Conversation
stintel
marked this pull request as draft
September 19, 2026 17:55
dsa_switch_shutdown() only disconnects the conduit during system shutdown. It does not remove the switch from the DSA tree before the driver releases its private state. Use dsa_unregister_switch() during probe unwind and driver removal so DSA tears down the tree first. Make the link-state IRQ device-managed so it cannot outlive the switch state. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Add RTL931x hardware programming for the tested two-member, one-link topology. Configure switch identities, device routes, the fabric port, stack trunks and non-unicast forwarding as one transaction. Snapshot every touched register, preserve unrelated shared fields and migrate local FDB device IDs when the silicon identity changes. Retain the snapshot until restoration succeeds so partial failures remain recoverable. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Expose a versioned generic netlink interface for configuring and inspecting the fixed two-member, one-link topology. Validate topology and monotonically increasing generations, serialize configuration with DSA register accesses and reject unsupported use of the active fabric port. Preserve explicit error state when hardware rollback cannot prove that the original configuration was restored. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
The RTL931x CPU tag identifies the original source device and port on RX and selects a destination switch unit on TX. Carry that information instead of treating every packet as local device zero. Extend the software trailer with a device field and cache the local switch identity in the tagger. Fence RX and drain conduit TX while that identity changes. Preserve DSA source metadata for exceptional local CPU-tag copies whose original source port differs from their physical ingress, while rejecting mismatches during identity changes. Retain fragmented-frame discard state across NAPI polls. Use the ratelimit helpers available in the OpenWrt 6.18 kernel when reporting rejected CPU-tag traffic. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Add the RTL931x CPU-tag modes needed for one-hop physical delivery and routed CPU-to-CPU unicast. Provide a consuming Ethernet helper which passes private destination metadata to the descriptor-based TX path. Intercept CPU2CPU receive reason 1 before the ordinary remote-device drop, preserve its classification across fragments and attach the local fabric ingress as metadata. Deliver only validated control frames to the stack protocol handler. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Provisionally reserve the candidate fabric port and exchange CPU Device Talk PING and PONG messages before changing either switch identity. Bind the proof and delayed replies to the carrier epoch and CPU-device generation so a link flap cannot authorize a stale peer. Apply the RTL931x stack link-up workaround around MAC release, preserve the pre-arm port-register image for rollback and leave the fabric port down whenever recovery is uncertain. Reject active LAG configuration before closing and restarting the selected port. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Separate the common CPU Device Talk fields from the probe-specific body so later routed messages can share the transport. Keep the version 1 PING and PONG layout byte-for-byte unchanged, including its fixed zero padding. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Split common and probe validation and teach the transmitter to carry bounded typed bodies. Enforce canonical lengths and zero padding so each message has one wire representation. Preserve the existing PING and PONG protocol and keep transmission serialized against CPU identity changes. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Allow the RTL931x CPU port to address a physical port on an explicit remote stack device. On RX, route a validated remote source device and port tuple through an RCU-published peer-netdevice map. Make the exported peer-port TX buffer contract explicit. Unshare and linearize each skb before appending the destination trailer because the MAC reads that trailer through skb->data and maps one contiguous DMA buffer. Retain the tuple across fragmented frames and resolve the destination only at the real packet tail. This lets map teardown synchronize against the complete receive operation and prevents stale netdevice access. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Add a typed CPU Device Talk RPC for querying a configured peer's switch inventory and per-port state. Fence every request with both boot nonces, the stack generation and the routed source and destination identities. Keep the initial operations read-only so retries with one transaction ID do not require a mutation replay cache. Expose only explicit administrative generic netlink queries instead of a raw register tunnel. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Some switch fabrics represent ports of a peer switch using netdevices owned by one control-plane switch. The peer's DSA user netdevice must stop being a data and configuration endpoint while phylink keeps the physical port running. Add an RTNL-serialized delegation API. Fence RX, TX, upper changes and DSA configuration while delegated, but preserve the host address and phylink state required by the replacement data path. On undelegation, restore normal DSA ownership before reattaching the user netdevice. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Add an absolute, master-only CPU Device Talk operation for transferring the follower user ports to the stack leader. Apply the DSA delegation API under RTNL and roll back every port if the full set cannot be transferred. Track the resulting mask for idempotent retries and lifecycle cleanup. Keep follower phylinks running while preventing their local DSA netdevices from carrying traffic or changing switch configuration. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Extend the peer switch query with user, administrative, carrier and delegated-port masks plus the peer topology and protocol capabilities. Return one coherent inventory snapshot so the leader can validate the entire follower before publishing replacement netdevices. Keep the per-port query for details which cannot be represented in the masks. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Add exact RTL931x SRAM and CAM lookup helpers for CPU host FDB entries. Claim or create each plain host entry, move its destination between the follower and leader CPU ports and restore or remove it when delegation ends. Handle both hash banks correctly, use the RTL931x maximum age of seven and disable background aging on CPU port 56. Recognize the exact aged static-entry shape produced by the hardware and canonicalize it before changing ownership, while rejecting duplicate or special entries. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Create ordinary Ethernet netdevices on the stack leader for the follower's user ports. Transmit through the leader conduit with an explicit remote device and physical port, and publish an immutable RCU map for the matching receive path. Delegate all follower ports before making the replacements visible. Fence and unpublish them before identity changes, link teardown or unregistration, and retain netdevice references until all RCU readers and conduit transmissions are drained. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Add a target-gated libnl-tiny client for the RTL931x stack generic netlink family. Support status, Device Talk arm and probe, two-member configuration, peer inventory and peer-netdevice control operations. Print typed replies as key-value records and include kernel extended ACK messages on failure. Enable GNU libc extensions because libnl-tiny's credential declarations require struct ucred from the system headers. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Remove the assumption that every follower user port has one common MAC address. Report each address through GET_PORT_STATE and have the leader query every advertised port before creating its replacement netdevice. Track a deduplicated set of follower host addresses for FDB ownership, so shared addresses are moved once while unique addresses retain their individual destinations. Bump the private and generic netlink protocol versions to reject peers and clients with the old address semantics. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Validate the complete follower port set before changing ownership. Require each transitioning DSA user netdevice to be administratively up, present and free of bridge, LAG, HSR or other upper devices. Reject any disagreement between the tracked and generic DSA delegation state. Report the failing port or host FDB operation on the follower and map permission conflicts separately from retryable busy conditions in the Device Talk result. This gives the leader an actionable extended ACK. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Fence leader-owned peer ports immediately when the fabric link goes down. Track the desired publication state and a monotonic link epoch so a down and up cycle during activation cannot publish stale inventory. After link recovery, rerun peer inventory and absolute delegation before reattaching the netdevices. Retry transient failures with bounded delayed work, and synchronize work cancellation with stack teardown so recovery cannot outlive its context. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Allow leader-owned peer ports to join one VLAN-aware Linux software bridge without claiming switchdev offload. Program only the hardware transport needed to carry those VLANs over the hidden fabric port while the leader bridge remains the forwarding authority. Add typed, absolute Device Talk operations for follower bridge-port and VLAN state. Keep peer learning disabled, redirect BPDUs to the master CPU, maintain tagged fabric membership on both switches and restore exact saved state during leave or rollback. Reject unsupported bridge flags, MST, VLAN protocols and multicast-offload modes. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Extend the stack status reply with the local user-port mask and the mask currently delegated to a remote leader. Exclude the selected fabric port while stacking is configured or provisionally armed. This lets an early boot manager distinguish hardware configuration from complete follower ownership transfer before releasing normal networking. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Replace the monolithic diagnostic source with a shared generic netlink client used by both the command-line utility and a new full-C daemon. Read the stack topology and boot policy from UCI. Bring up the fabric and required follower ports, verify the peer, configure both members and either publish leader peer netdevices or wait for follower delegation. Expose status and recovery methods through ubus and publish a separate readiness object only after the complete transaction is verified. Start the daemon before the network service and make the existing switch setup hook wait for that readiness object. Support indefinite waiting, fail-closed timeout and rollback-to-standalone policies, restoring the original administrative port state after a successful fallback. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
