Skip to contentSkip to conversation

clk: tegra: Use readl_relaxed_poll_timeout_atomic in tegra210_clock_init

PATCH Applied Nicolin Chen 3 participants 8 messages Sep 15, 2017 – Oct 20, 2017

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 clk: tegra: Use readl_relaxed_poll_timeout_atomic in tegra210_clock_init

Below is the call trace of tegra210_init_pllu() function:
start_kernel()
-> time_init()
--> of_clk_init()
---> tegra210_clock_init()
----> tegra210_pll_init()
-----> tegra210_init_pllu()

Because the preemption is disabled in the start_kernel before calling time_init, tegra210_init_pllu is actually in an atomic context while it includes a readl_relaxed_poll_timeout that might sleep.

So this patch just changes this readl_relaxed_poll_timeout() to its atomic version.

---
Changelog
v2:
* Corrected a typo in the commit log
* Added Peter's Acked-by.

 drivers/clk/tegra/clk-tegra210.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/clk/tegra/clk-tegra210.c +2 -2
index 0b9789a..ea695c4 100644
@@ -2587,8 +2587,8 @@ static int tegra210_enable_pllu(void)
 	reg |= PLL_ENABLE;
 	writel(reg, clk_base + PLLU_BASE);
 
-	readl_relaxed_poll_timeout(clk_base + PLLU_BASE, reg,
-				   reg & PLL_BASE_LOCK, 2, 1000);
+	readl_relaxed_poll_timeout_atomic(clk_base + PLLU_BASE, reg,
+					  reg & PLL_BASE_LOCK, 2, 1000);
 	if (!(reg & PLL_BASE_LOCK)) {
 		pr_err("Timed out waiting for PLL_U to lock\n");
 		return -ETIMEDOUT;
Signature · 1 line
2.1.4
#1 lore raw
Quoted 12 lines · Nicolin Chen

Below is the call trace of tegra210_init_pllu() function:
start_kernel()
-> time_init()
--> of_clk_init()
---> tegra210_clock_init()
----> tegra210_pll_init()
-----> tegra210_init_pllu()

Because the preemption is disabled in the start_kernel before calling time_init, tegra210_init_pllu is actually in an atomic context while it includes a readl_relaxed_poll_timeout that might sleep.

Quoted 5 lines

So this patch just changes this readl_relaxed_poll_timeout() to its atomic version.

Thierry, can you also take a look at this one? I sent a month ago. Thanks.

#2 lore raw
Quoted 13 lines · Nicolin Chen
Quoted 12 lines · Nicolin Chen

Below is the call trace of tegra210_init_pllu() function:
start_kernel()
-> time_init()
--> of_clk_init()
---> tegra210_clock_init()
----> tegra210_pll_init()
-----> tegra210_init_pllu()

Because the preemption is disabled in the start_kernel before calling time_init, tegra210_init_pllu is actually in an atomic context while it includes a readl_relaxed_poll_timeout that might sleep.

Quoted 7 lines
Quoted 5 lines

So this patch just changes this readl_relaxed_poll_timeout() to its atomic version.

Thierry, can you also take a look at this one? I sent a month ago. Thanks.

I'm wondering why we're not seeing a splat for this. Usually the kernel will warn if you sleep during atomic context. Does this mean we're just not hitting that case? readx_poll_timeout() has a might_sleep_if(), and therefore it should always cause the splat.

Any ideas why this has gone unnoticed for all this time?

Thierry

#3 lore raw
Quoted 11 lines · Thierry Reding
Quoted 7 lines
Quoted 7 lines

Below is the call trace of tegra210_init_pllu() function:
start_kernel()
-> time_init()
--> of_clk_init()
---> tegra210_clock_init()
----> tegra210_pll_init()
-----> tegra210_init_pllu()

I'm wondering why we're not seeing a splat for this. Usually the kernel will warn if you sleep during atomic context. Does this mean we're just not hitting that case?

Yes.

Quoted 2 lines

readx_poll_timeout() has a might_sleep_if(), and
therefore it should always cause the splat.

That's true as long as CONFIG_DEBUG_ATOMIC_SLEEP is enabled locally.

Quoted 1 line

Any ideas why this has gone unnoticed for all this time?

We can see in the tegra210_init_pllu() function that it'll not call tegra210_enable_pllu() if pllu is already enabled (by bootloader).

You can verify it by adding an irqs_disabled() in this routine. The function is called during system-boot and suspend-n-resume. And both cases should be irqs_disabled().

Thanks
Nicolin

#4 lore raw
Quoted 20 lines · Nicolin Chen
Quoted 11 lines · Thierry Reding
Quoted 7 lines
Quoted 7 lines

Below is the call trace of tegra210_init_pllu() function:
start_kernel()
-> time_init()
--> of_clk_init()
---> tegra210_clock_init()
----> tegra210_pll_init()
-----> tegra210_init_pllu()

I'm wondering why we're not seeing a splat for this. Usually the kernel will warn if you sleep during atomic context. Does this mean we're just not hitting that case?

