wifi: mt76: mt7925: fix rmmod hang and token idr race in unregister

mt7925e_unregister_device() had two independent bugs in its teardown
ordering, both in the RX NAPI/token-idr handling around
mt7925_tx_token_put(), plus a separate tasklet lifetime gap in
mt7925_pci_remove().

1. It called napi_disable() directly on every RX queue, then a few
   lines later mt792x_dma_cleanup() -> mt76_dma_cleanup() called
   napi_disable() on the same NAPI instances again, with no
   napi_enable() in between. napi_disable() leaves NAPI_STATE_SCHED
   set until a matching napi_enable(); a second, unpaired call on an
   already-disabled NAPI has nothing left to wait for that can ever
   clear that bit, so it hangs unconditionally.

   This is the same bug independently found and reported upstream by
   Mikhail Gavrilov ("wifi: mt76: mt792x: drop redundant napi_disable()
   in unregister path", 13b7e6a96a00) and by Devin Wittmayer
   (#70 review). Root cause: 4ab8f2122dcb added this same
   napi_disable() to mt76_dma_cleanup(), was reverted by f5f14a017454
   ("Needs to be fixed to avoid regression on mt762x"), then re-added
   by 332bbe9b2784 without removing the driver-side napi_disable()
   loops in mt7921/pci.c and mt7925/pci.c that predate it.

2. Once the redundant loop above was deleted without care, NAPI was no
   longer quiesced before mt7925_tx_token_put()'s idr_destroy(). A
   still in-flight RX NAPI poll can reach PKT_TYPE_TXRX_NOTIFY ->
   mt7925_mac_tx_free() -> mt76_token_release() -> idr_remove() on the
   same idr concurrently with that idr_destroy() -- a real
   use-after-free, not a redundant check. Confirmed independently
   twice: Sashiko (via Eric Biggers) against Mikhail's patch upstream,
   and Sashiko/Lucid-Duck against an earlier version of this fix
   (#70). tasklet_disable() alone is not sufficient: it
   only blocks *future* napi_schedule() calls, it does not wait out a
   poll already running. napi_disable() does wait for it, closing the
   race.

Fixed here by restoring a single napi_disable() loop in
mt7925e_unregister_device(), positioned before mt7925_tx_token_put()
as in bug 2, kept disabled afterward rather than re-enabled: upstream
PR #72 (following Mikhail's own proposed v2, reverting
332bbe9b2784) removed mt76_dma_cleanup()'s own napi_disable() entirely,
so this driver-side call is now the only one and there is nothing left
to double-disable. Re-enabling here would leave RX NAPI on going into
mt792x_dma_cleanup()'s netif_napi_del()/page_pool_destroy(), which is
exactly what those warn about if NAPI is still active. MT76_REMOVED
early-exits were added to mt792x_irq_tasklet(), mt792x_poll_tx() and
mt792x_poll_rx() (mt792x_dma.c) matching the pattern already used by
the non-PCI bus types in this driver family, so nothing can turn NAPI
back on in this window regardless.

Also: mt792x_dma_cleanup()/mt792x_wfsys_reset() do raw WFDMA/WFSYS
register I/O and need driver ownership of the chip.
mt76_unregister_device() above runs mac80211 vif teardown, which can
queue pm->ps_work and hand ownership back to firmware before
cancel_delayed_work_sync(&pm->ps_work) catches it. The pre-existing
__mt792x_mcu_drv_pmctrl(dev) before mt76_unregister_device() is kept
(vif teardown itself needs ownership too); a second call was added
right before mt792x_dma_cleanup() so ownership is guaranteed fresh at
the point it's actually needed, independent of what ps_work did during
teardown.

Separately: mt7925_pci_remove() called tasklet_disable() (inside
mt7925e_unregister_device()) but never tasklet_kill() before
mt76_free_device(). tasklet_disable() only blocks the tasklet from
running, it does not remove an already-scheduled instance from the
pending list; if one was still pending when the containing struct was
freed, that is a use-after-free waiting to happen. The only existing
tasklet_kill() in this file is in the suspend path, not remove. Also
independently flagged by Mikhail Gavrilov in the same upstream thread.
Added tasklet_kill(&dev->mt76.irq_tasklet) in mt7925_pci_remove(),
after devm_free_irq() and before mt76_free_device().

