| Message ID | 20191129115132.285988-1-lvivier@redhat.com |
|---|---|
| State | New |
| Headers | show |
| Series | runstate: ignore finishmigrate -> prelaunch transition | expand |
* Laurent Vivier (lvivier@redhat.com) wrote: > Commit 1bd71dce4bf2 tries to prevent a finishmigrate -> prelaunch > transition by exiting at the beginning of the main_loop_should_exit() > function if the state is already finishmigrate. > > As the finishmigrate state is set in the migration thread it can > happen concurrently to the function. The migration thread and the > function are normally protected by the iothread mutex and thus the > state should no evolve between the start of the function and its end. > > Unfortunately during the function life the lock is released by > pause_all_vcpus() just before the point we need to be sure we are > not in finishmigrate state and if the migration thread is waiting > for the lock it will take the opportunity to change the state > to finishmigrate. Ewww. I hate those short wakeups for pause_all_vcpus; I'm sure there are loads more corners that break. Still, I _think_ this is an improvement, so: Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com> > The only way to be sure we are not in the finishmigrate state when > we need is to check the state after the pause_all_vcpus() function. > > Fixes: 1bd71dce4bf2 ("runstate: ignore exit request in finish migrate state") > Signed-off-by: Laurent Vivier <lvivier@redhat.com> > --- > vl.c | 10 ++++++---- > 1 file changed, 6 insertions(+), 4 deletions(-) > > diff --git a/vl.c b/vl.c > index 6a65a64bfd64..bf0a6345d239 100644 > --- a/vl.c > +++ b/vl.c > @@ -1745,9 +1745,6 @@ static bool main_loop_should_exit(void) > RunState r; > ShutdownCause request; > > - if (runstate_check(RUN_STATE_FINISH_MIGRATE)) { > - return false; > - } > if (preconfig_exit_requested) { > if (runstate_check(RUN_STATE_PRECONFIG)) { > runstate_set(RUN_STATE_PRELAUNCH); > @@ -1776,8 +1773,13 @@ static bool main_loop_should_exit(void) > pause_all_vcpus(); > qemu_system_reset(request); > resume_all_vcpus(); > + /* > + * runstate can change in pause_all_vcpus() > + * as iothread mutex is unlocked > + */ > if (!runstate_check(RUN_STATE_RUNNING) && > - !runstate_check(RUN_STATE_INMIGRATE)) { > + !runstate_check(RUN_STATE_INMIGRATE) && > + !runstate_check(RUN_STATE_FINISH_MIGRATE)) { > runstate_set(RUN_STATE_PRELAUNCH); > } > } > -- > 2.23.0 > -- Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
On 06/12/2019 20:52, Dr. David Alan Gilbert wrote: > * Laurent Vivier (lvivier@redhat.com) wrote: >> Commit 1bd71dce4bf2 tries to prevent a finishmigrate -> prelaunch >> transition by exiting at the beginning of the main_loop_should_exit() >> function if the state is already finishmigrate. >> >> As the finishmigrate state is set in the migration thread it can >> happen concurrently to the function. The migration thread and the >> function are normally protected by the iothread mutex and thus the >> state should no evolve between the start of the function and its end. >> >> Unfortunately during the function life the lock is released by >> pause_all_vcpus() just before the point we need to be sure we are >> not in finishmigrate state and if the migration thread is waiting >> for the lock it will take the opportunity to change the state >> to finishmigrate. > > Ewww. > I hate those short wakeups for pause_all_vcpus; I'm sure there are loads > more corners that break. > > Still, I _think_ this is an improvement, so: > > Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com> > Who volunteers to take this in his queue? Thanks, Laurent
Ping? Thanks, Laurent On 12/12/2019 20:40, Laurent Vivier wrote: > On 06/12/2019 20:52, Dr. David Alan Gilbert wrote: >> * Laurent Vivier (lvivier@redhat.com) wrote: >>> Commit 1bd71dce4bf2 tries to prevent a finishmigrate -> prelaunch >>> transition by exiting at the beginning of the main_loop_should_exit() >>> function if the state is already finishmigrate. >>> >>> As the finishmigrate state is set in the migration thread it can >>> happen concurrently to the function. The migration thread and the >>> function are normally protected by the iothread mutex and thus the >>> state should no evolve between the start of the function and its end. >>> >>> Unfortunately during the function life the lock is released by >>> pause_all_vcpus() just before the point we need to be sure we are >>> not in finishmigrate state and if the migration thread is waiting >>> for the lock it will take the opportunity to change the state >>> to finishmigrate. >> >> Ewww. >> I hate those short wakeups for pause_all_vcpus; I'm sure there are loads >> more corners that break. >> >> Still, I _think_ this is an improvement, so: >> >> Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com> >> > > Who volunteers to take this in his queue? > > Thanks, > Laurent > >