A member can remain configured while its peer is rebooted and upgraded. The surviving member then retains the old peer boot nonce. Routed PING/PONG continues to work after the peer returns, but RPC requests are silently rejected because their session validation uses the stale nonce. This leaves stackd waiting for peer inventory indefinitely. Treat a successful routed probe as an explicit session re-establishment. Before accepting it, require the configured peer and master IDs, generation, physical stack port, and all configured/route status bits to match. If the boot nonce changed, invalidate the old session, reset mutation and replay state, and publish the new nonce. Keep the one-hop verification path using the same session publication helper. This allows independently rebooted members to converge again without disabling an otherwise operational stack. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Leader representor open and stop currently only change the local queue. The delegated follower netdevice remains physically enabled, so a down representor can continue to learn and emit traffic. Extend DSA delegation with an explicit physical administrative state. Keep the original user netdevice detached while the owner starts or stops the port, and restore its current state when delegation ends. Add an absolute full-bitmap Device Talk v9 mutation. Apply it transactionally on the follower, preserve down-port isolation through bridge replay, and verify the resulting inventory before activation. Drive the bitmap from representor IFF_UP state. Fence and recover all peer ports after an uncertain mutation. Replay the desired state during recovery, and detect later divergence during periodic inventory synchronization. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Peer netdevices expose a narrow remote bridge interface, but the current documentation does not clearly distinguish them from DSA user ports. Document the supported data, administrative, carrier and bridge surface. List unsupported PHY, ethtool and offload operations, and explain the existing Device Talk capability negotiation. Defer a permanent per-netdevice feature ABI unless representors survive the planned cascaded-DSA work. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
The generic delegation patch is a driver-shaped experimental API. Normal DSA cascading does not model two switches owned by separate kernels, but the current boolean is not yet a general ownership abstraction. Document the ownership, lifetime, phylink, suspend, removal and feature semantics that need an RFC before the API grows or is submitted. Also record that the independent RTL83xx switch-unregister correction is already a separate commit and must remain outside that series. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Automatic fallback must not reclaim follower ports while a former leader may still own them. That prevents split-brain forwarding, but leaves no supported recovery after an unrecoverable leader failure. Add an explicit local recovery command. Require the single fabric link to have no carrier and a newer generation. Close it before changing ownership, invalidate the peer session, restore delegated ports and saved switch state, and leave recovered ports down. Keep the operation out of Device Talk and stackd automatic fallback. Bump the generic netlink API to version 6, add matching CLI support and document the recovery procedure.
The proof-of-concept stack has exercised its principal data paths on hardware, but the protocol, recovery and rollback logic lacks automated coverage. Record that limitation explicitly rather than allowing successful bench tests to imply production readiness. List the unit, fault-injection and hardware validation areas that need coverage before wider deployment.
The Realtek Otto watchdog restart handler runs through an atomic notifier after MIPS has stopped the secondary CPUs. disable_irq() may sleep while synchronizing an active handler. If that handler was running on a stopped CPU, reboot deadlocks before the watchdog reset is armed. Backport a change to disable_irq_nosync(). Waiting for an old phase 1 handler is unnecessary when the watchdog is immediately reprogrammed to reset the SoC.
DSA assisted learning treats bridge ports outside its own tree as foreign and installs their learned addresses on the CPU port. This is incorrect for switchdev representors sharing the DSA switch parent ID: it overwrites the ASIC-learned remote destination and breaks hardware forwarding. Treat same-parent bridge ports as part of the local offload domain while retaining assisted learning for genuinely foreign ports. Include the RTL931x device ID in the debugfs L2 table output so distributed FDB destinations can be verified directly.
The current PCS attachment selection treats 10GBASE-R as fibre. Test the existing long-DAC analogue and TX settings against the observed one-sided and unstable links with FS SFPP-PC03 3 m passive cables. Override the attachment in the RTL931x configuration path only, before programming and recording it. Log the experimental override whenever the SerDes is reconfigured. Other families and interface modes retain their existing attachment selection, with no generic phylink changes. This deliberately forces DAC_LONG for all RTL931x 10GBASE-R links, including short DACs and optical modules. It does not identify the inserted medium or cable length and is not a general-purpose fix. DAC RX calibration remains unimplemented, so this selects the existing DAC path without running the fibre RX calibration. An absent fibre calibration warning is not evidence that the link is working correctly. No build was run. Hardware validation is pending with the 3 m DAC to Huawei and the 1 m DAC between the two Linksys switches.
