Re: [PATCH V13 12/12] PCI: tegra: Add Tegra194 PCIe support

From: Vidya Sagar
Date: Mon Aug 05 2019 - 12:54:56 EST

On 8/5/2019 7:31 PM, Lorenzo Pieralisi wrote:
On Fri, Aug 02, 2019 at 05:36:43PM +0530, Vidya Sagar wrote:
On 7/30/2019 9:19 PM, Lorenzo Pieralisi wrote:
On Tue, Jul 23, 2019 at 08:14:08PM +0530, Vidya Sagar wrote:
On 7/16/2019 4:52 PM, Lorenzo Pieralisi wrote:
On Sat, Jul 13, 2019 at 12:34:34PM +0530, Vidya Sagar wrote:

[...]

+static int tegra_pcie_bpmp_set_ctrl_state(struct tegra_pcie_dw *pcie,
+ bool enable)
+{
+ struct mrq_uphy_response resp;
+ struct tegra_bpmp_message msg;
+ struct mrq_uphy_request req;
+ int err;
+
+ if (pcie->cid == 5)
+ return 0;

What's wrong with cid == 5 ? Explain please.
Controller with ID=5 doesn't need any programming to enable it which is
done here through calling firmware API.


+ memset(&req, 0, sizeof(req));
+ memset(&resp, 0, sizeof(resp));
+
+ req.cmd = CMD_UPHY_PCIE_CONTROLLER_STATE;
+ req.controller_state.pcie_controller = pcie->cid;
+ req.controller_state.enable = enable;
+
+ memset(&msg, 0, sizeof(msg));
+ msg.mrq = MRQ_UPHY;
+ msg.tx.data = &req;
+ msg.tx.size = sizeof(req);
+ msg.rx.data = &resp;
+ msg.rx.size = sizeof(resp);
+
+ if (irqs_disabled())

Can you explain to me what this check is meant to achieve please ?
Firmware interface provides different APIs to be called when there are
no interrupts enabled in the system (noirq context) and otherwise
hence checking that situation here and calling appropriate API.

That's what I am questioning. Being called from {suspend/resume}_noirq()
callbacks (if that's the code path this check caters for) does not mean
irqs_disabled() == true.
Agree.
Actually, I got a hint of having this check from the following.
Both tegra_bpmp_transfer_atomic() and tegra_bpmp_transfer() are indirectly
called by APIs registered with .master_xfer() and .master_xfer_atomic() hooks of
struct i2c_algorithm and the decision to call which one of these is made using the
following check in i2c-core.h file.
static inline bool i2c_in_atomic_xfer_mode(void)
{
return system_state > SYSTEM_RUNNING && irqs_disabled();
}
I think I should use this condition as is IIUC.
Please let me know if there are any concerns with this.

It is not a concern, it is just that I don't understand how this code
can be called with IRQs disabled, if you can give me an execution path I
am happy to leave the check there. On top of that, when called from
suspend NOIRQ context, it is likely to use the blocking API (because
IRQs aren't disabled at CPU level) behind which there is most certainly
an IRQ required to wake the thread up and if the IRQ in question was
disabled in the suspend NOIRQ phase this code is likely to deadlock.

I want to make sure we can justify adding this check, I do not
want to add it because we think it can be needed when it may not
be needed at all (and it gets copy and pasted over and over again
in other drivers).
I had a discussion internally about this and the prescribed usage of these APIs
seem to be that
use tegra_bpmp_transfer() in .probe() and other paths where interrupts are
enabled as this API needs interrupts to be enabled for its working.
Use tegra_bpmp_transfer_atomic() surrounded by local_irq_save()/local_irq_restore()
in other paths where interrupt servicing is disabled.

Why tegra_bpmp_transfer_atomic() needs IRQs to be disabled ? And why
is it needed in this piece of code where IRQs are _never_ disabled
at CPU level ?

IRQs are enabled when you call a suspend_noirq() callback, so the
blocking API can be used as long as the IRQ descriptor backing
the IRQ that will wake-up the blocked call is marked as
IRQF_NO_SUSPEND.

