Skip to contentSkip to conversation

irqchip: riscv: Order normal writes and IPI writes

PATCH Superseded Xu Lu 4 participants 6 messages Jan 16, 2025 – Jan 20, 2025

Subsystems (3)
lore
Share & tools
lore

Xu Lu

author
#0 lore raw

PATCH 1/1 irqchip: riscv: Order normal writes and IPI writes

Replace writel_relaxed() with writel() when issuing IPI to ensure all previous write operations made by current CPU are visible to other CPUs.

---
 drivers/irqchip/irq-riscv-imsic-early.c      | 2 +-
 drivers/irqchip/irq-thead-c900-aclint-sswi.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/irqchip/irq-riscv-imsic-early.c +1 -1
index c5c2e6929a2f..275df5005705 100644
@@ -27,7 +27,7 @@ static void imsic_ipi_send(unsigned int cpu)
 {
 	struct imsic_local_config *local = per_cpu_ptr(imsic->global.local, cpu);
 
-	writel_relaxed(IMSIC_IPI_ID, local->msi_va);
+	writel(IMSIC_IPI_ID, local->msi_va);
 }
 
 static void imsic_ipi_starting_cpu(void)
drivers/irqchip/irq-thead-c900-aclint-sswi.c +1 -1
index b0e366ade427..8ff6e7a1363b 100644
@@ -31,7 +31,7 @@ static DEFINE_PER_CPU(void __iomem *, sswi_cpu_regs);
 
 static void thead_aclint_sswi_ipi_send(unsigned int cpu)
 {
-	writel_relaxed(0x1, per_cpu(sswi_cpu_regs, cpu));
+	writel(0x1, per_cpu(sswi_cpu_regs, cpu));
 }
 
 static void thead_aclint_sswi_ipi_clear(void)
Signature · 1 line
2.20.1
#1 lore raw
Quoted 2 lines · Xu Lu

Replace writel_relaxed() with writel() when issuing IPI to ensure all previous write operations made by current CPU are visible to other CPUs.

Did you experience an ordering issue from this?

- Charlie

Quoted 41 lines
---
 drivers/irqchip/irq-riscv-imsic-early.c      | 2 +-
 drivers/irqchip/irq-thead-c900-aclint-sswi.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/irqchip/irq-riscv-imsic-early.c +1 -1
index c5c2e6929a2f..275df5005705 100644
@@ -27,7 +27,7 @@ static void imsic_ipi_send(unsigned int cpu)
 {
 	struct imsic_local_config *local = per_cpu_ptr(imsic->global.local, cpu);
 
-	writel_relaxed(IMSIC_IPI_ID, local->msi_va);
+	writel(IMSIC_IPI_ID, local->msi_va);
 }
 
 static void imsic_ipi_starting_cpu(void)
drivers/irqchip/irq-thead-c900-aclint-sswi.c +1 -1
index b0e366ade427..8ff6e7a1363b 100644
@@ -31,7 +31,7 @@ static DEFINE_PER_CPU(void __iomem *, sswi_cpu_regs);
 
 static void thead_aclint_sswi_ipi_send(unsigned int cpu)
 {
-	writel_relaxed(0x1, per_cpu(sswi_cpu_regs, cpu));
+	writel(0x1, per_cpu(sswi_cpu_regs, cpu));
 }
 
 static void thead_aclint_sswi_ipi_clear(void)
Signature · 7 lines
2.20.1


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

Charlie!

Quoted 5 lines · Charlie Jenkins
Quoted 2 lines · Xu Lu

Replace writel_relaxed() with writel() when issuing IPI to ensure all previous write operations made by current CPU are visible to other CPUs.

Did you experience an ordering issue from this?

That's not the right question.

CPU 0 CPU 1
store A // data
store B // IPI
IPI handler
load A

The real question is whether the RISC-V memory model guarantees under all circumstances that A is globally visible before the IPI handler load. If so, then the writel_relaxed() is fine. If not, the fence is required.

That's not a question of observation. It's a question of facts defined by the architecture. People have wasted months to analyze such fails which tend to happen once every blue moon with no other trace than "something went wrong" ....

Quoted 1 line

- Charlie

Please trim your replies...

Thanks,

tglx

maintainer
#3 lore raw

Hi Thomas,

Quoted 23 lines · Thomas Gleixner

Charlie!

Quoted 5 lines · Charlie Jenkins
Quoted 2 lines · Xu Lu

Replace writel_relaxed() with writel() when issuing IPI to ensure all previous write operations made by current CPU are visible to other CPUs.

Did you experience an ordering issue from this?

That's not the right question.

CPU 0 CPU 1
store A // data
store B // IPI
IPI handler
load A

The real question is whether the RISC-V memory model guarantees under all circumstances that A is globally visible before the IPI handler load. If so, then the writel_relaxed() is fine. If not, the fence is required.

Quoted 4 lines

That's not a question of observation. It's a question of facts defined by the architecture. People have wasted months to analyze such fails which tend to happen once every blue moon with no other trace than "something went wrong" ....

The RISC-V FENCE instruction distinguishes between normal memory operations and I/O operations in its predecessor and successor sets where r = normal read, w = normal write, i = I/O read, and o = I/O write.

