| Message ID | 20230420101617.142225-4-krzysztof.kozlowski@linaro.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | ASoC/soundwire: qcom: correctly probe devices after link init | expand |
On 4/20/23 05:16, Krzysztof Kozlowski wrote: > Qualcomm WCD938x Soundwire codecs come as two Soundwire devices - TX > and RX - on two Soundwire buses. In DTS they are represented as three > device nodes: Soundwire TX, Soundwire RX and the platform codec node > (binding to this driver). > > Probing (and Soundwire enumeration) of all devices can happen in any > order, but only the Soundwire TX WCD938x device is used for accessing > actual WCD938x registers. It is possible that component bind() in the > platform driver will be called too early, before the Soundwire TX device > is fully enumerated. This might work or might not, but we cannot handle > it correctly from the codec driver. It's job for Soundwire master to > bring up devices in correct order. That last sentence isn't aligned with the way enumeration works in general for SoundWire. The Manager starts the clock, usually after a bus reset, and waits for Peripherals to signal their presence with Device0 Attached. If multiple Peripherals are attached as Device0, the enumeration will resolve conflicts at the hardware level, and the Manager *cannot* control the order of enumeration; the order is defined by the values in the devID registers, whichever Peripheral has the highest value in the DevID registers wins the enumeration, and others have to back-off and be enumerated later. Probing and enumeration are also different concepts. The SoundWire design allows for drivers to be probed even in the absence of any active hardware. This was added on purpose so that the driver could e.g. program a GPIO or talk to a power-management chip to allow SoundWire devices to start interacting with the bus. see also suggestion below... > Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org> > > --- > > Cc: Patrick Lai <quic_plai@quicinc.com> > --- > sound/soc/codecs/wcd938x.c | 11 +++++++++++ > 1 file changed, 11 insertions(+) > > diff --git a/sound/soc/codecs/wcd938x.c b/sound/soc/codecs/wcd938x.c > index 212667a7249c..e8e07e120fa1 100644 > --- a/sound/soc/codecs/wcd938x.c > +++ b/sound/soc/codecs/wcd938x.c > @@ -77,6 +77,8 @@ > #define WCD938X_MBHC_MOISTURE_RREF R_24_KOHM > #define WCD_MBHC_HS_V_MAX 1600 > > +#define WCD938X_ENUM_TIMEOUT_MS 500 > + > #define WCD938X_EAR_PA_GAIN_TLV(xname, reg, shift, max, invert, tlv_array) \ > { .iface = SNDRV_CTL_ELEM_IFACE_MIXER, .name = xname, \ > .access = SNDRV_CTL_ELEM_ACCESS_TLV_READ |\ > @@ -4425,6 +4427,15 @@ static int wcd938x_bind(struct device *dev) > wcd938x->sdw_priv[AIF1_PB]->slave_irq = wcd938x->virq; > wcd938x->sdw_priv[AIF1_CAP]->slave_irq = wcd938x->virq; > > + /* > + * Before any TX slave regmap usage, be sure the TX slave is actually > + * enumerated. > + */ ... the alternative is to move regmap to be cache-only in the probe and remove the cache-only property when the device is enumerated. That's a trick that's used for all resume cases in codecs in Intel platforms, and we need to extend it for the startup cases as well. > + ret = wait_for_completion_timeout(&wcd938x->tx_sdw_dev->enumeration_complete, > + msecs_to_jiffies(WCD938X_ENUM_TIMEOUT_MS)); > + if (!ret) > + dev_warn(dev, "Enumeration timeout in bind, possible failures in accessing registers\n"); > + > ret = wcd938x_set_micbias_data(wcd938x); > if (ret < 0) { > dev_err(dev, "%s: bad micbias pdata\n", __func__);
On Thu, Apr 20, 2023 at 09:18:15AM -0500, Pierre-Louis Bossart wrote: > On 4/20/23 05:16, Krzysztof Kozlowski wrote: > > + /* > > + * Before any TX slave regmap usage, be sure the TX slave is actually > > + * enumerated. > > + */ > the alternative is to move regmap to be cache-only in the probe and > remove the cache-only property when the device is enumerated. Right, it's generally a better approach unless you need the hardware to actually do something immediately - if you're just setting up what's needed whenver things actually get started then using cache only mode is much less complicated.
