| Message ID | 20241202220137.32584-3-farosas@suse.de (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | migration: Fix issues during qmp_migrate_cancel | expand |
On Mon, Dec 02, 2024 at 07:01:33PM -0300, Fabiano Rosas wrote: > Make sure postcopy threads are released when migrate_cancel is > issued. Kick the postcopy_pause semaphore and have the fault thread > read 'fault_thread_quit' when joining. > > While here fix the comment mentioning userfault_event_fd. > > Signed-off-by: Fabiano Rosas <farosas@suse.de> I remember when working on postcopy, I thought about failing migrate-cancel for postcopy in general, rejecting such request. And when working on the recover feature, there's no concern on having it being cancelled, because the user really shouldn't do that.. The problem is migrate-cancel means crashing the VM on both sides when QEMU already goes into postcopy stage. If the user wants to crash the VM anyway, an easier way to do is killing on both sides. If the user wished to cancel, we should tell them "postcopy cannot be cancelled, until complete". That's probably the major reason why people think postcopy is dangerous to use.. Or do we have any use case this could be a valid scenario?
Peter Xu <peterx@redhat.com> writes: > On Mon, Dec 02, 2024 at 07:01:33PM -0300, Fabiano Rosas wrote: >> Make sure postcopy threads are released when migrate_cancel is >> issued. Kick the postcopy_pause semaphore and have the fault thread >> read 'fault_thread_quit' when joining. >> >> While here fix the comment mentioning userfault_event_fd. >> >> Signed-off-by: Fabiano Rosas <farosas@suse.de> > > I remember when working on postcopy, I thought about failing migrate-cancel > for postcopy in general, rejecting such request. And when working on the > recover feature, there's no concern on having it being cancelled, because > the user really shouldn't do that.. > > The problem is migrate-cancel means crashing the VM on both sides when QEMU > already goes into postcopy stage. Well, that's the sillyness of having a cancel command, you can never know what "cancel" means. The "documentation" says: "Cancel the current executing migration process", it doesn't mention anything about the consequences of such action. > > If the user wants to crash the VM anyway, an easier way to do is killing on > both sides. I don't think this is fair. We expose a "cancel" command, we better do some cancelling or instead reject the command appropriately, not expect the user to "know better". > > If the user wished to cancel, we should tell them "postcopy cannot be > cancelled, until complete". That's probably the major reason why people > think postcopy is dangerous to use.. We could certainly add that restriction, I don't see a problem with it. That said, what is the actual use case for migrate_cancel? And how does that compare with yank? I feel like we've been kind of relying on nobody using those commands really. One truth that we have (because it's tested) is that the multifd migration allows migrate_cancel on the source and another migration to start from it. (btw, that reminds me that multifd+postcopy will probably break that test). > > Or do we have any use case this could be a valid scenario? Not that I know of. But you're the postcopy expert =)
On Wed, Dec 04, 2024 at 04:02:36PM -0300, Fabiano Rosas wrote: > Peter Xu <peterx@redhat.com> writes: > > > On Mon, Dec 02, 2024 at 07:01:33PM -0300, Fabiano Rosas wrote: > >> Make sure postcopy threads are released when migrate_cancel is > >> issued. Kick the postcopy_pause semaphore and have the fault thread > >> read 'fault_thread_quit' when joining. > >> > >> While here fix the comment mentioning userfault_event_fd. > >> > >> Signed-off-by: Fabiano Rosas <farosas@suse.de> > > > > I remember when working on postcopy, I thought about failing migrate-cancel > > for postcopy in general, rejecting such request. And when working on the > > recover feature, there's no concern on having it being cancelled, because > > the user really shouldn't do that.. > > > > The problem is migrate-cancel means crashing the VM on both sides when QEMU > > already goes into postcopy stage. > > Well, that's the sillyness of having a cancel command, you can never > know what "cancel" means. The "documentation" says: "Cancel the current > executing migration process", it doesn't mention anything about the > consequences of such action. We definitely need cancel. It was always used in precopy, and people use it a lot! > > > > > If the user wants to crash the VM anyway, an easier way to do is killing on > > both sides. > > I don't think this is fair. We expose a "cancel" command, we better do > some cancelling or instead reject the command appropriately, not expect > the user to "know better". That's exactly why we should fail it with a proper error message, IMHO. Because the user may not really understand the impact of postcopy. I remember I also tried to reuse migrate-cancel to work as what migrate-pause does currently (so as to not introduce yet another new QMP command). I don't remember why we came up with the new cmd, but with migrate-pause being there, I think the only sane way to respond to cancel request during postcopy is to fail it properly.. > > > > > If the user wished to cancel, we should tell them "postcopy cannot be > > cancelled, until complete". That's probably the major reason why people > > think postcopy is dangerous to use.. > > We could certainly add that restriction, I don't see a problem with > it. That said, what is the actual use case for migrate_cancel? And how > does that compare with yank? I feel like we've been kind of relying on > nobody using those commands really. We had "cancel" first, then "yank". I didn't remember who merged yank, especially for migration, and why it was ever needed. What I still remember is yank crash qemu migration quite frequently, though I can't blame it as it does help us to find quite a few mismatch on iochannel setup/cleanup or stuff like that. So it's sometimes "helpful" as kind of a program checker.. Migration users should have stick with "cancel" rather than "yank" - qmp "yank" would "FAIL" the migration instead of showing CANCELLED, definitely should avoid. I am not aware of anybody that uses "yank" for migration at all. So yeah, both commands are slightly duplicated, and if we want to throw one, it needs to be yank, not cancel. I'm fine keeping both.. > > One truth that we have (because it's tested) is that the multifd > migration allows migrate_cancel on the source and another migration to > start from it. > > (btw, that reminds me that multifd+postcopy will probably break that > test). When postcopy==true (even if multifd==true as well! Or not, it doesn't matter), we should reject the cancel request, IMHO.
