Re: [PATCH v3 2/4] PCI: of: Add API to retrieve equalization presets from device tree

From: Krishna Chaitanya Chundru
Date: Mon Dec 30 2024 - 23:38:43 EST



On 12/30/2024 7:11 PM, Konrad Dybcio wrote:
On 24.12.2024 11:57 AM, Dmitry Baryshkov wrote:
On Tue, 24 Dec 2024 at 12:36, Krishna Chaitanya Chundru
<quic_krichai@xxxxxxxxxxx> 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@xxxxxxxxxxxxxxxx>
---
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@xxxxxxxxxxx/

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.

Some(u8)/None would be fitting, but we're not there yet :(

Are all 0x00-0xff(ff) values valid for these presets?

currently 0xff are reserved not sure in future PCIe spec data rates
can use it or not.

- Krishna Chaitanya.
I think better to switch to have a
array for above 16gts.

Whichever way works for you.

Sorta-answering the earlier email, I have no concerns either

Konrad