Message-ID: <2b086263-6b42-b882-6d38-1c4dbc361773@linaro.org>
Date:   Mon, 6 Mar 2023 13:09:35 +0100
From:   Krzysztof Kozlowski <krzysztof.kozlowski@...aro.org>
To:     Konrad Dybcio <konrad.dybcio@...aro.org>,
        Andy Gross <agross@...nel.org>,
        Bjorn Andersson <andersson@...nel.org>,
        Georgi Djakov <djakov@...nel.org>,
        Rob Herring <robh+dt@...nel.org>,
        Thara Gopinath <thara.gopinath@...il.com>
Cc:     linux-arm-msm@...r.kernel.org, linux-pm@...r.kernel.org,
        devicetree@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 2/3] soc: qcom: icc-bwmon: Handle global registers
 correctly

On 06/03/2023 11:06, Konrad Dybcio wrote:
> 
> 
> On 5.03.2023 16:06, Krzysztof Kozlowski wrote:
>> On 04/03/2023 16:39, Konrad Dybcio wrote:
>>> The BWMON hardware has two sets of registers: one for the monitor itself
>>> and one called "global". It has what seems to be some kind of a head
>>> switch and an interrupt control register. It's usually 0x200 in size.
>>>
>>> On fairly recent SoCs (with the starting point seemingly being moving
>>> the OSM programming to the firmware) these two register sets are
>>> contiguous and overlapping, like this (on sm8450):
>>>
>>> /* notice how base.start == global_base.start+0x100 */
>>> reg = <0x90b6400 0x300>, <0x90b6300 0x200>;
>>> reg-names = "base", "global_base";
>>>
>>> Which led to some confusion and the assumption that since the
>>> "interesting" global registers begin right after global_base+0x100,
>>> there's no need to map two separate regions and one can simply subtract
>>> 0x100 from the offsets.
>>>
>>> This is however not the case for anything older than SDM845, as the
>>> global region can appear in seemingly random spots on the register map.
>>>
>>> Add support for it to let bwmon function on older SoCs like MSM8998 and
>>> allow operation with just one set of registers for newer platforms.
>>>
>>> Fixes: b9c2ae6cac40 ("soc: qcom: icc-bwmon: Add bandwidth monitoring driver")
>>
>> You did not describe any bug to be fixed. Adding support for different
>> devices with different memory layour is a feature, not bugfix. If this
>> is a bugfix, please share what exactly is broken on sdm845?
> The driver was completely broken for MSM8998, even though it
> had a compatible specifically for that SoC.

That's still nothing to fix - msm8998 was added here only for
compatibility reasons for bindings. It wasn't ever tested on MSM8998 and
also probably never written in a way that it should work for MSM8998. It
could work, but that was not the intention. The driver is supposed to
work on sdm845 and according to your description - there is nothing
wrong with that.

> 
>>
>>> Signed-off-by: Konrad Dybcio <konrad.dybcio@...aro.org>
>>> ---
>>>  drivers/soc/qcom/icc-bwmon.c | 136 +++++++++++++++++++++++++++++++++++++------
>>>  1 file changed, 118 insertions(+), 18 deletions(-)
>>>
>>> diff --git a/drivers/soc/qcom/icc-bwmon.c b/drivers/soc/qcom/icc-bwmon.c
>>> index d07be3700db6..9ef632d80ee3 100644
>>> --- a/drivers/soc/qcom/icc-bwmon.c
>>> +++ b/drivers/soc/qcom/icc-bwmon.c
>>> @@ -34,14 +34,27 @@
>>>  /* Internal sampling clock frequency */
>>>  #define HW_TIMER_HZ				19200000
>>>  
>>> -#define BWMON_V4_GLOBAL_IRQ_CLEAR		0x008
>>> -#define BWMON_V4_GLOBAL_IRQ_ENABLE		0x00c
>>> +#define BWMON_V4_GLOBAL_IRQ_CLEAR		0x108
>>> +#define BWMON_V4_GLOBAL_IRQ_ENABLE		0x10c
>>>  /*
>>>   * All values here and further are matching regmap fields, so without absolute
>>>   * register offsets.
>>>   */
>>>  #define BWMON_V4_GLOBAL_IRQ_ENABLE_ENABLE	BIT(0)
>>>  
>>> +/*
>>> + * Starting with SDM845, the BWMON4 register space has changed a bit:
>>> + * the global registers were jammed into the beginning of the monitor region.
>>> + * To keep the proper offsets, one would have to map <GLOBAL_BASE 0x200> and
>>> + * <GLOBAL_BASE+0x100 0x300>, which is straight up wrong.
>>> + * To facilitate for that, while allowing the older, arguably more proper
>>> + * implementations to work, offset the global registers by -0x100 to avoid
>>> + * having to map half of the global registers twice.
>>> + */
>>> +#define BWMON_V4_845_OFFSET			0x100
>>
>> MSM8998? It's a bit confusing to keep calling it 845 while it is for
>> MSM8998 variant... or it's not anymore for MSM8998?
> No, it's for 845 and up.