On Wed, Dec 04, 2024 at 02:39:12PM -0500, Peter Xu wrote: > On Wed, Dec 04, 2024 at 04:02:36PM -0300, Fabiano Rosas wrote: > > Peter Xu <peterx@redhat.com> writes: > > > > > On Mon, Dec 02, 2024 at 07:01:33PM -0300, Fabiano Rosas wrote: > > >> Make sure postcopy threads are released when migrate_cancel is > > >> issued. Kick the postcopy_pause semaphore and have the fault thread > > >> read 'fault_thread_quit' when joining. > > >> > > >> While here fix the comment mentioning userfault_event_fd. > > >> > > >> Signed-off-by: Fabiano Rosas <farosas@suse.de> > > > > > > I remember when working on postcopy, I thought about failing migrate-cancel > > > for postcopy in general, rejecting such request. And when working on the > > > recover feature, there's no concern on having it being cancelled, because > > > the user really shouldn't do that.. > > > > > > The problem is migrate-cancel means crashing the VM on both sides when QEMU > > > already goes into postcopy stage. > > > > Well, that's the sillyness of having a cancel command, you can never > > know what "cancel" means. The "documentation" says: "Cancel the current > > executing migration process", it doesn't mention anything about the > > consequences of such action. > > We definitely need cancel. It was always used in precopy, and people use > it a lot! Not a fair benchmark though. People use cancel alot because 'precopy' cannot complete in a predictable amount of time, any time guesstime can suddenly get much worse if the guest dirties memory differently. So people give up and cancel it after waiting ridiculously long and never being sure if it is nearing the finish. Once a migrate has been switched to 'postcopy' phase, however, we have what should be a highly predictable completion time, directly related to the amount of untransferred pages. That time should not get worse. The amount of time spent in the 'postcopy' phase will depend on how long you let the migrate run in 'precopy' before flipping to 'postcopy' IOW, I think there's a reasonable case to be made that NOT having the ability to cancel while in 'postcopy' phase would be mostly acceptable. You give up the ability to cancel for a while, in exchange for a clearly determined completion time. > > > If the user wants to crash the VM anyway, an easier way to do is killing on > > > both sides. > > > > I don't think this is fair. We expose a "cancel" command, we better do > > some cancelling or instead reject the command appropriately, not expect > > the user to "know better". > > That's exactly why we should fail it with a proper error message, IMHO. > Because the user may not really understand the impact of postcopy. Yep, I think users/apps expect "cancel" to be safe. So if it can't be safe at certain times, we should reject it in those time windows. > > > If the user wished to cancel, we should tell them "postcopy cannot be > > > cancelled, until complete". That's probably the major reason why people > > > think postcopy is dangerous to use.. > > > > We could certainly add that restriction, I don't see a problem with > > it. That said, what is the actual use case for migrate_cancel? And how > > does that compare with yank? I feel like we've been kind of relying on > > nobody using those commands really. > > We had "cancel" first, then "yank". I didn't remember who merged yank, > especially for migration, and why it was ever needed. yank is for the case where the network connections are completely stuck, causing QEMU to potentially get stalled in I/O operations until a TCP timeout is reached. yank force unwedges any stuck I/O by aggresively closing the connections. It is most useful in the non-migration use cases though. > Migration users should have stick with "cancel" rather than "yank" - qmp > "yank" would "FAIL" the migration instead of showing CANCELLED, definitely > should avoid. I am not aware of anybody that uses "yank" for migration at > all. > > So yeah, both commands are slightly duplicated, and if we want to throw > one, it needs to be yank, not cancel. I'm fine keeping both.. I would say the difference is like a graceful shutdown vs pulling the power plug in a bare metal machine 'cancel' is intended to be graceful. It should leave you with a functional QEMU (or refuse to run if unsafe). 'yank' is intended to be forceful, letting you get out of bad situations that would otherwise require you to kill the entire QEMU process, but still with possible associated risk data loss to the QEMU backends. They have overlap, but are none the less different. With regards, Daniel
Daniel P. Berrangé <berrange@redhat.com> writes: > On Wed, Dec 04, 2024 at 02:39:12PM -0500, Peter Xu wrote: >> On Wed, Dec 04, 2024 at 04:02:36PM -0300, Fabiano Rosas wrote: >> > Peter Xu <peterx@redhat.com> writes: >> > >> > > On Mon, Dec 02, 2024 at 07:01:33PM -0300, Fabiano Rosas wrote: >> > >> Make sure postcopy threads are released when migrate_cancel is >> > >> issued. Kick the postcopy_pause semaphore and have the fault thread >> > >> read 'fault_thread_quit' when joining. >> > >> >> > >> While here fix the comment mentioning userfault_event_fd. >> > >> >> > >> Signed-off-by: Fabiano Rosas <farosas@suse.de> >> > > >> > > I remember when working on postcopy, I thought about failing migrate-cancel >> > > for postcopy in general, rejecting such request. And when working on the >> > > recover feature, there's no concern on having it being cancelled, because >> > > the user really shouldn't do that.. >> > > >> > > The problem is migrate-cancel means crashing the VM on both sides when QEMU >> > > already goes into postcopy stage. >> > >> > Well, that's the sillyness of having a cancel command, you can never >> > know what "cancel" means. The "documentation" says: "Cancel the current >> > executing migration process", it doesn't mention anything about the >> > consequences of such action. >> >> We definitely need cancel. It was always used in precopy, and people use >> it a lot! To be clear, I'm not arguing against cancel. I'm just pointing out that it's silly because it's just like pressing C-c in the shell in the middle of something. What's the expected end state? Completely unspecified. I don't find it at all "elegant" that we treat cancel like error and just let the code carry on stumbling and exit eventually. Because then we have this C-c arriving