Skip to contentSkip to conversation

Revert "clocksource/drivers/riscv: Events are stopped during CPU suspend"

PATCH Applied Conor Dooley 3 participants 4 messages Nov 22, 2022 – Dec 1, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 Revert "clocksource/drivers/riscv: Events are stopped during CPU suspend"

This reverts commit 232ccac1bd9b5bfe73895f527c08623e7fa0752d.

On the subject of suspend, the RISC-V SBI spec states:

Quoted 5 lines

Request the SBI implementation to put the calling hart in a platform specific suspend (or low power) state specified by the suspend_type parameter. The hart will automatically come out of suspended state and resume normal execution when it receives an interrupt or platform specific hardware event.

This does not cover whether any given events actually reach the hart or not, just what the hart will do if it receives an event. On PolarFire SoC, and potentially other SiFive based implementations, events from the RISC-V timer do reach a hart during suspend. This is not the case for the implementation on the Allwinner D1 - there timer events are not received during suspend.

To fix this, the x86 C3STOP feature was enabled for the timer driver - but this has broken both RCU stall detection and timers generally on PolarFire SoC (and potentially other SiFive based implementations).

If an AXI read to the PCIe controller on PolarFire SoC times out, the system will stall, however, with this patch applied, the system just locks up without RCU stalling:
io scheduler mq-deadline registered
io scheduler kyber registered
microchip-pcie 2000000000.pcie: host bridge /soc/pcie@2000000000 ranges:
microchip-pcie 2000000000.pcie: MEM 0x2008000000..0x2087ffffff -> 0x0008000000
microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: axi read request error microchip-pcie 2000000000.pcie: axi read timeout microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer Freeing initrd memory: 7332K

Similarly issues were reported with clock_nanosleep() - with a test app that sleeps each cpu for 6, 5, 4, 3 ms respectively, HZ=250 & the blamed commit in place, the sleep times are rounded up to the next jiffy:

== CPU: 1 ==      == CPU: 2 ==      == CPU: 3 ==      == CPU: 4 ==
Mean: 7.974992    Mean: 7.976534    Mean: 7.962591    Mean: 3.952179

Std Dev: 0.154374 Std Dev: 0.156082 Std Dev: 0.171018 Std Dev: 0.076193
Hi: 9.472000 Hi: 10.495000 Hi: 8.864000 Hi: 4.736000
Lo: 6.087000 Lo: 6.380000 Lo: 4.872000 Lo: 3.403000
Samples: 521 Samples: 521 Samples: 521 Samples: 521

Fortunately, the D1 has a second timer, which is "currently used in preference to the RISC-V/SBI timer driver" so a revert here does not hurt operation of D1 in it's current form.

Ultimately, a DeviceTree property (or node) will be added to encode the behaviour of the timers, but until then revert the addition of CLOCK_EVT_FEAT_C3STOP.

---

For v2, I have re-worked the commit message to (hopefully) add improved context.

---
 drivers/clocksource/timer-riscv.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/clocksource/timer-riscv.c +1 -1
index 969a552da8d2..a0d66fabf073 100644
@@ -51,7 +51,7 @@ static int riscv_clock_next_event(unsigned long delta,
 static unsigned int riscv_clock_event_irq;
 static DEFINE_PER_CPU(struct clock_event_device, riscv_clock_event) = {
 	.name			= "riscv_timer_clockevent",
-	.features		= CLOCK_EVT_FEAT_ONESHOT | CLOCK_EVT_FEAT_C3STOP,
+	.features		= CLOCK_EVT_FEAT_ONESHOT,
 	.rating			= 100,
 	.set_next_event		= riscv_clock_next_event,
 };
Signature · 1 line
2.38.1
Acked-by on v2 1/1
#1 lore raw
Quoted 9 lines · Conor Dooley

This reverts commit 232ccac1bd9b5bfe73895f527c08623e7fa0752d.

On the subject of suspend, the RISC-V SBI spec states:

Quoted 5 lines

Request the SBI implementation to put the calling hart in a platform specific suspend (or low power) state specified by the suspend_type parameter. The hart will automatically come out of suspended state and resume normal execution when it receives an interrupt or platform specific hardware event.

