Re: [PATCH v2 5/6] soc: qcom: rpmh powerdomain driver

From: Rajendra Nayak
Date: Tue May 29 2018 - 06:19:51 EST


Hi David,

On 05/26/2018 06:38 AM, David Collins wrote:
> Hello Rajendra,
>
> On 05/25/2018 03:01 AM, Rajendra Nayak wrote:
>> The RPMh powerdomain driver aggregates the corner votes from various
>
> s/powerdomain/power domain/
>
> This applies to all instances in this patch. "Power domain" seems to be
> the preferred spelling in the kernel.

thanks, will change in all applicable places.

>
>
>> consumers for the ARC resources and communicates it to RPMh.
>>
>> We also add data for all powerdomains on sdm845 as part of the patch.
>> The driver can be extended to support other SoCs which support RPMh
>>
>> Signed-off-by: Rajendra Nayak <rnayak@xxxxxxxxxxxxxx>
>> ---
>> .../devicetree/bindings/power/qcom,rpmhpd.txt | 65 ++++
>> drivers/soc/qcom/Kconfig | 9 +
>> drivers/soc/qcom/Makefile | 1 +
>> drivers/soc/qcom/rpmhpd.c | 360 ++++++++++++++++++
>> 4 files changed, 435 insertions(+)
>> create mode 100644 Documentation/devicetree/bindings/power/qcom,rpmhpd.txt
>> create mode 100644 drivers/soc/qcom/rpmhpd.c
> ...
>> +++ b/Documentation/devicetree/bindings/power/qcom,rpmhpd.txt
>
> I think that this binding documentation should be in a patch separate from
> the driver.

yes, will split it when I repost

>
>
>> @@ -0,0 +1,65 @@
>> +Qualcomm RPMh Powerdomains
>
> s/Qualcomm/Qualcomm Technologies, Inc./
>
>
>> +
>> +* For RPMh powerdomains, we communicate a performance state to RPMh
>
> Does this line need to start with '*'?

No, will remove it

>
>
>> +which then translates it into a corresponding voltage on a rail
>> +
>> +Required Properties:
>> + - compatible: Should be one of the following
>> + * qcom,sdm845-rpmhpd: RPMh powerdomain for the sdm845 family of SoC
>> + - power-domain-cells: number of cells in power domain specifier
>> + must be 1
>> + - operating-points-v2: Phandle to the OPP table for the power-domain.
>> + Refer to Documentation/devicetree/bindings/power/power_domain.txt
>> + and Documentation/devicetree/bindings/opp/qcom-opp.txt for more details
>> +
>> +Example:
>> +
>> + rpmhpd: power-controller {
>> + compatible = "qcom,sdm845-rpmhpd";
>> + #power-domain-cells = <1>;
>> + operating-points-v2 = <&rpmhpd_opp_table>,
>> + <&rpmhpd_opp_table>,
>> + <&rpmhpd_opp_table>,
>> + <&rpmhpd_opp_table>,
>> + <&rpmhpd_opp_table>,
>> + <&rpmhpd_opp_table>,
>> + <&rpmhpd_opp_table>,
>> + <&rpmhpd_opp_table>,
>> + <&rpmhpd_opp_table>;
>
> Can this be changed to simply:
> operating-points-v2 = <&rpmhpd_opp_table>;
>
> The opp binding documentation [1] states that this should be ok:
>
> If only one phandle is available, then the same OPP table will be used
> for all power domains provided by the power domain provider.

thanks, I mentioned this to Viresh but didn't realize he fixed it up.
Will remove the redundant entries.