If it is for 845 and up, why then it is used for msm8998 in
msm8998_bwmon_reg_noread_ranges?

> 
>>
>>
>>> +#define BWMON_V4_GLOBAL_IRQ_CLEAR_845		(BWMON_V4_GLOBAL_IRQ_CLEAR - BWMON_V4_845_OFFSET)
>>> +#define BWMON_V4_GLOBAL_IRQ_ENABLE_845		(BWMON_V4_GLOBAL_IRQ_ENABLE - BWMON_V4_845_OFFSET)
>>> +
>>>  #define BWMON_V4_IRQ_STATUS			0x100
>>>  #define BWMON_V4_IRQ_CLEAR			0x108
>>>  
>>> @@ -118,8 +131,10 @@
>>>  #define BWMON_NEEDS_FORCE_CLEAR			BIT(1)
>>>  
>>>  enum bwmon_fields {
>>> -	F_GLOBAL_IRQ_CLEAR,
>>> -	F_GLOBAL_IRQ_ENABLE,
>>> +	/* Fields used only on >=SDM845 with BWMON_HAS_GLOBAL_IRQ */
>>> +	F_GLB_IRQ_CLEAR,
>>> +	F_GLB_IRQ_ENABLE,
>>
>> I am not sure what's the benefit of this rename.
> I need to somehow differentiate the members of enum bwmon_fields
> and enum bwmon_global_fields, keeping the "_GLOBAL" in global_fields
> seems to make the most sense, imo.

This complicates and makes code difficult to read. The entire driver was
designed for single address space to make everything a bit simpler and
easier to read, so any changes should not reduce this. Maybe just use
the same defines? Or name the other fields MSM8998?