at random moments in the middle of stuff. The way we do "exiting" in multifd is way more maintainable. If that flag is set, then let's exit, otherwise everything should work. Note that in some parts of this series I'm checking the CANCELLING state, which aside from being prone to be racy, is a step in a different direction than the way we've implemented cancel so far (shutdown and let the code exit vs. explicitly check against the CANCELLING state). Hope I'm being coherent, I'm not sure. > > Not a fair benchmark though. > > People use cancel alot because 'precopy' cannot complete in a > predictable amount of time, any time guesstime can suddenly > get much worse if the guest dirties memory differently. So > people give up and cancel it after waiting ridiculously long > and never being sure if it is nearing the finish. > > Once a migrate has been switched to 'postcopy' phase, however, > we have what should be a highly predictable completion time, > directly related to the amount of untransferred pages. That > time should not get worse. The amount of time spent in the > 'postcopy' phase will depend on how long you let the migrate > run in 'precopy' before flipping to 'postcopy' > > IOW, I think there's a reasonable case to be made that NOT > having the ability to cancel while in 'postcopy' phase would > be mostly acceptable. You give up the ability to cancel for > a while, in exchange for a clearly determined completion > time. Some of these words could be in the documentation for cancel to make it clear what the expectations are. > >> > > If the user wants to crash the VM anyway, an easier way to do is killing on >> > > both sides. >> > >> > I don't think this is fair. We expose a "cancel" command, we better do >> > some cancelling or instead reject the command appropriately, not expect >> > the user to "know better". >> >> That's exactly why we should fail it with a proper error message, IMHO. >> Because the user may not really understand the impact of postcopy. > > Yep, I think users/apps expect "cancel" to be safe. So if it can't be > safe at certain times, we should reject it in those time windows. Which kind of implies we should test this by migrate->cancel->migrate, like the multifd test does, and not like this series does. I've been focusing more on catching crashes/hangs. > >> > > If the user wished to cancel, we should tell them "postcopy cannot be >> > > cancelled, until complete". That's probably the major reason why people >> > > think postcopy is dangerous to use.. >> > >> > We could certainly add that restriction, I don't see a problem with >> > it. That said, what is the actual use case for migrate_cancel? And how >> > does that compare with yank? I feel like we've been kind of relying on >> > nobody using those commands really. >> >> We had "cancel" first, then "yank". I didn't remember who merged yank, >> especially for migration, and why it was ever needed. > > yank is for the case where the network connections are completely stuck, > causing QEMU to potentially get stalled in I/O operations until a TCP > timeout is reached. yank force unwedges any stuck I/O by aggresively > closing the connections. It is most useful in the non-migration use > cases though. > I asked because for socket at least yank and cancel do the same: shutdown(). It might be more trouble than it's worth it, but I wonder if we could have everything that can be stuck implement a yank routine and just have cancel call those. So for instance, when this series does sem_post on a bunch of semaphores explicitly, the cancel command would instead call whatever yank routine was registered by the code that can wait on the semaphore. With this there should be no surprises of a cancel arriving at a weird moment, because we'd require "code that sleeps" to implement a yank. >> Migration users should have stick with "cancel" rather than "yank" - qmp >> "yank" would "FAIL" the migration instead of showing CANCELLED, definitely >> should avoid. I am not aware of anybody that uses "yank" for migration at >> all. >> >> So yeah, both commands are slightly duplicated, and if we want to throw >> one, it needs to be yank, not cancel. I'm fine keeping both.. > > I would say the difference is like a graceful shutdown vs pulling the > power plug in a bare metal machine > > 'cancel' is intended to be graceful. It should leave you with a functional > QEMU (or refuse to run if unsafe). > > 'yank' is intended to be forceful, letting you get out of bad situations > that would otherwise require you to kill the entire QEMU process, but > still with possible associated risk data loss to the QEMU backends. For migration specifically I don't see much difference. Yank must leave QEMU in a usable state enough for a second migration to succeed, otherwise it's useless. > > They have overlap, but are none the less different. > > With regards, > Daniel
On Wed, Dec 04, 2024 at 08:02:31PM +0000, Daniel P. Berrangé wrote: > On Wed, Dec 04, 2024 at 02:39:12PM -0500, Peter Xu wrote: > > On Wed, Dec 04, 2024 at 04:02:36PM -0300, Fabiano Rosas wrote: > > > Peter Xu <peterx@redhat.com> writes: > > > > > > > On Mon, Dec 02, 2024 at 07:01:33PM -0300, Fabiano Rosas wrote: > > > >> Make sure postcopy threads are released when migrate_cancel is > > > >> issued. Kick the postcopy_pause semaphore and have the fault thread > > > >> read 'fault_thread_quit' when joining. > > > >> > > > >> While here fix the comment mentioning userfault_event_fd. > > > >> > > > >> Signed-off-by: Fabiano Rosas <farosas@suse.de> > > > > > > > > I remember when working on postcopy, I thought about failing migrate-cancel > > > > for postcopy in general, rejecting such request. And when working on the > > > > recover feature, there's no concern on having it being cancelled, because > > > > the user really shouldn't do that.. > > > > > > > > The problem is migrate-cancel means crashing the VM on both sides when QEMU > > > > already goes into postcopy stage. > > > > > > Well, that's the sillyness of having a cancel command, you can never > > > know what "cancel" means. The "documentation" says: "Cancel the current > > > executing migration process", it doesn't mention anything about the > > > consequences of such action. > > > > We definitely need cancel. It was always used in precopy, and people use > > it a lot! > > Not