On Wed, Mar 21, 2018 at 12:11:59AM +0000, Parav Pandit wrote:
> > > Right way to do ib core to have open() and close() callback.
> >
> > Probably.. But since we already accepted the patches that cause this bug for this
> > merge window we may be stuck accepting a hack.
> >
> > The hack is to let mlx5 opt out of resource tracking for it's internal objects.
> There are some patches from Steve that I didn't review deeply, but
> those patches let some internal data structure of the provider
> driver to be exposed to user for debugging purpose, which is very
> useful.
>
> Given UMR QP is one such core QP that might need debugging as well
> in future and some other internal UD QPs that Bodong is working on.
> Given that it might be useful to debug such internal QP as
> well. Having them resource tracked is useful instead of opting out.
> So not tracking them further creates blockers to not able to debug
> them in future.
Yes, that seems to be likely. Someone will ultimately have to add new
methods and revise mlx5. It really shouldn't be creating resources
prior to registering..
> > At least that way we still protect the ULPs on other non-broken drivers.
> Not sure if any other drivers have dangling resources after ib_unregister_device() is called.
> Otherwise Leon' patch would have notrack() API invoked in other provider drivers too.
It is more the ULPs that worry me. ULPs leaving dangling resources
after client unregister seems kind of likely.
> > Maybe restrack_init should be moved to the register as well in this patch?
> That alleast avoids the confusing code.
What if we add this hunk to the patch? Leon?
diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c
index 0ab99e62cc5ce0..247f2be0a1516b 100644
--- a/drivers/infiniband/core/device.c
+++ b/drivers/infiniband/core/device.c
@@ -464,6 +464,12 @@ int ib_register_device(struct ib_device *device,
struct ib_udata uhw = {.outlen = 0, .inlen = 0};
struct device *parent = device->dev.parent;
+ /*
+ * Nothing is permitted to create objects prior to calling
+ * register_device.
+ */
+ rdma_restrack_clean(&device->res);
+
WARN_ON_ONCE(device->dma_device);
if (device->dev.dma_ops) {
/*
Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
