Re: [openstack-dev] [Nova] nova-compute deadlock
Dear stackers, FYI. Eventually I report this problem to libguestfs. A workaround has been included into libguestfs code to fix it. Thanks for your supporting! https://bugzilla.redhat.com/show_bug.cgi?id=1123007 On Sat, Jun 7, 2014 at 3:27 AM, Qin Zhao chaoc...@gmail.com wrote: Yuriy, And I think if we use proxy object of multiprocessing, the green thread will not switch during we call libguestfs. Is that correct? On Fri, Jun 6, 2014 at 2:44 AM, Qin Zhao chaoc...@gmail.com wrote: Hi Yuriy, I read multiprocessing source code just now. Now I feel it may not solve this problem very easily. For example, let us assume that we will use the proxy object in Manager's process to call libguestfs. In manager.py, I see it needs to create a pipe, before fork the child process. The write end of this pipe is required by child process. http://sourcecodebrowser.com/python-multiprocessing/2.6.2.1/classmultiprocessing_1_1managers_1_1_base_manager.html#a57fe9abe7a3d281286556c4bf3fbf4d5 And in Process._bootstrp(), I think we will need to register a function to be called by _run_after_forkers(), in order to closed the fds inherited from Nova process. http://sourcecodebrowser.com/python-multiprocessing/2.6.2.1/classmultiprocessing_1_1process_1_1_process.html#ae594800e7bdef288d9bfbf8b79019d2e And we also can not close the write end fd created by Manager in _run_after_forkers(). One feasible way may be getting that fd from the 5th element of _args attribute of Process object, then skip to close this fd I have not investigate if or not Manager need to use other fds, besides this pipe. Personally, I feel such an implementation will be a little tricky and risky, because it tightly depends on Manager code. If Manager opens other files, or change the argument order, our code will fail to run. Am I wrong? Is there any other safer way? On Thu, Jun 5, 2014 at 11:40 PM, Yuriy Taraday yorik@gmail.com wrote: Please take a look at https://docs.python.org/2.7/library/multiprocessing.html#managers - everything is already implemented there. All you need is to start one manager that would serve all your requests to libguestfs. The implementation in stdlib will provide you with all exceptions and return values with minimum code changes on Nova side. Create a new Manager, register an libguestfs endpoint in it and call start(). It will spawn a separate process that will speak with calling process over very simple RPC. From the looks of it all you need to do is replace tpool.Proxy calls in VFSGuestFS.setup method to calls to this new Manager. On Thu, Jun 5, 2014 at 7:21 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Yuriy, Thanks for reading my bug! You are right. Python 3.3 or 3.4 should not have this issue, since they have can secure the file descriptor. Before OpenStack move to Python 3, we may still need a solution. Calling libguestfs in a separate process seems to be a way. This way, Nova code can close those fd by itself, not depending upon CLOEXEC. However, that will be an expensive solution, since it requires a lot of code change. At least we need to write code to pass the return value and exception between these two processes. That will make this solution very complex. Do you agree? On Thu, Jun 5, 2014 at 9:39 PM, Yuriy Taraday yorik@gmail.com wrote: This behavior of os.pipe() has changed in Python 3.x so it won't be an issue on newer Python (if only it was accessible for us). From the looks of it you can mitigate the problem by running libguestfs requests in a separate process (multiprocessing.managers comes to mind). This way the only descriptors child process could theoretically inherit would be long-lived pipes to main process although they won't leak because they should be marked with CLOEXEC before any libguestfs request is run. The other benefit is that this separate process won't be busy opening and closing tons of fds so the problem with inheriting will be avoided. On Thu, Jun 5, 2014 at 2:17 PM, laserjetyang laserjety...@gmail.com wrote: Will this patch of Python fix your problem? *http://bugs.python.org/issue7213 http://bugs.python.org/issue7213* On Wed, Jun 4, 2014 at 10:41 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Zhu Zhu, Thank you for reading my diagram! I need to clarify that this problem does not occur during data injection. Before creating the ISO, the driver code will extend the disk. Libguestfs is invoked in that time frame. And now I think this problem may occur at any time, if the code use tpool to invoke libguestfs, and one external commend is executed in another green thread simultaneously. Please correct me if I am wrong. I think one simple solution for this issue is to call libguestfs routine in greenthread, rather than another native thread. But it will impact the performance very much. So I do not think that is an acceptable solution. On Wed, Jun 4, 2014 at 12:00 PM,
