Skip to contentSkip to conversation

ptp_qoriq: Don't write system time into PHC at boot

PATCH In review Vladimir Oltean 3 participants 3 messages Sep 28, 2019 – Oct 1, 2019

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ptp_qoriq: Don't write system time into PHC at boot

Because ptp_qoriq_settime is being called prior to spin_lock_init, the following stack trace can be seen at driver probe time:

[    2.269117] the code is fine but needs lockdep annotation.
[    2.274569] turning off the locking correctness validator.
[    2.280027] CPU: 1 PID: 1 Comm: swapper/0 Not tainted 5.3.0-rc7-01478-g01eaa67a4797 #263
[    2.288073] Hardware name: Freescale LS1021A
[    2.292337] [<c0313cb4>] (unwind_backtrace) from [<c030e11c>] (show_stack+0x10/0x14)
[    2.300045] [<c030e11c>] (show_stack) from [<c1219440>] (dump_stack+0xcc/0xf8)
[    2.307235] [<c1219440>] (dump_stack) from [<c03b9b44>] (register_lock_class+0x730/0x73c)
[    2.315372] [<c03b9b44>] (register_lock_class) from [<c03b6190>] (__lock_acquire+0x78/0x270c)
[    2.323856] [<c03b6190>] (__lock_acquire) from [<c03b90cc>] (lock_acquire+0xe0/0x22c)
[    2.331649] [<c03b90cc>] (lock_acquire) from [<c123c310>] (_raw_spin_lock_irqsave+0x54/0x68)
[    2.340048] [<c123c310>] (_raw_spin_lock_irqsave) from [<c0e73fe4>] (ptp_qoriq_settime+0x38/0x80)
[    2.348878] [<c0e73fe4>] (ptp_qoriq_settime) from [<c0e746d4>] (ptp_qoriq_init+0x1f8/0x484)
[    2.357189] [<c0e746d4>] (ptp_qoriq_init) from [<c0e74aac>] (ptp_qoriq_probe+0xd0/0x184)
[    2.365243] [<c0e74aac>] (ptp_qoriq_probe) from [<c0b0a07c>] (platform_drv_probe+0x48/0x9c)
[    2.373555] [<c0b0a07c>] (platform_drv_probe) from [<c0b07a14>] (really_probe+0x1c4/0x400)
[    2.381779] [<c0b07a14>] (really_probe) from [<c0b07e28>] (driver_probe_device+0x78/0x1b8)
[    2.390003] [<c0b07e28>] (driver_probe_device) from [<c0b081d0>] (device_driver_attach+0x58/0x60)
[    2.398832] [<c0b081d0>] (device_driver_attach) from [<c0b082d4>] (__driver_attach+0xfc/0x160)
[    2.407402] [<c0b082d4>] (__driver_attach) from [<c0b05a84>] (bus_for_each_dev+0x68/0xb4)
[    2.415539] [<c0b05a84>] (bus_for_each_dev) from [<c0b06b68>] (bus_add_driver+0x104/0x20c)
[    2.423763] [<c0b06b68>] (bus_add_driver) from [<c0b0909c>] (driver_register+0x78/0x10c)
[    2.431815] [<c0b0909c>] (driver_register) from [<c030313c>] (do_one_initcall+0x8c/0x3ac)
[    2.439954] [<c030313c>] (do_one_initcall) from [<c1f013f4>] (kernel_init_freeable+0x468/0x548)
[    2.448610] [<c1f013f4>] (kernel_init_freeable) from [<c12344d8>] (kernel_init+0x8/0x10c)
[    2.456745] [<c12344d8>] (kernel_init) from [<c03010b4>] (ret_from_fork+0x14/0x20)
[    2.464273] Exception stack(0xea89ffb0 to 0xea89fff8)
[    2.469297] ffa0:                                     00000000 00000000 00000000 00000000
[    2.477432] ffc0: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
[    2.485566] ffe0: 00000000 00000000 00000000 00000000 00000013 00000000

This behavior has been introduced during the rework commit ff54571a747b ("ptp_qoriq: convert to use ptp_qoriq_init/free"), so technically it is a fix for that.

But I couldn't find a justification to just move the spin_lock_init a few lines above, because I don't see why the ptp_qoriq_settime call should be done at all at probe time. Writing a CLOCK_REALTIME value into a timer that is supposed to track CLOCK_TAI means there will be a 37 second offset that's going to render it useless for any practical purpose. Even furthermore, there may be concurrency with RTC drivers at probe time which is not handled (or there may be no RTC on the platform at all), so in practice even if the code is there, the ptp_qoriq clock may still tick in Jan 1st, 1970 until a better clock source becomes available.