> 
>>
>>> +
>>>  	F_IRQ_STATUS,
>>>  	F_IRQ_CLEAR,
>>>  	F_IRQ_ENABLE,
>>> @@ -145,6 +160,13 @@ enum bwmon_fields {
>>>  	F_NUM_FIELDS
>>>  };
>>>  
>>> +enum bwmon_global_fields {
>>> +	F_GLOBAL_IRQ_CLEAR,
>>> +	F_GLOBAL_IRQ_ENABLE,
>>> +
>>> +	F_NUM_GLOBAL_FIELDS
>>> +};
>>> +
>>>  struct icc_bwmon_data {
>>>  	unsigned int sample_ms;
>>>  	unsigned int count_unit_kb; /* kbytes */
>>> @@ -157,6 +179,9 @@ struct icc_bwmon_data {
>>>  
>>>  	const struct regmap_config *regmap_cfg;
>>>  	const struct reg_field *regmap_fields;
>>> +
>>> +	const struct regmap_config *global_regmap_cfg;
>>> +	const struct reg_field *global_regmap_fields;
>>>  };
>>>  
>>>  struct icc_bwmon {
>>> @@ -166,6 +191,7 @@ struct icc_bwmon {
>>>  
>>>  	struct regmap *regmap;
>>>  	struct regmap_field *regs[F_NUM_FIELDS];
>>> +	struct regmap_field *global_regs[F_NUM_FIELDS];
>>>  
>>>  	unsigned int max_bw_kbps;
>>>  	unsigned int min_bw_kbps;
>>> @@ -175,8 +201,8 @@ struct icc_bwmon {
>>>  
>>>  /* BWMON v4 */
>>>  static const struct reg_field msm8998_bwmon_reg_fields[] = {
>>> -	[F_GLOBAL_IRQ_CLEAR]	= REG_FIELD(BWMON_V4_GLOBAL_IRQ_CLEAR, 0, 0),
>>> -	[F_GLOBAL_IRQ_ENABLE]	= REG_FIELD(BWMON_V4_GLOBAL_IRQ_ENABLE, 0, 0),
>>> +	[F_GLB_IRQ_CLEAR]	= REG_FIELD(BWMON_V4_GLOBAL_IRQ_CLEAR_845, 0, 0),
>>> +	[F_GLB_IRQ_ENABLE]	= REG_FIELD(BWMON_V4_GLOBAL_IRQ_ENABLE_845, 0, 0),
>>>  	[F_IRQ_STATUS]		= REG_FIELD(BWMON_V4_IRQ_STATUS, 4, 7),
>>>  	[F_IRQ_CLEAR]		= REG_FIELD(BWMON_V4_IRQ_CLEAR, 4, 7),
>>>  	[F_IRQ_ENABLE]		= REG_FIELD(BWMON_V4_IRQ_ENABLE, 4, 7),
>>> @@ -202,7 +228,7 @@ static const struct reg_field msm8998_bwmon_reg_fields[] = {
>>>  };
>>>  
>>>  static const struct regmap_range msm8998_bwmon_reg_noread_ranges[] = {
>>> -	regmap_reg_range(BWMON_V4_GLOBAL_IRQ_CLEAR, BWMON_V4_GLOBAL_IRQ_CLEAR),
>>> +	regmap_reg_range(BWMON_V4_GLOBAL_IRQ_CLEAR_845, BWMON_V4_GLOBAL_IRQ_CLEAR_845),
>>>  	regmap_reg_range(BWMON_V4_IRQ_CLEAR, BWMON_V4_IRQ_CLEAR),
>>>  	regmap_reg_range(BWMON_V4_CLEAR, BWMON_V4_CLEAR),
>>>  };
>>> @@ -222,16 +248,34 @@ static const struct regmap_access_table msm8998_bwmon_reg_volatile_table = {
>>>  	.n_yes_ranges	= ARRAY_SIZE(msm8998_bwmon_reg_volatile_ranges),
>>>  };
>>>  
>>> +static const struct reg_field msm8998_bwmon_global_reg_fields[] = {
>>> +	[F_GLOBAL_IRQ_CLEAR]	= REG_FIELD(BWMON_V4_GLOBAL_IRQ_CLEAR, 0, 0),
>>> +	[F_GLOBAL_IRQ_ENABLE]	= REG_FIELD(BWMON_V4_GLOBAL_IRQ_ENABLE, 0, 0),
>>> +};
>>> +
>>> +static const struct regmap_range msm8998_bwmon_global_reg_noread_ranges[] = {
>>> +	regmap_reg_range(BWMON_V4_GLOBAL_IRQ_CLEAR, BWMON_V4_GLOBAL_IRQ_CLEAR),
>>> +};
>>> +
>>> +static const struct regmap_access_table msm8998_bwmon_global_reg_read_table = {
>>> +	.no_ranges	= msm8998_bwmon_global_reg_noread_ranges,
>>> +	.n_no_ranges	= ARRAY_SIZE(msm8998_bwmon_global_reg_noread_ranges),
>>> +};
>>> +
>>>  /*
>>>   * Fill the cache for non-readable registers only as rest does not really
>>>   * matter and can be read from the device.
>>>   */
>>>  static const struct reg_default msm8998_bwmon_reg_defaults[] = {