a fair benchmark though. > > People use cancel alot because 'precopy' cannot complete in a > predictable amount of time, any time guesstime can suddenly > get much worse if the guest dirties memory differently. So > people give up and cancel it after waiting ridiculously long > and never being sure if it is nearing the finish. > > Once a migrate has been switched to 'postcopy' phase, however, > we have what should be a highly predictable completion time, > directly related to the amount of untransferred pages. That > time should not get worse. The amount of time spent in the > 'postcopy' phase will depend on how long you let the migrate > run in 'precopy' before flipping to 'postcopy' Right. And if to be precise - the time doesn't matter, but "how many dirty page left". For extremely busy guests, longer precopy doesn't help, for example.. so time isn't very relevant, IMHO. It's just that postcopy always have an upper time limit, which is guest mem size / bw. > > IOW, I think there's a reasonable case to be made that NOT > having the ability to cancel while in 'postcopy' phase would > be mostly acceptable. You give up the ability to cancel for > a while, in exchange for a clearly determined completion > time. > > > > > If the user wants to crash the VM anyway, an easier way to do is killing on > > > > both sides. > > > > > > I don't think this is fair. We expose a "cancel" command, we better do > > > some cancelling or instead reject the command appropriately, not expect > > > the user to "know better". > > > > That's exactly why we should fail it with a proper error message, IMHO. > > Because the user may not really understand the impact of postcopy. > > Yep, I think users/apps expect "cancel" to be safe. So if it can't be > safe at certain times, we should reject it in those time windows. > > > > > If the user wished to cancel, we should tell them "postcopy cannot be > > > > cancelled, until complete". That's probably the major reason why people > > > > think postcopy is dangerous to use.. > > > > > > We could certainly add that restriction, I don't see a problem with > > > it. That said, what is the actual use case for migrate_cancel? And how > > > does that compare with yank? I feel like we've been kind of relying on > > > nobody using those commands really. > > > > We had "cancel" first, then "yank". I didn't remember who merged yank, > > especially for migration, and why it was ever needed. > > yank is for the case where the network connections are completely stuck, > causing QEMU to potentially get stalled in I/O operations until a TCP > timeout is reached. yank force unwedges any stuck I/O by aggresively > closing the connections. It is most useful in the non-migration use > cases though. > > > Migration users should have stick with "cancel" rather than "yank" - qmp > > "yank" would "FAIL" the migration instead of showing CANCELLED, definitely > > should avoid. I am not aware of anybody that uses "yank" for migration at > > all. > > > > So yeah, both commands are slightly duplicated, and if we want to throw > > one, it needs to be yank, not cancel. I'm fine keeping both.. > > I would say the difference is like a graceful shutdown vs pulling the > power plug in a bare metal machine > > 'cancel' is intended to be graceful. It should leave you with a functional > QEMU (or refuse to run if unsafe). > > 'yank' is intended to be forceful, letting you get out of bad situations > that would otherwise require you to kill the entire QEMU process, but > still with possible associated risk data loss to the QEMU backends. > > They have overlap, but are none the less different. The question is more about whether yank should be used at all for migration only, not about the rest instances. My answer is yank should never be used for migration, because "migrate_cancel" also unplugs the power plug.. It's not anything more enforced. It's only doing less always. E.g. migration_yank_iochannel() is exactly what we do with qmp_migrate_cancel() in the first place, only that migrate_cancel only does it on the main channel (on both qemufiles even if ioc is one), however it should be suffice, and behave the same way, as strong as "yank". Meanwhile, "yank" definitely lacks the graceful side of thing, not only on "cancelled", but also anything extra in migration_cancel() that wasn't about shutdown(). Examples, qmp_migrate_cancel has special care on: - migrate_dirty_limit(), where it'll shutdown the limit threads too alongside if migration cancels, - process the possible hang case with PRE_SWITCHOVER. - if the cancel happened right during switchover, and if at that time the disks are inactivated, it'll re-activate the disks In general, there's no reason to use yank on migration, IMHO, because it's not stronger either on shutting the NIC. Considering it's confusing to mostly everyone, and tons of people asked me about this.. maybe I should send a patch to remove yank from migration? Thanks,
On Wed, Dec 04, 2024 at 05:40:17PM -0300, Fabiano Rosas wrote: > To be clear, I'm not arguing against cancel. I'm just pointing out that > it's silly because it's just like pressing C-c in the shell in the > middle of something. What's the expected end state? Completely > unspecified. I don't find it at all "elegant" that we treat cancel like > error and just let the code carry on stumbling and exit > eventually. Because then we have this C-c arriving at random moments in > the middle of stuff. The way we do "exiting" in multifd is way more > maintainable. If that flag is set, then let's exit, otherwise everything > should work. If taking the example of C-c, then "migration during postcopy" is exactly "TASK_UNINTERRUPTIBLE".. :) IOW, "hanging death" the C-c is the correct and expected behavior for UNINTERRUPTIBLE tasks. [...] > > 'yank' is intended to be forceful, letting you get out of bad situations > > that would otherwise require you to kill the entire QEMU process, but > > still with possible associated risk data loss to the QEMU backends. > > For migration specifically I don't see much difference. Yank must leave > QEMU in a usable state enough for a second migration to succeed, > otherwise it's useless. Side note: when I said (in the other reply) that we should remove yank support on migration, I meant, we should probably deprecate that (and then remove it).
