[PATCH 21/23] wifi: mt76: mt7925: stop queueing resets once the device is being removed

[Sean Wang]

From: Jacobs Wu <jacobs.wu at mediatek.com>

mt7925e_unregister_device() cancels reset_work before it tears the device
down, but every source that can raise a reset stays live past that point:
the MCU command path, the system error recovery and interrupt handlers,
and the MAC watchdog all
call mt792x_reset(), and the interrupt tasklet is only disabled at the very
end of the function. A reset raised in that window is queued behind the
cancel and then runs while the device is being dismantled -
mt7925_mac_reset_work() sets hw_full_reset, stops the queues and
cancels the PM works before it can notice that the device is gone.

mt792x_reset() already bails out on !hw_init_done, and that flag has no
other consumer: it is set once during hardware init and read only there.
Clear it at the top of the teardown so no reset can be queued for its whole
duration.

Measured on rauru with kprobes on mt792x_reset() (queue), on
mt7925_mac_reset_work() (execution) and on mt76_unregister_device(), which
runs immediately after the cancel and so serves as the anchor, while
chip_reset was written in a loop across the module unload:

  before: 205 resets queued and 415 mt7925_mac_reset_work() runs after the
          anchor, that is after cancel_work_sync() had already returned
  after:  no run after the anchor; mt792x_reset() is still entered but
          returns early

A chip_reset on a running device still triggers a reset as before, and
unload/reload cycles stay clean.

Fixes: c948b5da6bbe ("wifi: mt76: mt7925: add Mediatek Wi-Fi7 driver for mt7925 chips")
Co-developed-by: Sean Wang <sean.wang at mediatek.com>
Signed-off-by: Sean Wang <sean.wang at mediatek.com>
Signed-off-by: Jacobs Wu <jacobs.wu at mediatek.com>
---
 drivers/net/wireless/mediatek/mt76/mt7925/pci.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/drivers/net/wireless/mediatek/mt76/mt7925/pci.c b/drivers/net/wireless/mediatek/mt76/mt7925/pci.c
index 09153d624fd5..f7b57a82f2d4 100644
--- a/drivers/net/wireless/mediatek/mt76/mt7925/pci.c
+++ b/drivers/net/wireless/mediatek/mt76/mt7925/pci.c
@@ -46,6 +46,16 @@ static void mt7925e_unregister_device(struct mt792x_dev *dev)
 	if (dev->phy.chip_cap & MT792x_CHIP_CAP_WF_RF_PIN_CTRL_EVT_EN)
 		wiphy_rfkill_stop_polling(hw->wiphy);
 
+	/* Stop new resets from being queued for the rest of the teardown.
+	 * mt792x_reset() bails out on !hw_init_done, which is otherwise only
+	 * set once at init, so clearing it here closes the window in which an
+	 * MCU timeout, a system error recovery interrupt or the watchdog
+	 * could still schedule
+	 * reset_work behind the cancel below and run it against a device that
+	 * is already being dismantled.
+	 */
+	dev->hw_init_done = false;
+
 	cancel_work_sync(&dev->reset_work);
 	cancel_work_sync(&dev->init_work);
 	mt76_unregister_device(&dev->mt76);
-- 
2.43.0