Re: [openstack-dev] [Nova] nova-compute deadlock
On Sat, May 31, 2014 at 01:25:04AM +0800, Qin Zhao wrote: Hi all, When I run Icehouse code, I encountered a strange problem. The nova-compute service becomes stuck, when I boot instances. I report this bug in https://bugs.launchpad.net/nova/+bug/1313477. After thinking several days, I feel I know its root cause. This bug should be a deadlock problem cause by pipe fd leaking. I draw a diagram to illustrate this problem. https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 However, I have not find a very good solution to prevent this deadlock. This problem is related with Python runtime, libguestfs, and eventlet. The situation is a little complicated. Is there any expert who can help me to look for a solution? I will appreciate for your help! Thanks for the useful diagram. libguestfs itself is very careful to open all file descriptors with O_CLOEXEC (atomically if the OS supports that), so I'm fairly confident that the bug is in Python 2, not in libguestfs. Another thing to say is that g.shutdown() sends a kill 9 signal to the subprocess. Furthermore you can obtain the qemu PID (g.get_pid()) and send any signal you want to the process. I wonder if a simpler way to fix this wouldn't be something like adding a tiny C extension to the Python code to use pipe2 to open the Python pipe with O_CLOEXEC atomically? Are we allowed Python extensions in OpenStack? BTW do feel free to CC libgues...@redhat.com on any libguestfs problems you have. You don't need to subscribe to the list. Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-p2v converts physical machines to virtual machines. Boot with a live CD or over the network (PXE) and turn machines into KVM guests. http://libguestfs.org/virt-v2v ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Nova] nova-compute deadlock
Yuriy, And I think if we use proxy object of multiprocessing, the green thread will not switch during we call libguestfs. Is that correct? On Fri, Jun 6, 2014 at 2:44 AM, Qin Zhao chaoc...@gmail.com wrote: Hi Yuriy, I read multiprocessing source code just now. Now I feel it may not solve this problem very easily. For example, let us assume that we will use the proxy object in Manager's process to call libguestfs. In manager.py, I see it needs to create a pipe, before fork the child process. The write end of this pipe is required by child process. http://sourcecodebrowser.com/python-multiprocessing/2.6.2.1/classmultiprocessing_1_1managers_1_1_base_manager.html#a57fe9abe7a3d281286556c4bf3fbf4d5 And in Process._bootstrp(), I think we will need to register a function to be called by _run_after_forkers(), in order to closed the fds inherited from Nova process. http://sourcecodebrowser.com/python-multiprocessing/2.6.2.1/classmultiprocessing_1_1process_1_1_process.html#ae594800e7bdef288d9bfbf8b79019d2e And we also can not close the write end fd created by Manager in _run_after_forkers(). One feasible way may be getting that fd from the 5th element of _args attribute of Process object, then skip to close this fd I have not investigate if or not Manager need to use other fds, besides this pipe. Personally, I feel such an implementation will be a little tricky and risky, because it tightly depends on Manager code. If Manager opens other files, or change the argument order, our code will fail to run. Am I wrong? Is there any other safer way? On Thu, Jun 5, 2014 at 11:40 PM, Yuriy Taraday yorik@gmail.com wrote: Please take a look at https://docs.python.org/2.7/library/multiprocessing.html#managers - everything is already implemented there. All you need is to start one manager that would serve all your requests to libguestfs. The implementation in stdlib will provide you with all exceptions and return values with minimum code changes on Nova side. Create a new Manager, register an libguestfs endpoint in it and call start(). It will spawn a separate process that will speak with calling process over very simple RPC. From the looks of it all you need to do is replace tpool.Proxy calls in VFSGuestFS.setup method to calls to this new Manager. On Thu, Jun 5, 2014 at 7:21 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Yuriy, Thanks for reading my bug! You are right. Python 3.3 or 3.4 should not have this issue, since they have can secure the file descriptor. Before OpenStack move to Python 