Peter Xu <peterx@redhat.com> writes: > On Wed, Dec 04, 2024 at 08:02:31PM +0000, Daniel P. Berrangé wrote: >> On Wed, Dec 04, 2024 at 02:39:12PM -0500, Peter Xu wrote: >> > On Wed, Dec 04, 2024 at 04:02:36PM -0300, Fabiano Rosas wrote: >> > > Peter Xu <peterx@redhat.com> writes: >> > > >> > > > On Mon, Dec 02, 2024 at 07:01:33PM -0300, Fabiano Rosas wrote: >> > > >> Make sure postcopy threads are released when migrate_cancel is >> > > >> issued. Kick the postcopy_pause semaphore and have the fault thread >> > > >> read 'fault_thread_quit' when joining. >> > > >> >> > > >> While here fix the comment mentioning userfault_event_fd. >> > > >> >> > > >> Signed-off-by: Fabiano Rosas <farosas@suse.de> >> > > > >> > > > I remember when working on postcopy, I thought about failing migrate-cancel >> > > > for postcopy in general, rejecting such request. And when working on the >> > > > recover feature, there's no concern on having it being cancelled, because >> > > > the user really shouldn't do that.. >> > > > >> > > > The problem is migrate-cancel means crashing the VM on both sides when QEMU >> > > > already goes into postcopy stage. >> > > >> > > Well, that's the sillyness of having a cancel command, you can never >> > > know what "cancel" means. The "documentation" says: "Cancel the current >> > > executing migration process", it doesn't mention anything about the >> > > consequences of such action. >> > >> > We definitely need cancel. It was always used in precopy, and people use >> > it a lot! >> >> Not a fair benchmark though. >> >> People use cancel alot because 'precopy' cannot complete in a >> predictable amount of time, any time guesstime can suddenly >> get much worse if the guest dirties memory differently. So >> people give up and cancel it after waiting ridiculously long >> and never being sure if it is nearing the finish. >> >> Once a migrate has been switched to 'postcopy' phase, however, >> we have what should be a highly predictable completion time, >> directly related to the amount of untransferred pages. That >> time should not get worse. The amount of time spent in the >> 'postcopy' phase will depend on how long you let the migrate >> run in 'precopy' before flipping to 'postcopy' > > Right. And if to be precise - the time doesn't matter, but "how many dirty > page left". For extremely busy guests, longer precopy doesn't help, for > example.. so time isn't very relevant, IMHO. It's just that postcopy > always have an upper time limit, which is guest mem size / bw. > >> >> IOW, I think there's a reasonable case to be made that NOT >> having the ability to cancel while in 'postcopy' phase would >> be mostly acceptable. You give up the ability to cancel for >> a while, in exchange for a clearly determined completion >> time. >> >> > > > If the user wants to crash the VM anyway, an easier way to do is killing on >> > > > both sides. >> > > >> > > I don't think this is fair. We expose a "cancel" command, we better do >> > > some cancelling or instead reject the command appropriately, not expect >> > > the user to "know better". >> > >> > That's exactly why we should fail it with a proper error message, IMHO. >> > Because the user may not really understand the impact of postcopy. >> >> Yep, I think users/apps expect "cancel" to be safe. So if it can't be >> safe at certain times, we should reject it in those time windows. >> >> > > > If the user wished to cancel, we should tell them "postcopy cannot be >> > > > cancelled, until complete". That's probably the major reason why people >> > > > think postcopy is dangerous to use.. >> > > >> > > We could certainly add that restriction, I don't see a problem with >> > > it. That said, what is the actual use case for migrate_cancel? And how >> > > does that compare with yank? I feel like we've been kind of relying on >> > > nobody using those commands really. >> > >> > We had "cancel" first, then "yank". I didn't remember who merged yank, >> > especially for migration, and why it was ever needed. >> >> yank is for the case where the network connections are completely stuck, >> causing QEMU to potentially get stalled in I/O operations until a TCP >> timeout is reached. yank force unwedges any stuck I/O by aggresively >> closing the connections. It is most useful in the non-migration use >> cases though. >> >> > Migration users should have stick with "cancel" rather than "yank" - qmp >> > "yank" would "FAIL" the migration instead of showing CANCELLED, definitely >> > should avoid. I am not aware of anybody that uses "yank" for migration at >> > all. >> > >> > So yeah, both commands are slightly duplicated, and if we want to throw >> > one, it needs to be yank, not cancel. I'm fine keeping both.. >> >> I would say the difference is like a graceful shutdown vs pulling the >> power plug in a bare metal machine >> >> 'cancel' is intended to be graceful. It should leave you with a functional >> QEMU (or refuse to run if unsafe). >> >> 'yank' is intended to be forceful, letting you get out of bad situations >> that would otherwise require you to kill the entire QEMU process, but >> still with possible associated risk data loss to the QEMU backends. >> >> They have overlap, but are none the less different. > > The question is more about whether yank should be used at all for > migration only, not about the rest instances. > > My answer is yank should never be used for migration, because > "migrate_cancel" also unplugs the power plug.. It's not anything more > enforced. It's only doing less always. > > E.g. migration_yank_iochannel() is exactly what we do with > qmp_migrate_cancel() in the first place, only that migrate_cancel only does > it on the main channel (on both qemufiles even if ioc is one), however it > should be suffice, and behave the same way, as strong as "yank". > > Meanwhile, "yank" definitely lacks the graceful side of thing, not only on > "cancelled", but also anything extra in migration_cancel() that wasn't > about shutdown(). Examples, qmp_migrate_cancel has special care on: > > - migrate_dirty_limit(), where it'll shutdown the limit threads too > alongside if migration cancels, > > - process the possible hang case with PRE_SWITCHOVER. > > - if the cancel happened right during switchover, and if at that time the > disks are inactivated, it'll re-activate the disks > > In general, there's no reason to use yank on migration, IMHO, because it's > not stronger either on shutting the NIC. > > Considering it's confusing to mostly everyone, and tons of people asked me > about this.. maybe I should send a patch to remove yank from migration? Take a look at my suggestion in the other thread, it might make yank make sense for migration after all. But I'm not against the removal, it's better than the current state IMO. > > Thanks,
On Wed, Dec 04, 2024 at 06:01:39PM -0300, Fabiano Rosas wrote: > > Considering it's confusing to mostly everyone, and tons of people asked me > > about this.. maybe I should send a patch to remove yank from migration? > > Take a look at my suggestion in the other thread, it might make yank > make sense for migration after all. But I'm not against the removal, > it's better than the current state IMO. I missed that. Copying that part over: > I asked because for socket at least yank and cancel do the same: > shutdown(). It might be more trouble than it's worth it, but I wonder if > we could have everything that can be stuck implement a yank routine and > just have cancel call those. So for instance, when this series does > sem_post on a bunch of semaphores explicitly, the cancel command would > instead call whatever yank routine was registered by the code that can > wait on the semaphore. With this there should be no surprises of a > cancel arriving at a weird moment, because we'd require "code that > sleeps" to implement a yank. If so, it means migration_cancel will still do the same as what yank would do (but it could cover more now, like processing sleeping semaphores). Actually, since it'll invoke a yank API it'll be the same from that regard. However as long as there's still something extra migrate_cancel would do after that yank API, then we should still suggest nobody using QMP "yank" on migration instance.. Unless we want to precisely define migrate_cancel exactly the same as the yank API would do. But then it means the two APIs are identical, then we should probably keep the one that people use the most, which is still, the migrate_cancel API..
