portal/devfs: share a dynamic char-major block instead of one major per device

[lacraig2]

Problem

Devices registered with major < 0 each call alloc_chrdev_region(), which allocates one whole char-device MAJOR per device. The kernel only has ~234 free majors, so firmware modeling many /dev nodes exhausts the pool.

Concretely on RouterOS6 (armel): nova opens /dev/log%d in a busy loop, so the config enumerates /dev/log0..255 (+ other nodes, ~339 total). Registrations past the cap fail with HYPER_RESP_WRITE_FAIL → the plugin logs Failed to register devfs device 'logNNN' (kernel returned 0) ×105. The guest still boots, but those /dev/logN nodes go unmodeled → ENOENT open-spin → CPU drain, defeating the very reason they're modeled.

Fix

igloo_alloc_dynamic_devt(): reserve majors in blocks of IGLOO_DYN_MINORS (256) and hand out one MINOR per device, allocating a fresh block only when the current fills. Effective cap goes from ~234 devices to ~234×256. The cdev_add()/device_create() failure paths only unregister_chrdev_region() for static-major registrations (a shared-block minor doesn't own its region).

On the prior revert

A prior shared-major attempt (driver 0.0.75, 6fff830) was reverted (a979a09, 0.0.76) for an unrelated mipsel do_ade unaligned-access crash — never documented at the time. I reproduced that crash on a period-accurate image and traced it to a separate unaligned-access bug in the era's driver/qemu, since fixed on the current base — not to major-sharing. Evidence: 0.0.75's exact allocator ported onto the current base boots mipsel clean.

Validation

• Real target — armel RouterOS6: 105 → 0 devfs registration failures; boots to READY, verifier all-pass.
• MIPS matrix — {mipsel, mipseb, mips64el, mips64eb} × {4.10, 6.13}: 8/8 clean boots, 0 do_ade (incl. big-endian, the worst case for unaligned access).
• Stress — 300-device /dev flood on mipsel: all register across the 256-minor block rollover, no crash (vs 93 failures per-device).

Follow-up

A regression guard (mipsel + dynamic-devfs flood boot test) will be added in penguin's basic_target CI as a tripwire once this is released and pinned.

[lacraig2]

Regression guard added in penguin: rehosting/penguin#891 (asserts no kernel oops during basic_target boot — the tripwire that was missing when 0.0.75 was reverted).