3, we may still need a solution. Calling libguestfs in a separate process seems to be a way. This way, Nova code can close those fd by itself, not depending upon CLOEXEC. However, that will be an expensive solution, since it requires a lot of code change. At least we need to write code to pass the return value and exception between these two processes. That will make this solution very complex. Do you agree? On Thu, Jun 5, 2014 at 9:39 PM, Yuriy Taraday yorik@gmail.com wrote: This behavior of os.pipe() has changed in Python 3.x so it won't be an issue on newer Python (if only it was accessible for us). From the looks of it you can mitigate the problem by running libguestfs requests in a separate process (multiprocessing.managers comes to mind). This way the only descriptors child process could theoretically inherit would be long-lived pipes to main process although they won't leak because they should be marked with CLOEXEC before any libguestfs request is run. The other benefit is that this separate process won't be busy opening and closing tons of fds so the problem with inheriting will be avoided. On Thu, Jun 5, 2014 at 2:17 PM, laserjetyang laserjety...@gmail.com wrote: Will this patch of Python fix your problem? *http://bugs.python.org/issue7213 http://bugs.python.org/issue7213* On Wed, Jun 4, 2014 at 10:41 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Zhu Zhu, Thank you for reading my diagram! I need to clarify that this problem does not occur during data injection. Before creating the ISO, the driver code will extend the disk. Libguestfs is invoked in that time frame. And now I think this problem may occur at any time, if the code use tpool to invoke libguestfs, and one external commend is executed in another green thread simultaneously. Please correct me if I am wrong. I think one simple solution for this issue is to call libguestfs routine in greenthread, rather than another native thread. But it will impact the performance very much. So I do not think that is an acceptable solution. On Wed, Jun 4, 2014 at 12:00 PM, Zhu Zhu bjzzu...@gmail.com wrote: Hi Qin Zhao, Thanks for raising this issue and analysis. According to the issue description and happen scenario( https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 ), if that's the case, concurrent
Re: [openstack-dev] [Nova] nova-compute deadlock
Will this patch of Python fix your problem? *http://bugs.python.org/issue7213 http://bugs.python.org/issue7213* On Wed, Jun 4, 2014 at 10:41 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Zhu Zhu, Thank you for reading my diagram! I need to clarify that this problem does not occur during data injection. Before creating the ISO, the driver code will extend the disk. Libguestfs is invoked in that time frame. And now I think this problem may occur at any time, if the code use tpool to invoke libguestfs, and one external commend is executed in another green thread simultaneously. Please correct me if I am wrong. I think one simple solution for this issue is to call libguestfs routine in greenthread, rather than another native thread. But it will impact the performance very much. So I do not think that is an acceptable solution. On Wed, Jun 4, 2014 at 12:00 PM, Zhu Zhu bjzzu...@gmail.com wrote: Hi Qin Zhao, Thanks for raising this issue and analysis. According to the issue description and happen scenario( https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 ), if that's the case, concurrent mutiple KVM spawn instances(*with both config drive and data injection enabled*) are triggered, the issue can be very likely to happen. As in libvirt/driver.py _create_image method, right after iso making cdb.make_drive, the driver will attempt data injection which will call the libguestfs launch in another thread. Looks there were also a couple of libguestfs hang issues from Launch pad as below. . I am not sure if libguestfs itself can have certain mechanism to free/close the fds that inherited from parent process instead of require explicitly calling the tear down. Maybe open a defect to libguestfs to see what their thoughts? https://bugs.launchpad.net/nova/+bug/1286256 https://bugs.launchpad.net/nova/+bug/1270304 -- Zhu Zhu Best Regards *From:* Qin Zhao chaoc...@gmail.com *Date:* 2014-05-31 01:25 *To:* OpenStack Development Mailing List (not for usage questions) openstack-dev@lists.openstack.org *Subject:* [openstack-dev] [Nova] nova-compute deadlock Hi all, When I run Icehouse code, I encountered a strange problem. The nova-compute service becomes stuck, when I boot instances. I report this bug in https://bugs.launchpad.net/nova/+bug/1313477. After thinking several days, I feel I know its root cause. This bug should be a deadlock problem cause by pipe fd leaking. I draw a diagram to illustrate this problem. https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 However, I have not find a very good solution to prevent this deadlock. This problem is related with Python runtime, libguestfs, and eventlet. The situation is a little complicated. Is there any expert who can help me to look for a solution? I will appreciate for your help! -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Nova] nova-compute deadlock