On Wed, Dec 04, 2024 at 03:51:27PM -0500, Peter Xu wrote: > On Wed, Dec 04, 2024 at 08:02:31PM +0000, Daniel P. Berrangé wrote: > > I would say the difference is like a graceful shutdown vs pulling the > > power plug in a bare metal machine > > > > 'cancel' is intended to be graceful. It should leave you with a functional > > QEMU (or refuse to run if unsafe). > > > > 'yank' is intended to be forceful, letting you get out of bad situations > > that would otherwise require you to kill the entire QEMU process, but > > still with possible associated risk data loss to the QEMU backends. > > > > They have overlap, but are none the less different. > > The question is more about whether yank should be used at all for > migration only, not about the rest instances. > > My answer is yank should never be used for migration, because > "migrate_cancel" also unplugs the power plug.. It's not anything more > enforced. It's only doing less always. > > E.g. migration_yank_iochannel() is exactly what we do with > qmp_migrate_cancel() in the first place, only that migrate_cancel only does > it on the main channel (on both qemufiles even if ioc is one), however it > should be suffice, and behave the same way, as strong as "yank". I recall at the time the yank stuff was introduced, one of the scenarios they were concerned about was related to locks held by QEMU code. eg that there are scenarios where migrate_cancel may not be processed promptly enough due to being stalled on mutexes held by other concurrently running threads. Now I would expect any such long duration stalls on migration mutexes to be bugs, but the intent of yank is to give a recovery mechanism that can workaround such bugs. The yank QMP command only interacts with its own local mutexes. So even though both migrate-cancel and yank end up calling the same qio_channel_shutdown() API, there can be practical differences in how they reach that qio_channel_shutdown() call. With regards, Daniel
Daniel P. Berrangé <berrange@redhat.com> writes: > On Wed, Dec 04, 2024 at 03:51:27PM -0500, Peter Xu wrote: >> On Wed, Dec 04, 2024 at 08:02:31PM +0000, Daniel P. Berrangé wrote: >> > I would say the difference is like a graceful shutdown vs pulling the >> > power plug in a bare metal machine >> > >> > 'cancel' is intended to be graceful. It should leave you with a functional >> > QEMU (or refuse to run if unsafe). >> > >> > 'yank' is intended to be forceful, letting you get out of bad situations >> > that would otherwise require you to kill the entire QEMU process, but >> > still with possible associated risk data loss to the QEMU backends. >> > >> > They have overlap, but are none the less different. >> >> The question is more about whether yank should be used at all for >> migration only, not about the rest instances. >> >> My answer is yank should never be used for migration, because >> "migrate_cancel" also unplugs the power plug.. It's not anything more >> enforced. It's only doing less always. >> >> E.g. migration_yank_iochannel() is exactly what we do with >> qmp_migrate_cancel() in the first place, only that migrate_cancel only does >> it on the main channel (on both qemufiles even if ioc is one), however it >> should be suffice, and behave the same way, as strong as "yank". > > I recall at the time the yank stuff was introduced, one of the scenarios > they were concerned about was related to locks held by QEMU code. eg that > there are scenarios where migrate_cancel may not be processed promptly > enough due to being stalled on mutexes held by other concurrently running > threads. Now I would expect any such long duration stalls on migration > mutexes to be bugs, but the intent of yank is to give a recovery mechanism > that can workaround such bugs. The yank QMP command only interacts with > its own local mutexes. Ok, so that could only mean a thread stuck in recv() while holding the BQL. I don't think we have any other locks which would stop migrate_cancel from making progress or other stall situations that could be helped by a shutdown(). Note that most of locks around qemu_file were a late addition. I don't think that scenario is possible today. I'll have to do some tests. On that note, how is yank supposed to be accessed? I don't see support in libvirt. Is there a way to hook into QMP after the fact somehow?
