On Fri Feb 9, 2024 at 6:19 PM AEST, Thomas Huth wrote:
> On 09/02/2024 08.01, Nicholas Piggin wrote:
> > Support multiple migrations by flipping dest file/socket variables to
> > source after the migration is complete, ready to start again. A new
> > destination is created if the test outputs the migrate line again.
> > Test cases may now switch to calling migrate() one or more times.
> >
> > Signed-off-by: Nicholas Piggin <npiggin@xxxxxxxxx>
> > ---
> ...
> > diff --git a/scripts/arch-run.bash b/scripts/arch-run.bash
> > index 3689d7c2..a914ba17 100644
> > --- a/scripts/arch-run.bash
> > +++ b/scripts/arch-run.bash
> > @@ -129,12 +129,16 @@ run_migration ()
> > return 77
> > fi
> >
> > + migcmdline=$@
> > +
> > trap 'trap - TERM ; kill 0 ; exit 2' INT TERM
> > - trap 'rm -f ${migout1} ${migout_fifo1} ${migsock} ${qmp1} ${qmp2} ${fifo}' RETURN EXIT
> > + trap 'rm -f ${migout1} ${migout2} ${migout_fifo1} ${migout_fifo2} ${migsock} ${qmp1} ${qmp2} ${fifo}' RETURN EXIT
> >
> > migsock=$(mktemp -u -t mig-helper-socket.XXXXXXXXXX)
> > migout1=$(mktemp -t mig-helper-stdout1.XXXXXXXXXX)
> > migout_fifo1=$(mktemp -u -t mig-helper-fifo-stdout1.XXXXXXXXXX)
> > + migout2=$(mktemp -t mig-helper-stdout2.XXXXXXXXXX)
> > + migout_fifo2=$(mktemp -u -t mig-helper-fifo-stdout2.XXXXXXXXXX)
> > qmp1=$(mktemp -u -t mig-helper-qmp1.XXXXXXXXXX)
> > qmp2=$(mktemp -u -t mig-helper-qmp2.XXXXXXXXXX)
> > fifo=$(mktemp -u -t mig-helper-fifo.XXXXXXXXXX)
> > @@ -142,18 +146,61 @@ run_migration ()
> > qmpout2=/dev/null
> >
> > mkfifo ${migout_fifo1}
> > - eval "$@" -chardev socket,id=mon1,path=${qmp1},server=on,wait=off \
> > + mkfifo ${migout_fifo2}
> > +
> > + eval "$migcmdline" \
> > + -chardev socket,id=mon1,path=${qmp1},server=on,wait=off \
> > -mon chardev=mon1,mode=control > ${migout_fifo1} &
> > live_pid=$!
> > cat ${migout_fifo1} | tee ${migout1} &
> >
> > - # We have to use cat to open the named FIFO, because named FIFO's, unlike
> > - # pipes, will block on open() until the other end is also opened, and that
> > - # totally breaks QEMU...
> > + # The test must prompt the user to migrate, so wait for the "migrate"
> > + # keyword
> > + while ! grep -q -i "Now migrate the VM" < ${migout1} ; do
> > + if ! ps -p ${live_pid} > /dev/null ; then
> > + echo "ERROR: Test exit before migration point." >&2
> > + qmp ${qmp1} '"quit"'> ${qmpout1} 2>/dev/null
> > + return 3
> > + fi
> > + sleep 0.1
> > + done
> > +
> > + # This starts the first source QEMU in advance of the test reaching the
> > + # migration point, since we expect at least one migration. Subsequent
> > + # sources are started as the test hits migrate keywords.
> > + do_migration || return $?
> > +
> > + while ps -p ${live_pid} > /dev/null ; do
> > + # Wait for EXIT or further migrations
> > + if ! grep -q -i "Now migrate the VM" < ${migout1} ; then
> > + sleep 0.1
> > + else
> > + do_migration || return $?
> > + fi
> > + done
> > +
> > + wait ${live_pid}
> > + ret=$?
> > +
> > + while (( $(jobs -r | wc -l) > 0 )); do
> > + sleep 0.1
> > + done
> > +
> > + return $ret
> > +}
> > +
> > +do_migration ()
> > +{
> > + # We have to use cat to open the named FIFO, because named FIFO's,
> > + # unlike pipes, will block on open() until the other end is also
> > + # opened, and that totally breaks QEMU...
> > mkfifo ${fifo}
> > - eval "$@" -chardev socket,id=mon2,path=${qmp2},server=on,wait=off \
> > - -mon chardev=mon2,mode=control -incoming unix:${migsock} < <(cat ${fifo}) &
> > + eval "$migcmdline" \
> > + -chardev socket,id=mon2,path=${qmp2},server=on,wait=off \
> > + -mon chardev=mon2,mode=control -incoming unix:${migsock} \
> > + < <(cat ${fifo}) > ${migout_fifo2} &
> > incoming_pid=$!
> > + cat ${migout_fifo2} | tee ${migout2} &
> >
> > # The test must prompt the user to migrate, so wait for the "migrate" keyword
> > while ! grep -q -i "Now migrate the VM" < ${migout1} ; do
>
> So the old check for the "migrate" keyword is also still around?
It's just the comment is staleish, it only checks "Now migrate...".
> Why do we
> need to wait on two spots for the "Now mirgrate..." string now?
So that the it ensures we do one migration, subsequent ones are
optional.
I was thinking we could just remove that, and possibly even
remove the MIGRATION=yes/no paths and always just use the same
code here. But that's for another time.
Actually there is some weirdness here. There are *three* spots
where it waits for migration. The first one in run_migration
can be removed, because it can call do_migration right away
to start up the destination qemu process ahead of the first
migration message as-per comment. I'll respin with that change.
Thanks,
Nick
