1 Commits

This Branch

This Branch

All Branches

Author	SHA1	Message	Date
MastaG	2c4ab4b421	fix: support Linux 7.3 cfg80211 API ... 7.3 lands the last part of a two-commit series that converts several cfg80211_ops cookie parameters from driver-invented output values to values cfg80211 pre-assigns and passes in as input, and renames probe_client to probe_peer. Building against 7.3 fails with 5 errors: - cfg80211_probe_status() gained an MLO link_id parameter between cookie and acked (kernel commit 010e955c203e). This driver has no MLO support; -1 is the documented value for non-MLO. - cfg80211_ops::probe_client was renamed to probe_peer (4ab9b637b94a). - cfg80211_ops::mgmt_tx, ::remain_on_channel and ::probe_peer all changed their cookie parameter from u64 * to u64 (914781c72813): cfg80211 now pre-assigns the cookie before calling the driver instead of the driver inventing one and returning it through the pointer. The cookie change is more than a type fix. This driver invents its mgmt_tx/remain_on_channel/probe_peer cookies from a skb pointer, an internal roc_cookie_cnt counter, and an internal probe_id counter respectively, and separately re-derives "the same" value later when reporting completion. Once cfg80211 assigns the cookie itself, the completion report must echo back the value cfg80211 gave the driver. Two of the three ops are thin wrappers around a lower-level helper that is also called from a purely-internal path with no cfg80211-assigned cookie to honor: rwnx_start_mgmt_xmit() is reused by TDLS, and rwnx_cfg80211_remain_on_channel_() starts the internal RoC from inside mgmt_tx. Each of those helpers gets a bool (use_given_cookie / the existing mgmt_roc_flag) so the internal callers keep inventing a cookie exactly as before on every kernel, while the cfg80211 op wrappers honor the pre-assigned value on 7.3+ and thread it through to the completion report. probe_peer has no helper and no internal caller; the op stores the cookie itself. - rwnx_start_mgmt_xmit() gains use_given_cookie; the resulting cookie is stored in a new rwnx_sw_txhdr::cookie and echoed back via cfg80211_mgmt_tx_status(). Because sw_txhdr comes from a kmem_cache and is never zeroed, every one of the five allocation sites initialises the field -- to the skb pointer, which is exactly what the completion path reported before the field existed -- so the TXU_CNTRL_MGMT filter on the reporting side is not what keeps it safe. - rwnx_cfg80211_remain_on_channel_()'s mgmt_roc_flag also selects the cookie source, stored in a new rwnx_roc_elem::cookie that the rwnx_msg_rx.c completion handlers read. - rwnx_cfg80211_probe_client() stores the given cookie in a new apm_probe_sta::cookie field for the async probe-status work. Every gate this adds uses AICWF_CFG80211_VERSION_CODE, so an OpenWrt backports build selects these paths from the wireless stack's version rather than the kernel's, matching the existing 7.2 gates. On kernels before 7.3 the compiled behaviour is identical to main with two exceptions, both requested in review and both changing only the order of a store relative to a queue/send: *cookie for probe_client is taken from probe_id before queue_work() rather than after it, and the RoC cookie is read from roc_cookie_cnt once, before rwnx_send_roc(), rather than after it returns. In each case the value differs from main's only if the worker/RX path has already completed and bumped the counter in between -- a window in which main returns N+1 while having already reported N to cfg80211. The reported and returned values now agree by construction. v2, addressing review: - RoC: the cookie is decided and stored in roc_elem before the element is published and rwnx_send_roc() is called, so an early channel-switch indication cannot report an uninitialised cookie. - probe_peer: the cookie is stored before queue_work(), so a worker running immediately on another CPU reports the right one. - Monitor injection: intended to initialise sw_txhdr->cookie in rwnx_start_monitor_if_xmit(), but the store landed in rwnx_start_xmit() instead, where it is never read. Fixed in v5. - Pre-3.14: use_given_cookie is part of that signature too and both old-signature callers pass false. - Rebased onto the OpenWrt backport-version change. v3: - rwnx_main.h gated the probe_client prototype on LINUX_VERSION_CODE while rwnx_main.c gated the definition on AICWF_CFG80211_VERSION_CODE. Identical on a native build, but a backports build with the two decoupled would see a u64 * prototype against a u64 definition. - rwnx_cfg80211_remain_on_channel() no longer duplicates the whole call to the shared helper just to change one argument; only the cookie pointer is selected by version. v4: - rwnx_cfg80211_remain_on_channel_() no longer writes *cookie before rwnx_send_roc(). The cookie is snapshotted into a local, roc_elem keeps its early assignment, and *cookie is written where it was before this patch: inside the error == 0 branch. v5: - The monitor-injection initialisation is now where it was always meant to be: rwnx_start_monitor_if_xmit(), between the skb assignment and raw_frame = 1, ungated, since that path runs on every kernel with CONFIG_RWNX_MON_XMIT. The value is (unsigned long)skb_mgmt, which is what rwnx_txdatacfm() reported for these frames before. - The remaining allocation sites that did not set the field -- intf_tx() and rwnx_probersp_work() -- now do, so all five do. - The internal RoC's block-local cookie in rwnx_cfg80211_mgmt_tx() is renamed roc_cookie: on 7.3 it had come to shadow a same-typed parameter with a different meaning. - Comment fixes: the rwnx_start_mgmt_xmit() kerneldoc documents use_given_cookie; the probe_peer comment states the 7.3 contract; the two #endif trailers that close chains now opening on the 7.3 gate say so; the unused rx_addr is explained on the 7.3 arm as well. - Rebased onto main (set_monitor_channel on 6.12 stable). Verified against real linux-cachyos-bc250 (7.2.6) and linux-cachyos-rc-bc250 (7.3-rc4) headers with LLVM=1, the way DKMS builds them, asserting on exit code and module count: upstream main @ 7.2.6 rc=0, 3 modules, 13 warnings upstream main @ 7.3-rc4 rc=2 - the errors this patch fixes this branch @ 7.2.6 rc=0, 3 modules, 13 warnings this branch @ 7.3-rc4 rc=0, 3 modules, 13 warnings warning sets on 7.2.6: identical, and none names a function this patch touches. The backports configuration is not something I can build here, so the header/definition agreement in v3 rests on both sites using the same macro rather than on a test.	2026-09-24 12:34:45 +02:00