This behavior of os.pipe() has changed in Python 3.x so it won't be an issue on newer Python (if only it was accessible for us). From the looks of it you can mitigate the problem by running libguestfs requests in a separate process (multiprocessing.managers comes to mind). This way the only descriptors child process could theoretically inherit would be long-lived pipes to main process although they won't leak because they should be marked with CLOEXEC before any libguestfs request is run. The other benefit is that this separate process won't be busy opening and closing tons of fds so the problem with inheriting will be avoided. On Thu, Jun 5, 2014 at 2:17 PM, laserjetyang laserjety...@gmail.com wrote: Will this patch of Python fix your problem? *http://bugs.python.org/issue7213 http://bugs.python.org/issue7213* On Wed, Jun 4, 2014 at 10:41 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Zhu Zhu, Thank you for reading my diagram! I need to clarify that this problem does not occur during data injection. Before creating the ISO, the driver code will extend the disk. Libguestfs is invoked in that time frame. And now I think this problem may occur at any time, if the code use tpool to invoke libguestfs, and one external commend is executed in another green thread simultaneously. Please correct me if I am wrong. I think one simple solution for this issue is to call libguestfs routine in greenthread, rather than another native thread. But it will impact the performance very much. So I do not think that is an acceptable solution. On Wed, Jun 4, 2014 at 12:00 PM, Zhu Zhu bjzzu...@gmail.com wrote: Hi Qin Zhao, Thanks for raising this issue and analysis. According to the issue description and happen scenario( https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 ), if that's the case, concurrent mutiple KVM spawn instances(*with both config drive and data injection enabled*) are triggered, the issue can be very likely to happen. As in libvirt/driver.py _create_image method, right after iso making cdb.make_drive, the driver will attempt data injection which will call the libguestfs launch in another thread. Looks there were also a couple of libguestfs hang issues from Launch pad as below. . I am not sure if libguestfs itself can have certain mechanism to free/close the fds that inherited from parent process instead of require explicitly calling the tear down. Maybe open a defect to libguestfs to see what their thoughts? https://bugs.launchpad.net/nova/+bug/1286256 https://bugs.launchpad.net/nova/+bug/1270304 -- Zhu Zhu Best Regards *From:* Qin Zhao chaoc...@gmail.com *Date:* 2014-05-31 01:25 *To:* OpenStack Development Mailing List (not for usage questions) openstack-dev@lists.openstack.org *Subject:* [openstack-dev] [Nova] nova-compute deadlock Hi all, When I run Icehouse code, I encountered a strange problem. The nova-compute service becomes stuck, when I boot instances. I report this bug in https://bugs.launchpad.net/nova/+bug/1313477. After thinking several days, I feel I know its root cause. This bug should be a deadlock problem cause by pipe fd leaking. I draw a diagram to illustrate this problem. https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 However, I have not find a very good solution to prevent this deadlock. This problem is related with Python runtime, libguestfs, and eventlet. The situation is a little complicated. Is there any expert who can help me to look for a solution? I will appreciate for your help! -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Kind regards, Yuriy. ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Nova] nova-compute deadlock
