep(0.2) with ssh, the do_testimage task output can be like this (which isn't too bad, just annoying): DEBUG: [Running]$ ssh -l root -o ServerAliveCountMax=2 -o ServerAliveInterval=30 -o UserKnownHostsFile=/dev/null -o StrictHostKeyChecking=no -o LogLevel=ERROR -p 2222 127.0.0.1 export PATH=/usr/sbin:/sbin:/usr/bin:/bin; cat /etc/controllerimage DEBUG: Waiting for process output: time: 1675942983.6053402, endtime: 1675943283.6033137 DEBUG: Partial data from SSH call: cat: DEBUG: Waiting for process output: time: 1675942986.2085373, endtime: 1675943286.208536 DEBUG: Partial data from SSH call: /etc/controllerimage DEBUG: Waiting for process output: time: 1675942986.2111514, endtime: 1675943286.2111504 DEBUG: Partial data from SSH call: : No such file or directory DEBUG: Waiting for process output: time: 1675942986.2140906, endtime: 1675943286.2140894 DEBUG: Partial data from SSH call: DEBUG: Waiting for process output: time: 1675942986.216261, endtime: 1675943286.21626 DEBUG: [Command returned '1' after 2.64 seconds] I think qemu is just slower than what anyone expects, including the ssh protocol which does some buffer but then gives up. It may also be the kernel and image running under qemu which busy starting up, e.g. generating entropy or keys. With the sleep(0.2) output is: DEBUG: [Running]$ ssh -l root -o ServerAliveCountMax=2 -o ServerAliveInterval=30 -o UserKnownHostsFile=/dev/null -o StrictHostKeyChecking=no -o LogLevel=ERROR -p 2222 127.0.0.1 export PATH=/usr/sbin:/sbin:/usr/bin:/bin; cat /etc/controllerimage DEBUG: Waiting for process output: time: 1675945065.6815689, endtime: 1675945365.6790204 DEBUG: Partial data from SSH call: cat: /etc/controllerimage: No such file or directory DEBUG: Waiting for process output: time: 1675945068.308055, endtime: 1675945368.3080513 DEBUG: [Command returned '1' after 2.83 seconds] DEBUG: Command: cat /etc/controllerimage Status: 1 Output: cat: /etc/controllerimage: No such file or directory NOTE: ... ok Similar thing happens then over serial console which is even slower. I know debugging the data read and written to serial are not wanted but I had the changes in place and saw bootloader and kernel messages being read initally one character at a time and later on 5-10 characters at a time which made debugging problems really annoying. The small sleep does improve things, doesn't solve everything but also doesn't really harm. No data is lost and less time is used looping around the file descriptor. Cheers, -Mikko
diff --git a/meta/lib/oeqa/core/target/ssh.py b/meta/lib/oeqa/core/target/ssh.py index 13fd5b2a49..466a795eb4 100644 --- a/meta/lib/oeqa/core/target/ssh.py +++ b/meta/lib/oeqa/core/target/ssh.py @@ -229,22 +229,28 @@ def SSHCall(command, logger, timeout=None, **opts): if timeout: endtime = starttime + timeout eof = False + os.set_blocking(process.stdout.fileno(), False) while time.time() < endtime and not eof: - logger.debug('time: %s, endtime: %s' % (time.time(), endtime)) try: + logger.debug('Waiting for process output: time: %s, endtime: %s' % (time.time(), endtime)) if select.select([process.stdout], [], [], 5)[0] != []: - reader = codecs.getreader('utf-8')(process.stdout, 'ignore') - data = reader.read(1024, 4096) + # wait a bit for more data, tries to avoid reading single characters + time.sleep(0.2) + data = process.stdout.read() if not data: - process.stdout.close() eof = True else: - output += data - logger.debug('Partial data from SSH call:\n%s' % data) + # ignore errors to capture as much as possible + string = data.decode('utf-8', errors='ignore') + output += string + logger.debug('Partial data from SSH call:\n%s' % string) endtime = time.time() + timeout except InterruptedError: + logger.debug('InterruptedError') continue + process.stdout.close() + # process hasn't returned yet if not eof: process.terminate() @@ -252,6 +258,7 @@ def SSHCall(command, logger, timeout=None, **opts): try: process.kill() except OSError: + logger.debug('OSError when killing process') pass endtime = time.time() - starttime lastline = ("\nProcess killed - no output for %d seconds. Total" @@ -263,6 +270,17 @@ def SSHCall(command, logger, timeout=None, **opts): output = process.communicate()[0].decode('utf-8', errors='ignore') logger.debug('Data from SSH call:\n%s' % output.rstrip()) + # timout or not, make sure process exits and is not hanging + if process.returncode == None: + try: + process.wait(timeout=5) + except TimeoutExpired: + try: + process.kill() + except OSError: + logger.debug('OSError') + pass + options = { "stdout": subprocess.PIPE, "stderr": subprocess.STDOUT, @@ -290,4 +308,5 @@ def SSHCall(command, logger, timeout=None, **opts): process.kill() logger.debug('Something went wrong, killing SSH process') raise - return (process.wait(), output.rstrip()) + + return (process.returncode, output.rstrip())
When qemu machine hangs, the ssh commands done by tests are not timing out. do_testimage() task has last logs like this: DEBUG: time: 1673531086.3155053, endtime: 1673531686.315502 The test process is stuck for hours, or for ever if the executing command or test case did not set a timeout correctly. The default 300 second timeout is not working when target hangs. Note that timeout is really a "inactive timeout" since data returned by the process will reset the timeout. Make the process stdout non-blocking so read() will always return right away using os.set_blocking() available in python 3.5 and later. Then change from python codec reader to plain read() and make the ssh subprocess stdout non-blocking. Even with select() making sure the file had input to be read, the codec reader was trying to find more stuff and blocking for ever when process hangs. While at it, add a small timeout to read data in larger chunks if possible. This avoids reading data one or few characters at a time and makes the debug logs more readable. close() the stdout file in all cases after read loop is complete. Then make sure to wait or kill the ssh subprocess in all cases. Just reading the output stream and receiving EOF there does not mean that the process exited, and wait() needs a timeout if the process is hanging. In the end kill the process and return the return value and captured output utf-8 encoded, just like before these changes. This fixes ssh run() related deadlocks when a qemu target hangs completely. Signed-off-by: Mikko Rapeli <mikko.rapeli@linaro.org> --- meta/lib/oeqa/core/target/ssh.py | 33 +++++++++++++++++++++++++------- 1 file changed, 26 insertions(+), 7 deletions(-)