The problem is not IRQs enabled/disabled at CPU level, the problem is
the IRQ descriptor of the IRQ required to handle the blocking BPMP call,
mark it as IRQF_NO_SUSPEND and remove the tegra_bpmp_transfer_atomic()
call from this code (or please give me a concrete example pinpointing
why it is needed).
Ideally, using tegra_bpmp_transfer() alone in all paths (.probe() as
well as .resume_noirq()) should have worked as the corresponding IRQ
is already flagged as IRQF_NO_SUSPEND, but, because of the way BPMP-FW
driver in kernel making its interface available through
.resume_early(), tegra_bpmp_transfer() wasn't working as expected and
I pushed a patch (CC'ing you) at
http://patchwork.ozlabs.org/patch/1140973/ to make it .resume_noirq()
from .resume_early(). With that in place, we can just use
tegra_bpmp_trasnfer(). I'll push a new patch with this change once my
BPMP-FW driver patch is approved.

Does this leave you with a resume_noirq() callbacks ordering issue to
sort out ?
Not really.


a.k.a How will you guarantee that the BPMP will resume before the host
bridge ?
It is already taken care of in the following way.
PCIe controller's device-tree node has an entry with a phandle of BPMP-FW's node
to get a handle of it and PCIe driver uses tegra_bpmp_get() API for that.
This API returns -EPROBE_DEFER if BPMP-FW's driver is not ready yet, which guarantees
that PCIe driver gets loaded only after BPMP-FW's driver and this order is followed
during noirq phase also.


Thanks,
Lorenzo

Thanks,
Vidya Sagar

Thanks,
Lorenzo

I'll go ahead and make next patch series with this if this looks fine to you.


Actually, if tegra_bpmp_transfer() requires IRQs to be enabled you may
even end up in a situation where that blocking call does not wake up
because the IRQ in question was disabled in the NOIRQ suspend/resume
phase.

[...]

+static int tegra_pcie_dw_probe(struct platform_device *pdev)
+{
+ const struct tegra_pcie_soc *data;
+ struct device *dev = &pdev->dev;
+ struct resource *atu_dma_res;
+ struct tegra_pcie_dw *pcie;
+ struct resource *dbi_res;
+ struct pcie_port *pp;
+ struct dw_pcie *pci;
+ struct phy **phys;
+ char *name;
+ int ret;
+ u32 i;
+
+ pcie = devm_kzalloc(dev, sizeof(*pcie), GFP_KERNEL);
+ if (!pcie)
+ return -ENOMEM;
+
+ pci = &pcie->pci;
+ pci->dev = &pdev->dev;
+ pci->ops = &tegra_dw_pcie_ops;
+ pp = &pci->pp;
+ pcie->dev = &pdev->dev;
+
+ data = (struct tegra_pcie_soc *)of_device_get_match_data(dev);
+ if (!data)
+ return -EINVAL;
+ pcie->mode = (enum dw_pcie_device_mode)data->mode;
+
+ ret = tegra_pcie_dw_parse_dt(pcie);
+ if (ret < 0) {
+ dev_err(dev, "Failed to parse device tree: %d\n", ret);
+ return ret;
+ }
+
+ pcie->pex_ctl_supply = devm_regulator_get(dev, "vddio-pex-ctl");
+ if (IS_ERR(pcie->pex_ctl_supply)) {
+ dev_err(dev, "Failed to get regulator: %ld\n",
+ PTR_ERR(pcie->pex_ctl_supply));
+ return PTR_ERR(pcie->pex_ctl_supply);
+ }
+
+ pcie->core_clk = devm_clk_get(dev, "core");
+ if (IS_ERR(pcie->core_clk)) {
+ dev_err(dev, "Failed to get core clock: %ld\n",
+ PTR_ERR(pcie->core_clk));
+ return PTR_ERR(pcie->core_clk);
+ }
+
+ pcie->appl_res = platform_get_resource_byname(pdev, IORESOURCE_MEM,
+ "appl");
+ if (!pcie->appl_res) {
+ dev_err(dev, "Failed to find \"appl\" region\n");
+ return PTR_ERR(pcie->appl_res);
+ }
+ pcie->appl_base = devm_ioremap_resource(dev, pcie->appl_res);
+ if (IS_ERR(pcie->appl_base))
+ return PTR_ERR(pcie->appl_base);
+
+ pcie->core_apb_rst = devm_reset_control_get(dev, "apb");
+ if (IS_ERR(pcie->core_apb_rst)) {
+ dev_err(dev, "Failed to get APB reset: %ld\n",
+ PTR_ERR(pcie->core_apb_rst));
+ return PTR_ERR(pcie->core_apb_rst);
+ }
+
+ phys = devm_kcalloc(dev, pcie->phy_count, sizeof(*phys), GFP_KERNEL);
+ if (!phys)
+ return PTR_ERR(phys);
+
+ for (i = 0; i < pcie->phy_count; i++) {
+ name = kasprintf(GFP_KERNEL, "p2u-%u", i);
+ if (!name) {
+ dev_err(dev, "Failed to create P2U string\n");
+ return -ENOMEM;
+ }
+ phys[i] = devm_phy_get(dev, name);
+ kfree(name);
+ if (IS_ERR(phys[i])) {
+ ret = PTR_ERR(phys[i]);
+ dev_err(dev, "Failed to get PHY: %d\n", ret);
+ return ret;
+ }
+ }
+
+ pcie->phys = phys;
+
+ dbi_res = platform_get_resource_byname(pdev, IORESOURCE_MEM, "dbi");
+ if (!dbi_res) {
+ dev_err(dev, "Failed to find \"dbi\" region\n");
+ return PTR_ERR(dbi_res);
+ }
+ pcie->dbi_res = dbi_res;
+
+ pci->dbi_base = devm_ioremap_resource(dev, dbi_res);
+ if (IS_ERR(pci->dbi_base))
+ return PTR_ERR(pci->dbi_base);
+
+ /* Tegra HW locates DBI2 at a fixed offset from DBI */
+ pci->dbi_base2 = pci->dbi_base + 0x1000;
+
+ atu_dma_res = platform_get_resource_byname(pdev, IORESOURCE_MEM,
+ "atu_dma");
+ if (!atu_dma_res) {
+ dev_err(dev, "Failed to find \"atu_dma\" region\n");
+ return PTR_ERR(atu_dma_res);
+ }
+ pcie->atu_dma_res = atu_dma_res;
+ pci->atu_base = devm_ioremap_resource(dev, atu_dma_res);
+ if (IS_ERR(pci->atu_base))
+ return PTR_ERR(pci->atu_base);
+
+ pcie->core_rst = devm_reset_control_get(dev, "core");
+ if (IS_ERR(pcie->core_rst)) {
+ dev_err(dev, "Failed to get core reset: %ld\n",
+ PTR_ERR(pcie->core_rst));
+ return PTR_ERR(pcie->core_rst);
+ }
+
+ pp->irq = platform_get_irq_byname(pdev, "intr");
+ if (!pp->irq) {
+ dev_err(dev, "Failed to get \"intr\" interrupt\n");
+ return -ENODEV;
+ }
+
+ ret = devm_request_irq(dev, pp->irq, tegra_pcie_irq_handler,
+ IRQF_SHARED, "tegra-pcie-intr", pcie);
+ if (ret) {
+ dev_err(dev, "Failed to request IRQ %d: %d\n", pp->irq, ret);
+ return ret;
+ }
+
+ pcie->bpmp = tegra_bpmp_get(dev);
+ if (IS_ERR(pcie->bpmp))
+ return PTR_ERR(pcie->bpmp);
+
+ platform_set_drvdata(pdev, pcie);
+
+ if (pcie->mode == DW_PCIE_RC_TYPE) {
+ ret = tegra_pcie_config_rp(pcie);
+ if (ret && ret != -ENOMEDIUM)
+ goto fail;
+ else
+ return 0;

