net: yusur: yk3: report RSS mailbox failures to callers · Pull Request !1876 · OpenCloudOS/OpenCloudOS-Kernel

A normal VF RSS query or configuration request can fail in the mailbox transport, yet `ethtool -x` / `ethtool -X` still receives success. The four RSS key/table helpers have void interfaces and discard failures. Their get/set_rxfh callbacks return zero regardless, so a query can present incomplete/zero data as valid, or a multi-chunk update can stop partway without notifying the administrator.

This fixes error reporting, not RSS performance or a security boundary. It is independent of the previously investigated `data_length` field, which is unchanged.

### Fix

Return status from the four RSS helpers and propagate it through both ethtool compatibility interfaces and the direct cdev RETA caller. Stop on the first failed operation and leave the existing state mutex through a shared unlock path. Successful PF register access, VF command construction, chunking, reply mode and wire format remain unchanged. No lock, capability or userspace data layout is added or changed.

The mailbox API uses positive internal status values: L2 failure is 1, missing L3 reply is 2. Passing those directly to ioctl would not be a valid negative Linux errno. A private RSS wrapper converts them to `-EIO` / `-ETIMEDOUT` and preserves existing negative errors. Table initialization uses the same conversion.

**This is not transactional rollback.** A key or earlier table chunks already written before a later error may remain changed. The fix makes the failure visible so the operator can detect and retry it; it does not claim all-or-nothing hardware updates or add acknowledgements to commands that previously did not wait for a reply.

### Origin and scope

The original driver addition [!985](https://gitee.com/OpenCloudOS/OpenCloudOS-Kernel/pulls/985), commit `b3fe1bfec4e23964dfd2592930690a4803056227`, already contains these void helpers and unconditional-success callbacks. Its parent tree has no YK3 driver directory. This is a downstream YK3 fix, not an upstream stable patch republished outside its synchronization path. All-state RSS and YK3 searches, including closed PRs, were checked before submission; no same-topic PR was found.

Target: `linux-6.6/devel`, base `85698632c91c5a933eafe70c4bf18e4c8aed6070`. Four production files change: RSS implementation/header, ethtool callbacks, and the single direct cdev call site. Current target selects the legacy get/set_rxfh signatures. The newer compatibility branches are updated and reviewed but were not compiled against a newer kernel.

### Before/after verification

**Component:** extract the complete target RSS helpers, PF receive callback, register packing/unpacking and ethtool get/set blocks. At 1/16/17/64 queues, cover normal PF/VF requests, transport refusal, missing reply on GET, existing negative errno, first/later chunk failure and combined-request short circuit. Across 84 request cases, baseline hides 52 modeled failures; the fixed version returns every expected result. Another 24 table-initialization cases verify success/readback, negative errors and conversion of four positive failure results. O0 and O2 repeat the same cases, not additional independent scenarios. Component locks and transport are models.

**Real kernel interfaces:** in a private networkless OCK QEMU guest, register a test netdev using the exact get/set_rxfh and RSS/PF callbacks, with real ethtool ioctl core and mutexes. Successful modeled mailbox deliveries dispatch through the exact PF callback and register helpers. Userspace makes real `SIOCETHTOOL` `ETHTOOL_GRSSH` / `ETHTOOL_SRSSH` requests. Separately, a private misc-device adapter invokes the exact cdev RETA handler with the target packed UAPI and real `copy_from_user` / ioctl return handling; this is not the whole production cdev dispatch/lifecycle.

| Final VM checks | Baseline | Fixed |
| --- | --- | --- |
| 152 ethtool ioctls, 16 timelines (PF/VF × four queue counts × two variants) | 24 modeled failures incorrectly return success | All 24 return the expected negative errno |
| 24 cdev ioctls, normal 1/16-queue requests | Four modeled failures incorrectly return success | All four return the expected negative errno |
| Initial/recovery RSS readback | Correct | Correct; 32 successful GET readbacks across both variants |
| Failure in a combined/chunked request | Unconditional success can conceal partial results | Stops after the failed operation and reports it; no rollback claim |

Machine checks validate errno, request counts, normal payload readback and recovery; the final guest loads/unloads both fixtures and shuts down without BUG/Oops/WARNING/panic. Underlying mailbox failures are deliberately selected model return values, not physical hardware faults. No real NIC, firmware, DMA, packet distribution, concurrency or performance test was performed.

**Build/gates:** real target Kbuild with `CONFIG_YK3=m` / `CONFIG_PCI_MSI=y` compiles every selected YK3 translation unit and relocatably links `yk3.o` before and after. This is not modpost, a loadable full-driver test or a full kernel build. The unchanged project commit gate and whitespace checks pass. Strict checkpatch: 0 errors, 2 conditional upstream-cherry-pick/vendor-sign-off reminders, 0 checks; this is downstream-only and uses the actual contributor's sign-off.

### Limitations and retained diagnostics

Full component UBSan also detects an existing unaligned u32 key access in the unchanged PF helper. That diagnostic is retained and is not attributed to this patch. An additional run explicitly named `ubsan-no-alignment` checks the remaining UBSan categories; **full UBSan clean is not claimed**. No separate alignment PR is presented without its own functional necessity evidence.

Earlier fixture attempts had a success-backend short-circuit, a missing sanitizer library, a userspace reference to a kernel-only macro, and an omitted core-required ring-count callback. They were corrected and the final matrix was rebuilt/rerun; these fixture failures are not product defects. Sources, exact-function hashes, object files and full logs are retained locally, not represented as public attachments. AI assisted investigation and testing. No merge, score or performance outcome is promised.

类型	指派人员	状态
审查	zoucao Huang Cun Jianping Liu	已完成 (0/0人)
测试	Jianping Liu	已完成 (0/0人)

GVPOpenCloudOS / OpenCloudOS-Kernel

搜索帮助

GVP OpenCloudOS / OpenCloudOS-Kernel