Re: [PATCH v2] RDMA/rtrs-srv: Incorporate ib_register_client into rtrs server init

[Danil Kipnis <danil.kipnis@cloud.ionos.com>]

On Tue, Aug 11, 2020 at 1:13 PM Jinpu Wang <jinpu.wang@cloud.ionos.com> wrote:
>
> On Tue, Aug 11, 2020 at 12:53 PM Haris Iqbal
> <haris.iqbal@cloud.ionos.com> wrote:
> >
> > On Tue, Aug 11, 2020 at 4:17 PM Leon Romanovsky <leon@kernel.org> wrote:
> > >
> > > On Tue, Aug 11, 2020 at 02:27:12PM +0530, Haris Iqbal wrote:
> > > > On Tue, Aug 11, 2020 at 2:15 PM Leon Romanovsky <leon@kernel.org> wrote:
> > > > >
> > > > > On Mon, Aug 10, 2020 at 05:20:49PM +0530, Md Haris Iqbal wrote:
> > > > > > The rnbd_server module's communication manager (cm) initialization depends
> > > > > > on the registration of the "network namespace subsystem" of the RDMA CM
> > > > > > agent module. As such, when the kernel is configured to load the
> > > > > > rnbd_server and the RDMA cma module during initialization; and if the
> > > > > > rnbd_server module is initialized before RDMA cma module, a null ptr
> > > > > > dereference occurs during the RDMA bind operation.
> > > > > >
> > > > > > Call trace below,
> > > > > >
> > > > > > [    1.904782] Call Trace:
> > > > > > [    1.904782]  ? xas_load+0xd/0x80
> > > > > > [    1.904782]  xa_load+0x47/0x80
> > > > > > [    1.904782]  cma_ps_find+0x44/0x70
> > > > > > [    1.904782]  rdma_bind_addr+0x782/0x8b0
> > > > > > [    1.904782]  ? get_random_bytes+0x35/0x40
> > > > > > [    1.904782]  rtrs_srv_cm_init+0x50/0x80
> > > > > > [    1.904782]  rtrs_srv_open+0x102/0x180
> > > > > > [    1.904782]  ? rnbd_client_init+0x6e/0x6e
> > > > > > [    1.904782]  rnbd_srv_init_module+0x34/0x84
> > > > > > [    1.904782]  ? rnbd_client_init+0x6e/0x6e
> > > > > > [    1.904782]  do_one_initcall+0x4a/0x200
> > > > > > [    1.904782]  kernel_init_freeable+0x1f1/0x26e
> > > > > > [    1.904782]  ? rest_init+0xb0/0xb0
> > > > > > [    1.904782]  kernel_init+0xe/0x100
> > > > > > [    1.904782]  ret_from_fork+0x22/0x30
> > > > > > [    1.904782] Modules linked in:
> > > > > > [    1.904782] CR2: 0000000000000015
> > > > > > [    1.904782] ---[ end trace c42df88d6c7b0a48 ]---
> > > > > >
> > > > > > All this happens cause the cm init is in the call chain of the module init,
> > > > > > which is not a preferred practice.
> > > > > >
> > > > > > So remove the call to rdma_create_id() from the module init call chain.
> > > > > > Instead register rtrs-srv as an ib client, which makes sure that the
> > > > > > rdma_create_id() is called only when an ib device is added.