Hi, Thanks for reading my bug! I think this patch can not fix this problem now, because pipe2() requires Python 3.3. On Thu, Jun 5, 2014 at 6:17 PM, laserjetyang laserjety...@gmail.com wrote: Will this patch of Python fix your problem? *http://bugs.python.org/issue7213 http://bugs.python.org/issue7213* On Wed, Jun 4, 2014 at 10:41 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Zhu Zhu, Thank you for reading my diagram! I need to clarify that this problem does not occur during data injection. Before creating the ISO, the driver code will extend the disk. Libguestfs is invoked in that time frame. And now I think this problem may occur at any time, if the code use tpool to invoke libguestfs, and one external commend is executed in another green thread simultaneously. Please correct me if I am wrong. I think one simple solution for this issue is to call libguestfs routine in greenthread, rather than another native thread. But it will impact the performance very much. So I do not think that is an acceptable solution. On Wed, Jun 4, 2014 at 12:00 PM, Zhu Zhu bjzzu...@gmail.com wrote: Hi Qin Zhao, Thanks for raising this issue and analysis. According to the issue description and happen scenario( https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 ), if that's the case, concurrent mutiple KVM spawn instances(*with both config drive and data injection enabled*) are triggered, the issue can be very likely to happen. As in libvirt/driver.py _create_image method, right after iso making cdb.make_drive, the driver will attempt data injection which will call the libguestfs launch in another thread. Looks there were also a couple of libguestfs hang issues from Launch pad as below. . I am not sure if libguestfs itself can have certain mechanism to free/close the fds that inherited from parent process instead of require explicitly calling the tear down. Maybe open a defect to libguestfs to see what their thoughts? https://bugs.launchpad.net/nova/+bug/1286256 https://bugs.launchpad.net/nova/+bug/1270304 -- Zhu Zhu Best Regards *From:* Qin Zhao chaoc...@gmail.com *Date:* 2014-05-31 01:25 *To:* OpenStack Development Mailing List (not for usage questions) openstack-dev@lists.openstack.org *Subject:* [openstack-dev] [Nova] nova-compute deadlock Hi all, When I run Icehouse code, I encountered a strange problem. The nova-compute service becomes stuck, when I boot instances. I report this bug in https://bugs.launchpad.net/nova/+bug/1313477. After thinking several days, I feel I know its root cause. This bug should be a deadlock problem cause by pipe fd leaking. I draw a diagram to illustrate this problem. https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 However, I have not find a very good solution to prevent this deadlock. This problem is related with Python runtime, libguestfs, and eventlet. The situation is a little complicated. Is there any expert who can help me to look for a solution? I will appreciate for your help! -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Nova] nova-compute deadlock
Hi Yuriy, Thanks for reading my bug! You are right. Python 3.3 or 3.4 should not have this issue, since they have can secure the file descriptor. Before OpenStack move to Python 3, we may still need a solution. Calling libguestfs in a separate process seems to be a way. This way, Nova code can close those fd by itself, not depending upon CLOEXEC. However, that will be an expensive solution, since it requires a lot of code change. At least we need to write code to pass the return value and exception between these two processes. That will make this solution very complex. Do you agree? On Thu, Jun 5, 2014 at 9:39 PM, Yuriy Taraday yorik@gmail.com wrote: This behavior of os.pipe() has changed in Python 3.x so it won't be an issue on newer Python (if only it was accessible for us). From the looks of it you can mitigate the problem by running libguestfs requests in a separate process (multiprocessing.managers comes to mind). This way the only descriptors child process could theoretically inherit would be long-lived pipes to main process although they won't leak because they should be marked with CLOEXEC before any libguestfs request is run. The other benefit is that this separate process won't be busy opening and closing tons of fds so the problem with inheriting will be avoided. On Thu, Jun 5, 2014 at 2:17 PM, laserjetyang laserjety...@gmail.com wrote: Will this patch of Python fix your problem? *http://bugs.python.org/issue7213 http://bugs.python.org/issue7213* On Wed, Jun 4, 2014 at 10:41 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Zhu Zhu, Thank you for reading my diagram! I need to clarify that this problem does not occur during data injection. Before creating the ISO, the driver code will extend the disk. Libguestfs is invoked in that time frame. And now I think this problem may occur at any time, if the code use tpool to invoke libguestfs, and one external commend is executed in another green thread simultaneously. Please correct me if I am wrong. I think one simple solution for this issue is to call libguestfs routine in greenthread, rather than another native thread. But it will impact the performance very much. So I do not think that is an acceptable solution. On Wed, Jun 4, 2014 at 12:00 PM, Zhu Zhu bjzzu...