axi_isolate: fix deadlock during drain

[dkimTT]

Fix

Follow up to #441: Two changes, both under TerminateTransaction = 1'b1, axi_isolate_inner is unchanged.

Registered selects

The demux selects are driven by a registered isolate_i instead of isolated_o. The register only updates on cycles where the channel has no request presented-and-unaccepted at a demux master port: once a request is presented, w_select_q is latched, so the select has to hold until the handshake completes. The same gate covers de-isolation: if isolate_i falls while a request sits unaccepted at the error slave, the select stays 1 until it is accepted; otherwise the AW would move to port 0 (inner module) while its W beats are already routed to the error slave.

Inner sizing

The inner now gets a larger NumPending than the demux:

localparam int unsigned DemuxCntWidth   = cf_math_pkg::idx_width(NumPending);
localparam int unsigned DemuxMaxPending = (32'd1 << DemuxLookBits) * ((32'd1 << DemuxCntWidth) - 32'd1);
localparam int unsigned InnerPending    = TerminateTransaction ? DemuxMaxPending + 32'd1 : NumPending;

DemuxMaxPending is the most transactions the demux's ID counters can have outstanding at once. With one more than that, the inner can never saturate and refuse a request the demux has already routed to port 0, so backpressure always happens at the demux's slave port, before w_select_q is latched.

Question: is the shared NumPending intentional?

The demux and inner module use NumPending differently: the demux only sizes its ID counters with it, so it admits up to 2 * (2**idx_width(NumPending) - 1) transactions (possible 5 for NumPending = 4), while the inner treats it as an exact limit and cuts the AW channel at pending_aw_q >= NumPending while staying in Normal, so the demux can route one more request to the inner than it will accept. If isolate_i is asserted in that state, the FSM enters Drain directly (Hold is skipped, since the parked request was never forwarded downstream) and the drain runs with a request already committed at port 0 and unaccepted. Was the demux admitting more than the inner's limit intentional, and is deriving the inner's capacity from the demux ceiling the right fix, or would you rather bound the demux to exactly NumPending?

Verification

Ran a cocotb bench covering the scenario from #441, terminations during a loaded drain (W and ID-hash interlocks, W-before-AW ordering), select-hold windows on both channels in both isolate_i directions, saturation at the demux counter ceiling, ATOP AR-credit draining, and seeded random traffic. The demux's slv_aw_select_stable / slv_ar_select_stable assertions were enabled in all runs, with no violations. Happy to contribute this bench if it's useful