>>> -	{ BWMON_V4_GLOBAL_IRQ_CLEAR, 0x0 },
>>> +	{ BWMON_V4_GLOBAL_IRQ_CLEAR_845, 0x0 },
>>>  	{ BWMON_V4_IRQ_CLEAR, 0x0 },
>>>  	{ BWMON_V4_CLEAR, 0x0 },
>>>  };
>>>  
>>> +static const struct reg_default msm8998_bwmon_global_reg_defaults[] = {
>>> +	{ BWMON_V4_GLOBAL_IRQ_CLEAR, 0x0 },
>>> +};
>>> +
>>>  static const struct regmap_config msm8998_bwmon_regmap_cfg = {
>>>  	.reg_bits		= 32,
>>>  	.reg_stride		= 4,
>>> @@ -252,10 +296,27 @@ static const struct regmap_config msm8998_bwmon_regmap_cfg = {
>>>  	.cache_type		= REGCACHE_RBTREE,
>>>  };
>>>  
>>> +static const struct regmap_config msm8998_bwmon_global_regmap_cfg = {
>>> +	.reg_bits		= 32,
>>> +	.reg_stride		= 4,
>>> +	.val_bits		= 32,
>>> +	/*
>>> +	 * No concurrent access expected - driver has one interrupt handler,
>>> +	 * regmap is not shared, no driver or user-space API.
>>> +	 */
>>> +	.disable_locking	= true,
>>> +	.rd_table		= &msm8998_bwmon_global_reg_read_table,
>>> +	.reg_defaults		= msm8998_bwmon_global_reg_defaults,
>>> +	.num_reg_defaults	= ARRAY_SIZE(msm8998_bwmon_global_reg_defaults),
>>> +	/*
>>> +	 * Cache is necessary for using regmap fields with non-readable
>>> +	 * registers.
>>> +	 */
>>> +	.cache_type		= REGCACHE_RBTREE,
>>> +};
>>> +
>>>  /* BWMON v5 */
>>>  static const struct reg_field sdm845_llcc_bwmon_reg_fields[] = {
>>> -	[F_GLOBAL_IRQ_CLEAR]	= {},
>>> -	[F_GLOBAL_IRQ_ENABLE]	= {},
>>>  	[F_IRQ_STATUS]		= REG_FIELD(BWMON_V5_IRQ_STATUS, 0, 3),
>>>  	[F_IRQ_CLEAR]		= REG_FIELD(BWMON_V5_IRQ_CLEAR, 0, 3),
>>>  	[F_IRQ_ENABLE]		= REG_FIELD(BWMON_V5_IRQ_ENABLE, 0, 3),
>>> @@ -369,16 +430,21 @@ static void bwmon_clear_irq(struct icc_bwmon *bwmon)
>>>  	regmap_field_force_write(bwmon->regs[F_IRQ_CLEAR], BWMON_IRQ_ENABLE_MASK);
>>>  	if (bwmon->data->quirks & BWMON_NEEDS_FORCE_CLEAR)
>>>  		regmap_field_force_write(bwmon->regs[F_IRQ_CLEAR], 0);
>>> -	if (bwmon->data->quirks & BWMON_HAS_GLOBAL_IRQ)
>>> -		regmap_field_force_write(bwmon->regs[F_GLOBAL_IRQ_CLEAR],
>>> +	if (bwmon->global_regs[0])
>>> +		regmap_field_force_write(bwmon->global_regs[F_GLOBAL_IRQ_CLEAR],
>>> +					 BWMON_V4_GLOBAL_IRQ_ENABLE_ENABLE);
>>> +	else
>>> +		regmap_field_force_write(bwmon->regs[F_GLB_IRQ_CLEAR],
>>>  					 BWMON_V4_GLOBAL_IRQ_ENABLE_ENABLE);
>>>  }
>>>  
>>>  static void bwmon_disable(struct icc_bwmon *bwmon)
>>>  {
>>>  	/* Disable interrupts. Strict ordering, see bwmon_clear_irq(). */
>>> -	if (bwmon->data->quirks & BWMON_HAS_GLOBAL_IRQ)
>>> -		regmap_field_write(bwmon->regs[F_GLOBAL_IRQ_ENABLE], 0x0);
>>> +	if (bwmon->global_regs[0])
>>> +		regmap_field_write(bwmon->global_regs[F_GLOBAL_IRQ_ENABLE], 0x0);
>>> +	else
>>> +		regmap_field_write(bwmon->regs[F_GLB_IRQ_ENABLE], 0x0);
>>>  	regmap_field_write(bwmon->regs[F_IRQ_ENABLE], 0x0);
>>>  
>>>  	/*
>>> @@ -391,9 +457,13 @@ static void bwmon_disable(struct icc_bwmon *bwmon)
>>>  static void bwmon_enable(struct icc_bwmon *bwmon, unsigned int irq_enable)
>>>  {
>>>  	/* Enable interrupts */
>>> -	if (bwmon->data->quirks & BWMON_HAS_GLOBAL_IRQ)
