bnxt_re: Fix witness reported calltrace while unloading driver

[sumit.saxena_broadcom.com]

Summary

bnxt_unregister_dev() called synchronize_rcu() while holding
bp->en_ops_lock, which witness flags as a sleeping-while-locked
violation on driver unload. Drop the lock around the RCU grace period
and reacquire it afterward.

Also gate the RoCE extended-stats hex dump in
bnxt_qplib_qext_stat() behind bootverbose instead of printing it
unconditionally.

[gallatin]

sys/dev/bnxt/bnxt_en/bnxt_ulp.c

114

Aren't you still sleeping while holding a lock? This isn't our msleep, which drops the lock and re-acquires it, but the linux kpi one, which does pause_sbt() under the hood.

again, I *HATE* that you mix linux and freebsd in the driver. it makes it almost impossible to reason about

[sumit.saxena_broadcom.com]

sys/dev/bnxt/bnxt_en/bnxt_ulp.c

114

Will fix up this in next revision.

[sumit.saxena_broadcom.com]

Handled review comments

[sumit.saxena_broadcom.com]

AI scan :

• The wait loop gave up after ~1 second and returned 0 (success) regardless of whether ulp->ref_count actually reached zero, telling callers unregistration completed when a ULP consumer might still hold a reference into the ulp_ops/async_events_bmap state this function just tore down. At least one caller (bnxt_re_dev_unreg() in bnxt_re/main.c) does check this return value and logs on failure. Return ETIMEDOUT when the loop exits without ref_count reaching zero.