>
>
>> + };
>> +
>> + rpmhpd_opp_table: opp-table {
>> + compatible = "operating-points-v2-qcom-level", "operating-points-v2";
>> +
>> + rpmhpd_opp1: opp@1 {
>
> Is there any significance to the 1 through 8 values in these OPP table
> nodes? If not, then could this be changed to something like:
>
> rpmhpd_retention: opp@16 {
> ...
> rpmhpd_turbo_l1: opp@416 {
>
>
>> + qcom-corner = <16>;
>
> s/qcom-corner/qcom,level/g

Thanks, I'll fix these up.
>
>
>> + };
>> +
>> + rpmhpd_opp2: opp@2 {
>> + qcom-corner = <48>;
>> + };
>> +
>> + rpmhpd_opp3: opp@3 {
>> + qcom-corner = <64>;
>> + };
>> +
>> + rpmhpd_opp4: opp@4 {
>> + qcom-corner = <128>;
>> + };
>> +
>> + rpmhpd_opp5: opp@5 {
>> + qcom-corner = <192>;
>> + };
>> +
>> + rpmhpd_opp6: opp@6 {
>> + qcom-corner = <256>;
>> + };
>> +
>> + rpmhpd_opp7: opp@7 {
>> + qcom-corner = <320>;
>> + };
>
> Can you please add 336 and 384 to your example? 384 at least should be
> present as it corresponds to the Turbo level which all supplies support.

will do.

>
>
>> + rpmhpd_opp8: opp@8 {
>> + qcom-corner = <416>;
>> + };
>> + };
>
> How are consumers of these power domains supposed to identify which domain
> within <&rpmhpd> to use (e.g. VDD_CX vs VDD_MX)? If the answer is a plain
> integer index, then could you please add per-platform #define constants in
> a DT header file which explicitly define the meaning for each index?
>
> How do consumers of these power domains identify which level they want to
> set for a specific power domain (e.g. Nominal vs Turbo)?
>
> Would it be helpful to provide a DT header file with #define constants for
> the cross-platform sparse level mapping? This is done in [2] for the
> downstream rpmh-regulator driver that handles ARC managed regulators.
>
>
> Would it be ok to add some consumer DT nodes in this binding file example
> so that it is clear how consumers interact with the rpmhpd?

yes, I'll add the header for indexes and performance levels.

>
>
>> diff --git a/drivers/soc/qcom/Kconfig b/drivers/soc/qcom/Kconfig
>> index a7a405178967..1faed239701d 100644
>> --- a/drivers/soc/qcom/Kconfig
>> +++ b/drivers/soc/qcom/Kconfig
>> @@ -74,6 +74,15 @@ config QCOM_RMTFS_MEM
>>
>> Say y here if you intend to boot the modem remoteproc.
>>
>> +config QCOM_RPMHPD
>> + tristate "Qualcomm RPMh Powerdomain driver"
>
> s/Qualcomm/Qualcomm Technologies, Inc./

All other config options in qcom/Kconfig use 'Qualcomm XYZ feature'
for the comment. Maybe I will leave it that way for consistency?

[]...

>> +
>> +struct rpmhpd {
>> + struct device *dev;
>> + struct generic_pm_domain pd;
>> + struct rpmhpd *peer;
>> + const bool active_only;
>> + unsigned long corner;
>
> Does this actually need to be unsigned long? It looks like unsigned int
> is being passed in from rpmhpd_set_performance(). Also, hlvl values sent
> to RPMh will only every be in the range 0 - 15.
>
> If you change the type here, then can you also please change long to int
> in to_active_sleep() and rpmhpd_aggregate_corner() below?

yes, there's no need for an unsigned long, will change it

>
>
>> + u32 level[RPMH_ARC_MAX_LEVELS];
>> + int level_count;
>> + bool enabled;
>> + const char *res_name;
>> + u32 addr;
>> +};
>
> Can you please indent the fields of this struct to the same column with tabs?

will do.

>
>
>> +
>> +struct rpmhpd_desc {
>> + struct rpmhpd **rpmhpds;
>> + size_t num_pds;
>> +};
>
> This struct could be removed and the per-platform arrays could instead be
> NULL terminated.

Yes, but I would prefer it this way unless you have strong objections.
Just makes it easier to do the allocations at probe for genpd_onecell_data structures.