Fixes: 332bbe9b2784 ("wifi: mt76: fix connac2/3 DMA queue cleanup")
Link: #70
Link: #71
Link: #72
Link: https://lore.kernel.org/all/CABXGCsO07SExb+Z0PeN6MZ1fKC24Tvn3ehSyeQc-3qFC7jM7dQ@mail.gmail.com/

Verified: all tests/mt7927/*.sh source contracts covering this
function pass, clean build against 7.2.0-rc1-mt7927-monitor-v3, and 4
consecutive real modprobe/modprobe -r cycles on physical MT7927
hardware immediately after a fresh reboot, zero hangs and zero dmesg
warnings.

Not in this change: mt7921e_unregister_device() (mt7921/pci.c) has the
same missing tasklet_kill() gap in mt7921_pci_remove(), and
mt7921_pci_remove() sets MT76_REMOVED *after* calling unregister
(mt7925_pci_remove() sets it before), so the MT76_REMOVED guards added
here don't protect mt7921e either. Left out of scope for this PR;
flagging so it isn't mistaken for an oversight.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Author: Ashcal9669

mt7925/pci.c

@@ -40,29 +40,56 @@ static int mt7925e_init_reset(struct mt792x_dev *dev)
 
 static void mt7925e_unregister_device(struct mt792x_dev *dev)
 {
-	int i;
 	struct mt76_connac_pm *pm = &dev->pm;
 	struct ieee80211_hw *hw = mt76_hw(dev);
+	int i;
 
 	if (dev->phy.chip_cap & MT792x_CHIP_CAP_WF_RF_PIN_CTRL_EVT_EN)
 		wiphy_rfkill_stop_polling(hw->wiphy);
 
 	cancel_work_sync(&dev->reset_work);
 	cancel_work_sync(&dev->init_work);
+	__mt792x_mcu_drv_pmctrl(dev);
 	mt76_unregister_device(&dev->mt76);
-	mt76_for_each_q_rx(&dev->mt76, i)
-		napi_disable(&dev->mt76.napi[i]);
 	cancel_delayed_work_sync(&pm->ps_work);
 	cancel_delayed_work_sync(&dev->mlo_pm_work);
 	cancel_work_sync(&pm->wake_work);
 
+	/* Quiesce RX NAPI before tx_token_put()'s idr_destroy(): a still
+	 * in-flight poll can reach PKT_TYPE_TXRX_NOTIFY -> mt76_token_release()
+	 * -> idr_remove() on the same idr, racing idr_destroy() below
+	 * (morrownr/mt76 PR #70 review, Sashiko/Lucid-Duck). tasklet_disable()
+	 * alone only stops *new* napi_schedule() calls, it does not wait out
+	 * a poll already running, so it is not sufficient by itself.
+	 *
+	 * Left disabled from here on, deliberately not re-enabled: PR #72
+	 * reverted mt76_dma_cleanup()'s own napi_disable(), so this is now
+	 * the only disable in the whole teardown and mt792x_dma_cleanup()'s
+	 * netif_napi_del()/page_pool_destroy() below expect NAPI to already
+	 * be off. Re-enabling here would leave it on for that, which is what
+	 * those warn on. MT76_REMOVED is already set by the caller, so
+	 * mt792x_poll_rx()/mt792x_poll_tx()/mt792x_irq_tasklet() can't turn
+	 * it back on in the meantime either.
+	 */
+	tasklet_disable(&dev->mt76.irq_tasklet);
+	mt76_for_each_q_rx(&dev->mt76, i)
+		napi_disable(&dev->mt76.napi[i]);
+
 	mt7925_tx_token_put(dev);
+
+	/* Re-assert driver ownership: mt76_unregister_device() above can run
+	 * vif teardown that queues pm->ps_work, which can hand ownership back
+	 * to firmware before cancel_delayed_work_sync(&pm->ps_work) catches
+	 * it. The first __mt792x_mcu_drv_pmctrl() call is still required
+	 * before mt76_unregister_device() itself (that teardown needs driver
+	 * ownership too); this second call guarantees mt792x_dma_cleanup()/
+	 * mt792x_wfsys_reset() below, which touch WFDMA/WFSYS registers
+	 * directly, don't run against a firmware-owned chip.
+	 */
 	__mt792x_mcu_drv_pmctrl(dev);
 	mt792x_dma_cleanup(dev);
 	mt792x_wfsys_reset(dev);
 	skb_queue_purge(&dev->mt76.mcu.res_q);
-
-	tasklet_disable(&dev->mt76.irq_tasklet);
 }
 