The ipi_mux_send_mask() does smp_mb__after_atomic() (equals to "fence rw,rw") before calling imsic_ipi_send(). This prevents ordering of normal memory writes in imsic_ipi_send() before smp_mb__after_atomic() in ipi_mux_send_mask() but it does not prevent I/O memory writes in imsic_ipi_send() to be ordered before smp_mb__after_atomic().

This means currently nothing prevents the I/O memory write in imsic_ipi_send() to be ordered before normal memory writes in ipi_mux_send_mask() hence there is a very unlikely possibility of an IPI handler on the target CPU seeing incorrect data.

The conversion of writel_relaxed() to writel() in imsic_ipi_send() adds a "fence w,o" before the actual I/O memory write which ensures that I/O memory write is not ordered before normal memory writes.

Based on the above, the conversion of writel_relaxed() to writel() in imsic_ipi_send() looks good to me.

Regards,
Anup

#4 lore raw
Quoted 38 lines · Anup Patel
Quoted 20 lines · Thomas Gleixner
Quoted 5 lines · Charlie Jenkins
Quoted 2 lines · Xu Lu

Replace writel_relaxed() with writel() when issuing IPI to ensure all previous write operations made by current CPU are visible to other CPUs.

Did you experience an ordering issue from this?

That's not the right question.

CPU 0 CPU 1
store A // data
store B // IPI
IPI handler
load A

The real question is whether the RISC-V memory model guarantees under all circumstances that A is globally visible before the IPI handler load. If so, then the writel_relaxed() is fine. If not, the fence is required.

Quoted 4 lines

That's not a question of observation. It's a question of facts defined by the architecture. People have wasted months to analyze such fails which tend to happen once every blue moon with no other trace than "something went wrong" ....

The RISC-V FENCE instruction distinguishes between normal memory operations and I/O operations in its predecessor and successor sets where r = normal read, w = normal write, i = I/O read, and o = I/O write.

The ipi_mux_send_mask() does smp_mb__after_atomic() (equals to "fence rw,rw") before calling imsic_ipi_send(). This prevents ordering of normal memory writes in imsic_ipi_send() before smp_mb__after_atomic() in ipi_mux_send_mask() but it does not prevent I/O memory writes in imsic_ipi_send() to be ordered before smp_mb__after_atomic().

Quoted 4 lines

This means currently nothing prevents the I/O memory write in imsic_ipi_send() to be ordered before normal memory writes in ipi_mux_send_mask() hence there is a very unlikely possibility of an IPI handler on the target CPU seeing incorrect data.

Very unlikely is a valid assumption for a single system, but at scale it becomes very likely :)

Quoted 7 lines

The conversion of writel_relaxed() to writel() in imsic_ipi_send() adds a "fence w,o" before the actual I/O memory write which ensures that I/O memory write is not ordered before normal memory writes.

Based on the above, the conversion of writel_relaxed() to writel() in imsic_ipi_send() looks good to me.

Can we please have something like the above in the change log so this is documented for posterity?

Thanks

tglx

author
#5 lore raw
Quoted 53 lines · Thomas Gleixner
Quoted 38 lines · Anup Patel
Quoted 20 lines · Thomas Gleixner
Quoted 5 lines · Charlie Jenkins
Quoted 2 lines · Xu Lu

Replace writel_relaxed() with writel() when issuing IPI to ensure all previous write operations made by current CPU are visible to other CPUs.

Did you experience an ordering issue from this?

That's not the right question.

CPU 0 CPU 1
store A // data
store B // IPI
IPI handler
load A

The real question is whether the RISC-V memory model guarantees under all circumstances that A is globally visible before the IPI handler load. If so, then the writel_relaxed() is fine. If not, the fence is required.

Quoted 4 lines

That's not a question of observation. It's a question of facts defined by the architecture. People have wasted months to analyze such fails which tend to happen once every blue moon with no other trace than "something went wrong" ....

The RISC-V FENCE instruction distinguishes between normal memory operations and I/O operations in its predecessor and successor sets where r = normal read, w = normal write, i = I/O read, and o = I/O write.

The ipi_mux_send_mask() does smp_mb__after_atomic() (equals to "fence rw,rw") before calling imsic_ipi_send(). This prevents ordering of normal memory writes in imsic_ipi_send() before smp_mb__after_atomic() in ipi_mux_send_mask() but it does not prevent I/O memory writes in imsic_ipi_send() to be ordered before smp_mb__after_atomic().

Quoted 4 lines

This means currently nothing prevents the I/O memory write in imsic_ipi_send() to be ordered before normal memory writes in ipi_mux_send_mask() hence there is a very unlikely possibility of an IPI handler on the target CPU seeing incorrect data.

Very unlikely is a valid assumption for a single system, but at scale it becomes very likely :)

Quoted 5 lines

The conversion of writel_relaxed() to writel() in imsic_ipi_send() adds a "fence w,o" before the actual I/O memory write which ensures that I/O memory write is not ordered before normal memory writes.

Quoted 5 lines
Quoted 2 lines

Based on the above, the conversion of writel_relaxed() to writel() in imsic_ipi_send() looks good to me.

Can we please have something like the above in the change log so this is documented for posterity?

Thanks for Anup's supplied information. I will refine the commit message and resend the patch.

Thanks,

Xu Lu

Quoted 4 lines

Thanks

tglx