>> +
>> +static DEFINE_MUTEX(rpmhpd_lock);
>> +
>> +/* sdm845 RPMh powerdomains */
>> +DEFINE_RPMHPD(sdm845, ebi);
>> +DEFINE_RPMHPD_AO(sdm845, mx, mx_ao);
>> +DEFINE_RPMHPD_AO(sdm845, cx, cx_ao);
>> +DEFINE_RPMHPD(sdm845, lmx);
>> +DEFINE_RPMHPD(sdm845, lcx);
>> +DEFINE_RPMHPD(sdm845, gfx);
>> +DEFINE_RPMHPD(sdm845, mss);
>> +
>> +static struct rpmhpd *sdm845_rpmhpds[] = {
>> + [0] = &sdm845_ebi,
>
> If you are going to explicitly index these elements, then can you please
> use #define constants from a DT header file that specifies meaningful
> names? The existing 0-8 indexing is no better than implicit indexing.

yes, I will use the macros from the header.

>
>
>> + [1] = &sdm845_mx,
>> + [2] = &sdm845_mx_ao,
>> + [3] = &sdm845_cx,
>> + [4] = &sdm845_cx_ao,
>> + [5] = &sdm845_lmx,
>> + [6] = &sdm845_lcx,
>> + [7] = &sdm845_gfx,
>> + [8] = &sdm845_mss,
>> +};
>> +
>> +static const struct rpmhpd_desc sdm845_desc = {
>> + .rpmhpds = sdm845_rpmhpds,
>> + .num_pds = ARRAY_SIZE(sdm845_rpmhpds),
>> +};
>> +
>> +static const struct of_device_id rpmhpd_match_table[] = {
>> + { .compatible = "qcom,sdm845-rpmhpd", .data = &sdm845_desc },
>> + { }
>> +};
>> +MODULE_DEVICE_TABLE(of, rpmhpd_match_table);
>> +
>> +static int rpmhpd_send_corner(struct rpmhpd *pd, int state, unsigned int corner)
>> +{
>> + struct tcs_cmd cmd = {
>> + .addr = pd->addr,
>> + .data = corner,
>> + };
>> +
>> + return rpmh_write(pd->dev, state, &cmd, 1);
>
> This can be optimized by calling rpmh_write_async() whenever the corner
> being sent is smaller than the last value sent. That way, no time is
> wasted waiting for an ACK when decreasing voltage. Would you mind adding
> the necessary check and previous request caching for this?

thanks, looks useful, I will do the necessary changes.

>
>
>> +};
>> +
>> +static void to_active_sleep(struct rpmhpd *pd, unsigned long corner,
>> + unsigned long *active, unsigned long *sleep)
>> +{
>> + *active = corner;
>> +
>> + if (pd->active_only)
>> + *sleep = 0;
>> + else
>> + *sleep = *active;
>> +}
>> +
>> +static int rpmhpd_aggregate_corner(struct rpmhpd *pd)
>> +{
>> + int ret;
>> + struct rpmhpd *peer = pd->peer;
>> + unsigned long active_corner, sleep_corner;
>> + unsigned long this_corner = 0, this_sleep_corner = 0;
>> + unsigned long peer_corner = 0, peer_sleep_corner = 0;
>
> s/this_corner/this_active_corner/
> s/peer_corner/peer_active_corner/
>
> This is more consistent and I think that it makes the code a little more
> readable.