So if the link is not up we still go ahead and make probe
succeed. What for ?
We may need root port to be available to support hot-plugging of
endpoint devices, so, we don't fail the probe.

We need it or we don't. If you do support hotplugging of endpoint
devices point me at the code, otherwise link up failure means
failure to probe.
Currently hotplugging of endpoint is not supported, but it is one of
the use cases that we may add support for in future.

You should elaborate on this, I do not understand what you mean,
either the root port(s) supports hotplug or it does not.

But, why should we fail probe if link up doesn't happen? As such,
nothing went wrong in terms of root port initialization right? I
checked other DWC based implementations and following are not failing
the probe pci-dra7xx.c, pcie-armada8k.c, pcie-artpec6.c, pcie-histb.c,
pcie-kirin.c, pcie-spear13xx.c, pci-exynos.c, pci-imx6.c,
pci-keystone.c, pci-layerscape.c

Although following do fail the probe if link is not up. pcie-qcom.c,
pcie-uniphier.c, pci-meson.c

So, to me, it looks more like a choice we can make whether to fail the
probe or not and in this case we are choosing not to fail.

I disagree. I had an offline chat with Bjorn and whether link-up should
fail the probe or not depends on whether the root port(s) is hotplug
capable or not and this in turn relies on the root port "Slot
implemented" bit in the PCI Express capabilities register.

It is a choice but it should be based on evidence.

Lorenzo
With Bjorn's latest comment on top of this, I think we are good not to fail
the probe here.

- Vidya Sagar