On Thu, Dec 05, 2024 at 10:18:53AM -0300, Fabiano Rosas wrote: > Daniel P. Berrangé <berrange@redhat.com> writes: > > > On Wed, Dec 04, 2024 at 03:51:27PM -0500, Peter Xu wrote: > >> On Wed, Dec 04, 2024 at 08:02:31PM +0000, Daniel P. Berrangé wrote: > >> > I would say the difference is like a graceful shutdown vs pulling the > >> > power plug in a bare metal machine > >> > > >> > 'cancel' is intended to be graceful. It should leave you with a functional > >> > QEMU (or refuse to run if unsafe). > >> > > >> > 'yank' is intended to be forceful, letting you get out of bad situations > >> > that would otherwise require you to kill the entire QEMU process, but > >> > still with possible associated risk data loss to the QEMU backends. > >> > > >> > They have overlap, but are none the less different. > >> > >> The question is more about whether yank should be used at all for > >> migration only, not about the rest instances. > >> > >> My answer is yank should never be used for migration, because > >> "migrate_cancel" also unplugs the power plug.. It's not anything more > >> enforced. It's only doing less always. > >> > >> E.g. migration_yank_iochannel() is exactly what we do with > >> qmp_migrate_cancel() in the first place, only that migrate_cancel only does > >> it on the main channel (on both qemufiles even if ioc is one), however it > >> should be suffice, and behave the same way, as strong as "yank". > > > > I recall at the time the yank stuff was introduced, one of the scenarios > > they were concerned about was related to locks held by QEMU code. eg that > > there are scenarios where migrate_cancel may not be processed promptly > > enough due to being stalled on mutexes held by other concurrently running > > threads. Now I would expect any such long duration stalls on migration > > mutexes to be bugs, but the intent of yank is to give a recovery mechanism > > that can workaround such bugs. The yank QMP command only interacts with > > its own local mutexes. > > Ok, so that could only mean a thread stuck in recv() while holding the > BQL. I don't think we have any other locks which would stop > migrate_cancel from making progress or other stall situations that could > be helped by a shutdown(). Note that most of locks around qemu_file were > a late addition. I don't think that scenario is possible today. I'll > have to do some tests. And if that is a real difference, I'd think whether we should simply make migrate_cancel be oob-capable too.. IOW, I still think it'll be good to stick with always one API to cancel a migration, no matter which it is. If we want to move over to yank then I think we should move all migrate_cancel operations into yank and deprecate "migrate_cancel', but that sounds overkill. There's only one thing that might not be oob-compatible there so far, which is bdrv_activate_all(). But I plan to remove it very soon (so that disks will be activated in the migration thread instead, just like failure cases). > > On that note, how is yank supposed to be accessed? I don't see support > in libvirt. Is there a way to hook into QMP after the fact somehow? >
On Thu, Dec 05, 2024 at 10:18:53AM -0300, Fabiano Rosas wrote: > Daniel P. Berrangé <berrange@redhat.com> writes: > > > On Wed, Dec 04, 2024 at 03:51:27PM -0500, Peter Xu wrote: > >> On Wed, Dec 04, 2024 at 08:02:31PM +0000, Daniel P. Berrangé wrote: > >> > I would say the difference is like a graceful shutdown vs pulling the > >> > power plug in a bare metal machine > >> > > >> > 'cancel' is intended to be graceful. It should leave you with a functional > >> > QEMU (or refuse to run if unsafe). > >> > > >> > 'yank' is intended to be forceful, letting you get out of bad situations > >> > that would otherwise require you to kill the entire QEMU process, but > >> > still with possible associated risk data loss to the QEMU backends. > >> > > >> > They have overlap, but are none the less different. > >> > >> The question is more about whether yank should be used at all for > >> migration only, not about the rest instances. > >> > >> My answer is yank should never be used for migration, because > >> "migrate_cancel" also unplugs the power plug.. It's not anything more > >> enforced. It's only doing less always. > >> > >> E.g. migration_yank_iochannel() is exactly what we do with > >> qmp_migrate_cancel() in the first place, only that migrate_cancel only does > >> it on the main channel (on both qemufiles even if ioc is one), however it > >> should be suffice, and behave the same way, as strong as "yank". > > > > I recall at the time the yank stuff was introduced, one of the scenarios > > they were concerned about was related to locks held by QEMU code. eg that > > there are scenarios where migrate_cancel may not be processed promptly > > enough due to being stalled on mutexes held by other concurrently running > > threads. Now I would expect any such long duration stalls on migration > > mutexes to be bugs, but the intent of yank is to give a recovery mechanism > > that can workaround such bugs. The yank QMP command only interacts with > > its own local mutexes. > > Ok, so that could only mean a thread stuck in recv() while holding the > BQL. I don't think we have any other locks which would stop > migrate_cancel from making progress or other stall situations that could > be helped by a shutdown(). Note that most of locks around qemu_file were > a late addition. I don't think that scenario is possible today. I'll > have to do some tests. Yes, in general there should never be a for "yank", *if* QMEU is implemented correctly. yank is there in case something unexpected happens. IOW, even if we think migration is perfect today, yank is still worth having there as a safety net. > On that note, how is yank supposed to be accessed? I don't see support > in libvirt. Is there a way to hook into QMP after the fact somehow? We've not wired up any API for this libvirt. I can be issued via libvirt's QMP passthrough API if desired though. With regards, Daniel