sure

>
>
>> +
>> + to_active_sleep(pd, pd->corner, &this_corner, &this_sleep_corner);
>> +
>> + if (peer && peer->enabled)
>> + to_active_sleep(peer, peer->corner, &peer_corner,
>> + &peer_sleep_corner);
>> +
>> + active_corner = max(this_corner, peer_corner);
>> +
>> + ret = rpmhpd_send_corner(pd, RPMH_ACTIVE_ONLY_STATE, active_corner);
>> + if (ret)
>> + return ret;
>> +
>> + sleep_corner = max(this_sleep_corner, peer_sleep_corner);
>> +
>> + return rpmhpd_send_corner(pd, RPMH_SLEEP_STATE, sleep_corner);
>> +}
>
> This aggregation function as well as the rpmhpd_send_corner() calls below
> are not sufficient for RPMh. There are 3 states that must all be used:
> RPMH_ACTIVE_ONLY_STATE, RPMH_WAKE_ONLY_STATE, and RPMH_SLEEP_STATE. The
> naming is somewhat confusing as rpmhpd is defining a different concept of
> active-only.
>
> For power domains without active-only or peers, only
> RPMH_ACTIVE_ONLY_STATE should be used. This instructs RPMh to issue the
> request immediately.
>
> For power domains with active-only, requests will need to be made for all
> three. active_corner would be sent for both RPMH_ACTIVE_ONLY_STATE (so
> that the request takes effect immediately) and RPMH_WAKE_ONLY_STATE (so
> that the request is inserted into the wake TCS). sleep_corner would be
> sent for RPMH_SLEEP_STATE (so that the request is inserted into the sleep
> TCS).
>
> You can see how this is handled in the RPMh clock driver in patch [3].

Thanks, I'll take another look at this and fix this up

>
> You may be able to get away with using only RPMH_SLEEP_STATE and
> RPMH_ACTIVE_ONLY_STATE assuming that you issue the RPMH_SLEEP_STATE
> request first due to the rpmh driver caching behavior added in the
> cache_rpm_request() function in [4]. However, could you please confirm
> with Lina that this usage will continue to work in the future? I'm not
> sure what guarantees are made at the rpmh API level.

sure, I'll check it up

>
>
>> +
>> +static int rpmhpd_power_on(struct generic_pm_domain *domain)
>> +{
>> + int ret = 0;
>> + struct rpmhpd *pd = domain_to_rpmhpd(domain);
>
> Minor: It might look a little nicer to list 'pd' definition first amongst
> the local variables in this function as well as those below.

okay

>
>
>> +
>> + mutex_lock(&rpmhpd_lock);
>> +
>> + pd->enabled = true;
>> +
>> + if (pd->corner)
>> + ret = rpmhpd_aggregate_corner(pd);
>> +
>> + mutex_unlock(&rpmhpd_lock);
>> +
>> + return ret;
>> +}
>> +
>> +static int rpmhpd_power_off(struct generic_pm_domain *domain)
>> +{
>> + int ret;
>> + struct rpmhpd *pd = domain_to_rpmhpd(domain);
>> +
>> + mutex_lock(&rpmhpd_lock);
>> +
>> + if (pd->level[0] == 0) {
>> + ret = rpmhpd_send_corner(pd, RPMH_ACTIVE_ONLY_STATE, 0);
>> + if (ret)
>> + return ret;
>> +
>> + ret = rpmhpd_send_corner(pd, RPMH_SLEEP_STATE, 0);
>> + if (ret)
>> + return ret;
>> + }
>> +
>> + pd->enabled = false;
>> +
>> + mutex_unlock(&rpmhpd_lock);
>> +
>> + return 0;
>> +}
>> +
>> +static int rpmhpd_set_performance(struct generic_pm_domain *domain,
>> + unsigned int state)
>> +{
>> + int ret = 0;
>> + struct rpmhpd *pd = domain_to_rpmhpd(domain);
>> +
>> + mutex_lock(&rpmhpd_lock);
>> +
>> + pd->corner = state;
>
> What is the numbering space for 'state'? I assume that it is a vlvl value
> corresponding to qcom,level in a DT OPP table node. If so, additional
> logic is required.
>
> When using RPMh, the platform and supply independent vlvl sparse numbering
> space is used by consumers so that they can always have consistent values.
> However, the actual requests sent to RPMh ARC must be in the hlvl
> numbering space (i.e. 0 - 15(max)). In the case of this driver, the
> acceptable hlvl values for a given power domain are 0 to pd->level_count - 1.
>
> I suspect that you need to add something like this here:
>
> int i;
>
> for (i = 0; i < pd->level_count; i++)
> if (state <= pd->level[i])
> break;
>
> if (i == pd->level_count) {
> ret = -EINVAL;
> dev_err(pd->dev, "invalid state=%u for domain %s",
> state, pd->pd.name);
> goto out;
> }
>
> pd->corner = i;
>
>
> Note that a given power domain will likely not support all of the vlvl
> values listed in the DT OPP table nodes.

yes indeed :/ I will add the translation bits here for vlvl to hlvl which is
needed before communicating with RPMh

>
>
>> +
>> + if (!pd->enabled)
>> + goto out;
>> +
>> + ret = rpmhpd_aggregate_corner(pd);
>> +out:
>> + mutex_unlock(&rpmhpd_lock);
>> +
>> + return ret;
>> +}
>> +
>> +static unsigned int rpmhpd_get_performance(struct generic_pm_domain *genpd,
>> + struct dev_pm_opp *opp)> +{
>> + struct device_node *np;
>> + unsigned int corner = 0;
>> +
>> + np = dev_pm_opp_get_of_node(opp);
>> + if (of_property_read_u32(np, "qcom,level", &corner)) {
>> + pr_err("%s: missing 'qcom,level' property\n", __func__);
>> + return 0;
>
> Why return 0 instead of an error?

The caller expects a 0 for failure

>
>
>> + }
>> +
>> + of_node_put(np);
>> +
>> + return corner;
>> +}
>
> Is there an API to determine the currently operating performance state of
> a given power domain? Is this information accessible from userspace? We
> will definitely need this for general debugging.

A quick look shows me its not. I agree its a necessary feature for debug.
I will add a patch to expose it via debugfs

>
>
>> +
>> +static int rpmhpd_update_level_mapping(struct rpmhpd *rpmhpd)
>> +{
>> + u8 *buf;
>
> This could be changed to the following in order to remove the need for
> kzalloc() and kfree() calls below:
>
> u8 buf[RPMH_ARC_MAX_LEVELS * RPMH_ARC_LEVEL_SIZE];
>
>
>> + int i, j, len, ret;
>> +
>> + len = cmd_db_read_aux_data_len(rpmhpd->res_name);
>> + if (len <= 0)
>> + return len;
>> +
>
> A check like this is needed here:
>
> if (len > RPMH_ARC_MAX_LEVELS * RPMH_ARC_LEVEL_SIZE)
> return -EINVAL;
>
>
>> + buf = kzalloc(len, GFP_KERNEL);
>> + if (!buf)
>> + return -ENOMEM;
>> +
>> + ret = cmd_db_read_aux_data(rpmhpd->res_name, buf, len);
>> + if (ret < 0)
>> + return ret;
>> +
>> + rpmhpd->level_count = len / RPMH_ARC_LEVEL_SIZE;
>> +
>> + for (i = 0; i < rpmhpd->level_count; i++) {
>
> If you make buf a fixed size array, then rpmhpd->level[i] = 0; is needed
> here or a memset() outside of the for loop.

Okay, I;ll move buf to a fixed size array instead

>
>
>> + for (j = 0; j < RPMH_ARC_LEVEL_SIZE; j++)
>> + rpmhpd->level[i] |=
>> + buf[i * RPMH_ARC_LEVEL_SIZE + j] << (8 * j);
>> +
>> + /*
>> + * The AUX data may be zero padded. These 0 valued entries at
>> + * the end of the map must be ignored.
>> + */
>> + if (i > 0 && rpmhpd->level[i] == 0) {
>> + rpmhpd->level_count = i;
>> + break;
>> + }
>> + pr_dbg("%s: ARC hlvl=%2d --> vlvl=%4u\n", rpmhpd->res_name, i,
>> + rpmhpd->level[i]);
>> + }
>> +
>> + kfree(buf);
>> + return 0;
>> +}
>> +
>> +static int rpmhpd_probe(struct platform_device *pdev)
>> +{
>> + int i, ret;
>> + size_t num;
>> + struct genpd_onecell_data *data;
>> + struct rpmhpd **rpmhpds;
>> + const struct rpmhpd_desc *desc;
>> +
>> + desc = of_device_get_match_data(&pdev->dev);
>> + if (!desc)
>> + return -EINVAL;
>> +
>> + rpmhpds = desc->rpmhpds;
>> + num = desc->num_pds;
>> +
>> + data = devm_kzalloc(&pdev->dev, sizeof(*data), GFP_KERNEL);
>> + if (!data)
>> + return -ENOMEM;
>> +
>> + data->domains = devm_kcalloc(&pdev->dev, num, sizeof(*data->domains),
>> + GFP_KERNEL);
>> + data->num_domains = num;
>> +
>> + ret = cmd_db_ready();
>> + if (ret) {
>> + if (ret != -EPROBE_DEFER)
>> + dev_err(&pdev->dev, "Command DB unavailable, ret=%d\n",
>> + ret);
>> + return ret;
>> + }
>> +
>> + for (i = 0; i < num; i++) {
>> + if (!rpmhpds[i])
>> + continue;
>
> Why is this check needed?

Just to check/ignore if there are any holes.
maybe I should atleast throw a warning instead of silently ignoring it?

>
>
>> +
>> + rpmhpds[i]->dev = &pdev->dev;
>> + rpmhpds[i]->addr = cmd_db_read_addr(rpmhpds[i]->res_name);
>> + if (!rpmhpds[i]->addr) {
>> + dev_err(&pdev->dev, "Could not find RPMh address for resource %s\n",
>> + rpmhpds[i]->res_name);
>> + return -ENODEV;
>> + }
>> +
>> + ret = cmd_db_read_slave_id(rpmhpds[i]->res_name);
>> + if (ret != CMD_DB_HW_ARC) {
>> + dev_err(&pdev->dev, "RPMh slave ID mismatch\n");
>> + return -EINVAL;
>> + }
>> +
>> + ret = rpmhpd_update_level_mapping(rpmhpds[i]);
>> + if (ret)
>> + return ret;
>> +
>> + rpmhpds[i]->pd.power_off = rpmhpd_power_off;
>> + rpmhpds[i]->pd.power_on = rpmhpd_power_on;
>> + rpmhpds[i]->pd.set_performance_state = rpmhpd_set_performance;
>> + rpmhpds[i]->pd.opp_to_performance_state = rpmhpd_get_performance;
>> + pm_genpd_init(&rpmhpds[i]->pd, NULL, true);
>> +
>> + data->domains[i] = &rpmhpds[i]->pd;
>> + }
>> +
>> + return of_genpd_add_provider_onecell(pdev->dev.of_node, data);
>> +}
>> +
>> +static int rpmhpd_remove(struct platform_device *pdev)
>> +{
>> + of_genpd_del_provider(pdev->dev.of_node);
>> + return 0;
>> +}
>> +
>> +static struct platform_driver rpmhpd_driver = {
>> + .driver = {
>> + .name = "qcom-rpmhpd",
>> + .of_match_table = rpmhpd_match_table,
>> + },
>> + .probe = rpmhpd_probe,
>> + .remove = rpmhpd_remove,
>> +};
>> +
>> +static int __init rpmhpd_init(void)
>> +{
>> + return platform_driver_register(&rpmhpd_driver);
>> +}
>> +core_initcall(rpmhpd_init);
>> +
>> +static void __exit rpmhpd_exit(void)
>> +{
>> + platform_driver_unregister(&rpmhpd_driver);
>> +}
>> +module_exit(rpmhpd_exit);
>> +
>> +MODULE_DESCRIPTION("Qualcomm RPMh Power Domain Driver");
>
> s/Qualcomm/Qualcomm Technologies, Inc./

will do

>
>
>> +MODULE_LICENSE("GPL v2");
>> +MODULE_ALIAS("platform:qcom-rpmhpd");
>
> Thanks,
> David

Thanks David for taking time to review.

--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation