Allocator + CMA settings: umbrella tracker for all platforms

[widgetii]

Overview

Tracking the unification of the MMZ allocator selection, CMA reservation, and OSAL module-loading conventions across every HiSilicon / Goke platform supported by openhisilicon + OpenIPC/firmware. Companion to #58 (ISP) and #82 (sensors).

The hi3516cv500 family was just brought in line with the hi3516ev200/ev300 reference shape — kernel-side CMA fallback, HIGHMEM disabled, 256 MiB CMA, open_* module naming, modprobe driven from load_hisilicon. The same modernisation applies to the rest of the family tree, with a couple of legacy V1/V2A platforms that need a different (mmz/) approach.

The pattern was established across these landed PRs:

• OpenIPC/linux#37 — kernel-side hi_cma.c fallback: when load_hisilicon doesn't pass an mmz= zone in cmdline, synthesise one from dma_contiguous_default_area. Lets stock OpenIPC firmware with no U-Boot tweaks sysupgrade cleanly into the new world.
• OpenIPC/firmware#2050 — cv500 osdrv: switch to DT-defined CMA-default allocator.
• OpenIPC/firmware#2058 — load open_hwrng early so userspace getrandom() doesn't block.
• OpenIPC/firmware#2060 — load_hisilicon: skip cmdline mem= when it equals total RAM (V4+CMA).
• OpenIPC/firmware#2064 — cv500: align module naming with EV300 (open_*); rewrite load_hisilicon to modprobe open_*.
• OpenIPC/firmware#2067 — cv500: disable HIGHMEM, bump default CMA to 256 MiB.

End-to-end verified on hi3516av300 + IMX415 (CV500 family) at master+101d778: 33 open_* modules loaded, IMX415 detected, RTSP/HTTP serve 4K JPEG snapshots.

Per-platform state

Numbers from current master (firmware) / upstream-patches+vendor branches (openipc/linux) / master (openhisilicon) on 2026-05-07.

Platform	Gen	Module naming (load_hisilicon)	HIGHMEM	CMA size	hi_cma fallback	mmz_allocator branches	Notes
hi3516cv500	V3.5	✅ modprobe open_* (40)	✅ off	✅ 256 MiB	✅ in vendor 4.9	hisi+cma	Reference shape (this PR series)
hi3516ev200	V4	✅ modprobe open_* (36)	✅ off	0 (DT-cma)	❌ needs port	hisi+cma	Original reference; no fallback patch yet
hi3516ev300	V4	(shares ev200 osdrv)	✅ off	0 (DT-cma)	❌ needs port	hisi+cma
hi3516cv300	V3	❌ insmod legacy (33)	❌ on	(none set)	❌ needs port	hisi+cma	9 sensors source-built, modules still vendor-named
hi3519v101	V3A	❌ insmod legacy (38)	❌ on	16 MiB	❌ needs port	hisi only
hi3516av200	V3A	(shares v101)	❌ on	16 MiB	❌ needs port	(n/a)
hi3516cv200	V2	❌ insmod legacy (28)	✅ off	(none set)	❌ needs port	hisi only
hi3516av100	V2A	❌ insmod legacy (26)	✅ off	(none set)	n/a — no hi_cma	hisi only	Standalone mmz/, no CMA path
hi3516cv100	V1	❌ insmod legacy (27)	✅ off	(none set)	n/a — no hi_cma	hisi only	Standalone mmz/, no CMA path
hi3536dv100	DVR	❌ insmod legacy (2)	✅ off	(none set)	n/a	(none)
hi3520dv200	DVR	❌ insmod legacy (1)	✅ off	(none set)	n/a	(none)	Just landed (#2065), uses minimal load script

Counts: "modprobe open_*" / "insmod legacy" reflect the count of active (uncommented) module-loading lines in general/package/hisilicon-osdrv-<chip>/files/script/load_hisilicon.

What "unified" means per dimension

1. Module naming → open_* everywhere

• Goal: hisilicon-opensdk.mk installs verbatim from $(@D)/kernel/open_*.ko into $(KMOD_DST); no rename to hi_*/hi3516cv500_*/sys_config.ko. Matches what was done for cv500 in firmware#2064.
• Why: load_hisilicon can drive everything via modprobe open_X (the upstream convention) instead of legacy insmod hi_X.ko filenames; the rmmod side already used open_* names in many scripts and silently no-op'd because nothing matched.
• Scope: cv200, cv300, v101, av100, cv100 (and dv100/dv200 small DVR scripts).

2. HIGHMEM disable + sized CMA reservation

• Goal: # CONFIG_HIGHMEM is not set + CONFIG_CMA_SIZE_MBYTES=256 in every *.generic.config for boards with ≥ 256 MiB RAM and the cma_allocator path. HIGHMEM creates a low/high boundary at 0xb0000000 that CMA reservations spanning it can't satisfy — atomic_pool_init panics on boot.
• Why: matches the cv500 fix from firmware#2067.
• Scope: cv300, v101, av200 still have CONFIG_HIGHMEM=y. cv200 and lower don't have the cma_allocator path so HIGHMEM is moot for them.

