[PATCH AUTOSEL 6.18-5.10] clk: mvebu: cp110 add CLK_IGNORE_UNUSED to pcie_x10, pcie_x11 & pcie_x4

[Sasha Levin]

From: Josua Mayer <josua at solid-run.com>

[ Upstream commit f0e6bc0c3ef4b4afb299bd6912586cafd5d864e9 ]

CP110 based platforms rely on the bootloader for pci port
initialization.
TF-A actively prevents non-uboot re-configuration of pci lanes, and many
boards do not have software control over the pci card reset.

If a pci port had link at boot-time and the clock is stopped at a later
point, the link fails and can not be recovered.

PCI controller driver probe - and by extension ownership of a driver for
the pci clocks - may be delayed especially on large modular kernels,
causing the clock core to start disabling unused clocks.

Add the CLK_IGNORE_UNUSED flag to the three pci port's clocks to ensure
they are not stopped before the pci controller driver has taken
ownership and tested for an existing link.

This fixes failed pci link detection when controller driver probes late,
e.g. with arm64 defconfig and CONFIG_PHY_MVEBU_CP110_COMPHY=m.

Closes: https://lore.kernel.org/r/b71596c7-461b-44b6-89ab-3cfbd492639f@solid-run.com
Signed-off-by: Josua Mayer <josua at solid-run.com>
Reviewed-by: Andrew Lunn <andrew at lunn.ch>
Signed-off-by: Gregory CLEMENT <gregory.clement at bootlin.com>
Signed-off-by: Sasha Levin <sashal at kernel.org>
---

LLM Generated explanations, may be completely bogus:

## Commit Analysis: clk: mvebu: cp110 add CLK_IGNORE_UNUSED to pcie_x10,
pcie_x11 & pcie_x4

### 1. COMMIT MESSAGE ANALYSIS

**Subject:** Adding CLK_IGNORE_UNUSED flag to three PCIe clock gates on
Marvell CP110 platforms.

**Key problem described:**
- CP110 platforms rely on bootloader for PCIe port initialization
- TF-A (Trusted Firmware-A) prevents non-U-Boot reconfiguration of PCIe
  lanes
- Many boards lack software control over PCIe card reset
- If a PCIe port had an active link at boot and the clock is stopped
  later, the link fails permanently and cannot be recovered
- PCIe controller driver probe may be delayed on large modular kernels,
  causing the clock framework to disable "unused" clocks before the
  driver takes ownership

**Important tags:**
- **No "Cc: stable at vger.kernel.org"** - maintainer didn't explicitly
  request backport
- **No "Fixes:" tag** - unclear when the issue was introduced
- **Reviewed-by: Andrew Lunn** - reputable ARM/networking kernel
  developer
- **Closes:** link to lore.kernel.org bug report - confirms real users
  hit this

### 2. CODE CHANGE ANALYSIS

The change adds a new `gate_flags()` helper function that returns
`CLK_IGNORE_UNUSED` for three specific clock gates:
- `CP110_GATE_PCIE_X1_0` (pcie_x10)
- `CP110_GATE_PCIE_X1_1` (pcie_x11)
- `CP110_GATE_PCIE_X4` (pcie_x4)

This flag is then passed to `init.flags` when registering gate clocks.

**Technical mechanism of the bug:**
1. Boot proceeds with PCIe link established by bootloader
2. Clock framework marks these PCIe clocks as "unused" (no driver
   claimed them yet)
3. Late in boot, clock framework garbage-collects unused clocks
4. PCIe clocks are disabled, breaking the active link
5. When PCIe driver finally probes (especially in modular configs), link
   is irrecoverably failed

**What CLK_IGNORE_UNUSED does:**
Tells the clock framework to never disable these clocks just because
they appear unclaimed. This is the standard mechanism for clocks that
must remain active until a driver explicitly takes ownership.

