Message ID | 20191122182943.4656-19-eric.auger@redhat.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | VIRTIO-IOMMU device | expand |
* Eric Auger (eric.auger@redhat.com) wrote: > Add Migration support. We rely on recently added gtree and qlist > migration. Besides, we have to fixup end point <-> domain link. > > Indeed each domain has a list of endpoints attached to it. And each > endpoint has a pointer to its domain. > > Raw gtree and qlist migration cannot handle this as it re-allocates > all the nodes while reconstructing the trees/lists. > > So in post_load we re-construct the relationship between endpoints > and domains. > > Signed-off-by: Eric Auger <eric.auger@redhat.com> From the migration side of things, Acked-by: Dr. David Alan Gilbert <dgilbert@redhat.com> > --- > hw/virtio/virtio-iommu.c | 127 ++++++++++++++++++++++++++++++++++++--- > 1 file changed, 117 insertions(+), 10 deletions(-) > > diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c > index c5b202fab7..4e92fc0c95 100644 > --- a/hw/virtio/virtio-iommu.c > +++ b/hw/virtio/virtio-iommu.c > @@ -692,16 +692,6 @@ static void virtio_iommu_set_features(VirtIODevice *vdev, uint64_t val) > trace_virtio_iommu_set_features(dev->acked_features); > } > > -/* > - * Migration is not yet supported: most of the state consists > - * of balanced binary trees which are not yet ready for getting > - * migrated > - */ > -static const VMStateDescription vmstate_virtio_iommu_device = { > - .name = "virtio-iommu-device", > - .unmigratable = 1, > -}; > - > static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data) > { > uint ua = GPOINTER_TO_UINT(a); > @@ -778,6 +768,123 @@ static void virtio_iommu_instance_init(Object *obj) > { > } > > +#define VMSTATE_INTERVAL \ > +{ \ > + .name = "interval", \ > + .version_id = 1, \ > + .minimum_version_id = 1, \ > + .fields = (VMStateField[]) { \ > + VMSTATE_UINT64(low, viommu_interval), \ > + VMSTATE_UINT64(high, viommu_interval), \ > + VMSTATE_END_OF_LIST() \ > + } \ > +} > + > +#define VMSTATE_MAPPING \ > +{ \ > + .name = "mapping", \ > + .version_id = 1, \ > + .minimum_version_id = 1, \ > + .fields = (VMStateField[]) { \ > + VMSTATE_UINT64(phys_addr, viommu_mapping), \ > + VMSTATE_UINT32(flags, viommu_mapping), \ > + VMSTATE_END_OF_LIST() \ > + }, \ > +} > + > +static const VMStateDescription vmstate_interval_mapping[2] = { > + VMSTATE_MAPPING, /* value */ > + VMSTATE_INTERVAL /* key */ > +}; > + > +static int domain_preload(void *opaque) > +{ > + viommu_domain *domain = opaque; > + > + domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp, > + NULL, g_free, g_free); > + return 0; > +} > + > +static const VMStateDescription vmstate_endpoint = { > + .name = "endpoint", > + .version_id = 1, > + .minimum_version_id = 1, > + .fields = (VMStateField[]) { > + VMSTATE_UINT32(id, viommu_endpoint), > + VMSTATE_END_OF_LIST() > + } > +}; > + > +static const VMStateDescription vmstate_domain = { > + .name = "domain", > + .version_id = 1, > + .minimum_version_id = 1, > + .pre_load = domain_preload, > + .fields = (VMStateField[]) { > + VMSTATE_UINT32(id, viommu_domain), > + VMSTATE_GTREE_V(mappings, viommu_domain, 1, > + vmstate_interval_mapping, > + viommu_interval, viommu_mapping), > + VMSTATE_QLIST_V(endpoint_list, viommu_domain, 1, > + vmstate_endpoint, viommu_endpoint, next), > + VMSTATE_END_OF_LIST() > + } > +}; > + > +static gboolean reconstruct_ep_domain_link(gpointer key, gpointer value, > + gpointer data) > +{ > + viommu_domain *d = (viommu_domain *)value; > + viommu_endpoint *iter, *tmp; > + viommu_endpoint *ep = (viommu_endpoint *)data; > + > + QLIST_FOREACH_SAFE(iter, &d->endpoint_list, next, tmp) { > + if (iter->id == ep->id) { > + /* remove the ep */ > + QLIST_REMOVE(iter, next); > + g_free(iter); > + /* replace it with the good one */ > + QLIST_INSERT_HEAD(&d->endpoint_list, ep, next); > + /* update the domain */ > + ep->domain = d; > + return true; /* stop the search */ > + } > + } > + return false; /* continue the traversal */ > +} > + > +static gboolean fix_endpoint(gpointer key, gpointer value, gpointer data) > +{ > + VirtIOIOMMU *s = (VirtIOIOMMU *)data; > + > + g_tree_foreach(s->domains, reconstruct_ep_domain_link, value); > + return false; > +} > + > +static int iommu_post_load(void *opaque, int version_id) > +{ > + VirtIOIOMMU *s = opaque; > + > + g_tree_foreach(s->endpoints, fix_endpoint, s); > + return 0; > +} > + > +static const VMStateDescription vmstate_virtio_iommu_device = { > + .name = "virtio-iommu-device", > + .minimum_version_id = 1, > + .version_id = 1, > + .post_load = iommu_post_load, > + .fields = (VMStateField[]) { > + VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 1, > + &vmstate_domain, viommu_domain), > + VMSTATE_GTREE_DIRECT_KEY_V(endpoints, VirtIOIOMMU, 1, > + &vmstate_endpoint, viommu_endpoint), > + VMSTATE_END_OF_LIST() > + }, > +}; > + > + > static const VMStateDescription vmstate_virtio_iommu = { > .name = "virtio-iommu", > .minimum_version_id = 1, > -- > 2.20.1 > -- Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
On Fri, Nov 22, 2019 at 07:29:41PM +0100, Eric Auger wrote: > +static const VMStateDescription vmstate_virtio_iommu_device = { > + .name = "virtio-iommu-device", > + .minimum_version_id = 1, > + .version_id = 1, > + .post_load = iommu_post_load, > + .fields = (VMStateField[]) { > + VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 1, > + &vmstate_domain, viommu_domain), > + VMSTATE_GTREE_DIRECT_KEY_V(endpoints, VirtIOIOMMU, 1, > + &vmstate_endpoint, viommu_endpoint), So if I understand correctly these fields are state that is modified by the guest? We don't need to save/load fields that cannot be modified by the guest, static information that is created from the QEMU command-line. I think the above covers everything we need to migrate in VirtIOIOMMU then, except for acked_features, which (as I pointed out on another patch) seems redundant anyway since there is vdev->guest_features. Reviewed-by: Jean-Philippe Brucker <jean-philippe@linaro.org> > + VMSTATE_END_OF_LIST() > + }, > +}; > + > + > static const VMStateDescription vmstate_virtio_iommu = { > .name = "virtio-iommu", > .minimum_version_id = 1, > -- > 2.20.1 > >
On Fri, Nov 22, 2019 at 07:29:41PM +0100, Eric Auger wrote: > +static const VMStateDescription vmstate_virtio_iommu_device = { > + .name = "virtio-iommu-device", > + .minimum_version_id = 1, > + .version_id = 1, > + .post_load = iommu_post_load, > + .fields = (VMStateField[]) { > + VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 1, > + &vmstate_domain, viommu_domain), > + VMSTATE_GTREE_DIRECT_KEY_V(endpoints, VirtIOIOMMU, 1, > + &vmstate_endpoint, viommu_endpoint), IIUC vmstate_domain already contains all the endpoint information (in endpoint_list of vmstate_domain), but here we migrate it twice. I suppose that's why now we need reconstruct_ep_domain_link() to fixup the duplicated migration? Then I'll instead ask whether we can skip migrating here? Then in post_load we simply: foreach(domain) foreach(endpoint in domain) g_tree_insert(s->endpoints); It might help to avoid the reconstruct_ep_domain_link ugliness? And besides, I also agree with Jean that the endpoint data structure could be reused with IOMMUDevice somehow. Thanks,
