Re: [PATCH 5/5] RDMA/uverbs: add UVERBS_METHOD_REG_REMOTE_MR

[Joel Nider]

linux-rdma-owner@xxxxxxxxxxxxxxx wrote on 01/29/2019 07:04:06 PM:

> On Tue, Jan 29, 2019 at 03:26:26PM +0200, Joel Nider wrote:
> > Add a new handler for new uverb reg_remote_mr. The purpose is to 
register
> > a memory region in a different address space (i.e. process) than the
> > caller.
> > 
> > The main use case which motivated this change is post-copy container
> > migration. When a migration manager (i.e. CRIU) starts a migration, it
> > must have an open connection for handling any page faults that occur
> > in the container after restoration on the target machine. Even though
> > CRIU establishes and maintains the connection, ultimately the memory
> > is copied from the container being migrated (i.e. a remote address
> > space). This container must remain passive -- meaning it cannot have
> > any knowledge of the RDMA connection; therefore the migration manager
> > must have the ability to register a remote memory region. This remote
> > memory region will serve as the source for any memory pages that must
> > be copied (on-demand or otherwise) during the migration.
> > 
> > Signed-off-by: Joel Nider <joeln@xxxxxxxxxx>
> >  drivers/infiniband/core/uverbs_std_types_mr.c | 129 
+++++++++++++++++++++++++-
> >  include/rdma/ib_verbs.h                       |   8 ++
> >  include/uapi/rdma/ib_user_ioctl_cmds.h        |  13 +++
> >  3 files changed, 149 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/infiniband/core/uverbs_std_types_mr.c b/drivers/
> infiniband/core/uverbs_std_types_mr.c
> > index 4d4be0c..bf7b4b2 100644
> > +++ b/drivers/infiniband/core/uverbs_std_types_mr.c
> > @@ -150,6 +150,99 @@ static int 
UVERBS_HANDLER(UVERBS_METHOD_DM_MR_REG)(
> >     return ret;
> >  }
> > 
> > +static int UVERBS_HANDLER(UVERBS_METHOD_REG_REMOTE_MR)(
> > +   struct uverbs_attr_bundle *attrs)
> > +{
> 
> I think this should just be REG_MR with an optional remote PID
> argument

Maybe I missed something.  Isn't REG_MR only implemented as a write() 
command? In our earlier conversation you told me all new commands must be 
implemented as ioctl() commands.


> >  DECLARE_UVERBS_NAMED_OBJECT(
> >     UVERBS_OBJECT_MR,
> >     UVERBS_TYPE_ALLOC_IDR(uverbs_free_mr),
> >     &UVERBS_METHOD(UVERBS_METHOD_DM_MR_REG),
> >     &UVERBS_METHOD(UVERBS_METHOD_MR_DESTROY),
> > -   &UVERBS_METHOD(UVERBS_METHOD_ADVISE_MR));
> > +   &UVERBS_METHOD(UVERBS_METHOD_ADVISE_MR),
> > +   &UVERBS_METHOD(UVERBS_METHOD_REG_REMOTE_MR),
> > +);
> 
> I'm kind of surprised this compiles with the trailing comma?
Personally, I think it is nicer with the trailing comma. Of course 
syntactically it makes no sense, but when adding a new entry, you don't 
have to touch the previous line, which makes the diff cleaner. If this is 
against standard practices I will remove the comma.