* Laurent Vivier (lvivier@redhat.com) wrote: > Ping? Juan says he's doing a migration pull soon - this seems like a good one for it. Dave > Thanks, > Laurent > > On 12/12/2019 20:40, Laurent Vivier wrote: > > On 06/12/2019 20:52, Dr. David Alan Gilbert wrote: > >> * Laurent Vivier (lvivier@redhat.com) wrote: > >>> Commit 1bd71dce4bf2 tries to prevent a finishmigrate -> prelaunch > >>> transition by exiting at the beginning of the main_loop_should_exit() > >>> function if the state is already finishmigrate. > >>> > >>> As the finishmigrate state is set in the migration thread it can > >>> happen concurrently to the function. The migration thread and the > >>> function are normally protected by the iothread mutex and thus the > >>> state should no evolve between the start of the function and its end. > >>> > >>> Unfortunately during the function life the lock is released by > >>> pause_all_vcpus() just before the point we need to be sure we are > >>> not in finishmigrate state and if the migration thread is waiting > >>> for the lock it will take the opportunity to change the state > >>> to finishmigrate. > >> > >> Ewww. > >> I hate those short wakeups for pause_all_vcpus; I'm sure there are loads > >> more corners that break. > >> > >> Still, I _think_ this is an improvement, so: > >> > >> Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com> > >> > > > > Who volunteers to take this in his queue? > > > > Thanks, > > Laurent > > > > > -- Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
Laurent Vivier <lvivier@redhat.com> wrote: > Ping? > > Thanks, > Laurent > > On 12/12/2019 20:40, Laurent Vivier wrote: >> On 06/12/2019 20:52, Dr. David Alan Gilbert wrote: >>> * Laurent Vivier (lvivier@redhat.com) wrote: >>>> Commit 1bd71dce4bf2 tries to prevent a finishmigrate -> prelaunch >>>> transition by exiting at the beginning of the main_loop_should_exit() >>>> function if the state is already finishmigrate. >>>> >>>> As the finishmigrate state is set in the migration thread it can >>>> happen concurrently to the function. The migration thread and the >>>> function are normally protected by the iothread mutex and thus the >>>> state should no evolve between the start of the function and its end. >>>> >>>> Unfortunately during the function life the lock is released by >>>> pause_all_vcpus() just before the point we need to be sure we are >>>> not in finishmigrate state and if the migration thread is waiting >>>> for the lock it will take the opportunity to change the state >>>> to finishmigrate. >>> >>> Ewww. >>> I hate those short wakeups for pause_all_vcpus; I'm sure there are loads >>> more corners that break. >>> >>> Still, I _think_ this is an improvement, so: >>> >>> Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com> >>> >> >> Who volunteers to take this in his queue? Reviewed-by: Juan Quintela <quintela@redhat.com> queued
diff --git a/vl.c b/vl.c index 6a65a64bfd64..bf0a6345d239 100644 --- a/vl.c +++ b/vl.c @@ -1745,9 +1745,6 @@ static bool main_loop_should_exit(void) RunState r; ShutdownCause request; - if (runstate_check(RUN_STATE_FINISH_MIGRATE)) { - return false; - } if (preconfig_exit_requested) { if (runstate_check(RUN_STATE_PRECONFIG)) { runstate_set(RUN_STATE_PRELAUNCH); @@ -1776,8 +1773,13 @@ static bool main_loop_should_exit(void) pause_all_vcpus(); qemu_system_reset(request); resume_all_vcpus(); + /* + * runstate can change in pause_all_vcpus() + * as iothread mutex is unlocked + */ if (!runstate_check(RUN_STATE_RUNNING) && - !runstate_check(RUN_STATE_INMIGRATE)) { + !runstate_check(RUN_STATE_INMIGRATE) && + !runstate_check(RUN_STATE_FINISH_MIGRATE)) { runstate_set(RUN_STATE_PRELAUNCH); } }
Commit 1bd71dce4bf2 tries to prevent a finishmigrate -> prelaunch transition by exiting at the beginning of the main_loop_should_exit() function if the state is already finishmigrate. As the finishmigrate state is set in the migration thread it can happen concurrently to the function. The migration thread and the function are normally protected by the iothread mutex and thus the state should no evolve between the start of the function and its end. Unfortunately during the function life the lock is released by pause_all_vcpus() just before the point we need to be sure we are not in finishmigrate state and if the migration thread is waiting for the lock it will take the opportunity to change the state to finishmigrate. The only way to be sure we are not in the finishmigrate state when we need is to check the state after the pause_all_vcpus() function. Fixes: 1bd71dce4bf2 ("runstate: ignore exit request in finish migrate state") Signed-off-by: Laurent Vivier <lvivier@redhat.com> --- vl.c | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-)