[PATCH AUTOSEL 5.4 031/130] mwifiex: pcie: skip cancel_work_sync() on reset failure path

From: Sasha Levin
Date: Tue Dec 22 2020 - 22:04:15 EST


From: Tsuchiya Yuto <kitakar@xxxxxxxxx>

[ Upstream commit 4add4d988f95f47493500a7a19c623827061589b ]

If a reset is performed, but even the reset fails for some reasons (e.g.,
on Surface devices, the fw reset requires another quirks),
cancel_work_sync() hangs in mwifiex_cleanup_pcie().

# firmware went into a bad state
[...]
[ 1608.281690] mwifiex_pcie 0000:03:00.0: info: shutdown mwifiex...
[ 1608.282724] mwifiex_pcie 0000:03:00.0: rx_pending=0, tx_pending=1, cmd_pending=0
[ 1608.292400] mwifiex_pcie 0000:03:00.0: PREP_CMD: card is removed
[ 1608.292405] mwifiex_pcie 0000:03:00.0: PREP_CMD: card is removed
# reset performed after firmware went into a bad state
[ 1609.394320] mwifiex_pcie 0000:03:00.0: WLAN FW already running! Skip FW dnld
[ 1609.394335] mwifiex_pcie 0000:03:00.0: WLAN FW is active
# but even the reset failed
[ 1619.499049] mwifiex_pcie 0000:03:00.0: mwifiex_cmd_timeout_func: Timeout cmd id = 0xfa, act = 0xe000
[ 1619.499094] mwifiex_pcie 0000:03:00.0: num_data_h2c_failure = 0
[ 1619.499103] mwifiex_pcie 0000:03:00.0: num_cmd_h2c_failure = 0
[ 1619.499110] mwifiex_pcie 0000:03:00.0: is_cmd_timedout = 1
[ 1619.499117] mwifiex_pcie 0000:03:00.0: num_tx_timeout = 0
[ 1619.499124] mwifiex_pcie 0000:03:00.0: last_cmd_index = 0
[ 1619.499133] mwifiex_pcie 0000:03:00.0: last_cmd_id: fa 00 07 01 07 01 07 01 07 01
[ 1619.499140] mwifiex_pcie 0000:03:00.0: last_cmd_act: 00 e0 00 00 00 00 00 00 00 00
[ 1619.499147] mwifiex_pcie 0000:03:00.0: last_cmd_resp_index = 3
[ 1619.499155] mwifiex_pcie 0000:03:00.0: last_cmd_resp_id: 07 81 07 81 07 81 07 81 07 81
[ 1619.499162] mwifiex_pcie 0000:03:00.0: last_event_index = 2
[ 1619.499169] mwifiex_pcie 0000:03:00.0: last_event: 58 00 58 00 58 00 58 00 58 00
[ 1619.499177] mwifiex_pcie 0000:03:00.0: data_sent=0 cmd_sent=1
[ 1619.499185] mwifiex_pcie 0000:03:00.0: ps_mode=0 ps_state=0
[ 1619.499215] mwifiex_pcie 0000:03:00.0: info: _mwifiex_fw_dpc: unregister device
# mwifiex_pcie_work hang happening
[ 1823.233923] INFO: task kworker/3:1:44 blocked for more than 122 seconds.
[ 1823.233932] Tainted: G WC OE 5.10.0-rc1-1-mainline #1
[ 1823.233935] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 1823.233940] task:kworker/3:1 state:D stack: 0 pid: 44 ppid: 2 flags:0x00004000
[ 1823.233960] Workqueue: events mwifiex_pcie_work [mwifiex_pcie]
[ 1823.233965] Call Trace:
[ 1823.233981] __schedule+0x292/0x820
[ 1823.233990] schedule+0x45/0xe0
[ 1823.233995] schedule_timeout+0x11c/0x160
[ 1823.234003] wait_for_completion+0x9e/0x100
[ 1823.234012] __flush_work.isra.0+0x156/0x210
[ 1823.234018] ? flush_workqueue_prep_pwqs+0x130/0x130
[ 1823.234026] __cancel_work_timer+0x11e/0x1a0
[ 1823.234035] mwifiex_cleanup_pcie+0x28/0xd0 [mwifiex_pcie]
[ 1823.234049] mwifiex_free_adapter+0x24/0xe0 [mwifiex]
[ 1823.234060] _mwifiex_fw_dpc+0x294/0x560 [mwifiex]
[ 1823.234074] mwifiex_reinit_sw+0x15d/0x300 [mwifiex]
[ 1823.234080] mwifiex_pcie_reset_done+0x50/0x80 [mwifiex_pcie]
[ 1823.234087] pci_try_reset_function+0x5c/0x90
[ 1823.234094] process_one_work+0x1d6/0x3a0
[ 1823.234100] worker_thread+0x4d/0x3d0
[ 1823.234107] ? rescuer_thread+0x410/0x410
[ 1823.234112] kthread+0x142/0x160
[ 1823.234117] ? __kthread_bind_mask+0x60/0x60
[ 1823.234124] ret_from_fork+0x22/0x30
[...]

