Hi Peter, > -----Original Message----- > From: Peter Xu [mailto:peterx@redhat.com] > Sent: Wednesday, June 5, 2024 3:32 AM > To: Gonglei (Arei) <arei.gonglei@huawei.com> > Cc: qemu-devel@nongnu.org; yu.zhang@ionos.com; mgalaxy@akamai.com; > elmar.gerdes@ionos.com; zhengchuan <zhengchuan@huawei.com>; > berrange@redhat.com; armbru@redhat.com; lizhijian@fujitsu.com; > pbonzini@redhat.com; mst@redhat.com; Xiexiangyou > <xiexiangyou@huawei.com>; linux-rdma@vger.kernel.org; lixiao (H) > <lixiao91@huawei.com>; jinpu.wang@ionos.com; Wangjialin > <wangjialin23@huawei.com>; Fabiano Rosas <farosas@suse.de> > Subject: Re: [PATCH 0/6] refactor RDMA live migration based on rsocket API > > Hi, Lei, Jialin, > > Thanks a lot for working on this! > > I think we'll need to wait a bit on feedbacks from Jinpu and his team on RDMA > side, also Daniel for iochannels. Also, please remember to copy Fabiano > Rosas in any relevant future posts. We'd also like to know whether he has any > comments too. I have him copied in this reply. > > On Tue, Jun 04, 2024 at 08:14:06PM +0800, Gonglei wrote: > > From: Jialin Wang <wangjialin23@huawei.com> > > > > Hi, > > > > This patch series attempts to refactor RDMA live migration by > > introducing a new QIOChannelRDMA class based on the rsocket API. > > > > The /usr/include/rdma/rsocket.h provides a higher level rsocket API > > that is a 1-1 match of the normal kernel 'sockets' API, which hides > > the detail of rdma protocol into rsocket and allows us to add support > > for some modern features like multifd more easily. > > > > Here is the previous discussion on refactoring RDMA live migration > > using the rsocket API: > > > > https://lore.kernel.org/qemu-devel/20240328130255.52257-1-philmd@linar > > o.org/ > > > > We have encountered some bugs when using rsocket and plan to submit > > them to the rdma-core community. > > > > In addition, the use of rsocket makes our programming more convenient, > > but it must be noted that this method introduces multiple memory > > copies, which can be imagined that there will be a certain performance > > degradation, hoping that friends with RDMA network cards can help verify, > thank you! > > It'll be good to elaborate if you tested it in-house. What people should expect > on the numbers exactly? Is that okay from Huawei's POV? > > Besides that, the code looks pretty good at a first glance to me. Before > others chim in, here're some high level comments.. > > Firstly, can we avoid using coroutine when listen()? Might be relevant when I > see that rdma_accept_incoming_migration() runs in a loop to do raccept(), but > would that also hang the qemu main loop even with the coroutine, before all > channels are ready? I'm not a coroutine person, but I think the hope is that > we can make dest QEMU run in a thread in the future just like the src QEMU, so > the less coroutine the better in this path. > Because rsocket is set to non-blocking, raccept will return EAGAIN when no connection is received, coroutine will yield, and will not hang the qemu main loop. > I think I also left a comment elsewhere on whether it would be possible to allow > iochannels implement their own poll() functions to avoid the per-channel poll > thread that is proposed in this series. > > https://lore.kernel.org/r/ZldY21xVExtiMddB@x1n > We noticed that, and it's a big operation. I'm not sure that's a better way. > Personally I think even with the thread proposal it's better than the old rdma > code, but I just still want to double check with you guys. E.g., maybe that just > won't work at all? Again, that'll also be based on the fact that we move > migration incoming into a thread first to keep the dest QEMU main loop intact, > I think, but I hope we will reach that irrelevant of rdma, IOW it'll be nice to > happen even earlier if possible. > Yep. This is a fairly big change, I wonder what other people's suggestions are? > Another nitpick is that qio_channel_rdma_listen_async() doesn't look used and > may prone to removal. > Yes. This is because when we wrote the test case, we wanted to test qio_channel_rdma_connect_async, and also I added qio_channel_rdma_listen_async. It is not used in the RDMA hot migration code. Regards, -Gonglei
On Wed, Jun 05, 2024 at 10:09:43AM +0000, Gonglei (Arei) wrote: > Hi Peter, > > > -----Original Message----- > > From: Peter Xu [mailto:peterx@redhat.com] > > Sent: Wednesday, June 5, 2024 3:32 AM > > To: Gonglei (Arei) <arei.gonglei@huawei.com> > > Cc: qemu-devel@nongnu.org; yu.zhang@ionos.com; mgalaxy@akamai.com; > > elmar.gerdes@ionos.com; zhengchuan <zhengchuan@huawei.com>; > > berrange@redhat.com; armbru@redhat.com; lizhijian@fujitsu.com; > > pbonzini@redhat.com; mst@redhat.com; Xiexiangyou > > <xiexiangyou@huawei.com>; linux-rdma@vger.kernel.org; lixiao (H) > > <lixiao91@huawei.com>; jinpu.wang@ionos.com; Wangjialin > > <wangjialin23@huawei.com>; Fabiano Rosas <farosas@suse.de> > > Subject: Re: [PATCH 0/6] refactor RDMA live migration based on rsocket API > > > > Hi, Lei, Jialin, > > > > Thanks a lot for working on this! > > > > I think we'll need to wait a bit on feedbacks from Jinpu and his team on RDMA > > side, also Daniel for iochannels. Also, please remember to copy Fabiano > > Rosas in any relevant future posts. We'd also like to know whether he has any > > comments too. I have him copied in this reply. > > > > On Tue, Jun 04, 2024 at 08:14:06PM +0800, Gonglei wrote: > > > From: Jialin Wang <wangjialin23@huawei.com> > > > > > > Hi, > > > > > > This patch series attempts to refactor RDMA live migration by > > > introducing a new QIOChannelRDMA class based on the rsocket API. > > > > > > The /usr/include/rdma/rsocket.h provides a higher level rsocket API > > > that is a 1-1 match of the normal kernel 'sockets' API, which hides > > > the detail of rdma protocol into rsocket and allows us to add support > > > for some modern features like multifd more easily. > > > > > > Here is the previous discussion on refactoring RDMA live migration > > > using the rsocket API: > > > > > > https://lore.kernel.org/qemu-devel/20240328130255.52257-1-philmd@linar > > > o.org/ > > > > > > We have encountered some bugs when using rsocket and plan to submit > > > them to the rdma-core community. > > > > > > In addition, the use of rsocket makes our programming more convenient, > > > but it must be noted that this method introduces multiple memory > > > copies, which can be imagined that there will be a certain performance > > > degradation, hoping that friends with RDMA network cards can help verify, > > thank you! > > > > It'll be good to elaborate if you tested it in-house. What people should expect > > on the numbers exactly? Is that okay from Huawei's POV? > > > > Besides that, the code looks pretty good at a first glance to me. Before > > others chim in, here're some high level comments.. > > > > Firstly, can we avoid using coroutine when listen()? Might be relevant when I > > see that rdma_accept_incoming_migration() runs in a loop to do raccept(), but > > would that also hang the qemu main loop even with the coroutine, before all > > channels are ready? I'm not a coroutine person, but I think the hope is that > > we can make dest QEMU run in a thread in the future just like the src QEMU, so > > the less coroutine the better in this path. > > > > Because rsocket is set to non-blocking, raccept will return EAGAIN when no connection > is received, coroutine will yield, and will not hang the qemu main loop. Ah that's ok. And also I just noticed it may not be a big deal either as long as we're before migration_incoming_process(). I'm wondering whether it can do it similarly like what we do with sockets in qio_net_listener_set_client_func_full(). After all, rsocket wants to mimic the socket API. It'll make sense if rsocket code tries to match with socket, or even reuse. > > > I think I also left a comment elsewhere on whether it would be possible to allow > > iochannels implement their own poll() functions to avoid the per-channel poll > > thread that is proposed in this series. > > > > https://lore.kernel.org/r/ZldY21xVExtiMddB@x1n > > > > We noticed that, and it's a big operation. I'm not sure that's a better way. > > > Personally I think even with the thread proposal it's better than the old rdma > > code, but