### 3. CLASSIFICATION

**Type:** Hardware workaround / quirk for platform-specific behavior

This falls into the "quirks and workarounds" exception category for
stable kernels. It's a workaround for the specific constraints of the
Marvell CP110 platform where:
- TF-A manages PCIe lane configuration
- Clock disable breaks PCIe links irreversibly
- Driver load timing varies across kernel configurations

### 4. SCOPE AND RISK ASSESSMENT

**Size:** ~21 lines added, 1 file changed, self-contained

**Risk level:** LOW
- Only affects Marvell CP110 platforms
- No changes to core clock framework
- Worst case: slightly higher power consumption (clocks stay on when
  could be off)
- No chance of breaking other subsystems

**Subsystem:** mvebu clock driver - mature platform-specific driver

### 5. USER IMPACT

**Affected users:** Marvell CP110-based platforms (Armada 7K/8K,
SolidRun products, etc.)

**Severity:** HIGH for affected users - PCIe devices completely fail to
be detected

**Real-world evidence:**
- Bug report on lore.kernel.org linked in commit
- Reproducible with "arm64 defconfig and
  CONFIG_PHY_MVEBU_CP110_COMPHY=m"

### 6. STABILITY INDICATORS

- **Reviewed-by:** Andrew Lunn (highly respected maintainer for this
  subsystem)
- The CP110 clock driver has existed for years - this isn't new code
- Change is isolated and uses standard clock framework mechanisms

### 7. DEPENDENCY CHECK

- Self-contained change, no dependencies on other commits
- The CP110 clock driver exists in stable trees
- No required prerequisite patches

### CONCERNS

1. **No explicit stable tag** - maintainer didn't mark for backport
2. **No Fixes: tag** - we don't know how far back this issue goes
3. **Workaround approach** - CLK_IGNORE_UNUSED is somewhat heavy-handed
   but appropriate here

### DECISION ANALYSIS

**For backporting:**
- Fixes real hardware failure (PCIe not working) on production hardware
- Small, surgical, self-contained fix
- Very low regression risk (only affects specific ARM platform)
- Follows the "hardware quirk" exception pattern for stable
- Uses standard clock framework mechanisms
- Has review from reputable subsystem maintainer

**Against backporting:**
- No Cc: stable from maintainer
- Affects relatively niche ARM SoC platform
- No indication of when bug was introduced

### Conclusion

This commit fixes a real, user-impacting bug where PCIe devices fail to
be detected on Marvell CP110-based platforms. The fix is a standard
hardware quirk/workaround pattern - adding CLK_IGNORE_UNUSED to prevent
premature clock disabling that breaks active PCIe links. The change is
small (~21 lines), isolated to a single platform-specific driver,
carries minimal regression risk, and has been reviewed by a reputable
kernel maintainer.

While there's no explicit stable tag, this fits the stable criteria for
hardware workarounds that fix real user-facing bugs. The worst case
(clocks stay on unnecessarily) is far preferable to the current bug
(PCIe completely broken).

**YES**

 drivers/clk/mvebu/cp110-system-controller.c | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/drivers/clk/mvebu/cp110-system-controller.c b/drivers/clk/mvebu/cp110-system-controller.c
index 03c59bf221060..b47c869060466 100644
--- a/drivers/clk/mvebu/cp110-system-controller.c
+++ b/drivers/clk/mvebu/cp110-system-controller.c
@@ -110,6 +110,25 @@ static const char * const gate_base_names[] = {
 	[CP110_GATE_EIP197]	= "eip197"
 };
 
+static unsigned long gate_flags(const u8 bit_idx)
+{
+	switch (bit_idx) {
+	case CP110_GATE_PCIE_X1_0:
+	case CP110_GATE_PCIE_X1_1:
+	case CP110_GATE_PCIE_X4:
+		/*
+		 * If a port had an active link at boot time, stopping
+		 * the clock creates a failed state from which controller
+		 * driver can not recover.
+		 * Prevent stopping this clock till after a driver has taken
+		 * ownership.
+		 */
+		return CLK_IGNORE_UNUSED;
+	default:
+		return 0;
+	}
+};
+
 struct cp110_gate_clk {
 	struct clk_hw hw;
 	struct regmap *regmap;
@@ -171,6 +190,7 @@ static struct clk_hw *cp110_register_gate(const char *name,
 	init.ops = &cp110_gate_ops;
 	init.parent_names = &parent_name;
 	init.num_parents = 1;
+	init.flags = gate_flags(bit_idx);
 
 	gate->regmap = regmap;
 	gate->bit_idx = bit_idx;
-- 
2.51.0