>>> -		regmap_field_write(bwmon->regs[F_GLOBAL_IRQ_ENABLE],
>>> -				   BWMON_V4_GLOBAL_IRQ_ENABLE_ENABLE);
>>> +	if (bwmon->global_regs[0])
>>> +		regmap_field_write(bwmon->global_regs[F_GLOBAL_IRQ_ENABLE],
>>> +					 BWMON_V4_GLOBAL_IRQ_ENABLE_ENABLE);
>>> +	else
>>> +		regmap_field_write(bwmon->regs[F_GLB_IRQ_ENABLE],
>>> +					 BWMON_V4_GLOBAL_IRQ_ENABLE_ENABLE);
>>
>> Probably this would be more readable if regmap_field_write() is called
>> only once and you parametrize the first argument (field) from the
>> 'struct bwmon'.
> Sounds good
> 
>>
>>> +
>>>  	regmap_field_write(bwmon->regs[F_IRQ_ENABLE], irq_enable);
>>>  
>>>  	/* Enable bwmon */
>>> @@ -556,7 +626,9 @@ static int bwmon_init_regmap(struct platform_device *pdev,
>>>  	struct device *dev = &pdev->dev;
>>>  	void __iomem *base;
>>>  	struct regmap *map;
>>> +	int ret;
>>>  
>>> +	/* Map the monitor base */
>>>  	base = devm_platform_ioremap_resource(pdev, 0);
>>>  	if (IS_ERR(base))
>>>  		return dev_err_probe(dev, PTR_ERR(base),
>>> @@ -567,12 +639,38 @@ static int bwmon_init_regmap(struct platform_device *pdev,
>>>  		return dev_err_probe(dev, PTR_ERR(map),
>>>  				     "failed to initialize regmap\n");
>>>  
>>> +	BUILD_BUG_ON(ARRAY_SIZE(msm8998_bwmon_global_reg_fields) != F_NUM_GLOBAL_FIELDS);
>>>  	BUILD_BUG_ON(ARRAY_SIZE(msm8998_bwmon_reg_fields) != F_NUM_FIELDS);
>>>  	BUILD_BUG_ON(ARRAY_SIZE(sdm845_llcc_bwmon_reg_fields) != F_NUM_FIELDS);
>>>  
>>> -	return devm_regmap_field_bulk_alloc(dev, map, bwmon->regs,
>>> +	ret = devm_regmap_field_bulk_alloc(dev, map, bwmon->regs,
>>>  					   bwmon->data->regmap_fields,
>>>  					   F_NUM_FIELDS);
>>
>> What exactly happens now on msm8998 (or updated sdm845 from your binding
>> example) for the "global" fields in this region? The regmap references
>> non-existing fields for the "monitor" region, doesn't it?
> That's correct, the regmap references 3 non-existstent (or at least
> not-what-we-think-they-are registers (ending with _845) for
> targets with two separate regions (such as MSM8998), but they're
> unaccessed if the "global" region has been mapped sucessfully
> [see `if (bwmon->global_regs[0])`].
> 
>>
>>> +	if (ret)
>>> +		return ret;
>>> +
>>> +	if (bwmon->data->quirks & BWMON_HAS_GLOBAL_IRQ) {
>>> +		/* Map the global base, if separate */
>>> +		base = devm_platform_ioremap_resource(pdev, 1);
>>
>> Wouldn't this now print errors for sdm845, thus introduce dmesg regression?
> I explicitly stated this in the cover letter and asked for opinions.
> 

Sorry, long time ago I stopped reading cover letters, maybe except it's
top few sentences. Just too many of them and too much of text usually
useless. Commits should describe everything as they go to the Git and
they should justify their own existence.

Anyway, above dmesg error regression is a no. The devices turn out not
to be compatible with each other so just adjust the bindings and match
each driver by proper compatible string.

Best regards,
Krzysztof

Powered by blists - more mailing lists