> > > > > >
> > > > > > Fixes: 9cb837480424 ("RDMA/rtrs: server: main functionality")
> > > > > > Reported-by: kernel test robot <rong.a.chen@intel.com>
> > > > > > Signed-off-by: Md Haris Iqbal <haris.iqbal@cloud.ionos.com>
> > > > > > ---
> > > > > > Change in v2:
> > > > > >         Use only single variable to track number of IB devices and failure
> > > > > >         Change according to kernel coding style
> > > > > >
> > > > > >  drivers/infiniband/ulp/rtrs/rtrs-srv.c | 79 +++++++++++++++++++++++++-
> > > > > >  drivers/infiniband/ulp/rtrs/rtrs-srv.h |  6 ++
> > > > > >  2 files changed, 82 insertions(+), 3 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/infiniband/ulp/rtrs/rtrs-srv.c b/drivers/infiniband/ulp/rtrs/rtrs-srv.c
> > > > > > index 0d9241f5d9e6..69a37ce73b0c 100644
> > > > > > --- a/drivers/infiniband/ulp/rtrs/rtrs-srv.c
> > > > > > +++ b/drivers/infiniband/ulp/rtrs/rtrs-srv.c
> > > > > > @@ -16,6 +16,7 @@
> > > > > >  #include "rtrs-srv.h"
> > > > > >  #include "rtrs-log.h"
> > > > > >  #include <rdma/ib_cm.h>
> > > > > > +#include <rdma/ib_verbs.h>
> > > > > >
> > > > > >  MODULE_DESCRIPTION("RDMA Transport Server");
> > > > > >  MODULE_LICENSE("GPL");
> > > > > > @@ -31,6 +32,7 @@ MODULE_LICENSE("GPL");
> > > > > >  static struct rtrs_rdma_dev_pd dev_pd;
> > > > > >  static mempool_t *chunk_pool;
> > > > > >  struct class *rtrs_dev_class;
> > > > > > +static struct rtrs_srv_ib_ctx ib_ctx;
> > > > > >
> > > > > >  static int __read_mostly max_chunk_size = DEFAULT_MAX_CHUNK_SIZE;
> > > > > >  static int __read_mostly sess_queue_depth = DEFAULT_SESS_QUEUE_DEPTH;
> > > > > > @@ -2033,6 +2035,64 @@ static void free_srv_ctx(struct rtrs_srv_ctx *ctx)
> > > > > >       kfree(ctx);
> > > > > >  }
> > > > > >
> > > > > > +static int rtrs_srv_add_one(struct ib_device *device)
> > > > > > +{
> > > > > > +     struct rtrs_srv_ctx *ctx;
> > > > > > +     int ret;
> > > > > > +
> > > > > > +     if (ib_ctx.ib_dev_count)
> > > > > > +             goto out;
> > > > > > +
> > > > > > +     /*
> > > > > > +      * Since our CM IDs are NOT bound to any ib device we will create them
> > > > > > +      * only once
> > > > > > +      */
> > > > > > +     ctx = ib_ctx.srv_ctx;
> > > > > > +     ret = rtrs_srv_rdma_init(ctx, ib_ctx.port);
> > > > > > +     if (ret) {
> > > > > > +             /*
> > > > > > +              * We errored out here.
> > > > > > +              * According to the ib code, if we encounter an error here then the
> > > > > > +              * error code is ignored, and no more calls to our ops are made.
> > > > > > +              */
> > > > > > +             pr_err("Failed to initialize RDMA connection");
> > > > > > +             ib_ctx.ib_dev_count = -1;
> > > > > > +             return ret;
> > > > > > +     }
> > > > > > +
> > > > > > +out:
> > > > > > +     /*
> > > > > > +      * Keep a track on the number of ib devices added
> > > > > > +      */
> > > > > > +     ib_ctx.ib_dev_count++;
> > > > > > +
> > > > > > +     return 0;
> > > > > > +}
> > > > > > +
> > > > > > +static void rtrs_srv_remove_one(struct ib_device *device, void *client_data)
> > > > > > +{
> > > > > > +     struct rtrs_srv_ctx *ctx;
> > > > > > +
> > > > > > +     ib_ctx.ib_dev_count--;
> > > > > > +
> > > > > > +     if (ib_ctx.ib_dev_count)
> > > > > > +             return;
> > > > > > +
> > > > > > +     /*
> > > > > > +      * Since our CM IDs are NOT bound to any ib device we will remove them
> > > > > > +      * only once, when the last device is removed
> > > > > > +      */
> > > > > > +     ctx = ib_ctx.srv_ctx;
> > > > > > +     rdma_destroy_id(ctx->cm_id_ip);
> > > > > > +     rdma_destroy_id(ctx->cm_id_ib);
> > > > > > +}
> > > > > > +
> > > > > > +static struct ib_client rtrs_srv_client = {
> > > > > > +     .name   = "rtrs_server",
> > > > > > +     .add    = rtrs_srv_add_one,
