Re: [PATCH 1/2] vfio/migration: Avoid savevm_vfio_handlers for cpr-exec and cpr-transfer

[Cédric Le Goater]

Hello,

On 8/26/26 00:49, Dongli Zhang wrote:
When the VFIO driver is vfio-pci, vfio_migration_query_flags() returns
-ENOTTY, so savevm_vfio_handlers is not registered or involved.

I was confused with this sentence. Here is a proposal :

  When the kernel VFIO driver does not support migration (e.g. vfio-pci),
  vfio_migration_query_flags() returns -ENOTTY and savevm_vfio_handlers
  is never registered.

However, VFIO drivers that support live migration, such as mlx5_vfio_pci,
do support vfio_migration_query_flags(). In that case, savevm_vfio_handlers
is registered and may be used to save and load the VFIO migration stream.
This is expected for VFIO live migration, but not for cpr-exec or
cpr-transfer.
ok

Even after migration converges with
"migrate_set_parameter downtime-limit 1000000", the target still reports
the following error:
This is unclear to me. Can you rephrase ?

failed (load of migration failed: Invalid argument: Failed to load vmstate 
version_id: 1, ret: -22)

Skip savevm_vfio_handlers for cpr-exec and cpr-transfer.

Signed-off-by: Dongli Zhang <dongli.zh...@oracle.com>
---
  hw/vfio/migration.c | 8 ++++++++
  1 file changed, 8 insertions(+)

diff --git a/hw/vfio/migration.c b/hw/vfio/migration.c
index 73c49d8c24..49aeaad516 100644
--- a/hw/vfio/migration.c
+++ b/hw/vfio/migration.c
@@ -723,6 +723,13 @@ static bool vfio_is_active_iterate(void *opaque)
      return vfio_device_state_is_precopy(vbasedev);
  }
  
+static bool vfio_is_active(void *opaque)
+{
+    MigMode mode = migrate_mode();
+
+    return mode != MIG_MODE_CPR_EXEC && mode != MIG_MODE_CPR_TRANSFER;
 
Let's introduce a new cpr helper :

   static inline bool cpr_is_active(void *opaque)
   {
       MigMode mode = migrate_mode();
       return mode == MIG_MODE_CPR_EXEC || mode == MIG_MODE_CPR_TRANSFER;
   }

cpr_incoming_needed() becomes :

   bool cpr_incoming_needed(void *opaque)
   {
       return cpr_is_active(opaque);
   }


migrate_ram_is_ignored() could be adjusted too :

   bool migrate_ram_is_ignored(RAMBlock *block)
   {
  -    MigMode mode = migrate_mode();
       return !qemu_ram_is_migratable(block) ||
  -           mode == MIG_MODE_CPR_TRANSFER ||
  -           mode == MIG_MODE_CPR_EXEC ||
  +           cpr_is_active() ||
              (migrate_ignore_shared() && qemu_ram_is_shared(block)
                                       && qemu_ram_is_named_file(block));
   }
 

+}
+
  /*
   * Note about migration rate limiting: VFIO migration buffer size is currently
   * limited to 1MB, so there is no need to check if migration rate exceeded (as
@@ -951,6 +958,7 @@ static const SaveVMHandlers savevm_vfio_handlers = {
      .save_cleanup = vfio_save_cleanup,
      .save_query_pending = vfio_state_pending,
      .is_active_iterate = vfio_is_active_iterate,
+    .is_active = vfio_is_active,

and

    .is_active = cpr_is_active,

Thanks,

C.

      .save_live_iterate = vfio_save_iterate,
      .save_complete = vfio_save_complete_precopy,
      .save_state = vfio_save_state,