On 20/04/2023 16:18, Pierre-Louis Bossart wrote: > > > On 4/20/23 05:16, Krzysztof Kozlowski wrote: >> Qualcomm WCD938x Soundwire codecs come as two Soundwire devices - TX >> and RX - on two Soundwire buses. In DTS they are represented as three >> device nodes: Soundwire TX, Soundwire RX and the platform codec node >> (binding to this driver). >> >> Probing (and Soundwire enumeration) of all devices can happen in any >> order, but only the Soundwire TX WCD938x device is used for accessing >> actual WCD938x registers. It is possible that component bind() in the >> platform driver will be called too early, before the Soundwire TX device >> is fully enumerated. This might work or might not, but we cannot handle >> it correctly from the codec driver. It's job for Soundwire master to >> bring up devices in correct order. > > That last sentence isn't aligned with the way enumeration works in > general for SoundWire. I was rather referring to driver point of view. The Qualcomm Soundwire should work, not expect devices to be powered off during their bind... > > The Manager starts the clock, usually after a bus reset, and waits for > Peripherals to signal their presence with Device0 Attached. > > If multiple Peripherals are attached as Device0, the enumeration will > resolve conflicts at the hardware level, and the Manager *cannot* > control the order of enumeration; the order is defined by the values in > the devID registers, whichever Peripheral has the highest value in the > DevID registers wins the enumeration, and others have to back-off and be > enumerated later. > > Probing and enumeration are also different concepts. The SoundWire > design allows for drivers to be probed even in the absence of any active > hardware. This was added on purpose so that the driver could e.g. > program a GPIO or talk to a power-management chip to allow SoundWire > devices to start interacting with the bus. > > see also suggestion below... > >> Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org> >> >> --- >> >> Cc: Patrick Lai <quic_plai@quicinc.com> >> --- >> sound/soc/codecs/wcd938x.c | 11 +++++++++++ >> 1 file changed, 11 insertions(+) >> >> diff --git a/sound/soc/codecs/wcd938x.c b/sound/soc/codecs/wcd938x.c >> index 212667a7249c..e8e07e120fa1 100644 >> --- a/sound/soc/codecs/wcd938x.c >> +++ b/sound/soc/codecs/wcd938x.c >> @@ -77,6 +77,8 @@ >> #define WCD938X_MBHC_MOISTURE_RREF R_24_KOHM >> #define WCD_MBHC_HS_V_MAX 1600 >> >> +#define WCD938X_ENUM_TIMEOUT_MS 500 >> + >> #define WCD938X_EAR_PA_GAIN_TLV(xname, reg, shift, max, invert, tlv_array) \ >> { .iface = SNDRV_CTL_ELEM_IFACE_MIXER, .name = xname, \ >> .access = SNDRV_CTL_ELEM_ACCESS_TLV_READ |\ >> @@ -4425,6 +4427,15 @@ static int wcd938x_bind(struct device *dev) >> wcd938x->sdw_priv[AIF1_PB]->slave_irq = wcd938x->virq; >> wcd938x->sdw_priv[AIF1_CAP]->slave_irq = wcd938x->virq; >> >> + /* >> + * Before any TX slave regmap usage, be sure the TX slave is actually >> + * enumerated. >> + */ > > ... > > the alternative is to move regmap to be cache-only in the probe and > remove the cache-only property when the device is enumerated. The driver wants already to use the regmap in RW just few lines below in wcd938x_set_micbias_data(). I guess I could move this entire piece of code to other place... > > That's a trick that's used for all resume cases in codecs in Intel > platforms, and