Implement the missing DAC LEQ/DFE receive-training sequence after selecting the DAC attachment profile. Validate the PCS link and symbol errors before reporting carrier instead of trusting the forced MAC status mirror. Advance calibration through PCS polling with bounded checks and per-lane retry backoff. Retrain after link loss or replacement and stop calibration when the PCS is disabled. Preserve the working sibling lane by avoiding shared PLL reconfiguration during retries. Keep the experimental DAC_LONG override and leave generic phylink and DTS unchanged. Validate with checkpatch, git diff --check, and interpreted C state-machine tests with mocked register access, including one-shot errors at 52 I/O/helper operations. No compilation was run; hardware validation remains pending.
Extend the fixed two-member RTL931x stack from one fabric link to a hardware stack trunk containing up to four physical links. This allows a 48-port switch pair to use the full four-port SFP+ block as a 40 Gbit/s fabric while retaining operation after individual member failures. The Realtek Mango hardware provides 16 stack-port slots and eight stack-trunk groups. Reserve one slot for each configured link. Place live slots in stack trunk 0 and route the peer over the same slot bitmap. Enable stack-hash recalculation and link-down avoidance only while the route has multiple members. Preserve the proven direct slot route when exactly one member is live. Arm every configured link but require only one successful one-hop Device Talk exchange before committing the stack identity. Keep unverified links out of the hardware route. Verify late or restored links directly before admitting them. Require the configured peer identity and boot nonce, and invalidate that verification whenever carrier is lost. This lets a stack boot with any non-empty subset of its configured fabric without allowing a cable moved to another switch into the trunk. Track configured, verified and active fabric masks separately. Program hardware routing from the live verified subset and refresh the stack trunk whenever that subset changes. Keep peer representors usable while at least one member remains active, but fence them when the final member goes down. Apply delegation baselines, distributed bridge isolation matrices and VLAN membership only to the routed fabric subset. Configured but unverified links remain available for one-hop Device Talk without carrying ordinary data. Apply data policy before admitting a link to the route, remove the route before withdrawing policy, and fail closed to an empty route after an uncertain update. Continue probing missing configured members after the daemon reaches ready state so late and restored links can join automatically. Make disabled-state daemon convergence replace the complete provisional topology before arming the configured UCI ports. This discards stale armed members and carrier snapshots after a failed attempt or topology change. Validate and prune the verified subset before administratively closing fabric netdevs for the hardware transition; preserve that snapshot while the ports are intentionally down. Otherwise stackd can enter a one-second close, stale-verification failure and reopen loop. Bump the private Device Talk protocol to version 10 and generic netlink to version 7 for the incompatible inventory and status changes. Accept a UCI list of one to four interfaces, reject duplicates, validate every configured interface against the active kernel trunk, and report all three masks through the CLI and ubus status. Add a read-only stack_routes debugfs dump covering stack identities, slot assignments, device routes and stack-trunk groups. It provides the register evidence needed for bench diagnosis without requiring /dev/mem. This remains a hardware-test WIP. No build was run while preparing this commit. Bench validation still needs to prove flow distribution, failed-member avoidance and restoration, degraded bootstrap, all-links-down recovery, bridge/VLAN state replay, and rollback while individual trunk members flap. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Reuse complete and partial provisional fabric sets instead of disarming them on every convergence attempt. Repeated disarm and reopen cycles restart PCS calibration and can prevent DAC links from reaching carrier while flooding the console. Only replace a disabled provisional set when it contains unwanted ports. Preserve armed ports after a daemon restart, partial setup or lost netlink reply, and propagate errors without resetting the fabric. Document the behavior and bump the package release. Validated with diff checks and interpreted C control-flow tests for partial setups, retries, lost arm replies and failed status reads. No compilation or hardware validation was performed. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Replace the blanket LAG/stack exclusion with checks for overlapping fabric ports and delegated ownership. A leader-local LACP uplink can be created before or after the native stack trunk without netifd ordering changes. Use the current hardware device ID in normal LAG entries and source-port mapping keys. Rebind existing groups when stack identity changes and when returning to standalone mode. Keep link-down avoidance enabled during fabric route changes. Correct LAG membership bookkeeping and primary-port reelection. Mirror leader-local LAG membership and selected TX candidates to the delegated follower using session-checked, sequenced Device Talk mutations. Preserve original table rows for cleanup, retain restoration state across failed updates, flush obsolete dynamic learning, and replay current groups before restoring peer bridge forwarding. Peer synchronization failures schedule representor recovery without rejecting the local uplink change. Bump the private Device Talk protocol to version 11, advertise local LAG synchronization, and document its scope and pending hardware tests. Both stack members require matching firmware; remote representor bonds and cross-chassis LACP remain unsupported. Validated with checkpatch, diff checks, C syntax parsing and interpreted tests using mocked tables, including membership updates, device-ID rebinding, peer-update failures and cleanup retries. No compilation or hardware validation was performed. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Name local ports and peer representors sw<member>s<slot>p<panel-port>. Use the configured member identity and front-panel numbering, reserving slot zero for built-in ports. Keep hardware indices and stack0 unchanged. Resolve local port locations from existing device-tree labels without DTS changes. Apply names before network startup, accept board labels as fabric configuration aliases, and refuse to rename live or enslaved ports. Roll back completed renames if a subsequent rename fails. Add an optional read-only Device Talk port-location request without a wire-version bump. Use peer-provided locations for representor names, falling back to legacy names when the peer lacks the request. Document the naming scheme, configuration migration and reboot requirements. Bump the package release. Validated with diff checks, shell syntax checks and interpreted C tests using mocked sysfs, ioctls and RPCs. No compilation or hardware validation was performed. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Member
Author
|
I will keep deleting openwrt[bot] comments as long as this is in draft status. It gravely pollutes the discussion. Could we maybe somehow make it ignoring this PR? Or draft PRs in general? |
Contributor
|
make "realtek: rtl83xx: unregister DSA switch on driver removal" and "realtek: watchdog: avoid synchronous IRQ disable during restart" a pre-PR |
Member
|
the PCS related stuff should also be moved out to a separate PR, needs to be discussed in my opinion. If I see correctly those target issues I also somewhere have on my ToDo, but solved quite complicated here. I'll have a closer look soon. EDIT: Overall, it really bad IMO that DSA starts to include any PCS header, this needs to be designed differently |
plappermaul
requested changes
Sep 19, 2026
Comment on lines
+214
to
+216
| if (priv->family_id == RTL9310_FAMILY_ID) | ||
| sw_w32_mask(BIT(priv->r->cpu_port & 0x1f), 0, | ||
| RTL931X_L2_PORT_AGE_CTRL_REG(priv->r->cpu_port)); |
Contributor
There was a problem hiding this comment.
After long and hard work we are close to decomission the family_id. Please use other implementation (maybe via config structure).
stintel
force-pushed
the
rtl931x/stacking
branch
from
September 19, 2026 20:20
e5cd741 to
3bb1241
Compare
The stacking support adds family_id checks while the driver is moving variant-specific behavior into its configuration structure. Add a stacking capability and a device-ID callback to the variant configuration. Select separate RTL930x and RTL931x MAC callbacks and reuse the configured force-mode mask when bringing links down. Move the CPU-port aging workaround into RTL931x learning setup. Keep the existing link-up sequence and failure recovery unchanged. Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
plappermaul
requested changes
Sep 19, 2026
Comment on lines
+298
to
+302
| static bool rteth_931x_decode_tag(struct rteth_frag *frag, | ||
| struct rteth_dsa_tag *t) | ||
| { | ||
| u16 spn = frag->cpu_tag[2] & RTETH_931X_TAG2_SPN; | ||
|
|
Contributor
There was a problem hiding this comment.
Mak the ethernet tx/rx header coding/decoding split and signature enhancment a separate PR.
openwrt-ai
reviewed
Sep 20, 2026
openwrt-ai
left a comment
Collaborator
There was a problem hiding this comment.
Commit checks
- Missing
Signed-off-by:on 10 commits:b69df4be,f1daa607,e269eede,63422d6c,e444ab08,1c4cc770,27485b4b,5fbf7d90,6488d719,144cb5d1. 2e1ee866"WIP: realtek: rtl83xx: add multi-link stack trunk" —WIP:subject on a 1020-line commit; either finish it or drop it from the series.5fbf7d90"net: dsa: preserve same-parent switchdev FDB entries" — uses the upstream kernel subject prefix; the sibling patch commit1e5f2c17correctly usesgeneric: 6.18: ...for the same directory.27485b4b"realtek: watchdog: avoid synchronous IRQ disable during restart" — body calls it a backport, but the change is authored in this series and is not in v6.18.52.