I just still want to double check with you guys. E.g., maybe that just > > won't work at all? Again, that'll also be based on the fact that we move > > migration incoming into a thread first to keep the dest QEMU main loop intact, > > I think, but I hope we will reach that irrelevant of rdma, IOW it'll be nice to > > happen even earlier if possible. > > > Yep. This is a fairly big change, I wonder what other people's suggestions are? Yes we can wait for others' opinions. And btw I'm not asking for it and I don't think it'll be a blocker for this approach to land, as I said this is better than the current code so it's definitely an improvement to me. I'm purely curious, because if you're not going to do it for rdma, maybe someday I'll try to do that, and I want to know what "big change" could be as I didn't dig further. It may help me by sharing what issues you've found. Thanks, -- Peter Xu
> -----Original Message----- > From: Peter Xu [mailto:peterx@redhat.com] > Sent: Wednesday, June 5, 2024 10:19 PM > To: Gonglei (Arei) <arei.gonglei@huawei.com> > Cc: qemu-devel@nongnu.org; yu.zhang@ionos.com; mgalaxy@akamai.com; > elmar.gerdes@ionos.com; zhengchuan <zhengchuan@huawei.com>; > berrange@redhat.com; armbru@redhat.com; lizhijian@fujitsu.com; > pbonzini@redhat.com; mst@redhat.com; Xiexiangyou > <xiexiangyou@huawei.com>; linux-rdma@vger.kernel.org; lixiao (H) > <lixiao91@huawei.com>; jinpu.wang@ionos.com; Wangjialin > <wangjialin23@huawei.com>; Fabiano Rosas <farosas@suse.de> > Subject: Re: [PATCH 0/6] refactor RDMA live migration based on rsocket API > > On Wed, Jun 05, 2024 at 10:09:43AM +0000, Gonglei (Arei) wrote: > > Hi Peter, > > > > > -----Original Message----- > > > From: Peter Xu [mailto:peterx@redhat.com] > > > Sent: Wednesday, June 5, 2024 3:32 AM > > > To: Gonglei (Arei) <arei.gonglei@huawei.com> > > > Cc: qemu-devel@nongnu.org; yu.zhang@ionos.com; > mgalaxy@akamai.com; > > > elmar.gerdes@ionos.com; zhengchuan <zhengchuan@huawei.com>; > > > berrange@redhat.com; armbru@redhat.com; lizhijian@fujitsu.com; > > > pbonzini@redhat.com; mst@redhat.com; Xiexiangyou > > > <xiexiangyou@huawei.com>; linux-rdma@vger.kernel.org; lixiao (H) > > > <lixiao91@huawei.com>; jinpu.wang@ionos.com; Wangjialin > > > <wangjialin23@huawei.com>; Fabiano Rosas <farosas@suse.de> > > > Subject: Re: [PATCH 0/6] refactor RDMA live migration based on > > > rsocket API > > > > > > Hi, Lei, Jialin, > > > > > > Thanks a lot for working on this! > > > > > > I think we'll need to wait a bit on feedbacks from Jinpu and his > > > team on RDMA side, also Daniel for iochannels. Also, please > > > remember to copy Fabiano Rosas in any relevant future posts. We'd > > > also like to know whether he has any comments too. I have him copied in > this reply. > > > > > > On Tue, Jun 04, 2024 at 08:14:06PM +0800, Gonglei wrote: > > > > From: Jialin Wang <wangjialin23@huawei.com> > > > > > > > > Hi, > > > > > > > > This patch series attempts to refactor RDMA live migration by > > > > introducing a new QIOChannelRDMA class based on the rsocket API. > > > > > > > > The /usr/include/rdma/rsocket.h provides a higher level rsocket > > > > API that is a 1-1 match of the normal kernel 'sockets' API, which > > > > hides the detail of rdma protocol into rsocket and allows us to > > > > add support for some modern features like multifd more easily. > > > > > > > > Here is the previous discussion on refactoring RDMA live migration > > > > using the rsocket API: > > > > > > > > https://lore.kernel.org/qemu-devel/20240328130255.52257-1-philmd@l > > > > inar > > > > o.org/ > > > > > > > > We have encountered some bugs when using rsocket and plan to > > > > submit them to the rdma-core community. > > > > > > > > In addition, the use of rsocket makes our programming more > > > > convenient, but it must be noted that this method introduces > > > > multiple memory copies, which can be imagined that there will be a > > > > certain performance degradation, hoping that friends with RDMA > > > > network cards can help verify, > > > thank you! > > > > > > It'll be good to elaborate if you tested it in-house. What people > > > should