From mboxrd@z Thu Jan 1 00:00:00 1970 From: Claude Code Review Bot To: dri-devel-reviews@example.com Subject: Claude review: accel: ethosu: Runtime PM refcounting and cmd stream validation fixes Date: Thu, 19 Feb 2026 09:57:30 +1000 Message-ID: In-Reply-To: <20260218-ethos-fixes-v1-0-be3fa3ea9a30@kernel.org> References: <20260218-ethos-fixes-v1-0-be3fa3ea9a30@kernel.org> X-Mailer: Claude Code Patch Reviewer Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: quoted-printable MIME-Version: 1.0 Overall Series Review Subject: accel: ethosu: Runtime PM refcounting and cmd stream validation fi= xes Author: "Rob Herring (Arm)" Patches: 4 Reviewed: 2026-02-19T09:57:30.759643 --- This is a three-patch fix series for the Arm Ethos-U NPU accelerator driver= , addressing runtime PM refcount underflow on job submit failure, incorrect= scalar mode detection for elementwise operations on the u65, and potential= integer underflow in IFM size calculations. The patches are from Rob Herri= ng and fix real bugs found during testing with additional models. The architectural approach in patch 1 =E2=80=94 splitting cleanup into pre-= queue and post-queue paths =E2=80=94 is the right idea. The two hardware-sp= ecific fixes in patches 2 and 3 are straightforward and well-motivated. One= concern in patch 1 is that `done_fence` (which is separately kzalloc'd) is= not freed by `ethosu_job_err_cleanup`, resulting in a memory leak on the e= rror paths that go through `out_cleanup_job` or `out_put_job`. This was als= o leaked in the original code (since `dma_fence_put` on a zero-initialized = fence triggers a kref WARN and doesn't actually free), but it would be good= to fix it properly while restructuring the cleanup code. --- Generated by Claude Code Patch Reviewer