sys/opencrypto/criov.c
247	The indentation of the continuing line is off, ditto below.
258	Would be nice to have a local var `pagelen = m_epg_pagelen(m, i, pgoff)`, ditto below.
261	I would KASSERT(offset < m_epg_trllen).
304	With commit 49f6925ca34 there is a good chance that the pages are physically contiguous, it may be profitable to check for that here.
312	The existence of this check implies that skip + len might be an offset past the end of the mbuf, but we appear to assume that skip alone is always within the bounds of the mbuf. It would be nice to have some assertions or comments documenting exactly what the assumptions are.
433	Just a general comment, I think it would be nicer and more efficient if we had a function which returned both the base address and length of the next segment. Many consumers call _segbase and _seglen back-to-back. It seems especially awkward with extpg buffers since we end up looping over the page array twice.
821	I don't really understand the ifdefs. On platforms without a direct map, PHYS_TO_DMAP() will panic. We don't have special ifdef guards around M_EXTPG handling elsewhere, so I'm not sure why it's required here.
sys/opencrypto/cryptodev.h
210 ↗	(On Diff #88709)	Drop extra whitespace while here?

sys/opencrypto/criov.c
312	I was probably coding this a bit defensively. Looking at the caller, we check skip and len against m_len first and only invoke this if skip is < m_len. We probably should assert that.
433	Yes, that was not as apparent for the previous types, but the double loop is a bit much, and in practice it is true that callers tend to fetch the pair.
821	The m_epg_contiguous_subsegment() function is conditional on the #ifdef. I could perhaps always define the function instead and have it's body be #ifdef'ed?

sys/opencrypto/criov.c
821	I see. It's more that in generic code we handle M_EXTPG mbufs unconditionally, implicitly assuming that they won't arise on platforms without a direct map. I'd maybe move m_epg_contiguous_subsegment() out of the ifdef as well. I guess that has the downside of compiling dead code on platforms which don't have a direct map, so I don't have a strong preference. It's ok as-is.