On 10 May 2021, at 19:04, Jason Gunthorpe [off-list ref] wrote:

On Wed, May 05, 2021 at 02:54:01PM +0200, Håkon Bugge wrote:

quoted

There are three conditions that must be fulfilled in order to consider
a partition match. Those are:

     1. Both P_Keys must valid
     2. At least one must be a full member
     3. The partitions (lower 15 bits) must match

In system employing both limited and full membership ports, we see
these false warning messages:

RDMA CMA: got different BTH P_Key (0x2a00) and primary path P_Key (0xaa00)
RDMA CMA: in the future this may cause the request to be dropped

even though the partition is the same.

See IBTA 10.9.1.2 Special P_Keys and 10.9.3 Partition Key Matching for
a reference.

Fixes: 84424a7fc793 ("IB/cma: Print warning on different inner and header P_Keys")
Signed-off-by: Håkon Bugge <redacted>
---
drivers/infiniband/core/cma.c | 22 ++++++++++++++++++++--
1 file changed, 20 insertions(+), 2 deletions(-)

What is this trying to fix?

The false warning messages. The wrong way though:-)

IMHO it is a bug on the sender side to send GMPs to use a pkey that
doesn't exactly match the data path pkey.

The active connector calls ib_addr_get_pkey(). This function extracts the pkey from byte 8/9 in the device's bcast address. However, RFC 4391 explicitly states:

4.1.  Broadcast-GID Parameters

   The broadcast-GID is set up with the following attributes:

       1. P_Key

          A "Full Membership" P_Key (high-order bit is set to 1) MUST be
          used so that all members may communicate with one another.


In other words, ib_addr_get_pkey() will sometimes wrongly return the full-member version of the partition, when the port is given only the limited member.

Let me do some post-coffee, pre-lunch work tomorrow to come up with a solution, aka ib_find_cached_pkey() followed by an ib_get_cached_pkey()?


Thxs, Håkon