On Thu, Dec 05, 2024 at 10:03:58AM -0500, Peter Xu wrote: > On Thu, Dec 05, 2024 at 10:18:53AM -0300, Fabiano Rosas wrote: > > Daniel P. Berrangé <berrange@redhat.com> writes: > > > > > On Wed, Dec 04, 2024 at 03:51:27PM -0500, Peter Xu wrote: > > >> On Wed, Dec 04, 2024 at 08:02:31PM +0000, Daniel P. Berrangé wrote: > > >> > I would say the difference is like a graceful shutdown vs pulling the > > >> > power plug in a bare metal machine > > >> > > > >> > 'cancel' is intended to be graceful. It should leave you with a functional > > >> > QEMU (or refuse to run if unsafe). > > >> > > > >> > 'yank' is intended to be forceful, letting you get out of bad situations > > >> > that would otherwise require you to kill the entire QEMU process, but > > >> > still with possible associated risk data loss to the QEMU backends. > > >> > > > >> > They have overlap, but are none the less different. > > >> > > >> The question is more about whether yank should be used at all for > > >> migration only, not about the rest instances. > > >> > > >> My answer is yank should never be used for migration, because > > >> "migrate_cancel" also unplugs the power plug.. It's not anything more > > >> enforced. It's only doing less always. > > >> > > >> E.g. migration_yank_iochannel() is exactly what we do with > > >> qmp_migrate_cancel() in the first place, only that migrate_cancel only does > > >> it on the main channel (on both qemufiles even if ioc is one), however it > > >> should be suffice, and behave the same way, as strong as "yank". > > > > > > I recall at the time the yank stuff was introduced, one of the scenarios > > > they were concerned about was related to locks held by QEMU code. eg that > > > there are scenarios where migrate_cancel may not be processed promptly > > > enough due to being stalled on mutexes held by other concurrently running > > > threads. Now I would expect any such long duration stalls on migration > > > mutexes to be bugs, but the intent of yank is to give a recovery mechanism > > > that can workaround such bugs. The yank QMP command only interacts with > > > its own local mutexes. > > > > Ok, so that could only mean a thread stuck in recv() while holding the > > BQL. I don't think we have any other locks which would stop > > migrate_cancel from making progress or other stall situations that could > > be helped by a shutdown(). Note that most of locks around qemu_file were > > a late addition. I don't think that scenario is possible today. I'll > > have to do some tests. > > And if that is a real difference, I'd think whether we should simply make > migrate_cancel be oob-capable too.. IOW, I still think it'll be good to > stick with always one API to cancel a migration, no matter which it is. If > we want to move over to yank then I think we should move all migrate_cancel > operations into yank and deprecate "migrate_cancel', but that sounds > overkill. Well migrate_cancel ought to be safer than yank. eg migrate_cancel (sh|c)ould refuse to run if issued during post-copy phase. Or even in precopy, if in the final vmstate copy & switchover phase we shouldn't need to cancel. yank meanwhile will always run, no matter what, because by design, it has no interaction with the migration code beyond knowing that a socket exists. I don't think we should combine them. They have alot of common, but there are subtle differences that are relevant to the scenarios in which thye are intended to be used. > There's only one thing that might not be oob-compatible there so far, which > is bdrv_activate_all(). But I plan to remove it very soon (so that disks > will be activated in the migration thread instead, just like failure cases). > > > > > On that note, how is yank supposed to be accessed? I don't see support > > in libvirt. Is there a way to hook into QMP after the fact somehow? > > > > -- > Peter Xu > With regards, Daniel
diff --git a/migration/migration.c b/migration/migration.c index 8c5bd0a75c..07fbb5c9f4 100644 --- a/migration/migration.c +++ b/migration/migration.c @@ -105,7 +105,7 @@ static bool migration_object_check(MigrationState *ms, Error **errp); static int migration_maybe_pause(MigrationState *s, int *current_active_state, int new_state); -static void migrate_fd_cancel(MigrationState *s); +static void migrate_fd_cancel(MigrationState *s, MigrationIncomingState *mis); static bool close_return_path_on_source(MigrationState *s); static void migration_completion_end(MigrationState *s); @@ -317,7 +317,7 @@ void migration_cancel(const Error *error) if (migrate_dirty_limit()) { qmp_cancel_vcpu_dirty_limit(false, -1, NULL); } - migrate_fd_cancel(current_migration); + migrate_fd_cancel(current_migration, current_incoming); } void migration_shutdown(void) @@ -1502,7 +1502,7 @@ static void migrate_fd_error(MigrationState *s, const Error *error) migrate_set_error(s, error); } -static void migrate_fd_cancel(MigrationState *s) +static void migrate_fd_cancel(MigrationState *s, MigrationIncomingState *mis) { int old_state ; @@ -1515,6 +1515,12 @@ static void migrate_fd_cancel(MigrationState *s) } } + if (mis->state == MIGRATION_STATUS_POSTCOPY_PAUSED) { + migrate_set_state(&mis->state, MIGRATION_STATUS_POSTCOPY_PAUSED, + MIGRATION_STATUS_CANCELLING); + qemu_sem_post(&mis->postcopy_pause_sem_dst); + } + do { old_state = s->state; if (!migration_is_running()) { @@ -1523,6 +1529,8 @@ static void migrate_fd_cancel(MigrationState *s) /* If the migration is paused, kick it out of the pause */ if (old_state == MIGRATION_STATUS_PRE_SWITCHOVER) { qemu_sem_post(&s->pause_sem); + } else if (old_state == MIGRATION_STATUS_POSTCOPY_PAUSED) { + qemu_sem_post(&s->postcopy_pause_sem); } migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING); } while (s->state != MIGRATION_STATUS_CANCELLING); diff --git a/migration/postcopy-ram.c b/migration/postcopy-ram.c index a535fd2e30..6882ef977d 100644 --- a/migration/postcopy-ram.c +++ b/migration/postcopy-ram.c @@ -634,6 +634,7 @@ int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) qatomic_set(&mis->fault_thread_quit, 1); postcopy_fault_thread_notify(mis); trace_postcopy_ram_incoming_cleanup_join(); + qemu_sem_post(&mis->postcopy_pause_sem_fault); qemu_thread_join(&mis->fault_thread); if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) { @@ -991,8 +992,7 @@ static void *postcopy_ram_fault_thread(void *opaque) /* * We're mainly waiting for the kernel to give us a faulting HVA, - * however we can be told to quit via userfault_quit_fd which is - * an eventfd + * however we can be told to quit via userfault_event_fd. */ poll_result = poll(pfd, pfd_len, -1 /* Wait forever */); @@ -1008,6 +1008,11 @@ static void *postcopy_ram_fault_thread(void *opaque) * the channel is rebuilt. */ postcopy_pause_fault_thread(mis); + + if (qatomic_read(&mis->fault_thread_quit)) { + trace_postcopy_ram_fault_thread_quit(); + break; + } } if (pfd[1].revents) { @@ -1082,6 +1087,11 @@ retry: if (ret) { /* May be network failure, try to wait for recovery */ postcopy_pause_fault_thread(mis); + + if (qatomic_read(&mis->fault_thread_quit)) { + trace_postcopy_ram_fault_thread_quit(); + break; + } goto retry; } }
Make sure postcopy threads are released when migrate_cancel is issued. Kick the postcopy_pause semaphore and have the fault thread read 'fault_thread_quit' when joining. While here fix the comment mentioning userfault_event_fd. Signed-off-by: Fabiano Rosas <farosas@suse.de> --- migration/migration.c | 14 +++++++++++--- migration/postcopy-ram.c | 14 ++++++++++++-- 2 files changed, 23 insertions(+), 5 deletions(-)