So just remove the ptp_qoriq_settime call and let the clock tick in 1970 until user space does something about it. Most other PTP drivers do the same, except chelsio cxgb4 and maybe a few others.

So while this is in fact changing the behavior that the ptp_qoriq (formerly gianfar_ptp) has had since day one, it is only fixing a recent rework commit.

Fixes ff54571a747b ("ptp_qoriq: convert to use ptp_qoriq_init/free")
---
 drivers/ptp/ptp_qoriq.c | 4 ----
 1 file changed, 4 deletions(-)
drivers/ptp/ptp_qoriq.c +0 -4
index c61f00b72e15..5dbcca2332d5 100644
@@ -450,7 +450,6 @@ int ptp_qoriq_init(struct ptp_qoriq *ptp_qoriq, void __iomem *base,
 {
 	struct device_node *node = ptp_qoriq->dev->of_node;
 	struct ptp_qoriq_registers *regs;
-	struct timespec64 now;
 	unsigned long flags;
 	u32 tmr_ctrl;
 
@@ -507,9 +506,6 @@ int ptp_qoriq_init(struct ptp_qoriq *ptp_qoriq, void __iomem *base,
 		ptp_qoriq->regs.etts_regs = base + ETTS_REGS_OFFSET;
 	}
 
-	ktime_get_real_ts64(&now);
-	ptp_qoriq_settime(&ptp_qoriq->caps, &now);
-
 	tmr_ctrl =
 	  (ptp_qoriq->tclk_period & TCLK_PERIOD_MASK) << TCLK_PERIOD_SHIFT |
 	  (ptp_qoriq->cksel & CKSEL_MASK) << CKSEL_SHIFT;
Signature · 1 line
2.17.1
maintainer
#1 lore raw

From: Vladimir Oltean <[email protected]>
Date: Sun, 29 Sep 2019 01:22:28 +0300

Quoted 52 lines

Because ptp_qoriq_settime is being called prior to spin_lock_init, the following stack trace can be seen at driver probe time:

[    2.269117] the code is fine but needs lockdep annotation.
[    2.274569] turning off the locking correctness validator.
[    2.280027] CPU: 1 PID: 1 Comm: swapper/0 Not tainted 5.3.0-rc7-01478-g01eaa67a4797 #263
[    2.288073] Hardware name: Freescale LS1021A
[    2.292337] [<c0313cb4>] (unwind_backtrace) from [<c030e11c>] (show_stack+0x10/0x14)
[    2.300045] [<c030e11c>] (show_stack) from [<c1219440>] (dump_stack+0xcc/0xf8)
[    2.307235] [<c1219440>] (dump_stack) from [<c03b9b44>] (register_lock_class+0x730/0x73c)
[    2.315372] [<c03b9b44>] (register_lock_class) from [<c03b6190>] (__lock_acquire+0x78/0x270c)
[    2.323856] [<c03b6190>] (__lock_acquire) from [<c03b90cc>] (lock_acquire+0xe0/0x22c)
[    2.331649] [<c03b90cc>] (lock_acquire) from [<c123c310>] (_raw_spin_lock_irqsave+0x54/0x68)
[    2.340048] [<c123c310>] (_raw_spin_lock_irqsave) from [<c0e73fe4>] (ptp_qoriq_settime+0x38/0x80)
[    2.348878] [<c0e73fe4>] (ptp_qoriq_settime) from [<c0e746d4>] (ptp_qoriq_init+0x1f8/0x484)
[    2.357189] [<c0e746d4>] (ptp_qoriq_init) from [<c0e74aac>] (ptp_qoriq_probe+0xd0/0x184)
[    2.365243] [<c0e74aac>] (ptp_qoriq_probe) from [<c0b0a07c>] (platform_drv_probe+0x48/0x9c)
[    2.373555] [<c0b0a07c>] (platform_drv_probe) from [<c0b07a14>] (really_probe+0x1c4/0x400)
[    2.381779] [<c0b07a14>] (really_probe) from [<c0b07e28>] (driver_probe_device+0x78/0x1b8)
[    2.390003] [<c0b07e28>] (driver_probe_device) from [<c0b081d0>] (device_driver_attach+0x58/0x60)
[    2.398832] [<c0b081d0>] (device_driver_attach) from [<c0b082d4>] (__driver_attach+0xfc/0x160)
[    2.407402] [<c0b082d4>] (__driver_attach) from [<c0b05a84>] (bus_for_each_dev+0x68/0xb4)
[    2.415539] [<c0b05a84>] (bus_for_each_dev) from [<c0b06b68>] (bus_add_driver+0x104/0x20c)
[    2.423763] [<c0b06b68>] (bus_add_driver) from [<c0b0909c>] (driver_register+0x78/0x10c)
[    2.431815] [<c0b0909c>] (driver_register) from [<c030313c>] (do_one_initcall+0x8c/0x3ac)
[    2.439954] [<c030313c>] (do_one_initcall) from [<c1f013f4>] (kernel_init_freeable+0x468/0x548)
[    2.448610] [<c1f013f4>] (kernel_init_freeable) from [<c12344d8>] (kernel_init+0x8/0x10c)
[    2.456745] [<c12344d8>] (kernel_init) from [<c03010b4>] (ret_from_fork+0x14/0x20)
[    2.464273] Exception stack(0xea89ffb0 to 0xea89fff8)
[    2.469297] ffa0:                                     00000000 00000000 00000000 00000000
[    2.477432] ffc0: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
[    2.485566] ffe0: 00000000 00000000 00000000 00000000 00000013 00000000

This behavior has been introduced during the rework commit ff54571a747b ("ptp_qoriq: convert to use ptp_qoriq_init/free"), so technically it is a fix for that.

But I couldn't find a justification to just move the spin_lock_init a few lines above, because I don't see why the ptp_qoriq_settime call should be done at all at probe time. Writing a CLOCK_REALTIME value into a timer that is supposed to track CLOCK_TAI means there will be a 37 second offset that's going to render it useless for any practical purpose. Even furthermore, there may be concurrency with RTC drivers at probe time which is not handled (or there may be no RTC on the platform at all), so in practice even if the code is there, the ptp_qoriq clock may still tick in Jan 1st, 1970 until a better clock source becomes available.

So just remove the ptp_qoriq_settime call and let the clock tick in 1970 until user space does something about it. Most other PTP drivers do the same, except chelsio cxgb4 and maybe a few others.

Quoted 6 lines

So while this is in fact changing the behavior that the ptp_qoriq (formerly gianfar_ptp) has had since day one, it is only fixing a recent rework commit.

Fixes ff54571a747b ("ptp_qoriq: convert to use ptp_qoriq_init/free")

Richard, please review.

Quoted 29 lines
---
 drivers/ptp/ptp_qoriq.c | 4 ----
 1 file changed, 4 deletions(-)
drivers/ptp/ptp_qoriq.c +0 -4
index c61f00b72e15..5dbcca2332d5 100644
@@ -450,7 +450,6 @@ int ptp_qoriq_init(struct ptp_qoriq *ptp_qoriq, void __iomem *base,
 {
 	struct device_node *node = ptp_qoriq->dev->of_node;
 	struct ptp_qoriq_registers *regs;
-	struct timespec64 now;
 	unsigned long flags;
 	u32 tmr_ctrl;
 
@@ -507,9 +506,6 @@ int ptp_qoriq_init(struct ptp_qoriq *ptp_qoriq, void __iomem *base,
 		ptp_qoriq->regs.etts_regs = base + ETTS_REGS_OFFSET;
 	}
 
-	ktime_get_real_ts64(&now);
-	ptp_qoriq_settime(&ptp_qoriq->caps, &now);
-
 	tmr_ctrl =
 	  (ptp_qoriq->tclk_period & TCLK_PERIOD_MASK) << TCLK_PERIOD_SHIFT |
 	  (ptp_qoriq->cksel & CKSEL_MASK) << CKSEL_SHIFT;
Signature · 1 line
2.17.1
maintainer
#2 lore raw
In reply to Vladimir Oltean
Quoted 6 lines · Vladimir Oltean

But I couldn't find a justification to just move the spin_lock_init a few lines above, because I don't see why the ptp_qoriq_settime call should be done at all at probe time. Writing a CLOCK_REALTIME value into a timer that is supposed to track CLOCK_TAI means there will be a 37 second offset that's going to render it useless for any practical purpose.

Rhetorical question: Is off-by-37 more wrong than 1970?

Some PHC drivers start counting from zero (1970), and some from the system time (which might be any value at all). It really doesn't matter what the initial value is, IMHO, because it will always be wrong.

Quoted 3 lines

So just remove the ptp_qoriq_settime call and let the clock tick in 1970 until user space does something about it. Most other PTP drivers do the same, except chelsio cxgb4 and maybe a few others.

Not sure about "most".

Please fix the actual bug, the spin lock issue, and don't worry about changing the initial value.

Thanks,
Richard