Quoted 8 lines

This does not cover whether any given events actually reach the hart or not, just what the hart will do if it receives an event. On PolarFire SoC, and potentially other SiFive based implementations, events from the RISC-V timer do reach a hart during suspend. This is not the case for the implementation on the Allwinner D1 - there timer events are not received during suspend.

To fix this, the x86 C3STOP feature was enabled for the timer driver -

C3STOP isn't inherently x86-specific.

Quoted 28 lines

but this has broken both RCU stall detection and timers generally on PolarFire SoC (and potentially other SiFive based implementations).

If an AXI read to the PCIe controller on PolarFire SoC times out, the system will stall, however, with this patch applied, the system just locks up without RCU stalling:
io scheduler mq-deadline registered
io scheduler kyber registered
microchip-pcie 2000000000.pcie: host bridge /soc/pcie@2000000000 ranges:
microchip-pcie 2000000000.pcie: MEM 0x2008000000..0x2087ffffff -> 0x0008000000
microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: axi read request error microchip-pcie 2000000000.pcie: axi read timeout microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer Freeing initrd memory: 7332K

Similarly issues were reported with clock_nanosleep() - with a test app that sleeps each cpu for 6, 5, 4, 3 ms respectively, HZ=250 & the blamed commit in place, the sleep times are rounded up to the next jiffy:

== CPU: 1 ==      == CPU: 2 ==      == CPU: 3 ==      == CPU: 4 ==
Mean: 7.974992    Mean: 7.976534    Mean: 7.962591    Mean: 3.952179
Quoted 8 lines

Std Dev: 0.154374 Std Dev: 0.156082 Std Dev: 0.171018 Std Dev: 0.076193
Hi: 9.472000 Hi: 10.495000 Hi: 8.864000 Hi: 4.736000
Lo: 6.087000 Lo: 6.380000 Lo: 4.872000 Lo: 3.403000
Samples: 521 Samples: 521 Samples: 521 Samples: 521

Fortunately, the D1 has a second timer, which is "currently used in preference to the RISC-V/SBI timer driver" so a revert here does not hurt operation of D1 in it's current form.

typo: its

Quoted 36 lines

Ultimately, a DeviceTree property (or node) will be added to encode the behaviour of the timers, but until then revert the addition of CLOCK_EVT_FEAT_C3STOP.

---

For v2, I have re-worked the commit message to (hopefully) add improved context.

---
 drivers/clocksource/timer-riscv.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/clocksource/timer-riscv.c +0 -0
index 969a552da8d2..a0d66fabf073 100644

drivers/clocksource/timer-riscv.c static int riscv_clock_next_event(unsigned long delta, · line 57

Quoted 8 lines
drivers/clocksource/timer-riscv.c +1 -1
@@ -51,7 +51,7 @@ static int riscv_clock_next_event(unsigned long delta,
 static unsigned int riscv_clock_event_irq;
 static DEFINE_PER_CPU(struct clock_event_device, riscv_clock_event) = {
 	.name			= "riscv_timer_clockevent",
-	.features		= CLOCK_EVT_FEAT_ONESHOT | CLOCK_EVT_FEAT_C3STOP,
+	.features		= CLOCK_EVT_FEAT_ONESHOT,
 	.rating			= 100,
 	.set_next_event		= riscv_clock_next_event,
 };
#2 lore raw

Hey Samuel,

Quoted 6 lines · Samuel Holland
Quoted 3 lines · Conor Dooley

This reverts commit 232ccac1bd9b5bfe73895f527c08623e7fa0752d.

To fix this, the x86 C3STOP feature was enabled for the timer driver -

C3STOP isn't inherently x86-specific.

I think I originally had feature with "s around it & meant this as a tongue-in-cheek reference to the header, which describes it as an "x86 (mis)feature" or something like that. Think I decided against that but forgot to drop the x86 bit..
Could easily do s/x86// and it'd still make sense.

Quoted 5 lines
Quoted 3 lines

Fortunately, the D1 has a second timer, which is "currently used in preference to the RISC-V/SBI timer driver" so a revert here does not hurt operation of D1 in it's current form.

typo: its

Good spot :)

Quoted 1 line

Thanks!

