When I tried to fix the watchdog of rt2800, I found that sometimes
the watchdog could not reset the hung device. This is because the
queue did not completely stop, it just became very slow. The Mediatek
vendor driver for the new chips (MT7603/MT7612) has a DMA busy
watchdog to detect device hangs by checking DMA busy status. This
implementation is something similar to it. To reduce unnecessary
watchdog reset, we can check the INT_STATUS register together as I
found that when the radio hung, the RX/TX coherent interrupt will
always stuck at triggered state.
This patch also changes the watchdog module parameters to the new
'hang_watchdog' and 'dma_busy_watchdog' so that we can control them
separately. That's because they may have different behavior on
specific chip.
This watchdog function is a slight schedule and it won't affect the
WiFi transmission speed. Watchdog can help the driver automatically
recover from the abnormal state. So I think it should be default on.
Anyway it can be disabled by module parameter 'dma_busy_watchdog=0'.
Tested on MT7620 and RT5350.
Signed-off-by: Shiji Yang <[email protected]>
---
.../net/wireless/ralink/rt2x00/rt2800lib.c | 81 ++++++++++++++++---
drivers/net/wireless/ralink/rt2x00/rt2x00.h | 3 +
2 files changed, 72 insertions(+), 12 deletions(-)
diff --git a/drivers/net/wireless/ralink/rt2x00/rt2800lib.c b/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
index 594dd3d9f..6ca2f2c23 100644
--- a/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
+++ b/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
@@ -30,9 +30,15 @@
#include "rt2800lib.h"
#include "rt2800.h"
-static bool modparam_watchdog;
-module_param_named(watchdog, modparam_watchdog, bool, S_IRUGO);
-MODULE_PARM_DESC(watchdog, "Enable watchdog to detect tx/rx hangs and reset hardware if detected");
+static bool modparam_dma_wdt = true;
+module_param_named(dma_busy_watchdog, modparam_dma_wdt, bool, 0444);
+MODULE_PARM_DESC(dma_busy_watchdog, "Enable watchdog to detect tx/rx"
+ " DMA busy and reset hardware if detected");
+
+static bool modparam_hang_wdt;
+module_param_named(hang_watchdog, modparam_hang_wdt, bool, 0444);
+MODULE_PARM_DESC(hang_watchdog, "Enable watchdog to detect tx/rx hangs"
+ " and reset hardware if detected");
/*
* Register access.
@@ -1260,15 +1266,12 @@ static void rt2800_update_survey(struct rt2x00_dev *rt2x00dev)
chan_survey->time_ext_busy += rt2800_register_read(rt2x00dev, CH_BUSY_STA_SEC);
}
-void rt2800_watchdog(struct rt2x00_dev *rt2x00dev)
+static bool rt2800_watchdog_hung(struct rt2x00_dev *rt2x00dev)
{
struct data_queue *queue;
bool hung_tx = false;
bool hung_rx = false;
- if (test_bit(DEVICE_STATE_SCANNING, &rt2x00dev->flags))
- return;
-
rt2800_update_survey(rt2x00dev);
queue_for_each(rt2x00dev, queue) {
@@ -1296,18 +1299,72 @@ void rt2800_watchdog(struct rt2x00_dev *rt2x00dev)
}
}
+ if (!hung_tx && !hung_rx)
+ return false;
+
if (hung_tx)
rt2x00_warn(rt2x00dev, "Watchdog TX hung detected\n");
if (hung_rx)
rt2x00_warn(rt2x00dev, "Watchdog RX hung detected\n");
- if (hung_tx || hung_rx) {
- queue_for_each(rt2x00dev, queue)
- queue->wd_count = 0;
+ queue_for_each(rt2x00dev, queue)
+ queue->wd_count = 0;
+
+ return true;
+}
+
+static bool rt2800_watchdog_dma_busy(struct rt2x00_dev *rt2x00dev)
+{
+ bool busy_rx, busy_tx;
+ u32 reg_cfg = rt2800_register_read(rt2x00dev, WPDMA_GLO_CFG);
+ u32 reg_int = rt2800_register_read(rt2x00dev, INT_SOURCE_CSR);
+
+ if (rt2x00_get_field32(reg_cfg, WPDMA_GLO_CFG_RX_DMA_BUSY) &&
+ rt2x00_get_field32(reg_int, INT_SOURCE_CSR_RX_COHERENT))
+ rt2x00dev->rxdma_busy++;
+ else
+ rt2x00dev->rxdma_busy = 0;
+
+ if (rt2x00_get_field32(reg_cfg, WPDMA_GLO_CFG_TX_DMA_BUSY) &&
+ rt2x00_get_field32(reg_int, INT_SOURCE_CSR_TX_COHERENT))
+ rt2x00dev->txdma_busy++;
+ else
+ rt2x00dev->txdma_busy = 0;
+
+ busy_rx = rt2x00dev->rxdma_busy > 30 ? true : false;
+ busy_tx = rt2x00dev->txdma_busy > 30 ? true : false;
+ if (!busy_rx && !busy_tx)
+ return false;
+
+ if (busy_rx)
+ rt2x00_warn(rt2x00dev, "Watchdog RX DMA busy detected\n");