@gmail.com wrote: Hi Qin Zhao, Thanks for raising this issue and analysis. According to the issue description and happen scenario( https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 ), if that's the case, concurrent mutiple KVM spawn instances(*with both config drive and data injection enabled*) are triggered, the issue can be very likely to happen. As in libvirt/driver.py _create_image method, right after iso making cdb.make_drive, the driver will attempt data injection which will call the libguestfs launch in another thread. Looks there were also a couple of libguestfs hang issues from Launch pad as below. . I am not sure if libguestfs itself can have certain mechanism to free/close the fds that inherited from parent process instead of require explicitly calling the tear down. Maybe open a defect to libguestfs to see what their thoughts? https://bugs.launchpad.net/nova/+bug/1286256 https://bugs.launchpad.net/nova/+bug/1270304 -- Zhu Zhu Best Regards *From:* Qin Zhao chaoc...@gmail.com *Date:* 2014-05-31 01:25 *To:* OpenStack Development Mailing List (not for usage questions) openstack-dev@lists.openstack.org *Subject:* [openstack-dev] [Nova] nova-compute deadlock Hi all, When I run Icehouse code, I encountered a strange problem. The nova-compute service becomes stuck, when I boot instances. I report this bug in https://bugs.launchpad.net/nova/+bug/1313477. After thinking several days, I feel I know its root cause. This bug should be a deadlock problem cause by pipe fd leaking. I draw a diagram to illustrate this problem. https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 However, I have not find a very good solution to prevent this deadlock. This problem is related with Python runtime, libguestfs, and eventlet. The situation is a little complicated. Is there any expert who can help me to look for a solution? I will appreciate for your help! -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Kind regards, Yuriy.
Re: [openstack-dev] [Nova] nova-compute deadlock
Please take a look at https://docs.python.org/2.7/library/multiprocessing.html#managers - everything is already implemented there. All you need is to start one manager that would serve all your requests to libguestfs. The implementation in stdlib will provide you with all exceptions and return values with minimum code changes on Nova side. Create a new Manager, register an libguestfs endpoint in it and call start(). It will spawn a separate process that will speak with calling process over very simple RPC. From the looks of it all you need to do is replace tpool.Proxy calls in VFSGuestFS.setup method to calls to this new Manager. On Thu, Jun 5, 2014 at 7:21 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Yuriy, Thanks for reading my bug! You are right. Python 3.3 or 3.4 should not have this issue, since they have can secure the file descriptor. Before OpenStack move to Python 3, we may still need a solution. Calling libguestfs in a separate process seems to be a way. This way, Nova code can close those fd by itself, not depending upon CLOEXEC. However, that will be an expensive solution, since it requires a lot of code change. At least we need to write code to pass the return value and exception between these two processes. That will make this solution very complex. Do you agree? On Thu, Jun 5, 2014 at 9:39 PM, Yuriy Taraday yorik@gmail.com wrote: This behavior of os.pipe() has changed in Python 3.x so it won't be an issue on newer Python (if only it was accessible for us). From the looks of it you can mitigate the problem by running libguestfs requests in a separate process (multiprocessing.managers comes to mind). This way the only descriptors child process could theoretically inherit would be long-lived pipes to main process although they won't leak because they should be marked with CLOEXEC before any libguestfs request is run. The other benefit is that this separate process won't be busy opening and closing tons of fds so the problem with inheriting will be avoided. On Thu, Jun 5, 2014 at 2:17 PM, laserjetyang laserjety...