Message ID | 20241223-preset_v2-v3-2-a339f475caf5@oss.qualcomm.com (mailing list archive) |
---|---|
State | Not Applicable |
Headers | show |
Series | PCI: dwc: Add support for configuring lane equalization presets | expand |
On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: > PCIe equalization presets are predefined settings used to optimize > signal integrity by compensating for signal loss and distortion in > high-speed data transmission. > > As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates > of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to > configure lane equalization presets for each lane to enhance the PCIe > link reliability. Each preset value represents a different combination > of pre-shoot and de-emphasis values. For each data rate, different > registers are defined: for 8.0 GT/s, registers are defined in section > 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has > an extra receiver preset hint, requiring 16 bits per lane, while the > remaining data rates use 8 bits per lane. > > Based on the number of lanes and the supported data rate, this function > reads the device tree property and stores in the presets structure. > > Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> > --- > drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ > drivers/pci/pci.h | 17 +++++++++++++++-- > 2 files changed, 60 insertions(+), 2 deletions(-) > > diff --git a/drivers/pci/of.c b/drivers/pci/of.c > index dacea3fc5128..99e0e7ae12e9 100644 > --- a/drivers/pci/of.c > +++ b/drivers/pci/of.c > @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, > return slot_power_limit_mw; > } > EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); > + kerneldoc? Define who should free the memory and how. > +int of_pci_get_equalization_presets(struct device *dev, > + struct pci_eq_presets *presets, > + int num_lanes) > +{ > + char name[20]; > + void **preset; > + void *temp; > + int ret; > + > + if (of_property_present(dev->of_node, "eq-presets-8gts")) { > + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); > + if (!presets->eq_presets_8gts) > + return -ENOMEM; > + > + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", > + presets->eq_presets_8gts, num_lanes); > + if (ret) { > + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); > + return ret; > + } > + } > + > + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { > + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); > + if (of_property_present(dev->of_node, name)) { > + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); > + if (!temp) > + return -ENOMEM; > + > + ret = of_property_read_u8_array(dev->of_node, name, > + temp, num_lanes); > + if (ret) { > + dev_err(dev, "Error %s %d\n", name, ret); > + return ret; > + } > + > + preset = (void **)((u8 *)presets + i * sizeof(void *)); Ugh. > + *preset = temp; > + } > + } > + > + return 0; > +} > +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); > diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h > index 14d00ce45bfa..82362d58bedc 100644 > --- a/drivers/pci/pci.h > +++ b/drivers/pci/pci.h > @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) > } > > struct device_node; > - > +struct pci_eq_presets { > + void *eq_presets_8gts; > + void *eq_presets_16gts; > + void *eq_presets_32gts; > + void *eq_presets_64gts; Why are all of those void*? 8gts is u16*, all other are u8*. > +}; Empty lines before and after the struct definition. > #ifdef CONFIG_OF > int of_pci_parse_bus_range(struct device_node *node, struct resource *res); > int of_get_pci_domain_nr(struct device_node *node); > @@ -746,7 +751,9 @@ void pci_set_bus_of_node(struct pci_bus *bus); > void pci_release_bus_of_node(struct pci_bus *bus); > > int devm_of_pci_bridge_init(struct device *dev, struct pci_host_bridge *bridge); > - > +int of_pci_get_equalization_presets(struct device *dev, > + struct pci_eq_presets *presets, > + int num_lanes); Keep the empty line. > #else > static inline int > of_pci_parse_bus_range(struct device_node *node, struct resource *res) > @@ -793,6 +800,12 @@ static inline int devm_of_pci_bridge_init(struct device *dev, struct pci_host_br > return 0; > } > > +static inline int of_pci_get_equalization_presets(struct device *dev, > + struct pci_eq_presets *presets, > + int num_lanes) > +{ > + return 0; > +} > #endif /* CONFIG_OF */ > > struct of_changeset; > > -- > 2.34.1 >
On 12/23/2024 5:17 PM, Dmitry Baryshkov wrote: > On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: >> PCIe equalization presets are predefined settings used to optimize >> signal integrity by compensating for signal loss and distortion in >> high-speed data transmission. >> >> As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates >> of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to >> configure lane equalization presets for each lane to enhance the PCIe >> link reliability. Each preset value represents a different combination >> of pre-shoot and de-emphasis values. For each data rate, different >> registers are defined: for 8.0 GT/s, registers are defined in section >> 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has >> an extra receiver preset hint, requiring 16 bits per lane, while the >> remaining data rates use 8 bits per lane. >> >> Based on the number of lanes and the supported data rate, this function >> reads the device tree property and stores in the presets structure. >> >> Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> >> --- >> drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ >> drivers/pci/pci.h | 17 +++++++++++++++-- >> 2 files changed, 60 insertions(+), 2 deletions(-) >> >> diff --git a/drivers/pci/of.c b/drivers/pci/of.c >> index dacea3fc5128..99e0e7ae12e9 100644 >> --- a/drivers/pci/of.c >> +++ b/drivers/pci/of.c >> @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, >> return slot_power_limit_mw; >> } >> EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); >> + > > kerneldoc? Define who should free the memory and how. > I will update this in next series. as we are allocating using devm_kzalloc it should be freed on driver detach, as no special freeing is required. >> +int of_pci_get_equalization_presets(struct device *dev, >> + struct pci_eq_presets *presets, >> + int num_lanes) >> +{ >> + char name[20]; >> + void **preset; >> + void *temp; >> + int ret; >> + >> + if (of_property_present(dev->of_node, "eq-presets-8gts")) { >> + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); >> + if (!presets->eq_presets_8gts) >> + return -ENOMEM; >> + >> + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", >> + presets->eq_presets_8gts, num_lanes); >> + if (ret) { >> + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); >> + return ret; >> + } >> + } >> + >> + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { >> + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); >> + if (of_property_present(dev->of_node, name)) { >> + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); >> + if (!temp) >> + return -ENOMEM; >> + >> + ret = of_property_read_u8_array(dev->of_node, name, >> + temp, num_lanes); >> + if (ret) { >> + dev_err(dev, "Error %s %d\n", name, ret); >> + return ret; >> + } >> + >> + preset = (void **)((u8 *)presets + i * sizeof(void *)); > > Ugh. > I was trying iterate over each element on the structure as presets holds the starting address of the structure and to that we are adding size of the void * point to go to each element. I did this way to reduce the redundant code to read all the gts which has same way of storing the data from the device tree. I will add comments here in the next series. >> + *preset = temp; >> + } >> + } >> + >> + return 0; >> +} >> +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); >> diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h >> index 14d00ce45bfa..82362d58bedc 100644 >> --- a/drivers/pci/pci.h >> +++ b/drivers/pci/pci.h >> @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) >> } >> >> struct device_node; >> - >> +struct pci_eq_presets { >> + void *eq_presets_8gts; >> + void *eq_presets_16gts; >> + void *eq_presets_32gts; >> + void *eq_presets_64gts; > > Why are all of those void*? 8gts is u16*, all other are u8*. > To have common parsing logic I moved them to void*, as these are pointers actual memory is allocated by of_pci_get_equalization_presets() based upon the gts these should not give any issues. >> +}; > > Empty lines before and after the struct definition. > ack. - Krishna Chaitanya. >> #ifdef CONFIG_OF >> int of_pci_parse_bus_range(struct device_node *node, struct resource *res); >> int of_get_pci_domain_nr(struct device_node *node); >> @@ -746,7 +751,9 @@ void pci_set_bus_of_node(struct pci_bus *bus); >> void pci_release_bus_of_node(struct pci_bus *bus); >> >> int devm_of_pci_bridge_init(struct device *dev, struct pci_host_bridge *bridge); >> - >> +int of_pci_get_equalization_presets(struct device *dev, >> + struct pci_eq_presets *presets, >> + int num_lanes); > > Keep the empty line. > >> #else >> static inline int >> of_pci_parse_bus_range(struct device_node *node, struct resource *res) >> @@ -793,6 +800,12 @@ static inline int devm_of_pci_bridge_init(struct device *dev, struct pci_host_br >> return 0; >> } >> >> +static inline int of_pci_get_equalization_presets(struct device *dev, >> + struct pci_eq_presets *presets, >> + int num_lanes) >> +{ >> + return 0; >> +} >> #endif /* CONFIG_OF */ >> >> struct of_changeset; >> >> -- >> 2.34.1 >> >
On Mon, Dec 23, 2024 at 08:02:23PM +0530, Krishna Chaitanya Chundru wrote: > > > On 12/23/2024 5:17 PM, Dmitry Baryshkov wrote: > > On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: > > > PCIe equalization presets are predefined settings used to optimize > > > signal integrity by compensating for signal loss and distortion in > > > high-speed data transmission. > > > > > > As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates > > > of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to > > > configure lane equalization presets for each lane to enhance the PCIe > > > link reliability. Each preset value represents a different combination > > > of pre-shoot and de-emphasis values. For each data rate, different > > > registers are defined: for 8.0 GT/s, registers are defined in section > > > 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has > > > an extra receiver preset hint, requiring 16 bits per lane, while the > > > remaining data rates use 8 bits per lane. > > > > > > Based on the number of lanes and the supported data rate, this function > > > reads the device tree property and stores in the presets structure. > > > > > > Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> > > > --- > > > drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ > > > drivers/pci/pci.h | 17 +++++++++++++++-- > > > 2 files changed, 60 insertions(+), 2 deletions(-) > > > > > > diff --git a/drivers/pci/of.c b/drivers/pci/of.c > > > index dacea3fc5128..99e0e7ae12e9 100644 > > > --- a/drivers/pci/of.c > > > +++ b/drivers/pci/of.c > > > @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, > > > return slot_power_limit_mw; > > > } > > > EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); > > > + > > > > kerneldoc? Define who should free the memory and how. > > > I will update this in next series. > as we are allocating using devm_kzalloc it should be freed on driver > detach, as no special freeing is required. > > > +int of_pci_get_equalization_presets(struct device *dev, > > > + struct pci_eq_presets *presets, > > > + int num_lanes) > > > +{ > > > + char name[20]; > > > + void **preset; > > > + void *temp; > > > + int ret; > > > + > > > + if (of_property_present(dev->of_node, "eq-presets-8gts")) { > > > + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); > > > + if (!presets->eq_presets_8gts) > > > + return -ENOMEM; > > > + > > > + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", > > > + presets->eq_presets_8gts, num_lanes); > > > + if (ret) { > > > + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); > > > + return ret; > > > + } > > > + } > > > + > > > + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { > > > + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); > > > + if (of_property_present(dev->of_node, name)) { > > > + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); > > > + if (!temp) > > > + return -ENOMEM; > > > + > > > + ret = of_property_read_u8_array(dev->of_node, name, > > > + temp, num_lanes); > > > + if (ret) { > > > + dev_err(dev, "Error %s %d\n", name, ret); > > > + return ret; > > > + } > > > + > > > + preset = (void **)((u8 *)presets + i * sizeof(void *)); > > > > Ugh. > > > I was trying iterate over each element on the structure as presets holds the > starting address of the structure and to that we are adding size of the void > * point to go to each element. I did this way to reduce the > redundant code to read all the gts which has same way of storing the data > from the device tree. I will add comments here in the next series. Please rewrite this in a cleaner way. The code shouldn't raise questions. > > > + *preset = temp; > > > + } > > > + } > > > + > > > + return 0; > > > +} > > > +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); > > > diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h > > > index 14d00ce45bfa..82362d58bedc 100644 > > > --- a/drivers/pci/pci.h > > > +++ b/drivers/pci/pci.h > > > @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) > > > } > > > struct device_node; > > > - > > > +struct pci_eq_presets { > > > + void *eq_presets_8gts; > > > + void *eq_presets_16gts; > > > + void *eq_presets_32gts; > > > + void *eq_presets_64gts; > > > > Why are all of those void*? 8gts is u16*, all other are u8*. > > > To have common parsing logic I moved them to void*, as these are pointers > actual memory is allocated by of_pci_get_equalization_presets() > based upon the gts these should not give any issues. Please, don't. They have types. void pointers are for the opaque data. > > > +}; > > > > Empty lines before and after the struct definition. > > > ack. > > - Krishna Chaitanya. > > > #ifdef CONFIG_OF > > > int of_pci_parse_bus_range(struct device_node *node, struct resource *res); > > > int of_get_pci_domain_nr(struct device_node *node); > > > @@ -746,7 +751,9 @@ void pci_set_bus_of_node(struct pci_bus *bus); > > > void pci_release_bus_of_node(struct pci_bus *bus); > > > int devm_of_pci_bridge_init(struct device *dev, struct pci_host_bridge *bridge); > > > - > > > +int of_pci_get_equalization_presets(struct device *dev, > > > + struct pci_eq_presets *presets, > > > + int num_lanes); > > > > Keep the empty line. > > > > > #else > > > static inline int > > > of_pci_parse_bus_range(struct device_node *node, struct resource *res) > > > @@ -793,6 +800,12 @@ static inline int devm_of_pci_bridge_init(struct device *dev, struct pci_host_br > > > return 0; > > > } > > > +static inline int of_pci_get_equalization_presets(struct device *dev, > > > + struct pci_eq_presets *presets, > > > + int num_lanes) > > > +{ > > > + return 0; > > > +} > > > #endif /* CONFIG_OF */ > > > struct of_changeset; > > > > > > -- > > > 2.34.1 > > > > >
On 12/23/2024 8:56 PM, Dmitry Baryshkov wrote: > On Mon, Dec 23, 2024 at 08:02:23PM +0530, Krishna Chaitanya Chundru wrote: >> >> >> On 12/23/2024 5:17 PM, Dmitry Baryshkov wrote: >>> On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: >>>> PCIe equalization presets are predefined settings used to optimize >>>> signal integrity by compensating for signal loss and distortion in >>>> high-speed data transmission. >>>> >>>> As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates >>>> of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to >>>> configure lane equalization presets for each lane to enhance the PCIe >>>> link reliability. Each preset value represents a different combination >>>> of pre-shoot and de-emphasis values. For each data rate, different >>>> registers are defined: for 8.0 GT/s, registers are defined in section >>>> 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has >>>> an extra receiver preset hint, requiring 16 bits per lane, while the >>>> remaining data rates use 8 bits per lane. >>>> >>>> Based on the number of lanes and the supported data rate, this function >>>> reads the device tree property and stores in the presets structure. >>>> >>>> Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> >>>> --- >>>> drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ >>>> drivers/pci/pci.h | 17 +++++++++++++++-- >>>> 2 files changed, 60 insertions(+), 2 deletions(-) >>>> >>>> diff --git a/drivers/pci/of.c b/drivers/pci/of.c >>>> index dacea3fc5128..99e0e7ae12e9 100644 >>>> --- a/drivers/pci/of.c >>>> +++ b/drivers/pci/of.c >>>> @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, >>>> return slot_power_limit_mw; >>>> } >>>> EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); >>>> + >>> >>> kerneldoc? Define who should free the memory and how. >>> >> I will update this in next series. >> as we are allocating using devm_kzalloc it should be freed on driver >> detach, as no special freeing is required. >>>> +int of_pci_get_equalization_presets(struct device *dev, >>>> + struct pci_eq_presets *presets, >>>> + int num_lanes) >>>> +{ >>>> + char name[20]; >>>> + void **preset; >>>> + void *temp; >>>> + int ret; >>>> + >>>> + if (of_property_present(dev->of_node, "eq-presets-8gts")) { >>>> + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); >>>> + if (!presets->eq_presets_8gts) >>>> + return -ENOMEM; >>>> + >>>> + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", >>>> + presets->eq_presets_8gts, num_lanes); >>>> + if (ret) { >>>> + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); >>>> + return ret; >>>> + } >>>> + } >>>> + >>>> + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { >>>> + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); >>>> + if (of_property_present(dev->of_node, name)) { >>>> + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); >>>> + if (!temp) >>>> + return -ENOMEM; >>>> + >>>> + ret = of_property_read_u8_array(dev->of_node, name, >>>> + temp, num_lanes); >>>> + if (ret) { >>>> + dev_err(dev, "Error %s %d\n", name, ret); >>>> + return ret; >>>> + } >>>> + >>>> + preset = (void **)((u8 *)presets + i * sizeof(void *)); >>> >>> Ugh. >>> >> I was trying iterate over each element on the structure as presets holds the >> starting address of the structure and to that we are adding size of the void >> * point to go to each element. I did this way to reduce the >> redundant code to read all the gts which has same way of storing the data >> from the device tree. I will add comments here in the next series. > > Please rewrite this in a cleaner way. The code shouldn't raise > questions. > >>>> + *preset = temp; >>>> + } >>>> + } >>>> + >>>> + return 0; >>>> +} >>>> +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); >>>> diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h >>>> index 14d00ce45bfa..82362d58bedc 100644 >>>> --- a/drivers/pci/pci.h >>>> +++ b/drivers/pci/pci.h >>>> @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) >>>> } >>>> struct device_node; >>>> - >>>> +struct pci_eq_presets { >>>> + void *eq_presets_8gts; >>>> + void *eq_presets_16gts; >>>> + void *eq_presets_32gts; >>>> + void *eq_presets_64gts; >>> >>> Why are all of those void*? 8gts is u16*, all other are u8*. >>> >> To have common parsing logic I moved them to void*, as these are pointers >> actual memory is allocated by of_pci_get_equalization_presets() >> based upon the gts these should not give any issues. > > Please, don't. They have types. void pointers are for the opaque data. > ok. I think then better to use v1 patch https://lore.kernel.org/all/20241116-presets-v1-2-878a837a4fee@quicinc.com/ konrad, any objection on using v1 as that will be cleaner way even if we have some repetitive code. - Krishna Chaitanya. >>>> +}; >>> >>> Empty lines before and after the struct definition. >>> >> ack. >> >> - Krishna Chaitanya. >>>> #ifdef CONFIG_OF >>>> int of_pci_parse_bus_range(struct device_node *node, struct resource *res); >>>> int of_get_pci_domain_nr(struct device_node *node); >>>> @@ -746,7 +751,9 @@ void pci_set_bus_of_node(struct pci_bus *bus); >>>> void pci_release_bus_of_node(struct pci_bus *bus); >>>> int devm_of_pci_bridge_init(struct device *dev, struct pci_host_bridge *bridge); >>>> - >>>> +int of_pci_get_equalization_presets(struct device *dev, >>>> + struct pci_eq_presets *presets, >>>> + int num_lanes); >>> >>> Keep the empty line. >>> >>>> #else >>>> static inline int >>>> of_pci_parse_bus_range(struct device_node *node, struct resource *res) >>>> @@ -793,6 +800,12 @@ static inline int devm_of_pci_bridge_init(struct device *dev, struct pci_host_br >>>> return 0; >>>> } >>>> +static inline int of_pci_get_equalization_presets(struct device *dev, >>>> + struct pci_eq_presets *presets, >>>> + int num_lanes) >>>> +{ >>>> + return 0; >>>> +} >>>> #endif /* CONFIG_OF */ >>>> struct of_changeset; >>>> >>>> -- >>>> 2.34.1 >>>> >>> >
On Mon, Dec 23, 2024 at 10:13:29PM +0530, Krishna Chaitanya Chundru wrote: > > > On 12/23/2024 8:56 PM, Dmitry Baryshkov wrote: > > On Mon, Dec 23, 2024 at 08:02:23PM +0530, Krishna Chaitanya Chundru wrote: > > > > > > > > > On 12/23/2024 5:17 PM, Dmitry Baryshkov wrote: > > > > On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: > > > > > PCIe equalization presets are predefined settings used to optimize > > > > > signal integrity by compensating for signal loss and distortion in > > > > > high-speed data transmission. > > > > > > > > > > As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates > > > > > of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to > > > > > configure lane equalization presets for each lane to enhance the PCIe > > > > > link reliability. Each preset value represents a different combination > > > > > of pre-shoot and de-emphasis values. For each data rate, different > > > > > registers are defined: for 8.0 GT/s, registers are defined in section > > > > > 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has > > > > > an extra receiver preset hint, requiring 16 bits per lane, while the > > > > > remaining data rates use 8 bits per lane. > > > > > > > > > > Based on the number of lanes and the supported data rate, this function > > > > > reads the device tree property and stores in the presets structure. > > > > > > > > > > Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> > > > > > --- > > > > > drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ > > > > > drivers/pci/pci.h | 17 +++++++++++++++-- > > > > > 2 files changed, 60 insertions(+), 2 deletions(-) > > > > > > > > > > diff --git a/drivers/pci/of.c b/drivers/pci/of.c > > > > > index dacea3fc5128..99e0e7ae12e9 100644 > > > > > --- a/drivers/pci/of.c > > > > > +++ b/drivers/pci/of.c > > > > > @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, > > > > > return slot_power_limit_mw; > > > > > } > > > > > EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); > > > > > + > > > > > > > > kerneldoc? Define who should free the memory and how. > > > > > > > I will update this in next series. > > > as we are allocating using devm_kzalloc it should be freed on driver > > > detach, as no special freeing is required. > > > > > +int of_pci_get_equalization_presets(struct device *dev, > > > > > + struct pci_eq_presets *presets, > > > > > + int num_lanes) > > > > > +{ > > > > > + char name[20]; > > > > > + void **preset; > > > > > + void *temp; > > > > > + int ret; > > > > > + > > > > > + if (of_property_present(dev->of_node, "eq-presets-8gts")) { > > > > > + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); > > > > > + if (!presets->eq_presets_8gts) > > > > > + return -ENOMEM; > > > > > + > > > > > + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", > > > > > + presets->eq_presets_8gts, num_lanes); > > > > > + if (ret) { > > > > > + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); > > > > > + return ret; > > > > > + } > > > > > + } > > > > > + > > > > > + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { > > > > > + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); > > > > > + if (of_property_present(dev->of_node, name)) { > > > > > + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); > > > > > + if (!temp) > > > > > + return -ENOMEM; > > > > > + > > > > > + ret = of_property_read_u8_array(dev->of_node, name, > > > > > + temp, num_lanes); > > > > > + if (ret) { > > > > > + dev_err(dev, "Error %s %d\n", name, ret); > > > > > + return ret; > > > > > + } > > > > > + > > > > > + preset = (void **)((u8 *)presets + i * sizeof(void *)); > > > > > > > > Ugh. > > > > > > > I was trying iterate over each element on the structure as presets holds the > > > starting address of the structure and to that we are adding size of the void > > > * point to go to each element. I did this way to reduce the > > > redundant code to read all the gts which has same way of storing the data > > > from the device tree. I will add comments here in the next series. > > > > Please rewrite this in a cleaner way. The code shouldn't raise > > questions. > > > > > > > + *preset = temp; > > > > > + } > > > > > + } > > > > > + > > > > > + return 0; > > > > > +} > > > > > +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); > > > > > diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h > > > > > index 14d00ce45bfa..82362d58bedc 100644 > > > > > --- a/drivers/pci/pci.h > > > > > +++ b/drivers/pci/pci.h > > > > > @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) > > > > > } > > > > > struct device_node; > > > > > - > > > > > +struct pci_eq_presets { > > > > > + void *eq_presets_8gts; > > > > > + void *eq_presets_16gts; > > > > > + void *eq_presets_32gts; > > > > > + void *eq_presets_64gts; > > > > > > > > Why are all of those void*? 8gts is u16*, all other are u8*. > > > > > > > To have common parsing logic I moved them to void*, as these are pointers > > > actual memory is allocated by of_pci_get_equalization_presets() > > > based upon the gts these should not give any issues. > > > > Please, don't. They have types. void pointers are for the opaque data. > > > ok. > > I think then better to use v1 patch > https://lore.kernel.org/all/20241116-presets-v1-2-878a837a4fee@quicinc.com/ > > konrad, any objection on using v1 as that will be cleaner way even if we > have some repetitive code. Konrad had a nice suggestion about using the array of values. Please use such an array for 16gts and above. This removes most of repetitive code.
On 12/24/2024 12:00 AM, Dmitry Baryshkov wrote: > On Mon, Dec 23, 2024 at 10:13:29PM +0530, Krishna Chaitanya Chundru wrote: >> >> >> On 12/23/2024 8:56 PM, Dmitry Baryshkov wrote: >>> On Mon, Dec 23, 2024 at 08:02:23PM +0530, Krishna Chaitanya Chundru wrote: >>>> >>>> >>>> On 12/23/2024 5:17 PM, Dmitry Baryshkov wrote: >>>>> On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: >>>>>> PCIe equalization presets are predefined settings used to optimize >>>>>> signal integrity by compensating for signal loss and distortion in >>>>>> high-speed data transmission. >>>>>> >>>>>> As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates >>>>>> of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to >>>>>> configure lane equalization presets for each lane to enhance the PCIe >>>>>> link reliability. Each preset value represents a different combination >>>>>> of pre-shoot and de-emphasis values. For each data rate, different >>>>>> registers are defined: for 8.0 GT/s, registers are defined in section >>>>>> 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has >>>>>> an extra receiver preset hint, requiring 16 bits per lane, while the >>>>>> remaining data rates use 8 bits per lane. >>>>>> >>>>>> Based on the number of lanes and the supported data rate, this function >>>>>> reads the device tree property and stores in the presets structure. >>>>>> >>>>>> Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> >>>>>> --- >>>>>> drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ >>>>>> drivers/pci/pci.h | 17 +++++++++++++++-- >>>>>> 2 files changed, 60 insertions(+), 2 deletions(-) >>>>>> >>>>>> diff --git a/drivers/pci/of.c b/drivers/pci/of.c >>>>>> index dacea3fc5128..99e0e7ae12e9 100644 >>>>>> --- a/drivers/pci/of.c >>>>>> +++ b/drivers/pci/of.c >>>>>> @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, >>>>>> return slot_power_limit_mw; >>>>>> } >>>>>> EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); >>>>>> + >>>>> >>>>> kerneldoc? Define who should free the memory and how. >>>>> >>>> I will update this in next series. >>>> as we are allocating using devm_kzalloc it should be freed on driver >>>> detach, as no special freeing is required. >>>>>> +int of_pci_get_equalization_presets(struct device *dev, >>>>>> + struct pci_eq_presets *presets, >>>>>> + int num_lanes) >>>>>> +{ >>>>>> + char name[20]; >>>>>> + void **preset; >>>>>> + void *temp; >>>>>> + int ret; >>>>>> + >>>>>> + if (of_property_present(dev->of_node, "eq-presets-8gts")) { >>>>>> + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); >>>>>> + if (!presets->eq_presets_8gts) >>>>>> + return -ENOMEM; >>>>>> + >>>>>> + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", >>>>>> + presets->eq_presets_8gts, num_lanes); >>>>>> + if (ret) { >>>>>> + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); >>>>>> + return ret; >>>>>> + } >>>>>> + } >>>>>> + >>>>>> + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { >>>>>> + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); >>>>>> + if (of_property_present(dev->of_node, name)) { >>>>>> + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); >>>>>> + if (!temp) >>>>>> + return -ENOMEM; >>>>>> + >>>>>> + ret = of_property_read_u8_array(dev->of_node, name, >>>>>> + temp, num_lanes); >>>>>> + if (ret) { >>>>>> + dev_err(dev, "Error %s %d\n", name, ret); >>>>>> + return ret; >>>>>> + } >>>>>> + >>>>>> + preset = (void **)((u8 *)presets + i * sizeof(void *)); >>>>> >>>>> Ugh. >>>>> >>>> I was trying iterate over each element on the structure as presets holds the >>>> starting address of the structure and to that we are adding size of the void >>>> * point to go to each element. I did this way to reduce the >>>> redundant code to read all the gts which has same way of storing the data >>>> from the device tree. I will add comments here in the next series. >>> >>> Please rewrite this in a cleaner way. The code shouldn't raise >>> questions. >>> >>>>>> + *preset = temp; >>>>>> + } >>>>>> + } >>>>>> + >>>>>> + return 0; >>>>>> +} >>>>>> +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); >>>>>> diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h >>>>>> index 14d00ce45bfa..82362d58bedc 100644 >>>>>> --- a/drivers/pci/pci.h >>>>>> +++ b/drivers/pci/pci.h >>>>>> @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) >>>>>> } >>>>>> struct device_node; >>>>>> - >>>>>> +struct pci_eq_presets { >>>>>> + void *eq_presets_8gts; >>>>>> + void *eq_presets_16gts; >>>>>> + void *eq_presets_32gts; >>>>>> + void *eq_presets_64gts; >>>>> >>>>> Why are all of those void*? 8gts is u16*, all other are u8*. >>>>> >>>> To have common parsing logic I moved them to void*, as these are pointers >>>> actual memory is allocated by of_pci_get_equalization_presets() >>>> based upon the gts these should not give any issues. >>> >>> Please, don't. They have types. void pointers are for the opaque data. >>> >> ok. >> >> I think then better to use v1 patch >> https://lore.kernel.org/all/20241116-presets-v1-2-878a837a4fee@quicinc.com/ >> >> konrad, any objection on using v1 as that will be cleaner way even if we >> have some repetitive code. > > Konrad had a nice suggestion about using the array of values. Please use > such an array for 16gts and above. This removes most of repetitive code. > I don't feel having array in the preset structure looks good, I have come up with this logic if you feel it is not so good I will go to the suggested way by having array for 16gts and above. if (of_property_present(dev->of_node, "eq-presets-8gts")) { presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); if (!presets->eq_presets_8gts) return -ENOMEM; ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", presets->eq_presets_8gts, num_lanes); if (ret) { dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); return ret; } } for (int i = EQ_PRESET_TYPE_16GTS; i < EQ_PRESET_TYPE_64GTS; i++) { snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); if (of_property_present(dev->of_node, name)) { temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); if (!temp) return -ENOMEM; ret = of_property_read_u8_array(dev->of_node, name, temp, num_lanes); if (ret) { dev_err(dev, "Error %s %d\n", name, ret); return ret; } switch (i) { case EQ_PRESET_TYPE_16GTS: presets->eq_presets_16gts = temp; break; case EQ_PRESET_TYPE_32GTS: presets->eq_presets_32gts = temp; break; case EQ_PRESET_TYPE_64GTS: presets->eq_presets_64gts = temp; break; } } } - Krishna Chaitanya. >
On Tue, Dec 24, 2024 at 02:47:00PM +0530, Krishna Chaitanya Chundru wrote: > > > On 12/24/2024 12:00 AM, Dmitry Baryshkov wrote: > > On Mon, Dec 23, 2024 at 10:13:29PM +0530, Krishna Chaitanya Chundru wrote: > > > > > > > > > On 12/23/2024 8:56 PM, Dmitry Baryshkov wrote: > > > > On Mon, Dec 23, 2024 at 08:02:23PM +0530, Krishna Chaitanya Chundru wrote: > > > > > > > > > > > > > > > On 12/23/2024 5:17 PM, Dmitry Baryshkov wrote: > > > > > > On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: > > > > > > > PCIe equalization presets are predefined settings used to optimize > > > > > > > signal integrity by compensating for signal loss and distortion in > > > > > > > high-speed data transmission. > > > > > > > > > > > > > > As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates > > > > > > > of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to > > > > > > > configure lane equalization presets for each lane to enhance the PCIe > > > > > > > link reliability. Each preset value represents a different combination > > > > > > > of pre-shoot and de-emphasis values. For each data rate, different > > > > > > > registers are defined: for 8.0 GT/s, registers are defined in section > > > > > > > 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has > > > > > > > an extra receiver preset hint, requiring 16 bits per lane, while the > > > > > > > remaining data rates use 8 bits per lane. > > > > > > > > > > > > > > Based on the number of lanes and the supported data rate, this function > > > > > > > reads the device tree property and stores in the presets structure. > > > > > > > > > > > > > > Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> > > > > > > > --- > > > > > > > drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ > > > > > > > drivers/pci/pci.h | 17 +++++++++++++++-- > > > > > > > 2 files changed, 60 insertions(+), 2 deletions(-) > > > > > > > > > > > > > > diff --git a/drivers/pci/of.c b/drivers/pci/of.c > > > > > > > index dacea3fc5128..99e0e7ae12e9 100644 > > > > > > > --- a/drivers/pci/of.c > > > > > > > +++ b/drivers/pci/of.c > > > > > > > @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, > > > > > > > return slot_power_limit_mw; > > > > > > > } > > > > > > > EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); > > > > > > > + > > > > > > > > > > > > kerneldoc? Define who should free the memory and how. > > > > > > > > > > > I will update this in next series. > > > > > as we are allocating using devm_kzalloc it should be freed on driver > > > > > detach, as no special freeing is required. > > > > > > > +int of_pci_get_equalization_presets(struct device *dev, > > > > > > > + struct pci_eq_presets *presets, > > > > > > > + int num_lanes) > > > > > > > +{ > > > > > > > + char name[20]; > > > > > > > + void **preset; > > > > > > > + void *temp; > > > > > > > + int ret; > > > > > > > + > > > > > > > + if (of_property_present(dev->of_node, "eq-presets-8gts")) { > > > > > > > + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); > > > > > > > + if (!presets->eq_presets_8gts) > > > > > > > + return -ENOMEM; > > > > > > > + > > > > > > > + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", > > > > > > > + presets->eq_presets_8gts, num_lanes); > > > > > > > + if (ret) { > > > > > > > + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); > > > > > > > + return ret; > > > > > > > + } > > > > > > > + } > > > > > > > + > > > > > > > + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { > > > > > > > + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); > > > > > > > + if (of_property_present(dev->of_node, name)) { > > > > > > > + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); > > > > > > > + if (!temp) > > > > > > > + return -ENOMEM; > > > > > > > + > > > > > > > + ret = of_property_read_u8_array(dev->of_node, name, > > > > > > > + temp, num_lanes); > > > > > > > + if (ret) { > > > > > > > + dev_err(dev, "Error %s %d\n", name, ret); > > > > > > > + return ret; > > > > > > > + } > > > > > > > + > > > > > > > + preset = (void **)((u8 *)presets + i * sizeof(void *)); > > > > > > > > > > > > Ugh. > > > > > > > > > > > I was trying iterate over each element on the structure as presets holds the > > > > > starting address of the structure and to that we are adding size of the void > > > > > * point to go to each element. I did this way to reduce the > > > > > redundant code to read all the gts which has same way of storing the data > > > > > from the device tree. I will add comments here in the next series. > > > > > > > > Please rewrite this in a cleaner way. The code shouldn't raise > > > > questions. > > > > > > > > > > > + *preset = temp; > > > > > > > + } > > > > > > > + } > > > > > > > + > > > > > > > + return 0; > > > > > > > +} > > > > > > > +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); > > > > > > > diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h > > > > > > > index 14d00ce45bfa..82362d58bedc 100644 > > > > > > > --- a/drivers/pci/pci.h > > > > > > > +++ b/drivers/pci/pci.h > > > > > > > @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) > > > > > > > } > > > > > > > struct device_node; > > > > > > > - > > > > > > > +struct pci_eq_presets { > > > > > > > + void *eq_presets_8gts; > > > > > > > + void *eq_presets_16gts; > > > > > > > + void *eq_presets_32gts; > > > > > > > + void *eq_presets_64gts; > > > > > > > > > > > > Why are all of those void*? 8gts is u16*, all other are u8*. > > > > > > > > > > > To have common parsing logic I moved them to void*, as these are pointers > > > > > actual memory is allocated by of_pci_get_equalization_presets() > > > > > based upon the gts these should not give any issues. > > > > > > > > Please, don't. They have types. void pointers are for the opaque data. > > > > > > > ok. > > > > > > I think then better to use v1 patch > > > https://lore.kernel.org/all/20241116-presets-v1-2-878a837a4fee@quicinc.com/ > > > > > > konrad, any objection on using v1 as that will be cleaner way even if we > > > have some repetitive code. > > > > Konrad had a nice suggestion about using the array of values. Please use > > such an array for 16gts and above. This removes most of repetitive code. > > > I don't feel having array in the preset structure looks good, I have > come up with this logic if you feel it is not so good I will go to the > suggested way by having array for 16gts and above. > > if (of_property_present(dev->of_node, "eq-presets-8gts")) { > presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * > num_lanes, GFP_KERNEL); > if (!presets->eq_presets_8gts) > return -ENOMEM; > > ret = of_property_read_u16_array(dev->of_node, > "eq-presets-8gts", > > presets->eq_presets_8gts, num_lanes); > if (ret) { > dev_err(dev, "Error reading eq-presets-8gts %d\n", > ret); > return ret; > } > } > > for (int i = EQ_PRESET_TYPE_16GTS; i < EQ_PRESET_TYPE_64GTS; i++) { > snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); > if (of_property_present(dev->of_node, name)) { > temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, > GFP_KERNEL); > if (!temp) > return -ENOMEM; > > ret = of_property_read_u8_array(dev->of_node, name, > temp, num_lanes); > if (ret) { > dev_err(dev, "Error %s %d\n", name, ret); > return ret; > } > > switch (i) { > case EQ_PRESET_TYPE_16GTS: > presets->eq_presets_16gts = temp; > break; > case EQ_PRESET_TYPE_32GTS: > presets->eq_presets_32gts = temp; > break; > case EQ_PRESET_TYPE_64GTS: > presets->eq_presets_64gts = temp; > break; > } This looks like 'presets->eq_presets[i] = temp;', but I won't insist on that. Also, a strange thought came to my mind: we know that there won't be more than 16 lanes. Can we have the following structure instead: #define MAX_LANES 16 enum pcie_gts { PCIE_GTS_16GTS, PCIE_GTS_32GTS, PCIE_GTS_64GTS, PCIE_GTS_MAX, }; struct pci_eq_presets { u16 eq_presets_8gts[MAX_LANES]; u8 eq_presets_Ngts[PCIE_GTS_MAX][MAX_LANES]; }; This should allow you to drop the of_property_present() and devm_kzalloc(). Just read DT data into a corresponding array. > } > } > - Krishna Chaitanya. > > >
On 12/24/2024 3:25 PM, Dmitry Baryshkov wrote: > On Tue, Dec 24, 2024 at 02:47:00PM +0530, Krishna Chaitanya Chundru wrote: >> >> >> On 12/24/2024 12:00 AM, Dmitry Baryshkov wrote: >>> On Mon, Dec 23, 2024 at 10:13:29PM +0530, Krishna Chaitanya Chundru wrote: >>>> >>>> >>>> On 12/23/2024 8:56 PM, Dmitry Baryshkov wrote: >>>>> On Mon, Dec 23, 2024 at 08:02:23PM +0530, Krishna Chaitanya Chundru wrote: >>>>>> >>>>>> >>>>>> On 12/23/2024 5:17 PM, Dmitry Baryshkov wrote: >>>>>>> On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: >>>>>>>> PCIe equalization presets are predefined settings used to optimize >>>>>>>> signal integrity by compensating for signal loss and distortion in >>>>>>>> high-speed data transmission. >>>>>>>> >>>>>>>> As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates >>>>>>>> of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to >>>>>>>> configure lane equalization presets for each lane to enhance the PCIe >>>>>>>> link reliability. Each preset value represents a different combination >>>>>>>> of pre-shoot and de-emphasis values. For each data rate, different >>>>>>>> registers are defined: for 8.0 GT/s, registers are defined in section >>>>>>>> 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has >>>>>>>> an extra receiver preset hint, requiring 16 bits per lane, while the >>>>>>>> remaining data rates use 8 bits per lane. >>>>>>>> >>>>>>>> Based on the number of lanes and the supported data rate, this function >>>>>>>> reads the device tree property and stores in the presets structure. >>>>>>>> >>>>>>>> Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> >>>>>>>> --- >>>>>>>> drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ >>>>>>>> drivers/pci/pci.h | 17 +++++++++++++++-- >>>>>>>> 2 files changed, 60 insertions(+), 2 deletions(-) >>>>>>>> >>>>>>>> diff --git a/drivers/pci/of.c b/drivers/pci/of.c >>>>>>>> index dacea3fc5128..99e0e7ae12e9 100644 >>>>>>>> --- a/drivers/pci/of.c >>>>>>>> +++ b/drivers/pci/of.c >>>>>>>> @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, >>>>>>>> return slot_power_limit_mw; >>>>>>>> } >>>>>>>> EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); >>>>>>>> + >>>>>>> >>>>>>> kerneldoc? Define who should free the memory and how. >>>>>>> >>>>>> I will update this in next series. >>>>>> as we are allocating using devm_kzalloc it should be freed on driver >>>>>> detach, as no special freeing is required. >>>>>>>> +int of_pci_get_equalization_presets(struct device *dev, >>>>>>>> + struct pci_eq_presets *presets, >>>>>>>> + int num_lanes) >>>>>>>> +{ >>>>>>>> + char name[20]; >>>>>>>> + void **preset; >>>>>>>> + void *temp; >>>>>>>> + int ret; >>>>>>>> + >>>>>>>> + if (of_property_present(dev->of_node, "eq-presets-8gts")) { >>>>>>>> + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); >>>>>>>> + if (!presets->eq_presets_8gts) >>>>>>>> + return -ENOMEM; >>>>>>>> + >>>>>>>> + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", >>>>>>>> + presets->eq_presets_8gts, num_lanes); >>>>>>>> + if (ret) { >>>>>>>> + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); >>>>>>>> + return ret; >>>>>>>> + } >>>>>>>> + } >>>>>>>> + >>>>>>>> + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { >>>>>>>> + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); >>>>>>>> + if (of_property_present(dev->of_node, name)) { >>>>>>>> + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); >>>>>>>> + if (!temp) >>>>>>>> + return -ENOMEM; >>>>>>>> + >>>>>>>> + ret = of_property_read_u8_array(dev->of_node, name, >>>>>>>> + temp, num_lanes); >>>>>>>> + if (ret) { >>>>>>>> + dev_err(dev, "Error %s %d\n", name, ret); >>>>>>>> + return ret; >>>>>>>> + } >>>>>>>> + >>>>>>>> + preset = (void **)((u8 *)presets + i * sizeof(void *)); >>>>>>> >>>>>>> Ugh. >>>>>>> >>>>>> I was trying iterate over each element on the structure as presets holds the >>>>>> starting address of the structure and to that we are adding size of the void >>>>>> * point to go to each element. I did this way to reduce the >>>>>> redundant code to read all the gts which has same way of storing the data >>>>>> from the device tree. I will add comments here in the next series. >>>>> >>>>> Please rewrite this in a cleaner way. The code shouldn't raise >>>>> questions. >>>>> >>>>>>>> + *preset = temp; >>>>>>>> + } >>>>>>>> + } >>>>>>>> + >>>>>>>> + return 0; >>>>>>>> +} >>>>>>>> +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); >>>>>>>> diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h >>>>>>>> index 14d00ce45bfa..82362d58bedc 100644 >>>>>>>> --- a/drivers/pci/pci.h >>>>>>>> +++ b/drivers/pci/pci.h >>>>>>>> @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) >>>>>>>> } >>>>>>>> struct device_node; >>>>>>>> - >>>>>>>> +struct pci_eq_presets { >>>>>>>> + void *eq_presets_8gts; >>>>>>>> + void *eq_presets_16gts; >>>>>>>> + void *eq_presets_32gts; >>>>>>>> + void *eq_presets_64gts; >>>>>>> >>>>>>> Why are all of those void*? 8gts is u16*, all other are u8*. >>>>>>> >>>>>> To have common parsing logic I moved them to void*, as these are pointers >>>>>> actual memory is allocated by of_pci_get_equalization_presets() >>>>>> based upon the gts these should not give any issues. >>>>> >>>>> Please, don't. They have types. void pointers are for the opaque data. >>>>> >>>> ok. >>>> >>>> I think then better to use v1 patch >>>> https://lore.kernel.org/all/20241116-presets-v1-2-878a837a4fee@quicinc.com/ >>>> >>>> konrad, any objection on using v1 as that will be cleaner way even if we >>>> have some repetitive code. >>> >>> Konrad had a nice suggestion about using the array of values. Please use >>> such an array for 16gts and above. This removes most of repetitive code. >>> >> I don't feel having array in the preset structure looks good, I have >> come up with this logic if you feel it is not so good I will go to the >> suggested way by having array for 16gts and above. >> >> if (of_property_present(dev->of_node, "eq-presets-8gts")) { >> presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * >> num_lanes, GFP_KERNEL); >> if (!presets->eq_presets_8gts) >> return -ENOMEM; >> >> ret = of_property_read_u16_array(dev->of_node, >> "eq-presets-8gts", >> >> presets->eq_presets_8gts, num_lanes); >> if (ret) { >> dev_err(dev, "Error reading eq-presets-8gts %d\n", >> ret); >> return ret; >> } >> } >> >> for (int i = EQ_PRESET_TYPE_16GTS; i < EQ_PRESET_TYPE_64GTS; i++) { >> snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); >> if (of_property_present(dev->of_node, name)) { >> temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, >> GFP_KERNEL); >> if (!temp) >> return -ENOMEM; >> >> ret = of_property_read_u8_array(dev->of_node, name, >> temp, num_lanes); >> if (ret) { >> dev_err(dev, "Error %s %d\n", name, ret); >> return ret; >> } >> >> switch (i) { >> case EQ_PRESET_TYPE_16GTS: >> presets->eq_presets_16gts = temp; >> break; >> case EQ_PRESET_TYPE_32GTS: >> presets->eq_presets_32gts = temp; >> break; >> case EQ_PRESET_TYPE_64GTS: >> presets->eq_presets_64gts = temp; >> break; >> } > > This looks like 'presets->eq_presets[i] = temp;', but I won't insist on > that. > > Also, a strange thought came to my mind: we know that there won't be > more than 16 lanes. Can we have the following structure instead: > > #define MAX_LANES 16 > enum pcie_gts { > PCIE_GTS_16GTS, > PCIE_GTS_32GTS, > PCIE_GTS_64GTS, > PCIE_GTS_MAX, > }; > struct pci_eq_presets { > u16 eq_presets_8gts[MAX_LANES]; > u8 eq_presets_Ngts[PCIE_GTS_MAX][MAX_LANES]; > }; > > This should allow you to drop the of_property_present() and > devm_kzalloc(). Just read DT data into a corresponding array. > in the dwc driver patch I was using pointers and memory allocation to known if the property is present or not. If I use this way I might end up reading dt property again. I think better to switch to have a array for above 16gts. - Krishna Chaitanya. >> } >> } >> - Krishna Chaitanya. >> >>> >
On Tue, 24 Dec 2024 at 12:36, Krishna Chaitanya Chundru <quic_krichai@quicinc.com> wrote: > > > > On 12/24/2024 3:25 PM, Dmitry Baryshkov wrote: > > On Tue, Dec 24, 2024 at 02:47:00PM +0530, Krishna Chaitanya Chundru wrote: > >> > >> > >> On 12/24/2024 12:00 AM, Dmitry Baryshkov wrote: > >>> On Mon, Dec 23, 2024 at 10:13:29PM +0530, Krishna Chaitanya Chundru wrote: > >>>> > >>>> > >>>> On 12/23/2024 8:56 PM, Dmitry Baryshkov wrote: > >>>>> On Mon, Dec 23, 2024 at 08:02:23PM +0530, Krishna Chaitanya Chundru wrote: > >>>>>> > >>>>>> > >>>>>> On 12/23/2024 5:17 PM, Dmitry Baryshkov wrote: > >>>>>>> On Mon, Dec 23, 2024 at 12:21:15PM +0530, Krishna Chaitanya Chundru wrote: > >>>>>>>> PCIe equalization presets are predefined settings used to optimize > >>>>>>>> signal integrity by compensating for signal loss and distortion in > >>>>>>>> high-speed data transmission. > >>>>>>>> > >>>>>>>> As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates > >>>>>>>> of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to > >>>>>>>> configure lane equalization presets for each lane to enhance the PCIe > >>>>>>>> link reliability. Each preset value represents a different combination > >>>>>>>> of pre-shoot and de-emphasis values. For each data rate, different > >>>>>>>> registers are defined: for 8.0 GT/s, registers are defined in section > >>>>>>>> 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has > >>>>>>>> an extra receiver preset hint, requiring 16 bits per lane, while the > >>>>>>>> remaining data rates use 8 bits per lane. > >>>>>>>> > >>>>>>>> Based on the number of lanes and the supported data rate, this function > >>>>>>>> reads the device tree property and stores in the presets structure. > >>>>>>>> > >>>>>>>> Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> > >>>>>>>> --- > >>>>>>>> drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ > >>>>>>>> drivers/pci/pci.h | 17 +++++++++++++++-- > >>>>>>>> 2 files changed, 60 insertions(+), 2 deletions(-) > >>>>>>>> > >>>>>>>> diff --git a/drivers/pci/of.c b/drivers/pci/of.c > >>>>>>>> index dacea3fc5128..99e0e7ae12e9 100644 > >>>>>>>> --- a/drivers/pci/of.c > >>>>>>>> +++ b/drivers/pci/of.c > >>>>>>>> @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, > >>>>>>>> return slot_power_limit_mw; > >>>>>>>> } > >>>>>>>> EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); > >>>>>>>> + > >>>>>>> > >>>>>>> kerneldoc? Define who should free the memory and how. > >>>>>>> > >>>>>> I will update this in next series. > >>>>>> as we are allocating using devm_kzalloc it should be freed on driver > >>>>>> detach, as no special freeing is required. > >>>>>>>> +int of_pci_get_equalization_presets(struct device *dev, > >>>>>>>> + struct pci_eq_presets *presets, > >>>>>>>> + int num_lanes) > >>>>>>>> +{ > >>>>>>>> + char name[20]; > >>>>>>>> + void **preset; > >>>>>>>> + void *temp; > >>>>>>>> + int ret; > >>>>>>>> + > >>>>>>>> + if (of_property_present(dev->of_node, "eq-presets-8gts")) { > >>>>>>>> + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); > >>>>>>>> + if (!presets->eq_presets_8gts) > >>>>>>>> + return -ENOMEM; > >>>>>>>> + > >>>>>>>> + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", > >>>>>>>> + presets->eq_presets_8gts, num_lanes); > >>>>>>>> + if (ret) { > >>>>>>>> + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); > >>>>>>>> + return ret; > >>>>>>>> + } > >>>>>>>> + } > >>>>>>>> + > >>>>>>>> + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { > >>>>>>>> + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); > >>>>>>>> + if (of_property_present(dev->of_node, name)) { > >>>>>>>> + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); > >>>>>>>> + if (!temp) > >>>>>>>> + return -ENOMEM; > >>>>>>>> + > >>>>>>>> + ret = of_property_read_u8_array(dev->of_node, name, > >>>>>>>> + temp, num_lanes); > >>>>>>>> + if (ret) { > >>>>>>>> + dev_err(dev, "Error %s %d\n", name, ret); > >>>>>>>> + return ret; > >>>>>>>> + } > >>>>>>>> + > >>>>>>>> + preset = (void **)((u8 *)presets + i * sizeof(void *)); > >>>>>>> > >>>>>>> Ugh. > >>>>>>> > >>>>>> I was trying iterate over each element on the structure as presets holds the > >>>>>> starting address of the structure and to that we are adding size of the void > >>>>>> * point to go to each element. I did this way to reduce the > >>>>>> redundant code to read all the gts which has same way of storing the data > >>>>>> from the device tree. I will add comments here in the next series. > >>>>> > >>>>> Please rewrite this in a cleaner way. The code shouldn't raise > >>>>> questions. > >>>>> > >>>>>>>> + *preset = temp; > >>>>>>>> + } > >>>>>>>> + } > >>>>>>>> + > >>>>>>>> + return 0; > >>>>>>>> +} > >>>>>>>> +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); > >>>>>>>> diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h > >>>>>>>> index 14d00ce45bfa..82362d58bedc 100644 > >>>>>>>> --- a/drivers/pci/pci.h > >>>>>>>> +++ b/drivers/pci/pci.h > >>>>>>>> @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) > >>>>>>>> } > >>>>>>>> struct device_node; > >>>>>>>> - > >>>>>>>> +struct pci_eq_presets { > >>>>>>>> + void *eq_presets_8gts; > >>>>>>>> + void *eq_presets_16gts; > >>>>>>>> + void *eq_presets_32gts; > >>>>>>>> + void *eq_presets_64gts; > >>>>>>> > >>>>>>> Why are all of those void*? 8gts is u16*, all other are u8*. > >>>>>>> > >>>>>> To have common parsing logic I moved them to void*, as these are pointers > >>>>>> actual memory is allocated by of_pci_get_equalization_presets() > >>>>>> based upon the gts these should not give any issues. > >>>>> > >>>>> Please, don't. They have types. void pointers are for the opaque data. > >>>>> > >>>> ok. > >>>> > >>>> I think then better to use v1 patch > >>>> https://lore.kernel.org/all/20241116-presets-v1-2-878a837a4fee@quicinc.com/ > >>>> > >>>> konrad, any objection on using v1 as that will be cleaner way even if we > >>>> have some repetitive code. > >>> > >>> Konrad had a nice suggestion about using the array of values. Please use > >>> such an array for 16gts and above. This removes most of repetitive code. > >>> > >> I don't feel having array in the preset structure looks good, I have > >> come up with this logic if you feel it is not so good I will go to the > >> suggested way by having array for 16gts and above. > >> > >> if (of_property_present(dev->of_node, "eq-presets-8gts")) { > >> presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * > >> num_lanes, GFP_KERNEL); > >> if (!presets->eq_presets_8gts) > >> return -ENOMEM; > >> > >> ret = of_property_read_u16_array(dev->of_node, > >> "eq-presets-8gts", > >> > >> presets->eq_presets_8gts, num_lanes); > >> if (ret) { > >> dev_err(dev, "Error reading eq-presets-8gts %d\n", > >> ret); > >> return ret; > >> } > >> } > >> > >> for (int i = EQ_PRESET_TYPE_16GTS; i < EQ_PRESET_TYPE_64GTS; i++) { > >> snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); > >> if (of_property_present(dev->of_node, name)) { > >> temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, > >> GFP_KERNEL); > >> if (!temp) > >> return -ENOMEM; > >> > >> ret = of_property_read_u8_array(dev->of_node, name, > >> temp, num_lanes); > >> if (ret) { > >> dev_err(dev, "Error %s %d\n", name, ret); > >> return ret; > >> } > >> > >> switch (i) { > >> case EQ_PRESET_TYPE_16GTS: > >> presets->eq_presets_16gts = temp; > >> break; > >> case EQ_PRESET_TYPE_32GTS: > >> presets->eq_presets_32gts = temp; > >> break; > >> case EQ_PRESET_TYPE_64GTS: > >> presets->eq_presets_64gts = temp; > >> break; > >> } > > > > This looks like 'presets->eq_presets[i] = temp;', but I won't insist on > > that. > > > > Also, a strange thought came to my mind: we know that there won't be > > more than 16 lanes. Can we have the following structure instead: > > > > #define MAX_LANES 16 > > enum pcie_gts { > > PCIE_GTS_16GTS, > > PCIE_GTS_32GTS, > > PCIE_GTS_64GTS, > > PCIE_GTS_MAX, > > }; > > struct pci_eq_presets { > > u16 eq_presets_8gts[MAX_LANES]; > > u8 eq_presets_Ngts[PCIE_GTS_MAX][MAX_LANES]; > > }; > > > > This should allow you to drop the of_property_present() and > > devm_kzalloc(). Just read DT data into a corresponding array. > > > in the dwc driver patch I was using pointers and memory allocation > to known if the property is present or not. If I use this way I might > end up reading dt property again. Add foo_valid flags to the struct. > I think better to switch to have a > array for above 16gts. Whichever way works for you. > > - Krishna Chaitanya. > >> } > >> } > >> - Krishna Chaitanya. > >> > >>> > >
diff --git a/drivers/pci/of.c b/drivers/pci/of.c index dacea3fc5128..99e0e7ae12e9 100644 --- a/drivers/pci/of.c +++ b/drivers/pci/of.c @@ -826,3 +826,48 @@ u32 of_pci_get_slot_power_limit(struct device_node *node, return slot_power_limit_mw; } EXPORT_SYMBOL_GPL(of_pci_get_slot_power_limit); + +int of_pci_get_equalization_presets(struct device *dev, + struct pci_eq_presets *presets, + int num_lanes) +{ + char name[20]; + void **preset; + void *temp; + int ret; + + if (of_property_present(dev->of_node, "eq-presets-8gts")) { + presets->eq_presets_8gts = devm_kzalloc(dev, sizeof(u16) * num_lanes, GFP_KERNEL); + if (!presets->eq_presets_8gts) + return -ENOMEM; + + ret = of_property_read_u16_array(dev->of_node, "eq-presets-8gts", + presets->eq_presets_8gts, num_lanes); + if (ret) { + dev_err(dev, "Error reading eq-presets-8gts %d\n", ret); + return ret; + } + } + + for (int i = 1; i < sizeof(struct pci_eq_presets) / sizeof(void *); i++) { + snprintf(name, sizeof(name), "eq-presets-%dgts", 8 << i); + if (of_property_present(dev->of_node, name)) { + temp = devm_kzalloc(dev, sizeof(u8) * num_lanes, GFP_KERNEL); + if (!temp) + return -ENOMEM; + + ret = of_property_read_u8_array(dev->of_node, name, + temp, num_lanes); + if (ret) { + dev_err(dev, "Error %s %d\n", name, ret); + return ret; + } + + preset = (void **)((u8 *)presets + i * sizeof(void *)); + *preset = temp; + } + } + + return 0; +} +EXPORT_SYMBOL_GPL(of_pci_get_equalization_presets); diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h index 14d00ce45bfa..82362d58bedc 100644 --- a/drivers/pci/pci.h +++ b/drivers/pci/pci.h @@ -731,7 +731,12 @@ static inline u64 pci_rebar_size_to_bytes(int size) } struct device_node; - +struct pci_eq_presets { + void *eq_presets_8gts; + void *eq_presets_16gts; + void *eq_presets_32gts; + void *eq_presets_64gts; +}; #ifdef CONFIG_OF int of_pci_parse_bus_range(struct device_node *node, struct resource *res); int of_get_pci_domain_nr(struct device_node *node); @@ -746,7 +751,9 @@ void pci_set_bus_of_node(struct pci_bus *bus); void pci_release_bus_of_node(struct pci_bus *bus); int devm_of_pci_bridge_init(struct device *dev, struct pci_host_bridge *bridge); - +int of_pci_get_equalization_presets(struct device *dev, + struct pci_eq_presets *presets, + int num_lanes); #else static inline int of_pci_parse_bus_range(struct device_node *node, struct resource *res) @@ -793,6 +800,12 @@ static inline int devm_of_pci_bridge_init(struct device *dev, struct pci_host_br return 0; } +static inline int of_pci_get_equalization_presets(struct device *dev, + struct pci_eq_presets *presets, + int num_lanes) +{ + return 0; +} #endif /* CONFIG_OF */ struct of_changeset;
PCIe equalization presets are predefined settings used to optimize signal integrity by compensating for signal loss and distortion in high-speed data transmission. As per PCIe spec 6.0.1 revision section 8.3.3.3 & 4.2.4 for data rates of 8.0 GT/s, 16.0 GT/s, 32.0 GT/s, and 64.0 GT/s, there is a way to configure lane equalization presets for each lane to enhance the PCIe link reliability. Each preset value represents a different combination of pre-shoot and de-emphasis values. For each data rate, different registers are defined: for 8.0 GT/s, registers are defined in section 7.7.3.4; for 16.0 GT/s, in section 7.7.5.9, etc. The 8.0 GT/s rate has an extra receiver preset hint, requiring 16 bits per lane, while the remaining data rates use 8 bits per lane. Based on the number of lanes and the supported data rate, this function reads the device tree property and stores in the presets structure. Signed-off-by: Krishna Chaitanya Chundru <krishna.chundru@oss.qualcomm.com> --- drivers/pci/of.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ drivers/pci/pci.h | 17 +++++++++++++++-- 2 files changed, 60 insertions(+), 2 deletions(-)