+
+ if (busy_tx)
+ rt2x00_warn(rt2x00dev, "Watchdog TX DMA busy detected\n");
+
+ rt2x00dev->rxdma_busy = 0;
+ rt2x00dev->txdma_busy = 0;
+
+ return true;
+}
+
+void rt2800_watchdog(struct rt2x00_dev *rt2x00dev)
+{
+ bool reset = false;
+
+ if (test_bit(DEVICE_STATE_SCANNING, &rt2x00dev->flags))
+ return;
+
+ if (modparam_dma_wdt)
+ reset = rt2800_watchdog_dma_busy(rt2x00dev);
+
+ if (modparam_hang_wdt)
+ reset = rt2800_watchdog_hung(rt2x00dev) || reset;
+
+ if (reset)
ieee80211_restart_hw(rt2x00dev->hw);
- }
}
EXPORT_SYMBOL_GPL(rt2800_watchdog);
@@ -12015,7 +12072,7 @@ int rt2800_probe_hw(struct rt2x00_dev *rt2x00dev)
__set_bit(REQUIRE_TASKLET_CONTEXT, &rt2x00dev->cap_flags);
}
- if (modparam_watchdog) {
+ if (modparam_hang_wdt || modparam_dma_wdt) {
__set_bit(CAPABILITY_RESTART_HW, &rt2x00dev->cap_flags);
rt2x00dev->link.watchdog_interval = msecs_to_jiffies(100);
} else {
diff --git a/drivers/net/wireless/ralink/rt2x00/rt2x00.h b/drivers/net/wireless/ralink/rt2x00/rt2x00.h
index aaaf99331..62fed38f4 100644
--- a/drivers/net/wireless/ralink/rt2x00/rt2x00.h
+++ b/drivers/net/wireless/ralink/rt2x00/rt2x00.h
@@ -926,6 +926,9 @@ struct rt2x00_dev {
*/
u16 beacon_int;
+ /* Rx/Tx DMA busy watchdog counter */
+ u16 rxdma_busy, txdma_busy;
+
/**
* Timestamp of last received beacon
*/
--
2.39.2
On Sat, Oct 28, 2023 at 08:15:30PM +0800, Shiji Yang wrote:
> When I tried to fix the watchdog of rt2800, I found that sometimes
> the watchdog could not reset the hung device. This is because the
> queue did not completely stop, it just became very slow. The Mediatek
> vendor driver for the new chips (MT7603/MT7612) has a DMA busy
> watchdog to detect device hangs by checking DMA busy status. This
> implementation is something similar to it. To reduce unnecessary
> watchdog reset, we can check the INT_STATUS register together as I
> found that when the radio hung, the RX/TX coherent interrupt will
> always stuck at triggered state.
>
> This patch also changes the watchdog module parameters to the new
> 'hang_watchdog' and 'dma_busy_watchdog' so that we can control them
> separately. That's because they may have different behavior on
> specific chip.
>
> This watchdog function is a slight schedule and it won't affect the
> WiFi transmission speed. Watchdog can help the driver automatically
> recover from the abnormal state. So I think it should be default on.
> Anyway it can be disabled by module parameter 'dma_busy_watchdog=0'.
>
> Tested on MT7620 and RT5350.
I think this will not work on USB as INT_SOURCE_CSR is mmio/pci
specific. Did you tested on USB ? Or this is disabled for USB by
default?
> Signed-off-by: Shiji Yang <[email protected]>
> ---
> .../net/wireless/ralink/rt2x00/rt2800lib.c | 81 ++++++++++++++++---
> drivers/net/wireless/ralink/rt2x00/rt2x00.h | 3 +
> 2 files changed, 72 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/net/wireless/ralink/rt2x00/rt2800lib.c b/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
> index 594dd3d9f..6ca2f2c23 100644
> --- a/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
> +++ b/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
> @@ -30,9 +30,15 @@
> #include "rt2800lib.h"
> #include "rt2800.h"
>
> -static bool modparam_watchdog;
> -module_param_named(watchdog, modparam_watchdog, bool, S_IRUGO);
> -MODULE_PARM_DESC(watchdog, "Enable watchdog to detect tx/rx hangs and reset hardware if detected");
> +static bool modparam_dma_wdt = true;
> +module_param_named(dma_busy_watchdog, modparam_dma_wdt, bool, 0444);
> +MODULE_PARM_DESC(dma_busy_watchdog, "Enable watchdog to detect tx/rx"
> + " DMA busy and reset hardware if detected");
> +
> +static bool modparam_hang_wdt;
> +module_param_named(hang_watchdog, modparam_hang_wdt, bool, 0444);
> +MODULE_PARM_DESC(hang_watchdog, "Enable watchdog to detect tx/rx hangs"
> + " and reset hardware if detected");
Do not have strong opinion here. But please consider to keep old module
parameter name and make it bitmask, 1 - hang_wdt, 2 dma_wdt, 3 - both.
