let's align with other transports which use uct_base_ep_is_connected:

ep_addr is optional, so cannot use:

    if (!(params->field_mask & UCT_EP_IS_CONNECTED_FIELD_EP_ADDR)) {
        return 0;
    }

is not addr_qp == 0 means false for is_connected state?

it means we skip the comparison between addr_qp and qp_num (and only compare device address)

ok

there is some code duplication with uct_rc_mlx5_base_ep_is_connected
to eliminate it we can create new parent struct for uct_gga_mlx5_ep_address_t and uct_rc_mlx5_ep_address_t - that will contain only qp_num. This struct can be called uct_rc_mlx5_base_ep_address_t.
uct_rc_mlx5_base_ep_is_connected will cast params->ep_addr to uct_rc_mlx5_base_ep_address_t*, if the flag UCT_EP_IS_CONNECTED_FIELD_EP_ADDR is specified.

it looks like braking of backward compatibility...

is there currently a scenario where gga transport is selected by default by UCP?

I think it can be selected for rma_bw lane

why need uct_base_ep_is_connected? seems rc_mlx5 does not need it

to compare iface address (only required for GGA)

both will compare IB device address which is redundant
can we reuse common code with uct_gga_mlx5_iface_is_reachable_v2 and compare iface address directly?
BTW @evgeny-leksikov @amaslenn the gga transport does not currently implement returning info string from uct_gga_mlx5_iface_is_reachable_v2

BTW @evgeny-leksikov @amaslenn the gga transport does not currently implement returning info string from uct_gga_mlx5_iface_is_reachable_v2

thanks, I'll add

minor: swap order (call uct_gga_mlx5_iface_is_same_device first)

uct_iface_is_reachable_params_t looks redundant here, maybe just const char* str, size_t length and use ucs_snprintf_safe?

if we drop uct_iface_is_reachable_params_t then there will be code duplication:

    char *info_str_buf      = UCS_PARAM_VALUE(UCT_IFACE_IS_REACHABLE_FIELD,
                                              params, info_string, INFO_STRING,
                                              NULL);
    size_t info_str_buf_len = UCS_PARAM_VALUE(UCT_IFACE_IS_REACHABLE_FIELD,
                                              params, info_string_length,
                                              INFO_STRING_LENGTH, 0);

We can to move most of logic of this function up, to uct_gga_mlx5_iface_is_reachable_v2().
So uct_gga_mlx5_iface_is_same_device() will get only tl_iface and iface_addr, and return 0/1. It will contain only the logic in lines 627-629. Then no need to define dummy params inside uct_gga_mlx5_ep_is_connected()

there will still be code duplication:

    if (gga_addr == NULL) {
        uct_iface_fill_info_str_buf(params, "iface address is empty");
        return 0;
    }

uct_iface_fill_info_str_buf(params, "iface address is empty"); should be only in uct_gga_mlx5_iface_is_reachable_v2; uct_gga_mlx5_ep_is_connected can just check if (gga_addr == NULL)

if (gga_addr == NULL) will be the duplication

IMO it's small enough (1 line) duplication and it can avoid extra initialization of dummy struct so it's worth it

We can to move most of logic of this function up, to uct_gga_mlx5_iface_is_reachable_v2().
So uct_gga_mlx5_iface_is_same_device() will get only tl_iface and iface_addr, and return 0/1. It will contain only the logic in lines 627-629. Then no need to define dummy params inside uct_gga_mlx5_ep_is_connected()