 static void mt7925_reg_remap_restore(struct mt792x_dev *dev)
@@ -725,6 +752,14 @@ static void mt7925_pci_remove(struct pci_dev *pdev)
 	set_bit(MT76_REMOVED, &mdev->phy.state);
 	mt7925e_unregister_device(dev);
 	devm_free_irq(&pdev->dev, pdev->irq, dev);
+	/* mt7925e_unregister_device() only tasklet_disable()s irq_tasklet,
+	 * which blocks it from running but does not remove it from the
+	 * pending list if it was already scheduled. tasklet_kill() here
+	 * makes sure nothing is left pending against dev before
+	 * mt76_free_device() frees the memory it points into (reported
+	 * upstream: lore.kernel.org/all/CABXGCsO07SExb+Z0PeN6MZ1fKC24Tvn3ehSyeQc-3qFC7jM7dQ@mail.gmail.com/).
+	 */
+	tasklet_kill(&dev->mt76.irq_tasklet);
 	mt76_free_device(&dev->mt76);
 	pci_free_irq_vectors(pdev);
 }

mt792x_dma.c

@@ -31,6 +31,17 @@ void mt792x_irq_tasklet(unsigned long data)
 	const struct mt792x_irq_map *irq_map = dev->irq_map;
 	u32 intr, mask = 0;
 
+	/* This can already be queued by mt792x_irq_handler() before
+	 * MT76_REMOVED is observed there. Bail out here too instead of
+	 * touching hardware registers and calling napi_schedule() during
+	 * teardown -- every other bus type in this driver family already
+	 * makes this check (usb.c, mt76x02_usb_mcu.c); PCI was missing it,
+	 * and napi_disable() during remove has no bound on how long it
+	 * waits for a NAPI poll that keeps getting rescheduled.
+	 */
+	if (test_bit(MT76_REMOVED, &dev->mphy.state))
+		return;
+
 	mt76_wr(dev, irq_map->host_irq_enable, 0);
 
 	intr = mt76_rr(dev, MT_WFDMA0_HOST_INT_STA);
@@ -515,6 +526,11 @@ int mt792x_poll_tx(struct napi_struct *napi, int budget)
 
 	dev = container_of(napi, struct mt792x_dev, mt76.tx_napi);
 
+	if (test_bit(MT76_REMOVED, &dev->mphy.state)) {
+		napi_complete(napi);
+		return 0;
+	}
+
 	if (!mt76_connac_pm_ref(&dev->mphy, &dev->pm)) {
 		napi_complete(napi);
 		queue_work(dev->mt76.wq, &dev->pm.wake_work);
@@ -538,6 +554,11 @@ int mt792x_poll_rx(struct napi_struct *napi, int budget)
 
 	dev = mt76_priv(napi->dev);
 
+	if (test_bit(MT76_REMOVED, &dev->mphy.state)) {
+		napi_complete(napi);
+		return 0;
+	}
+
 	if (!mt76_connac_pm_ref(&dev->mphy, &dev->pm)) {
 		napi_complete(napi);
 		queue_work(dev->mt76.wq, &dev->pm.wake_work);
@@ -620,4 +641,3 @@ int mt792x_wfsys_reset(struct mt792x_dev *dev)
 	return mt792x_wfsys_reset_default(dev);
 }
 EXPORT_SYMBOL_GPL(mt792x_wfsys_reset);
-