Such way, it will keep old meaning if someone is using the parameter in
script/config.
I also wandering if we need two implementations. If dma
hang detection is superior, it should replace the old one IMHO.
Or queue hang should be used for USB and dma hang for pci/mmio ?
Otherwise code looks fine/correct to me.
Regards
Stanislaw
On Wed, 1 Nov 2023 16:56:55 +0100, Stanislaw Gruszka wrote:
>On Sat, Oct 28, 2023 at 08:15:30PM +0800, Shiji Yang wrote:
>> When I tried to fix the watchdog of rt2800, I found that sometimes
>> the watchdog could not reset the hung device. This is because the
>> queue did not completely stop, it just became very slow. The Mediatek
>> vendor driver for the new chips (MT7603/MT7612) has a DMA busy
>> watchdog to detect device hangs by checking DMA busy status. This
>> implementation is something similar to it. To reduce unnecessary
>> watchdog reset, we can check the INT_STATUS register together as I
>> found that when the radio hung, the RX/TX coherent interrupt will
>> always stuck at triggered state.
>>
>> This patch also changes the watchdog module parameters to the new
>> 'hang_watchdog' and 'dma_busy_watchdog' so that we can control them
>> separately. That's because they may have different behavior on
>> specific chip.
>>
>> This watchdog function is a slight schedule and it won't affect the
>> WiFi transmission speed. Watchdog can help the driver automatically
>> recover from the abnormal state. So I think it should be default on.
>> Anyway it can be disabled by module parameter 'dma_busy_watchdog=0'.
>>
>> Tested on MT7620 and RT5350.
>
>I think this will not work on USB as INT_SOURCE_CSR is mmio/pci
>specific. Did you tested on USB ? Or this is disabled for USB by
>default?
Hi! Thanks for your information. I didn't realize that they have such
difference. I don't have Ralink USB NIC so I only test it on RT5350
and MT7620. I'll disable it for USB devices in v2.
>
>> Signed-off-by: Shiji Yang <[email protected]>
>> ---
>> .../net/wireless/ralink/rt2x00/rt2800lib.c | 81 ++++++++++++++++---
>> drivers/net/wireless/ralink/rt2x00/rt2x00.h | 3 +
>> 2 files changed, 72 insertions(+), 12 deletions(-)
>>
>> diff --git a/drivers/net/wireless/ralink/rt2x00/rt2800lib.c b/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
>> index 594dd3d9f..6ca2f2c23 100644
>> --- a/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
>> +++ b/drivers/net/wireless/ralink/rt2x00/rt2800lib.c
>> @@ -30,9 +30,15 @@
>> #include "rt2800lib.h"
>> #include "rt2800.h"
>>
>> -static bool modparam_watchdog;
>> -module_param_named(watchdog, modparam_watchdog, bool, S_IRUGO);
>> -MODULE_PARM_DESC(watchdog, "Enable watchdog to detect tx/rx hangs and reset hardware if detected");
>> +static bool modparam_dma_wdt = true;
>> +module_param_named(dma_busy_watchdog, modparam_dma_wdt, bool, 0444);
>> +MODULE_PARM_DESC(dma_busy_watchdog, "Enable watchdog to detect tx/rx"
>> + " DMA busy and reset hardware if detected");
>> +
>> +static bool modparam_hang_wdt;
>> +module_param_named(hang_watchdog, modparam_hang_wdt, bool, 0444);
>> +MODULE_PARM_DESC(hang_watchdog, "Enable watchdog to detect tx/rx hangs"
>> + " and reset hardware if detected");
>
>Do not have strong opinion here. But please consider to keep old module
>parameter name and make it bitmask, 1 - hang_wdt, 2 dma_wdt, 3 - both.
>Such way, it will keep old meaning if someone is using the parameter in
>script/config.
I'll update them in v2. Thanks.
>
>I also wandering if we need two implementations. If dma
>hang detection is superior, it should replace the old one IMHO.
>Or queue hang should be used for USB and dma hang for pci/mmio ?
Since the DMA busy watchdog will be disabled for USB NICs in v2,
so I think we still need both of them. As for queue hang issue,
To be honest, I never really encounter it on MT7620 in past half
year. Enabling queue hang watchdog just report some wrong resets
for me even when I increased the threshold from 16 to 64. And the
large threshold will make it unable to catch the DMA hung issue.
Anyway, it should be still valuable for USB device.
>
>Otherwise code looks fine/correct to me.
>
>Regards
>Stanislaw