Hi Jean, On 12/10/19 5:50 PM, Jean-Philippe Brucker wrote: > On Fri, Nov 22, 2019 at 07:29:41PM +0100, Eric Auger wrote: >> +static const VMStateDescription vmstate_virtio_iommu_device = { >> + .name = "virtio-iommu-device", >> + .minimum_version_id = 1, >> + .version_id = 1, >> + .post_load = iommu_post_load, >> + .fields = (VMStateField[]) { >> + VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 1, >> + &vmstate_domain, viommu_domain), >> + VMSTATE_GTREE_DIRECT_KEY_V(endpoints, VirtIOIOMMU, 1, >> + &vmstate_endpoint, viommu_endpoint), > > So if I understand correctly these fields are state that is modified by > the guest? We don't need to save/load fields that cannot be modified by > the guest, static information that is created from the QEMU command-line. Yes that's correct. > > I think the above covers everything we need to migrate in VirtIOIOMMU > then, except for acked_features, which (as I pointed out on another patch) > seems redundant anyway since there is vdev->guest_features. you're right, acked features were not properly migrated. > > Reviewed-by: Jean-Philippe Brucker <jean-philippe@linaro.org> Thanks! Eric > >> + VMSTATE_END_OF_LIST() >> + }, >> +}; >> + >> + >> static const VMStateDescription vmstate_virtio_iommu = { >> .name = "virtio-iommu", >> .minimum_version_id = 1, >> -- >> 2.20.1 >> >> >
Hi Peter, On 12/10/19 9:01 PM, Peter Xu wrote: > On Fri, Nov 22, 2019 at 07:29:41PM +0100, Eric Auger wrote: >> +static const VMStateDescription vmstate_virtio_iommu_device = { >> + .name = "virtio-iommu-device", >> + .minimum_version_id = 1, >> + .version_id = 1, >> + .post_load = iommu_post_load, >> + .fields = (VMStateField[]) { >> + VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 1, >> + &vmstate_domain, viommu_domain), >> + VMSTATE_GTREE_DIRECT_KEY_V(endpoints, VirtIOIOMMU, 1, >> + &vmstate_endpoint, viommu_endpoint), > > IIUC vmstate_domain already contains all the endpoint information (in > endpoint_list of vmstate_domain), but here we migrate it twice. I migrated both because at that time I considered we could have endpoints not attached to any domains but I think I can now simplify based on the fact any EP is attached. I > suppose that's why now we need reconstruct_ep_domain_link() to fixup > the duplicated migration? Even if I only migrate the domain gtree, I need to reconstruct the ep->domain which was not migrated, on purpose, as it pointed to the old domain in the origin. > > Then I'll instead ask whether we can skip migrating here? Then in > post_load we simply: > > foreach(domain) > foreach(endpoint in domain) > g_tree_insert(s->endpoints); > > It might help to avoid the reconstruct_ep_domain_link ugliness? I agree that it is simpler. Also need to update the ep->domain as mentionned above. Thank you for the suggestion. > > And besides, I also agree with Jean that the endpoint data structure > could be reused with IOMMUDevice somehow. As I replied to Jean, I think it makes sense to keep both structures as endpoints are not indexed by the same key and the bus number is resolved later. Thanks Eric > > Thanks, >
diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c index c5b202fab7..4e92fc0c95 100644 --- a/hw/virtio/virtio-iommu.c +++ b/hw/virtio/virtio-iommu.c @@ -692,16 +692,6 @@ static void virtio_iommu_set_features(VirtIODevice *vdev, uint64_t val) trace_virtio_iommu_set_features(dev->acked_features); } -/* - * Migration is not yet supported: most of the state consists - * of balanced binary trees which are not yet ready for getting - * migrated - */ -static const VMStateDescription vmstate_virtio_iommu_device = { - .name = "virtio-iommu-device", - .unmigratable = 1, -}; - static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data) { uint ua = GPOINTER_TO_UINT(a); @@ -778,6 +768,123 @@ static