This is a deadlock caused by calling cancel_work_sync() in
mwifiex_cleanup_pcie():

- Device resets are done via mwifiex_pcie_card_reset()
- which schedules card->work to call mwifiex_pcie_card_reset_work()
- which calls pci_try_reset_function().
- This leads to mwifiex_pcie_reset_done() be called on the same workqueue,
which in turn calls
- mwifiex_reinit_sw() and that calls
- _mwifiex_fw_dpc().

The problem is now that _mwifiex_fw_dpc() calls mwifiex_free_adapter()
in case firmware initialization fails. That ends up calling
mwifiex_cleanup_pcie().

Note that all those calls are still running on the workqueue. So when
mwifiex_cleanup_pcie() now calls cancel_work_sync(), it's really waiting
on itself to complete, causing a deadlock.

This commit fixes the deadlock by skipping cancel_work_sync() on a reset
failure path.

After this commit, when reset fails, the following output is
expected to be shown:

kernel: mwifiex_pcie 0000:03:00.0: info: _mwifiex_fw_dpc: unregister device
kernel: mwifiex: Failed to bring up adapter: -5
kernel: mwifiex_pcie 0000:03:00.0: reinit failed: -5

To reproduce this issue, for example, try putting the root port of wifi
into D3 (replace "00:1d.3" with your setup).

# put into D3 (root port)
sudo setpci -v -s 00:1d.3 CAP_PM+4.b=0b

Cc: Maximilian Luz <luzmaximilian@xxxxxxxxx>
Signed-off-by: Tsuchiya Yuto <kitakar@xxxxxxxxx>
Signed-off-by: Kalle Valo <kvalo@xxxxxxxxxxxxxx>
Link: https://lore.kernel.org/r/20201028142346.18355-1-kitakar@xxxxxxxxx
Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>
---
drivers/net/wireless/marvell/mwifiex/pcie.c | 18 +++++++++++++++++-
drivers/net/wireless/marvell/mwifiex/pcie.h | 2 ++
2 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/drivers/net/wireless/marvell/mwifiex/pcie.c b/drivers/net/wireless/marvell/mwifiex/pcie.c
index fc1706d0647d7..58c9623c3a916 100644
--- a/drivers/net/wireless/marvell/mwifiex/pcie.c
+++ b/drivers/net/wireless/marvell/mwifiex/pcie.c
@@ -377,6 +377,8 @@ static void mwifiex_pcie_reset_prepare(struct pci_dev *pdev)
clear_bit(MWIFIEX_IFACE_WORK_DEVICE_DUMP, &card->work_flags);
clear_bit(MWIFIEX_IFACE_WORK_CARD_RESET, &card->work_flags);
mwifiex_dbg(adapter, INFO, "%s, successful\n", __func__);
+
+ card->pci_reset_ongoing = true;
}

/*
@@ -405,6 +407,8 @@ static void mwifiex_pcie_reset_done(struct pci_dev *pdev)
dev_err(&pdev->dev, "reinit failed: %d\n", ret);
else
mwifiex_dbg(adapter, INFO, "%s, successful\n", __func__);
+
+ card->pci_reset_ongoing = false;
}

static const struct pci_error_handlers mwifiex_pcie_err_handler = {
@@ -2995,7 +2999,19 @@ static void mwifiex_cleanup_pcie(struct mwifiex_adapter *adapter)
int ret;
u32 fw_status;

- cancel_work_sync(&card->work);
+ /* Perform the cancel_work_sync() only when we're not resetting
+ * the card. It's because that function never returns if we're
+ * in reset path. If we're here when resetting the card, it means
+ * that we failed to reset the card (reset failure path).
+ */
+ if (!card->pci_reset_ongoing) {
+ mwifiex_dbg(adapter, MSG, "performing cancel_work_sync()...\n");
+ cancel_work_sync(&card->work);
+ mwifiex_dbg(adapter, MSG, "cancel_work_sync() done\n");
+ } else {
+ mwifiex_dbg(adapter, MSG,
+ "skipped cancel_work_sync() because we're in card reset failure path\n");
+ }

ret = mwifiex_read_reg(adapter, reg->fw_status, &fw_status);
if (fw_status == FIRMWARE_READY_PCIE) {
diff --git a/drivers/net/wireless/marvell/mwifiex/pcie.h b/drivers/net/wireless/marvell/mwifiex/pcie.h
index f7ce9b6db6b41..72d0c01ff3592 100644
--- a/drivers/net/wireless/marvell/mwifiex/pcie.h
+++ b/drivers/net/wireless/marvell/mwifiex/pcie.h
@@ -391,6 +391,8 @@ struct pcie_service_card {
struct mwifiex_msix_context share_irq_ctx;
struct work_struct work;
unsigned long work_flags;
+
+ bool pci_reset_ongoing;
};

static inline int
--
2.27.0