Perhaps the two minor commit message bits could be fixed on application? Otherwise, I will send a reworded one in a few days.

Conor.

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

merge bot Applied
#3 lore raw
In reply to Conor Dooley
The following commit has been merged into the timers/urgent branch of tip:

The following commit has been merged into the timers/urgent branch of tip:

Commit-ID:     d9f15a9de44affe733e34f93bc184945ba277e6d
Gitweb:        https://git.kernel.org/tip/d9f15a9de44affe733e34f93bc184945ba277e6d
Author:        Conor Dooley <[email protected]>
AuthorDate:    Tue, 22 Nov 2022 12:16:21 
Committer:     Thomas Gleixner <[email protected]>
CommitterDate: Thu, 01 Dec 2022 12:05:29 +01:00

Revert "clocksource/drivers/riscv: Events are stopped during CPU suspend"

This reverts commit 232ccac1bd9b5bfe73895f527c08623e7fa0752d.

On the subject of suspend, the RISC-V SBI spec states:

This does not cover whether any given events actually reach the hart or not, just what the hart will do if it receives an event. On PolarFire SoC, and potentially other SiFive based implementations, events from the RISC-V timer do reach a hart during suspend. This is not the case for the implementation on the Allwinner D1 - there timer events are not received during suspend.

To fix this, the CLOCK_EVT_FEAT_C3STOP (mis)feature was enabled for the timer driver - but this has broken both RCU stall detection and timers generally on PolarFire SoC and potentially other SiFive based implementations.

If an AXI read to the PCIe controller on PolarFire SoC times out, the system will stall, however, with CLOCK_EVT_FEAT_C3STOP active, the system just locks up without RCU stalling:

io scheduler mq-deadline registered
io scheduler kyber registered
microchip-pcie 2000000000.pcie: host bridge /soc/pcie@2000000000 ranges:
microchip-pcie 2000000000.pcie: MEM 0x2008000000..0x2087ffffff -> 0x0008000000
microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: axi read request error microchip-pcie 2000000000.pcie: axi read timeout microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer microchip-pcie 2000000000.pcie: sec error in pcie2axi buffer microchip-pcie 2000000000.pcie: ded error in pcie2axi buffer Freeing initrd memory: 7332K

Similarly issues were reported with clock_nanosleep() - with a test app that sleeps each cpu for 6, 5, 4, 3 ms respectively, HZ=250 & the blamed commit in place, the sleep times are rounded up to the next jiffy:

== CPU: 1 ==      == CPU: 2 ==      == CPU: 3 ==      == CPU: 4 ==
Mean: 7.974992    Mean: 7.976534    Mean: 7.962591    Mean: 3.952179

Std Dev: 0.154374 Std Dev: 0.156082 Std Dev: 0.171018 Std Dev: 0.076193
Hi: 9.472000 Hi: 10.495000 Hi: 8.864000 Hi: 4.736000
Lo: 6.087000 Lo: 6.380000 Lo: 4.872000 Lo: 3.403000
Samples: 521 Samples: 521 Samples: 521 Samples: 521

Fortunately, the D1 has a second timer, which is "currently used in preference to the RISC-V/SBI timer driver" so a revert here does not hurt operation of D1 in its current form.

Ultimately, a DeviceTree property (or node) will be added to encode the behaviour of the timers, but until then revert the addition of CLOCK_EVT_FEAT_C3STOP.

---
 drivers/clocksource/timer-riscv.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/clocksource/timer-riscv.c +1 -1
index 969a552..a0d66fa 100644
@@ -51,7 +51,7 @@ static int riscv_clock_next_event(unsigned long delta,
 static unsigned int riscv_clock_event_irq;
 static DEFINE_PER_CPU(struct clock_event_device, riscv_clock_event) = {
 	.name			= "riscv_timer_clockevent",
-	.features		= CLOCK_EVT_FEAT_ONESHOT | CLOCK_EVT_FEAT_C3STOP,
+	.features		= CLOCK_EVT_FEAT_ONESHOT,
 	.rating			= 100,
 	.set_next_event		= riscv_clock_next_event,
 };