@gmail.com wrote: Will this patch of Python fix your problem? *http://bugs.python.org/issue7213 http://bugs.python.org/issue7213* On Wed, Jun 4, 2014 at 10:41 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Zhu Zhu, Thank you for reading my diagram! I need to clarify that this problem does not occur during data injection. Before creating the ISO, the driver code will extend the disk. Libguestfs is invoked in that time frame. And now I think this problem may occur at any time, if the code use tpool to invoke libguestfs, and one external commend is executed in another green thread simultaneously. Please correct me if I am wrong. I think one simple solution for this issue is to call libguestfs routine in greenthread, rather than another native thread. But it will impact the performance very much. So I do not think that is an acceptable solution. On Wed, Jun 4, 2014 at 12:00 PM, Zhu Zhu bjzzu...@gmail.com wrote: Hi Qin Zhao, Thanks for raising this issue and analysis. According to the issue description and happen scenario( https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 ), if that's the case, concurrent mutiple KVM spawn instances(*with both config drive and data injection enabled*) are triggered, the issue can be very likely to happen. As in libvirt/driver.py _create_image method, right after iso making cdb.make_drive, the driver will attempt data injection which will call the libguestfs launch in another thread. Looks there were also a couple of libguestfs hang issues from Launch pad as below. . I am not sure if libguestfs itself can have certain mechanism to free/close the fds that inherited from parent process instead of require explicitly calling the tear down. Maybe open a defect to libguestfs to see what their thoughts? https://bugs.launchpad.net/nova/+bug/1286256 https://bugs.launchpad.net/nova/+bug/1270304 -- Zhu Zhu Best Regards *From:* Qin Zhao chaoc...@gmail.com *Date:* 2014-05-31 01:25 *To:* OpenStack Development Mailing List (not for usage questions) openstack-dev@lists.openstack.org *Subject:* [openstack-dev] [Nova] nova-compute deadlock Hi all, When I run Icehouse code, I encountered a strange problem. The nova-compute service becomes stuck, when I boot instances. I report this bug in https://bugs.launchpad.net/nova/+bug/1313477. After thinking several days, I feel I know its root cause. This bug should be a deadlock problem cause by pipe fd leaking. I draw a diagram to illustrate this problem. https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 However, I have not find a very good solution to prevent this deadlock. This problem is related with Python runtime, libguestfs, and eventlet. The situation is a little complicated. Is there any
Re: [openstack-dev] [Nova] nova-compute deadlock
Hi Yuriy, I read multiprocessing source code just now. Now I feel it may not solve this problem very easily. For example, let us assume that we will use the proxy object in Manager's process to call libguestfs. In manager.py, I see it needs to create a pipe, before fork the child process. The write end of this pipe is required by child process. http://sourcecodebrowser.com/python-multiprocessing/2.6.2.1/classmultiprocessing_1_1managers_1_1_base_manager.html#a57fe9abe7a3d281286556c4bf3fbf4d5 And in Process._bootstrp(), I think we will need to register a function to be called by _run_after_forkers(), in order to closed the fds inherited from Nova process. http://sourcecodebrowser.com/python-multiprocessing/2.6.2.1/classmultiprocessing_1_1process_1_1_process.html#ae594800e7bdef288d9bfbf8b79019d2e And we also can not close the write end fd created by Manager in _run_after_forkers(). One feasible way may be getting that fd from the 5th element of _args attribute of Process object, then skip to close this fd I have not investigate if or not Manager need to use other fds, besides this pipe. Personally, I feel such an implementation will be a little tricky and risky, because it tightly depends on Manager code. If Manager opens other files, or change the argument order, our code will fail to run. Am I wrong? Is there any other safer way? On Thu, Jun 5, 2014 at 11:40 PM, Yuriy Taraday yorik@gmail.com wrote: Please take a look at https://docs.python.org/2.7/library/multiprocessing.html#managers - everything is already implemented there. All you need is to start one manager that would serve all your requests to libguestfs. The implementation in stdlib will provide you with all exceptions and return values with minimum code changes on Nova side. Create a new Manager, register an libguestfs endpoint in it and call start(). It will spawn a separate process that will speak with calling process over very simple RPC. From the looks of it all you need to do is replace tpool.Proxy calls in VFSGuestFS.setup method to calls to this new Manager. On Thu, Jun 5, 2014 at 7:21 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Yuriy, Thanks for reading