we need to extend it for the startup cases as well. Can you point me to some specific piece of driver, so I could see how it is done? It might help me to prepare a better patch for this. > >> + ret = wait_for_completion_timeout(&wcd938x->tx_sdw_dev->enumeration_complete, >> + msecs_to_jiffies(WCD938X_ENUM_TIMEOUT_MS)); >> + if (!ret) >> + dev_warn(dev, "Enumeration timeout in bind, possible failures in accessing registers\n"); >> + >> ret = wcd938x_set_micbias_data(wcd938x); >> if (ret < 0) { >> dev_err(dev, "%s: bad micbias pdata\n", __func__); Best regards, Krzysztof
On Thu, Apr 20, 2023 at 07:57:11PM +0200, Krzysztof Kozlowski wrote: > On 20/04/2023 16:18, Pierre-Louis Bossart wrote: > > the alternative is to move regmap to be cache-only in the probe and > > remove the cache-only property when the device is enumerated. > The driver wants already to use the regmap in RW just few lines below in > wcd938x_set_micbias_data(). > I guess I could move this entire piece of code to other place... The register map is fully writeable (modulo any volatile registers of course) in cache only mode, it's just that hardware access is suppressed. The driver needs to resync the cache when transitioning out of cache only mode (often in a runtime PM resume or something), this will ensure any pending changes make it to the hardware.
diff --git a/sound/soc/codecs/wcd938x.c b/sound/soc/codecs/wcd938x.c index 212667a7249c..e8e07e120fa1 100644 --- a/sound/soc/codecs/wcd938x.c +++ b/sound/soc/codecs/wcd938x.c @@ -77,6 +77,8 @@ #define WCD938X_MBHC_MOISTURE_RREF R_24_KOHM #define WCD_MBHC_HS_V_MAX 1600 +#define WCD938X_ENUM_TIMEOUT_MS 500 + #define WCD938X_EAR_PA_GAIN_TLV(xname, reg, shift, max, invert, tlv_array) \ { .iface = SNDRV_CTL_ELEM_IFACE_MIXER, .name = xname, \ .access = SNDRV_CTL_ELEM_ACCESS_TLV_READ |\ @@ -4425,6 +4427,15 @@ static int wcd938x_bind(struct device *dev) wcd938x->sdw_priv[AIF1_PB]->slave_irq = wcd938x->virq; wcd938x->sdw_priv[AIF1_CAP]->slave_irq = wcd938x->virq; + /* + * Before any TX slave regmap usage, be sure the TX slave is actually + * enumerated. + */ + ret = wait_for_completion_timeout(&wcd938x->tx_sdw_dev->enumeration_complete, + msecs_to_jiffies(WCD938X_ENUM_TIMEOUT_MS)); + if (!ret) + dev_warn(dev, "Enumeration timeout in bind, possible failures in accessing registers\n"); + ret = wcd938x_set_micbias_data(wcd938x); if (ret < 0) { dev_err(dev, "%s: bad micbias pdata\n", __func__);
Qualcomm WCD938x Soundwire codecs come as two Soundwire devices - TX and RX - on two Soundwire buses. In DTS they are represented as three device nodes: Soundwire TX, Soundwire RX and the platform codec node (binding to this driver). Probing (and Soundwire enumeration) of all devices can happen in any order, but only the Soundwire TX WCD938x device is used for accessing actual WCD938x registers. It is possible that component bind() in the platform driver will be called too early, before the Soundwire TX device is fully enumerated. This might work or might not, but we cannot handle it correctly from the codec driver. It's job for Soundwire master to bring up devices in correct order. At least add some simple warning, so such condition will not be undetected. Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org> --- Cc: Patrick Lai <quic_plai@quicinc.com> --- sound/soc/codecs/wcd938x.c | 11 +++++++++++ 1 file changed, 11 insertions(+)