[PATCH v6 0/4] Add Aspeed AST2700 SDRAM EDAC support

From: Ryan Chen

Date: Tue Nov 17 2026 - 04:26:59 EST


Add ECC error reporting for the Aspeed AST2700 SoC to the existing
aspeed_edac driver.

The AST2700 memory controller keeps the same overall EDAC programming
model as the earlier Aspeed BMC SoCs, but uses a different register
layout, a split interrupt status/clear/mask scheme and DDR4/DDR5 memory.
Rather than fork the driver, the existing code is first generalised,
then the AST2700 is added as one more per-SoC variant.

The bug fixes, cleanups and AST2700 binding from v5 (v5 patches 1-5)
have been applied to the edac-drivers-aspeed branch of
git://git.kernel.org/pub/scm/linux/kernel/git/bp/bp.git, and this
series is based on that branch.

Patches 1 and 2 rework the implementation to accommodate the AST2700:
patch 1 replaces the driver's private regmap with direct register
access, and patch 2 moves the SoC-specific details behind per-SoC chip
data. Patch 3 adds the AST2700 support itself. Patch 4 is Stefan's
step-down patch, amended to leave the entry in place with me as its
reviewer.

Claude, an AI coding assistant, was used in developing patches 1-3.
It drafted parts of the code and of the changelogs, and proposed the
locking changes in this revision. I reviewed every change, and the
AST2600 and AST2700 test results are in patches 2 and 3.

Signed-off-by: Ryan Chen <ryan_chen@xxxxxxxxxxxxxx>
---
Changes in v6:
- Rebase onto the edac-drivers-aspeed branch of bp.git and drop the five
v5 patches applied there (v5 patches 1-5). v5 patches 6-9 are now
patches 1-4.
- Patch 1: use a spinlock_t rather than a raw_spinlock_t for the
register lock, as regmap did; the interrupt handler is not
IRQF_NO_THREAD. Patches 2 and 3 follow.
- Patch 1: drop the __guarded_by() annotations and the context analysis
opt-in, and assign the register base in probe without taking the
lock, so aspeed_probe() no longer mixes scoped_guard() with goto
unwinding.
- Link to v5: https://lore.kernel.org/r/20261007-edac-v5-0-5c31fc5df4cf@xxxxxxxxxxxxxx

Changes in v5:
- v5 patch 2: register the ECC interrupt with request_irq() instead of
devm_request_irq() followed by an explicit free, and keep the
interrupt number alongside the driver's other state so
aspeed_remove() does not have to call platform_get_irq() again.
- Link to v4: https://lore.kernel.org/r/20260930-edac-v4-0-c2e526f3ed79@xxxxxxxxxxxxxx

Changes in v4:
- Put the bug fixes first: the two fixes became v4 patches 1 and 2, the
cleanups v4 patches 3 and 4, and the bindings patch moved to v4
patch 5 ahead of the AST2700 work.
- Explain in v4 patch 1 where the grain value comes from, rather than
the warning it silences.
- Drop the Cc: stable tags from the two fixes. Neither meets the bar in
Documentation/process/stable-kernel-rules.rst; the Fixes: tags stay.
- Merge Stefan's step-down patch into v4 patch 9 (now patch 4),
keeping the MAINTAINERS entry alive with a reviewer instead of
removing it.
- Link to v3: https://lore.kernel.org/r/20260916-edac-v3-0-23e9ed2fbdf9@xxxxxxxxxxxxxx

Changes in v3:
- Establish the AST2700 interrupt mask in the driver rather than
deriving it from the register's contents: ast2700_set_irq() now
writes the mask outright, masking every source and unmasking only
the ECC ones, and the mask is established before the interrupt is
requested.
- New patch: add myself as a reviewer for the driver in MAINTAINERS.
- Tested on an AST2600 with CONFIG_PROVE_LOCKING=y; injected correctable
errors are reported from the interrupt handler with no lockdep reports.
- Link to v2: https://lore.kernel.org/r/20260824-edac-v2-0-c8d8bb693586@xxxxxxxxxxxxxx

Changes in v2:
- New patch (v5 patch 2, now applied): free the ECC interrupt before
the mem_ctl_info in aspeed_remove(). The interrupt is requested with
devm_request_irq() and is only released after .remove() has returned,
so edac_mc_free() could free the handler's context while the handler
was still running on another CPU.
- Acquire the register lock with the irqsave variant in init_csrows() and
in aspeed_probe(). The interrupt handler takes the same lock in hardirq
context.
- Move the dev_dbg() of the interrupt status register out of the
raw_spinlock critical section in the interrupt handler.
- Opt aspeed_edac.o into context analysis in drivers/edac/Makefile. The
check is opt-in per object, so without it the __guarded_by() annotation
added by the same patch was never actually verified.
- Return IRQ_NONE from ast2700_dramc_isr() when no ECC interrupt status
bit is set; the handler clears only the ECC bits, so unconditionally
claiming the interrupt could livelock the level-triggered line.
- Tested v2 on both AST2600 and AST2700 by injecting a correctable
error with the memory controller's ECC error injection; each is
reported as a CE with the expected failure address.
- Link to v1: https://lore.kernel.org/r/20260812-edac-v1-0-03992edea297@xxxxxxxxxxxxxx

---
Ryan Chen (3):
EDAC/aspeed: Replace regmap with direct register access
EDAC/aspeed: Abstract SoC differences behind chip data
EDAC/aspeed: Add AST2700 support

Stefan M Schaeckeler (1):
MAINTAINERS: Step down as Aspeed AST2500 EDAC driver maintainer

CREDITS | 8 +
.../bindings/edac/aspeed,ast2400-sdram-edac.yaml | 2 +-
MAINTAINERS | 2 +-
drivers/edac/aspeed_edac.c | 385 ++++++++++++++-------
4 files changed, 273 insertions(+), 124 deletions(-)
---
base-commit: 65d0a904696bae7e99a2c6eef61d4df03bac025f
change-id: 20260625-edac-8e960e02e7f7

Best regards,
--
Ryan Chen <ryan_chen@xxxxxxxxxxxxxx>