my bug! You are right. Python 3.3 or 3.4 should not have this issue, since they have can secure the file descriptor. Before OpenStack move to Python 3, we may still need a solution. Calling libguestfs in a separate process seems to be a way. This way, Nova code can close those fd by itself, not depending upon CLOEXEC. However, that will be an expensive solution, since it requires a lot of code change. At least we need to write code to pass the return value and exception between these two processes. That will make this solution very complex. Do you agree? On Thu, Jun 5, 2014 at 9:39 PM, Yuriy Taraday yorik@gmail.com wrote: This behavior of os.pipe() has changed in Python 3.x so it won't be an issue on newer Python (if only it was accessible for us). From the looks of it you can mitigate the problem by running libguestfs requests in a separate process (multiprocessing.managers comes to mind). This way the only descriptors child process could theoretically inherit would be long-lived pipes to main process although they won't leak because they should be marked with CLOEXEC before any libguestfs request is run. The other benefit is that this separate process won't be busy opening and closing tons of fds so the problem with inheriting will be avoided. On Thu, Jun 5, 2014 at 2:17 PM, laserjetyang laserjety...@gmail.com wrote: Will this patch of Python fix your problem? *http://bugs.python.org/issue7213 http://bugs.python.org/issue7213* On Wed, Jun 4, 2014 at 10:41 PM, Qin Zhao chaoc...@gmail.com wrote: Hi Zhu Zhu, Thank you for reading my diagram! I need to clarify that this problem does not occur during data injection. Before creating the ISO, the driver code will extend the disk. Libguestfs is invoked in that time frame. And now I think this problem may occur at any time, if the code use tpool to invoke libguestfs, and one external commend is executed in another green thread simultaneously. Please correct me if I am wrong. I think one simple solution for this issue is to call libguestfs routine in greenthread, rather than another native thread. But it will impact the performance very much. So I do not think that is an acceptable solution. On Wed, Jun 4, 2014 at 12:00 PM, Zhu Zhu bjzzu...@gmail.com wrote: Hi Qin Zhao, Thanks for raising this issue and analysis. According to the issue description and happen scenario( https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 ), if that's the case, concurrent mutiple KVM spawn instances(*with both config drive and data injection enabled*) are triggered, the issue can be very likely to happen. As in libvirt/driver.py _create_image method, right after iso making cdb.make_drive, the driver will
Re: [openstack-dev] [Nova] nova-compute deadlock
Hi Qin Zhao, Thanks for raising this issue and analysis. According to the issue description and happen scenario(https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720), if that's the case, concurrent mutiple KVM spawn instances(with both config drive and data injection enabled) are triggered, the issue can be very likely to happen. As in libvirt/driver.py _create_image method, right after iso making cdb.make_drive, the driver will attempt data injection which will call the libguestfs launch in another thread. Looks there were also a couple of libguestfs hang issues from Launch pad as below. . I am not sure if libguestfs itself can have certain mechanism to free/close the fds that inherited from parent process instead of require explicitly calling the tear down. Maybe open a defect to libguestfs to see what their thoughts? https://bugs.launchpad.net/nova/+bug/1286256 https://bugs.launchpad.net/nova/+bug/1270304 Zhu Zhu Best Regards From: Qin Zhao Date: 2014-05-31 01:25 To: OpenStack Development Mailing List (not for usage questions) Subject: [openstack-dev] [Nova] nova-compute deadlock Hi all, When I run Icehouse code, I encountered a strange problem. The nova-compute service becomes stuck, when I boot instances. I report this bug in https://bugs.launchpad.net/nova/+bug/1313477. After thinking several days, I feel I know its root cause. This bug should be a deadlock problem cause by pipe fd leaking. I draw a diagram to illustrate this problem. https://docs.google.com/drawings/d/1pItX9urLd6fmjws3BVovXQvRg_qMdTHS-0JhYfSkkVc/pub?w=960h=720 However, I have not find a very good solution to prevent this deadlock. This problem is related with Python runtime, libguestfs, and eventlet. The situation is a little complicated. Is there any expert who can help me to look for a solution? I will appreciate for your help! -- Qin Zhao ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev