Re: [PATCH v2 3/3] PCI: qcom-ep: Add ICC bandwidth voting support

From: Krishna Chaitanya Chundru
Date: Fri Jun 09 2023 - 07:54:35 EST



On 6/7/2023 10:13 PM, Bjorn Helgaas wrote:
On Wed, Jun 07, 2023 at 09:48:07PM +0530, Krishna chaitanya chundru wrote:
Add support to vote for ICC bandwidth based on the link
speed and width.

This patch is inspired from pcie-qcom driver to add basic
interconnect support.
Wrap to fill 75 columns like the rest of the git history.

Link:https://patchwork.kernel.org/project/linux-pci/patch/20221102090705.23634-3-johan+linaro@xxxxxxxxxx/
Add space after "Link:"

Probably better to use a lore link
(https://lore.kernel.org/r/20221102090705.23634-3-johan+linaro@xxxxxxxxxx/)
because I think lore is more general-purpose than patchwork and we
don't get any benefit from the patchwork features here.
done
Signed-off-by: Krishna chaitanya chundru <quic_krichai@xxxxxxxxxxx>
---
drivers/pci/controller/dwc/pcie-qcom-ep.c | 68 +++++++++++++++++++++++++++++++
1 file changed, 68 insertions(+)

diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
index 19b3283..5f9139d 100644
--- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
+++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
@@ -13,6 +13,7 @@
#include <linux/debugfs.h>
#include <linux/delay.h>
#include <linux/gpio/consumer.h>
+#include <linux/interconnect.h>
#include <linux/mfd/syscon.h>
#include <linux/phy/pcie.h>
#include <linux/phy/phy.h>
@@ -28,6 +29,7 @@
#define PARF_SYS_CTRL 0x00
#define PARF_DB_CTRL 0x10
#define PARF_PM_CTRL 0x20
+#define PARF_PM_STTS 0x24
#define PARF_MHI_CLOCK_RESET_CTRL 0x174
#define PARF_MHI_BASE_ADDR_LOWER 0x178
#define PARF_MHI_BASE_ADDR_UPPER 0x17c
@@ -128,6 +130,9 @@
/* DBI register fields */
#define DBI_CON_STATUS_POWER_STATE_MASK GENMASK(1, 0)
+#define DBI_LINKCTRLSTATUS 0x80
+#define DBI_LINKCTRLSTATUS_SHIFT 16
+
#define XMLH_LINK_UP 0x400
#define CORE_RESET_TIME_US_MIN 1000
#define CORE_RESET_TIME_US_MAX 1005
@@ -178,6 +183,8 @@ struct qcom_pcie_ep {
struct phy *phy;
struct dentry *debugfs;
+ struct icc_path *icc;
Seems gratuitously different from the "icc_mem" name used in
pcie-qcom. Use the same name unless there's a reason to be different.
done
struct clk_bulk_data *clks;
int num_clks;
@@ -253,9 +260,51 @@ static void qcom_pcie_dw_stop_link(struct dw_pcie *pci)
disable_irq(pcie_ep->perst_irq);
}
+static void qcom_pcie_ep_icc_update(struct qcom_pcie_ep *pcie_ep)
+{
+ struct dw_pcie *pci = &pcie_ep->pci;
+ int speed, width;
+ u32 val, bw;
+ int ret;
+
+ if (!pcie_ep->icc)
+ return;
+
+ val = dw_pcie_readl_dbi(pci, DBI_LINKCTRLSTATUS);
+ val = val >> DBI_LINKCTRLSTATUS_SHIFT;
+
+ speed = FIELD_GET(PCI_EXP_LNKSTA_CLS, val);
+ width = FIELD_GET(PCI_EXP_LNKSTA_NLW, val);
This whole function is basically identical to qcom_pcie_icc_update().
Could a single implementation be shared? qcom_pcie_icc_update() uses
dw_pcie_find_capability(pci, PCI_CAP_ID_EXP) instead of the hard-coded
DBI_LINKCTRLSTATUS, but there are other instances of
dw_pcie_find_capability() in pcie-qcom-ep.c, so it seems possible that
the same code could be used both places.
done
+ switch (speed) {
+ case 1:
+ bw = MBps_to_icc(250); /* BW for GEN1 per lane: 250MBps */
+ break;
+ case 2:
+ bw = MBps_to_icc(500); /* BW for GEN2 per lane: 500MBps */
+ break;
+ case 3:
+ bw = MBps_to_icc(985); /* BW for GEN3 per lane: 985MBps */
+ break;
+ default:
+ WARN_ON_ONCE(1);
+ fallthrough;
+ case 4:
+ bw = MBps_to_icc(1969); /* BW for GEN4 per lane: 985MBps */
+ break;
+ }
+
+ ret = icc_set_bw(pcie_ep->icc, 0, width * bw);
+ if (ret) {
+ dev_err(pci->dev, "failed to set interconnect bandwidth: %d\n",
+ ret);
+ }
+}
+
static int qcom_pcie_enable_resources(struct qcom_pcie_ep *pcie_ep)
{
int ret;
+ struct dw_pcie *pci = &pcie_ep->pci;
ret = clk_bulk_prepare_enable(pcie_ep->num_clks, pcie_ep->clks);
if (ret)
@@ -277,6 +326,20 @@ static int qcom_pcie_enable_resources(struct qcom_pcie_ep *pcie_ep)
if (ret)
goto err_phy_exit;
+ /*
+ * Some Qualcomm platforms require interconnect bandwidth constraints
+ * to be set before enabling interconnect clocks.
+ *
+ * Set an initial average bandwidth corresponding to single-lane Gen 1
+ * for the pcie to mem path.
+ */
+ ret = icc_set_bw(pcie_ep->icc, 0, MBps_to_icc(250)); /* BW for GEN1 per lane: 250MBps */
Reformat the comment so this all fits in 80 columns like the rest of
the file.
done
+ if (ret) {
+ dev_err(pci->dev, "failed to set interconnect bandwidth: %d\n",
+ ret);
+ goto err_phy_exit;
+ }
This plus the pcie_ep->icc init below is basically identical to
qcom_pcie_icc_init() in pcie_qcom.c. Why not use the same structure
here, with a qcom_pcie_icc_init() function? It's better to be the
same than different (when possible, of course).

In pcie_qcom.c driver the resources will be turned on the probe itself.

But in pcie-qcom-ep.c driver will wait for the host to toggle the perst to enable all the resources

that is reason it is different here.

- Krishna chaitanya.

return 0;
err_phy_exit:
@@ -550,6 +613,10 @@ static int qcom_pcie_ep_get_resources(struct platform_device *pdev,
if (IS_ERR(pcie_ep->phy))
ret = PTR_ERR(pcie_ep->phy);
+ pcie_ep->icc = devm_of_icc_get(dev, "pcie-mem");
+ if (IS_ERR(pcie_ep->icc))
+ ret = PTR_ERR(pcie_ep->icc);
+
return ret;
}
@@ -572,6 +639,7 @@ static irqreturn_t qcom_pcie_ep_global_irq_thread(int irq, void *data)
} else if (FIELD_GET(PARF_INT_ALL_BME, status)) {
dev_dbg(dev, "Received BME event. Link is enabled!\n");
pcie_ep->link_status = QCOM_PCIE_EP_LINK_ENABLED;
+ qcom_pcie_ep_icc_update(pcie_ep);
} else if (FIELD_GET(PARF_INT_ALL_PM_TURNOFF, status)) {
dev_dbg(dev, "Received PM Turn-off event! Entering L23\n");
val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
--
2.7.4