On Wed, Oct 23, 2024 at 11:30:14AM +0300, Avihai Horon wrote:
>
> On 22/10/2024 19:07, Peter Xu wrote:
> > External email: Use caution opening links or attachments
> >
> >
> > Migration object can be freed before some other device codes run, while we
> > do have a bunch of migration helpers exported in migration/misc.h that
> > logically can be invoked at any time of QEMU, even during destruction of a
> > VM.
> >
> > Make all these functions safe to be called, especially, not crashing after
> > the migration object is freed.
> >
> > Add a rich comment in the header explaining how to guarantee thread safe on
> > using these functions, and we choose BQL because fundamentally that's how
> > it's working now. We can move to other things (e.g. RCU) whenever
> > necessary in the future but it's an overkill if we have BQL anyway in
> > most/all existing callers.
> >
> > When at it, update some comments, e.g. migrate_announce_params() is
> > exported from options.c now.
> >
> > Cc: Cédric Le Goater <c...@redhat.com>
> > Cc: Avihai Horon <avih...@nvidia.com>
> > Cc: Fabiano Rosas <faro...@suse.de>
> > Cc: Dr. David Alan Gilbert <d...@treblig.org>
> > Signed-off-by: Peter Xu <pet...@redhat.com>
> > ---
> > include/migration/misc.h | 33 ++++++++++++++++++++++++++++-----
> > migration/migration.c | 22 +++++++++++++++++++++-
> > 2 files changed, 49 insertions(+), 6 deletions(-)
> >
> > diff --git a/include/migration/misc.h b/include/migration/misc.h
> > index bfadc5613b..8d6812b8c7 100644
> > --- a/include/migration/misc.h
> > +++ b/include/migration/misc.h
> > @@ -19,8 +19,26 @@
> > #include "qapi/qapi-types-net.h"
> > #include "migration/client-options.h"
> >
> > -/* migration/ram.c */
> > +/*
> > + * Misc migration functions exported to be used in QEMU generic system
> > + * code outside migration/.
> > + *
> > + * By default, BQL is required to use below functions to avoid race
> > + * conditions (e.g. concurrent free of the migration object). It's
> > + * caller's responsibility to make sure it's thread safe otherwise when
> > + * below helpers are used without BQL held.
> > + *
> > + * One example of the special case is migration_thread(), who will take a
> > + * refcount of the migration object. The refcount will make sure the
> > + * migration object will not be freed concurrently when accessing through
> > + * below helpers.
> > + *
> > + * When unsure, always take BQL first before using the helpers.
> > + */
> >
> > +/*
> > + * migration/ram.c
> > + */
> > typedef enum PrecopyNotifyReason {
> > PRECOPY_NOTIFY_SETUP = 0,
> > PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC = 1,
> > @@ -43,14 +61,19 @@ void ram_mig_init(void);
> > void qemu_guest_free_page_hint(void *addr, size_t len);
> > bool migrate_ram_is_ignored(RAMBlock *block);
> >
> > -/* migration/block.c */
> > -
> > +/*
> > + * migration/options.c
> > + */
> > AnnounceParameters *migrate_announce_params(void);
> > -/* migration/savevm.c */
> >
> > +/*
> > + * migration/savevm.c
> > + */
> > void dump_vmstate_json_to_file(FILE *out_fp);
> >
> > -/* migration/migration.c */
> > +/*
> > + * migration/migration.c
> > + */
> > void migration_object_init(void);
> > void migration_shutdown(void);
> > bool migration_is_idle(void);
> > diff --git a/migration/migration.c b/migration/migration.c
> > index bcb735869b..27341eed50 100644
> > --- a/migration/migration.c
> > +++ b/migration/migration.c
> > @@ -1121,6 +1121,10 @@ bool migration_is_setup_or_active(void)
> > {
> > MigrationState *s = current_migration;
> >
> > + if (!s) {
> > + return false;
> > + }
> > +
> > switch (s->state) {
> > case MIGRATION_STATUS_ACTIVE:
> > case MIGRATION_STATUS_POSTCOPY_ACTIVE:
> > @@ -1136,7 +1140,6 @@ bool migration_is_setup_or_active(void)
> >
> > default:
> > return false;
> > -
> > }
> > }
> >
> > @@ -1685,6 +1688,10 @@ bool migration_is_active(void)
> > {
> > MigrationState *s = current_migration;
> >
> > + if (!s) {
> > + return false;
> > + }
> > +
> > return (s->state == MIGRATION_STATUS_ACTIVE ||
> > s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
> > }
> > @@ -1693,6 +1700,10 @@ bool migration_is_device(void)
> > {
> > MigrationState *s = current_migration;
> >
> > + if (!s) {
> > + return false;
> > + }
> > +
> > return s->state == MIGRATION_STATUS_DEVICE;
> > }
> >
> > @@ -1700,6 +1711,11 @@ bool migration_thread_is_self(void)
> > {
> > MigrationState *s = current_migration;
> >
> > + /* If no migration object, must not be the migration thread */
> > + if (!s) {
> > + return false;
> > + }
> > +
> > return qemu_thread_is_self(&s->thread);
> > }
> >
> > @@ -3077,6 +3093,10 @@ void migration_file_set_error(int ret, Error *err)
> > {
> > MigrationState *s = current_migration;
> >
> > + if (!s) {
> > + return;
> > + }
> > +
>
> I think this is not enough because current_migration is never set to NULL
> after it's destroyed.
>
> Can we add "current_migration = NULL;" to migration_instance_finalize()?
Good point..
I thought it was cleared already in migration_shutdown(), but now I just
noticed why it can't - we have too many dangling references in migration/
so that it needs to be there even if migration threads holds one refcount.
I'll add one more patch as you suggested as of now and repost soon. It's
not the cleanest that we clear that global pointer in a finalize(), but
looks like that's the only simple way forward.
Will repost soon, thanks.
--
Peter Xu
