On Fri, Aug 9, 2024 at 5:53 AM Stanimir Varbanov [off-list ref] wrote:

Hi Jim,

On 8/1/24 01:28, Jim Quinlan wrote:

quoted

The 7712 SOC adds a software init reset device for the PCIe HW.
If found in the DT node, use it.

Signed-off-by: Jim Quinlan <redacted>
---
 drivers/pci/controller/pcie-brcmstb.c | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/drivers/pci/controller/pcie-brcmstb.c b/drivers/pci/controller/pcie-brcmstb.c
index 4d68fe318178..948fd4d176bc 100644
--- a/drivers/pci/controller/pcie-brcmstb.c
+++ b/drivers/pci/controller/pcie-brcmstb.c

@@ -266,6 +266,7 @@ struct brcm_pcie {
      struct reset_control    *rescal;
      struct reset_control    *perst_reset;
      struct reset_control    *bridge_reset;
+     struct reset_control    *swinit_reset;
      int                     num_memc;
      u64                     memc_size[PCIE_BRCM_MAX_MEMC];
      u32                     hw_rev;

@@ -1633,12 +1634,30 @@ static int brcm_pcie_probe(struct platform_device *pdev)
      if (IS_ERR(pcie->bridge_reset))
              return PTR_ERR(pcie->bridge_reset);

+     pcie->swinit_reset = devm_reset_control_get_optional_exclusive(&pdev->dev, "swinit");
+     if (IS_ERR(pcie->swinit_reset))
+             return PTR_ERR(pcie->swinit_reset);
+
      ret = clk_prepare_enable(pcie->clk);
      if (ret)
              return dev_err_probe(&pdev->dev, ret, "could not enable clock\n");

      pcie->bridge_sw_init_set(pcie, 0);

+     if (pcie->swinit_reset) {
+             ret = reset_control_assert(pcie->swinit_reset);
+             if (dev_err_probe(&pdev->dev, ret, "could not assert reset 'swinit'\n"))
+                     goto clk_disable_unprepare;
+
+             /* HW team recommends 1us for proper sync and propagation of reset */
+             udelay(1);

Hmm, shouldn't this delay be part of .assert/.deassert reset_control
driver?  I think this detail is reset-control hw specific and the
consumers does not need to know it.

This was discussed previously.  I pointed out that we use a reset
provider that governs dozens of devices.  The only thing that the
provider could do is to employ a  worst case delay used for all
resets.  This is unacceptable; we have certain devices that may have
to invoke
reset often and require timely action, and we do not want them having
to wait the same amount of worst case delay as for example, a UART device reset.

Further, if I do a "grep reset_control_assert -A 10 drivers"  I see
plenty of existing drivers that use usleep/msleep/udelay after the call to
reset_control_assert, just as I am doing now.

As far as my opinion goes (FWIW) I think the delay is more apt to
be present in the consumer driver and not the provider driver.  To
ascertain this specific delay I had to consult with the PCIe HW team,
not the HW team that implemented the reset controller.

Regards,
Jim Quinlan
Broadcom

quoted

+
+             ret = reset_control_deassert(pcie->swinit_reset);
+             if (dev_err_probe(&pdev->dev, ret,
+                               "could not de-assert reset 'swinit' after asserting\n"))
+                     goto clk_disable_unprepare;
+     }
+
      ret = reset_control_reset(pcie->rescal);
      if (dev_err_probe(&pdev->dev, ret, "failed to deassert 'rescal'\n"))
              goto clk_disable_unprepare;

~Stan