3. Kernel-side CMA fallback default pool

• Goal: port the hi_cma.c fallback — when no mmz= cmdline zone is registered, synthesise one from dma_contiguous_default_area so OSAL allocators don't return NULL. Lets sysupgrade work without U-Boot bootargs surgery.
• Scope: ev200 (V4), cv300 (V3), v101 (V3A), cv200 (V2) vendor kernel branches in OpenIPC/linux. cv500 already done. av100/cv100 have no hi_cma.c (they use the standalone mmz/ path).

4. CMA-default reserved-memory in DT (V3.5+ kernels)

• Already in cv500 / dv300 mainline-port DTSI via firmware#2050.
• Will need to land for any future *_neo variants of other families when they materialise (cv300_neo, v101_neo, etc.).

5. set_allocator script (boot-time MMZ size negotiation)

• ev200 + av100 ship a set_allocator helper that calls into U-Boot env to derive the MMZ zone string. Only those two; cv500 doesn't have one and works without (uses osmem/totalmem env vars directly in load_hisilicon).
• Probably not worth unifying — cv500 simpler model is fine.

Recommended PR order

The work is mostly per-platform and parallelisable. Suggested order — easiest first, biggest blast radius last:

1. hi3516cv300 — port hi_cma fallback + opensdk install (drop renames) + load_hisilicon → modprobe open_* + disable HIGHMEM. Closest analog to the cv500 work; well-tested userspace.
2. hi3519v101 — same shape, fewer sensors, smaller module set.
3. hi3516cv200 — needs hi_cma fallback (vendor kernel) + load_hisilicon rewrite. HIGHMEM already off.
4. hi3516ev200 + hi3516ev300 — already have open_* naming; port hi_cma fallback into hisilicon-hi3516ev200 linux branch so first-boot after firmware update doesn't need bootargs change.
5. hi3516av100, hi3516cv100 — different allocator (standalone mmz/, no hi_cma.c). Just the load_hisilicon → modprobe open_* rewrite + opensdk install simplification. Lower priority; legacy hardware.

Per-PR shape

Each platform's PR series follows the same template (mirror of cv500):

• OpenIPC/linux (vendor branch): port hi_cma_get_zone() fallback patch — single commit, ~40 LOC. (feat: add hi3516cv200 SoC family support #37 for cv500 as reference.)
• OpenIPC/firmware PR A: hisilicon-opensdk.mk — drop open_*.ko → hi_*/<chip>_* rename loop in the per-family install block; install verbatim. (#2064 cv500 example.)
• OpenIPC/firmware PR B: general/package/hisilicon-osdrv-<chip>/files/script/load_hisilicon — rewrite insert_* chains from insmod hi_X.ko → modprobe open_X. The remove_* side typically already uses open_* names (silent no-op pre-fix). Bundled in same PR as A.
• OpenIPC/firmware PR C (where HIGHMEM=y today): kernel config — flip HIGHMEM off, bump CONFIG_CMA_SIZE_MBYTES. (#2067 cv500 example.)

Sub-issue / sub-PR table

Platform	hi_cma fallback (linux)	naming + load_hisilicon (firmware)	HIGHMEM/CMA (firmware)	Status
hi3516cv500	linux#37 ✅	firmware#2064 ✅	firmware#2067 ✅	✅ DONE
hi3516ev200	not yet	already open_*	already off / DT-cma	partial — needs hi_cma fallback port
hi3516ev300	not yet	(shares ev200)	already off / DT-cma	partial — needs hi_cma fallback port
hi3516cv300	not yet	not yet	needs HIGHMEM off	not started
hi3519v101	not yet	not yet	needs HIGHMEM off + CMA bump	not started
hi3516av200	not yet	(shares v101)	needs HIGHMEM off + CMA bump	not started
hi3516cv200	not yet	not yet	already off	not started
hi3516av100	n/a (no hi_cma)	not yet	already off	not started
hi3516cv100	n/a (no hi_cma)	not yet	already off	not started

Issues/PRs will be linked here as they're opened.

Out of scope (intentionally)

• Sensor source builds — tracked in Sensor source compilation: umbrella tracker for all platforms #82.
• ISP source builds — tracked in ISP source compilation: umbrella tracker for all platforms #58.
• Kernel modernisation (vendor 4.9 → mainline 7.0+) — tracked per-platform in *_neo variants. cv500_neo is the reference; cv300/v101/cv100/av100 don't have neo builds yet.
• Set-allocator script — only ev200 + av100 have it; works as-is, no need to propagate to families that don't need it.

🤖 Generated with Claude Code