Yes.

Quoted 2 lines

readx_poll_timeout() has a might_sleep_if(), and
therefore it should always cause the splat.

That's true as long as CONFIG_DEBUG_ATOMIC_SLEEP is enabled locally.

Quoted 7 lines
Quoted 1 line

Any ideas why this has gone unnoticed for all this time?

We can see in the tegra210_init_pllu() function that it'll not call tegra210_enable_pllu() if pllu is already enabled (by bootloader).

You can verify it by adding an irqs_disabled() in this routine. The function is called during system-boot and suspend-n-resume. And both

A correction: using mainline kernel, only system-boot as the commit log describes.

Quoted 4 lines

cases should be irqs_disabled().

Thanks
Nicolin

maintainer
#5 lore raw
In reply to Nicolin Chen
Quoted 18 lines · Nicolin Chen
Quoted 11 lines · Thierry Reding
Quoted 7 lines
Quoted 7 lines

Below is the call trace of tegra210_init_pllu() function:
start_kernel()
-> time_init()
--> of_clk_init()
---> tegra210_clock_init()
----> tegra210_pll_init()
-----> tegra210_init_pllu()

I'm wondering why we're not seeing a splat for this. Usually the kernel will warn if you sleep during atomic context. Does this mean we're just not hitting that case?

Yes.

Quoted 2 lines

readx_poll_timeout() has a might_sleep_if(), and
therefore it should always cause the splat.

Quoted 6 lines

That's true as long as CONFIG_DEBUG_ATOMIC_SLEEP is enabled locally.

Quoted 1 line

Any ideas why this has gone unnoticed for all this time?

We can see in the tegra210_init_pllu() function that it'll not call tegra210_enable_pllu() if pllu is already enabled (by bootloader).

I was thinking that same and so I clobbered the PLLU enable bit with u-boot, however, then the kernel appears to hang on boot when enabling the PLL. So although this is probably a separate issue, I am curious if you have booted the mainline with the PLLU disabled?

Cheers
Jon

Signature · 1 line
nvpublic
#6 lore raw
Quoted 24 lines · Jon Hunter
Quoted 22 lines · Nicolin Chen
Quoted 11 lines · Thierry Reding
Quoted 7 lines
Quoted 7 lines

Below is the call trace of tegra210_init_pllu() function:
start_kernel()
-> time_init()
--> of_clk_init()
---> tegra210_clock_init()
----> tegra210_pll_init()
-----> tegra210_init_pllu()

I'm wondering why we're not seeing a splat for this. Usually the kernel will warn if you sleep during atomic context. Does this mean we're just not hitting that case?

Yes.

Quoted 2 lines

readx_poll_timeout() has a might_sleep_if(), and
therefore it should always cause the splat.

That's true as long as CONFIG_DEBUG_ATOMIC_SLEEP is enabled locally.

Quoted 1 line

Any ideas why this has gone unnoticed for all this time?

Quoted 7 lines
Quoted 2 lines

We can see in the tegra210_init_pllu() function that it'll not call tegra210_enable_pllu() if pllu is already enabled (by bootloader).

I was thinking that same and so I clobbered the PLLU enable bit with u-boot, however, then the kernel appears to hang on boot when enabling the PLL. So although this is probably a separate issue, I am curious if you have booted the mainline with the PLLU disabled?

I am not sure if clearing the enable bit only would work. But the test below should be able to verify the situation -- setting the PLLU_BASE register to its reset value.

drivers/clk/tegra/clk-tegra210.c +1 -0
index ea695c4..2279373 100644
@@ -2602,6 +2602,7 @@ static int tegra210_init_pllu(void)
 	u32 reg;
 	int err;
 
+	writel_relaxed(0x1011902, clk_base + PLLU_BASE);
	tegra210_pllu_set_defaults(&pll_u_vco_params);
	/* skip initialization when pllu is in hw controlled mode */
	reg = readl_relaxed(clk_base + PLLU_BASE);
#7 lore raw
In reply to Nicolin Chen
Quoted 17 lines · Nicolin Chen

Below is the call trace of tegra210_init_pllu() function:
start_kernel()
-> time_init()
--> of_clk_init()
---> tegra210_clock_init()
----> tegra210_pll_init()
-----> tegra210_init_pllu()

Because the preemption is disabled in the start_kernel before calling time_init, tegra210_init_pllu is actually in an atomic context while it includes a readl_relaxed_poll_timeout that might sleep.

So this patch just changes this readl_relaxed_poll_timeout() to its atomic version.

Quoted 8 lines

---
Changelog
v2:
* Corrected a typo in the commit log
* Added Peter's Acked-by.

 drivers/clk/tegra/clk-tegra210.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

Applied, thanks.

Thierry