expect on the numbers exactly? Is that okay from Huawei's POV? > > > > > > Besides that, the code looks pretty good at a first glance to me. > > > Before others chim in, here're some high level comments.. > > > > > > Firstly, can we avoid using coroutine when listen()? Might be > > > relevant when I see that rdma_accept_incoming_migration() runs in a > > > loop to do raccept(), but would that also hang the qemu main loop > > > even with the coroutine, before all channels are ready? I'm not a > > > coroutine person, but I think the hope is that we can make dest QEMU > > > run in a thread in the future just like the src QEMU, so the less coroutine > the better in this path. > > > > > > > Because rsocket is set to non-blocking, raccept will return EAGAIN > > when no connection is received, coroutine will yield, and will not hang the > qemu main loop. > > Ah that's ok. And also I just noticed it may not be a big deal either as long as > we're before migration_incoming_process(). > > I'm wondering whether it can do it similarly like what we do with sockets in > qio_net_listener_set_client_func_full(). After all, rsocket wants to mimic the > socket API. It'll make sense if rsocket code tries to match with socket, or > even reuse. > Actually we tried this solution, but it didn't work. Pls see patch 3/6 Known limitations: For a blocking rsocket fd, if we use io_create_watch to wait for POLLIN or POLLOUT events, since the rsocket fd is blocking, we cannot determine when it is not ready to read/write as we can with non-blocking fds. Therefore, when an event occurs, it will occurs always, potentially leave the qemu hanging. So we need be cautious to avoid hanging when using io_create_watch . Regards, -Gonglei > > > > > I think I also left a comment elsewhere on whether it would be > > > possible to allow iochannels implement their own poll() functions to > > > avoid the per-channel poll thread that is proposed in this series. > > > > > > https://lore.kernel.org/r/ZldY21xVExtiMddB@x1n > > > > > > > We noticed that, and it's a big operation. I'm not sure that's a better way. > > > > > Personally I think even with the thread proposal it's better than > > > the old rdma code, but I just still want to double check with you > > > guys. E.g., maybe that just won't work at all? Again, that'll also > > > be based on the fact that we move migration incoming into a thread > > > first to keep the dest QEMU main loop intact, I think, but I hope we > > > will reach that irrelevant of rdma, IOW it'll be nice to happen even earlier if > possible. > > > > > Yep. This is a fairly big change, I wonder what other people's suggestions > are? > > Yes we can wait for others' opinions. And btw I'm not asking for it and I don't > think it'll be a blocker for this approach to land, as I said this is better than the > current code so it's definitely an improvement to me. > > I'm purely curious, because if you're not going to do it for rdma, maybe > someday I'll try to do that, and I want to know what "big change" could be as I > didn't dig further. It may help me by sharing what issues you've found. > > Thanks, > > -- > Peter Xu
On Fri, Jun 07, 2024 at 08:49:01AM +0000, Gonglei (Arei) wrote: > Actually we tried this solution, but it didn't work. Pls see patch 3/6 > > Known limitations: > For a blocking rsocket fd, if we use io_create_watch to wait for > POLLIN or POLLOUT events, since the rsocket fd is blocking, we > cannot determine when it is not ready to read/write as we can with > non-blocking fds. Therefore, when an event occurs, it will occurs > always, potentially leave the qemu hanging. So we need be cautious > to avoid hanging when using io_create_watch . I'm not sure I fully get that part, though. In: https://lore.kernel.org/all/ZldY21xVExtiMddB@x1n/ I was thinking of iochannel implements its own poll with the _POLL flag, so in that case it'll call qio_channel_poll() which should call rpoll() directly. So I didn't expect using qio_channel_create_watch(). I thought the context was gmainloop won't work with rsocket fds in general, but maybe I missed something. Thanks, -- Peter Xu
© 2016 - 2024 Red Hat, Inc.