Hey there, >From IBTA: C17-19: For deadlock prevention, the CA shall not continuously and permanently assert backpressure (i.e. fail to grant link credits). My understanding is that CX-5, when an EQ becomes full, asserts back-pressure into the fabric, thus violating C17-19. For a system with N CQs, if all of them is associated with the same EQ, at most N EQEs may be (attempted) written to the EQ. I see that the mlx4 driver handles this "correctly" by sizing the synchronous, affiliated completion EQs, proportional with the number of CQs. From mlx4_init_eq_table(): err = mlx4_create_eq(dev, dev->quotas.cq + MLX4_NUM_SPARE_EQE, [] But, the mlx5 driver doesn't. It does in create_comp_eqs(): nent = MLX5_COMP_EQ_SIZE; which is hard-coded to 1024. So, a CX-5 system with more that 1K CQs may violate C17-19. What is the reason this is so different from the mlx4 driver? Thxs, Håkon