Generated by Claude Code
| @@ -0,0 +1,514 @@ | |||
| From ffd3885e66b0f65bd123ecd46c4154543149b3de Mon Sep 17 00:00:00 2001 | |||
Collaborator
There was a problem hiding this comment.
netifd carries no patches today and is still at PKG_RELEASE:=1, which this PR leaves untouched. Bump it so the rebuilt package is actually reinstalled on upgrade.
Generated by Claude Code
Comment on lines
+106
to
+108
| + while ((len = read(fd, buf, sizeof(buf))) != 0) { | ||
| + if (len < 0) | ||
| + return -1; |
Collaborator
There was a problem hiding this comment.
No EINTR retry here, although the waitpid() below has one. netifd reaps proto-shell helpers constantly, so a SIGCHLD landing mid-read() makes an otherwise healthy selector look failed, and config_select_profile() then aborts the entire config_init_all(). Continue on EINTR instead of returning -1 (and re-count the hunk when regenerating).
Generated by Claude Code
Comment on lines
+183
to
+186
| + read_error = config_selector_read(pfds[0], profile, size); | ||
| + close(pfds[0]); | ||
| + if (read_error) | ||
| + kill(pid, SIGKILL); |
Collaborator
There was a problem hiding this comment.
There is no deadline on this blocking read/wait pair, and config_select_profile() runs synchronously from config_init_all() — which is also reached from the reload ubus method inside uloop_run(). A selector that hangs, or one that forks a child keeping the pipe write end open, stalls netifd's whole event loop forever, which contradicts the commit message's "without ... delaying process startup". An alarm/deadline with SIGKILL on expiry would make the fail-closed contract hold.
Generated by Claude Code
| config_free_vlans(); | ||
| config_init_ip(); | ||
| config_init_rules(); | ||
| @@ -839,7 +900,8 @@ config_init_all(void) |
Collaborator
There was a problem hiding this comment.
This hunk's old range 839–845 overlaps the previous hunk's 834–841 (and 900–907 overlaps 894–902) — no valid diff can do that, so these headers were hand-edited. Against the pinned netifd the two netifd_ucode_config_load() calls are 14 lines apart, not 5, and the whole config.c part currently applies only with offsets up to 9. Regenerate with make package/netifd/refresh.
Generated by Claude Code
| @@ -0,0 +1,547 @@ | |||
| From: Stijn Tintel <stijn@linux-ipv6.be> | |||
Collaborator
There was a problem hiding this comment.
704- is already used in this directory by 704-net-dsa-support-EPROBE_DEFER-when-reading-the-port-M.patch.`` Renumber to a free slot — 709 is unused.
Generated by Claude Code
| @@ -0,0 +1,34 @@ | |||
| From: Stijn Tintel <stijn@linux-ipv6.be> | |||
Collaborator
There was a problem hiding this comment.
Upstream still has disable_irq(ctrl->irq_phase1) at realtek_otto_wdt.c:186 in v6.18.52 and this patch is authored here, so the commit body's "Backport a change to disable_irq_nosync()" is inaccurate. Every other 0xx- patch in this directory carries a -vX.Y- upstream marker; a not-yet-accepted change belongs in the 3xx/7xx/8xx range the other local patches use.
Generated by Claude Code
Contributor
|
Please provide the ethernet commits as separate PR. Thanks |
Formality Check: FailedWe checked this pull request against the contribution guidelines. Here is what needs your attention: 🛑 CRITICAL ERRORS
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A month ago I acquired 2 Linksys LGS352C units for EUR100 each. Then discover the rtl931x silicon has stacking support. So I launched codex and I now have a single switch stack combining those 2 units.
I am posting this work early, so people know this kind of enterprise feature is possible. It's a very first attempt, and while it works, it's not working how I would want it to work, and will need a lot of work before this can land upstream.
As most active realtek contributors are not active on IRC, this PR also serves as a place to discuss, in public.
Some snippets of the code in action:
`ubus call rtl931x.stack status` on stack leader
`ip link` on stack leader
Details