void virtio_iommu_instance_init(Object *obj) { } +#define VMSTATE_INTERVAL \ +{ \ + .name = "interval", \ + .version_id = 1, \ + .minimum_version_id = 1, \ + .fields = (VMStateField[]) { \ + VMSTATE_UINT64(low, viommu_interval), \ + VMSTATE_UINT64(high, viommu_interval), \ + VMSTATE_END_OF_LIST() \ + } \ +} + +#define VMSTATE_MAPPING \ +{ \ + .name = "mapping", \ + .version_id = 1, \ + .minimum_version_id = 1, \ + .fields = (VMStateField[]) { \ + VMSTATE_UINT64(phys_addr, viommu_mapping), \ + VMSTATE_UINT32(flags, viommu_mapping), \ + VMSTATE_END_OF_LIST() \ + }, \ +} + +static const VMStateDescription vmstate_interval_mapping[2] = { + VMSTATE_MAPPING, /* value */ + VMSTATE_INTERVAL /* key */ +}; + +static int domain_preload(void *opaque) +{ + viommu_domain *domain = opaque; + + domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp, + NULL, g_free, g_free); + return 0; +} + +static const VMStateDescription vmstate_endpoint = { + .name = "endpoint", + .version_id = 1, + .minimum_version_id = 1, + .fields = (VMStateField[]) { + VMSTATE_UINT32(id, viommu_endpoint), + VMSTATE_END_OF_LIST() + } +}; + +static const VMStateDescription vmstate_domain = { + .name = "domain", + .version_id = 1, + .minimum_version_id = 1, + .pre_load = domain_preload, + .fields = (VMStateField[]) { + VMSTATE_UINT32(id, viommu_domain), + VMSTATE_GTREE_V(mappings, viommu_domain, 1, + vmstate_interval_mapping, + viommu_interval, viommu_mapping), + VMSTATE_QLIST_V(endpoint_list, viommu_domain, 1, + vmstate_endpoint, viommu_endpoint, next), + VMSTATE_END_OF_LIST() + } +}; + +static gboolean reconstruct_ep_domain_link(gpointer key, gpointer value, + gpointer data) +{ + viommu_domain *d = (viommu_domain *)value; + viommu_endpoint *iter, *tmp; + viommu_endpoint *ep = (viommu_endpoint *)data; + + QLIST_FOREACH_SAFE(iter, &d->endpoint_list, next, tmp) { + if (iter->id == ep->id) { + /* remove the ep */ + QLIST_REMOVE(iter, next); + g_free(iter); + /* replace it with the good one */ + QLIST_INSERT_HEAD(&d->endpoint_list, ep, next); + /* update the domain */ + ep->domain = d; + return true; /* stop the search */ + } + } + return false; /* continue the traversal */ +} + +static gboolean fix_endpoint(gpointer key, gpointer value, gpointer data) +{ + VirtIOIOMMU *s = (VirtIOIOMMU *)data; + + g_tree_foreach(s->domains, reconstruct_ep_domain_link, value); + return false; +} + +static int iommu_post_load(void *opaque, int version_id) +{ + VirtIOIOMMU *s = opaque; + + g_tree_foreach(s->endpoints, fix_endpoint, s); + return 0; +} + +static const VMStateDescription vmstate_virtio_iommu_device = { + .name = "virtio-iommu-device", + .minimum_version_id = 1, + .version_id = 1, + .post_load = iommu_post_load, + .fields = (VMStateField[]) { + VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 1, + &vmstate_domain, viommu_domain), + VMSTATE_GTREE_DIRECT_KEY_V(endpoints, VirtIOIOMMU, 1, + &vmstate_endpoint, viommu_endpoint), + VMSTATE_END_OF_LIST() + }, +}; + + static const VMStateDescription vmstate_virtio_iommu = { .name = "virtio-iommu", .minimum_version_id = 1,
Add Migration support. We rely on recently added gtree and qlist migration. Besides, we have to fixup end point <-> domain link. Indeed each domain has a list of endpoints attached to it. And each endpoint has a pointer to its domain. Raw gtree and qlist migration cannot handle this as it re-allocates all the nodes while reconstructing the trees/lists. So in post_load we re-construct the relationship between endpoints and domains. Signed-off-by: Eric Auger <eric.auger@redhat.com> --- hw/virtio/virtio-iommu.c | 127 ++++++++++++++++++++++++++++++++++++--- 1 file changed, 117 insertions(+), 10 deletions(-)