A fork of the Linux MediaTek Wi-Fi driver (openwrt/mt76) with one fix: injected Wi-Fi frames on MT7920/MT7921/MT7922 PCIe cards now go out at the rate that was requested. Without this fix, AirDrop between an iPhone and a Linux laptop over omdrop-owl is unreliable on 2.4 GHz and can fail entirely on 5 GHz.
Status: experimental. The fix is not in upstream mt76 or any released kernel yet.
AirDrop on non-Apple hardware relies on injecting raw Wi-Fi frames at a chosen legacy rate. On these MediaTek cards, the PCIe driver loses the requested rate before it builds the hardware descriptor. Every injected unicast frame then goes out at the 1 Mbps CCK default. That is too slow for reliable 2.4 GHz transfers, and in the reported 5 GHz failure, injected unicast never reached the air (omdrop-owl issue 14).
The fix saves the requested rate before TX status bookkeeping overwrites it, then sets only the descriptor's rate field. Everything else the driver normally does, including antenna selection and ACK policy, is unchanged. HT/VHT rates and hardware encapsulation are left alone.
Check with lspci -nnk | grep -A3 -i mediatek. The fix applies to cards using the mt7921e driver:
| Chip | PCI ID seen in testing | Status |
|---|---|---|
| MT7922 | 14c3:7922, 14c3:0616 |
Works: transfers in both directions, active-monitor and shared modes |
| MT7921 | 14c3:7961 |
Works: independent tester, transfers in both directions in active-monitor mode |
| MT7920 | — | Included because it shares the same code path; not tested on hardware |
| MT7902 | — | Deliberately excluded |
USB (mt7921u) and SDIO (mt7921s) versions are not affected by this change. The hardware results above come from the earlier revisions of the fix. The current revision only widens which chips are enabled and has not yet been confirmed on hardware. The full record is in docs/VALIDATION.md.
The usual approach is to patch the mt76 source that matches your running kernel and rebuild only mt7921e, leaving every other driver module stock. Don't build this fork's master against a distro kernel: it tracks newer kernel APIs and won't link.
# 1. Fetch the mt76 sources for your kernel version (here 7.2.5; check `uname -r`).
# For an x.y.0 kernel the tag is vX.Y, e.g. v7.2.
git clone --depth 1 --branch v7.2.5 --filter=blob:none --sparse \
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git linux-src
cd linux-src
git sparse-checkout set drivers/net/wireless/mediatek/mt76
cd drivers/net/wireless/mediatek/mt76
# 2. Apply the fix.
curl -fsSL https://github.com/shaecodes-wilde/mt76/compare/467ae2e3f190...fix/mt7921-family-driver.diff | patch -p1
# 3. Build only mt7921e against your installed kernel headers.
make -C /lib/modules/$(uname -r)/build M="$PWD/mt7921" \
CONFIG_MT7921_COMMON= CONFIG_MT7921S= CONFIG_MT7921U= modules
# 4. Swap it in until the next reboot.
sudo rmmod mt7921e
sudo insmod mt7921/mt7921e.koTo go back to the stock driver, reboot or run sudo rmmod mt7921e && sudo modprobe mt7921e.
Notes:
- These steps were checked against Linux 7.2.5: the patch applies cleanly and
mt7921e.kobuilds. If your distro carries its own kernel patches and step 2 fails, use your distro's kernel source package. - With Secure Boot enforcing module signatures, an unsigned module is refused. Sign it or use a test machine.
- This repository provides no installer or DKMS package. Making the module persistent is up to you.
- The driver fix doesn't configure omdrop-owl. Shared mode needs Wi-Fi on an AirDrop-compatible channel (6, 44, or 149).
The change is confined to mt7921e_tx_prepare_skb() in mt7921/pci_mac.c. Root cause: mt76_tx_status_skb_add() clears status_driver_data, which overlaps info->control.flags, so IEEE80211_TX_CTRL_RATE_INJECT has already been erased when the common TXWI builder checks it.
The clean patch series is on its own branches, based on openwrt/mt76 467ae2e3f190, without this README:
fix/mt7922-injected-rate-upstream: the original MT7922-only fix (02b6c9a1)fix/mt7921-family-driver: the follow-up that extends the fix to MT7920/MT7921 (2a2f085d)
Offline regression testing: a 24-case kernel test module exercises the real PCI/status/TXWI routines. Each unpatched base failed 11 cases and each patched version passed all 24, against openwrt/mt76, Linux 7.2.5 and the maintainer tree at 2e6968e9286f. The expanded 72-case family suite builds but has not been executed yet. Both patches pass checkpatch.
Before mailing, the series needs to be squashed with a revised description, given a verified earliest Fixes: reference (the current tag only covers MT7922) and a human DCO sign-off. The reproducer and source hashes are in the omdrop-owl findings.
Same as upstream mt76; see LICENSE.