> > > > > > +     .remove = rtrs_srv_remove_one
> > > > > > +};
> > > > > > +
> > > > > >  /**
> > > > > >   * rtrs_srv_open() - open RTRS server context
> > > > > >   * @ops:             callback functions
> > > > > > @@ -2051,12 +2111,26 @@ struct rtrs_srv_ctx *rtrs_srv_open(struct rtrs_srv_ops *ops, u16 port)
> > > > > >       if (!ctx)
> > > > > >               return ERR_PTR(-ENOMEM);
> > > > > >
> > > > > > -     err = rtrs_srv_rdma_init(ctx, port);
> > > > > > +     ib_ctx = (struct rtrs_srv_ib_ctx) {
> > > > > > +             .srv_ctx        = ctx,
> > > > > > +             .port           = port,
> > > > > > +     };
> > > > > > +
> > > > > > +     err = ib_register_client(&rtrs_srv_client);
> > > > > >       if (err) {
> > > > > >               free_srv_ctx(ctx);
> > > > > >               return ERR_PTR(err);
> > > > > >       }
> > > > > >
> > > > > > +     /*
> > > > > > +      * Since ib_register_client does not propagate the device add error
> > > > > > +      * we check if .add was called and the RDMA connection init failed
> > > > > > +      */
> > > > > > +     if (ib_ctx.ib_dev_count < 0) {
> > > > > > +             free_srv_ctx(ctx);
> > > > > > +             return ERR_PTR(-ENODEV);
> > > > > > +     }
> > > > >
> > > > > I afraid that you overcomplicated here, ib_register_client() doesn't
> > > > > return error if ->add() for specific device failed, it doesn't mean
> > > > > that ->add won't be called again for another device.
> > > > >
> > > > > So you don't need to use ib_dev_count == -1, just keep it to be 0 and
> > > > > leave to  rtrs_srv_close() to free srv_ctx.
> > > >
> > > > Leaving it 0 when there is an error is not gonna work. Since when the
> > > > modules are all built-in, a call to ib_register_client() will not
> > > > result in a call to ->add() then and there. So ib_register_client()
> > > > will return after registering the client, but without calling ->add().
> > > > Which means, ib_dev_count would be 0.
> > >
> > > If ib_dev_count == 0 => rtrs_srv_rdma_init() didn't success => nothing
> > > to release.
> >
> > True, but we have to send a failure back to the caller of
> > "rtrs_srv_open()" (and user of this ulp); which in our case is
> > rnbd-srv's function rnbd_srv_init_module().
> > In our case, the rnbd-drv module init would fail if "rtrs_srv_open()"
> > fails, meaning rtrs_srv_rdma_init() had failed.
> >
> > Even if we are talking in generic terms, any module calling the
> > "rtrs_srv_open()" of the rtrs ulp, would want to know if the server
> > open failed or succeeded right?
> I think Leon is right, any success of call to ->add, we have something
> to rtrs_srv_rdma_init,
> we can consider rtrs_srv_open is success instead of an error.

I don't think Leon is right. In case when all modules are built in,
add is not called in place when ib_client_register is called (cause
there are no registered devices and add_client_context will be called
at some later point in time from enable_device_and_get instead). So we
have two cases: built in and not built in. Then there is a possibility
that rtrs_srv_rdma_init fails (create_id, bind_addr, etc.) or doesn't
fail. Particularly we need to separate the case where add hasn't been
called at all yet (i.e. modules are built in, we just need to wait
until add gets called and then can start listening) and the case where
it did get called but failed (i.e. modules are not built in it got
called but failed - in that case we need to refuse to load module and
return error since we can't start listening).
The latter case is indicated by ib_dev_count = -1. I think it would
make code easier to read if instead of setting ib_dev_count to -1 and
explicitly checking whether it's below 0, one would introduce an
additional variable for the error code returned by rtrs_srv_rdma_init
and check it instead.

>
> Thanks Leon for catching this.
>
> Regards!