[PATCH v4 0/3] Optimize tlbflush path

[PATCH v4 0/3] Optimize tlbflush path

[Atish Patra]

This series adds few optimizations to reduce the trap cost in the tlb
flush path. We should only make SBI calls to remote tlb flush only if
absolutely required. 

This series is based on Christoph's series:

http://lists.infradead.org/pipermail/linux-riscv/2019-August/006148.html

Changes from v3->v4
1. Simplify the local vs remote usecase.
2. Reorder the patches in the series.

Changes from v2->v3:
1. Split the patches into smaller one per optimization.

Atish Patra (3):
RISC-V: Do not invoke SBI call if cpumask is empty
RISC-V: Issue a local tlbflush if possible.
RISC-V: Issue a tlb page flush if possible

arch/riscv/mm/tlbflush.c | 25 +++++++++++++++++++++++--
1 file changed, 23 insertions(+), 2 deletions(-)

--
2.21.0

[PATCH v4 1/3] RISC-V: Do not invoke SBI call if cpumask is empty

[Atish Patra]

SBI calls are expensive. If cpumask is empty, there is no need to
trap via SBI as no remote tlb flushing is required.

Signed-off-by: Atish Patra <atish.patra@wdc.com>
---
 arch/riscv/mm/tlbflush.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/arch/riscv/mm/tlbflush.c b/arch/riscv/mm/tlbflush.c
index df93b26f1b9d..1293b8017ee0 100644
--- a/arch/riscv/mm/tlbflush.c
+++ b/arch/riscv/mm/tlbflush.c
@@ -14,6 +14,9 @@ static void __sbi_tlb_flush_range(struct cpumask *cmask, unsigned long start,
 {
 	struct cpumask hmask;
 
+	if (cpumask_empty(cmask))
+		return;
+
 	riscv_cpuid_to_hartid_mask(cmask, &hmask);
 	sbi_remote_sfence_vma(hmask.bits, start, size);
 }
-- 
2.21.0

[PATCH v4 2/3] RISC-V: Issue a local tlbflush if possible.

[Atish Patra]

In RISC-V, tlb flush happens via SBI which is expensive. If the local
cpu is the only cpu in cpumask, there is no need to invoke a SBI call.

Just do a local flush and return.

Signed-off-by: Atish Patra <atish.patra@wdc.com>
---
 arch/riscv/mm/tlbflush.c | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)

diff --git a/arch/riscv/mm/tlbflush.c b/arch/riscv/mm/tlbflush.c
index 1293b8017ee0..8172fbf46123 100644
--- a/arch/riscv/mm/tlbflush.c
+++ b/arch/riscv/mm/tlbflush.c
@@ -2,6 +2,7 @@
 
 #include <linux/mm.h>
 #include <linux/smp.h>
+#include <linux/sched.h>
 #include <asm/sbi.h>
 
 void flush_tlb_all(void)
@@ -9,16 +10,30 @@ void flush_tlb_all(void)
 	sbi_remote_sfence_vma(NULL, 0, -1);
 }
 
+/*
+ * This function must not be called with cmask being null.
+ * Kernel may panic if cmask is NULL.
+ */
 static void __sbi_tlb_flush_range(struct cpumask *cmask, unsigned long start,
 		unsigned long size)
 {
 	struct cpumask hmask;
+	unsigned int cpuid;
 
 	if (cpumask_empty(cmask))
 		return;
 
-	riscv_cpuid_to_hartid_mask(cmask, &hmask);
-	sbi_remote_sfence_vma(hmask.bits, start, size);
+	cpuid = get_cpu();
+
+	if (cpumask_any_but(cmask, cpuid) >= nr_cpu_ids) {
+		/* local cpu is the only cpu present in cpumask */
+		local_flush_tlb_all();
+	} else {
+		riscv_cpuid_to_hartid_mask(cmask, &hmask);
+		sbi_remote_sfence_vma(cpumask_bits(&hmask), start, size);
+	}
+
+	put_cpu();
 }
 
 void flush_tlb_mm(struct mm_struct *mm)
-- 
2.21.0

[PATCH v4 3/3] RISC-V: Issue a tlb page flush if possible

[Atish Patra]

If tlbflush request is for page only, there is no need to do a
complete local tlb shootdown.

Just do a local tlb flush for the given address.

Signed-off-by: Atish Patra <atish.patra@wdc.com>
---
 arch/riscv/mm/tlbflush.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/arch/riscv/mm/tlbflush.c b/arch/riscv/mm/tlbflush.c
index 8172fbf46123..b1c04751bcf1 100644
--- a/arch/riscv/mm/tlbflush.c
+++ b/arch/riscv/mm/tlbflush.c
@@ -27,7 +27,10 @@ static void __sbi_tlb_flush_range(struct cpumask *cmask, unsigned long start,
 
 	if (cpumask_any_but(cmask, cpuid) >= nr_cpu_ids) {
 		/* local cpu is the only cpu present in cpumask */
-		local_flush_tlb_all();
+		if (size <= PAGE_SIZE)
+			local_flush_tlb_page(start);
+		else
+			local_flush_tlb_all();
 	} else {
 		riscv_cpuid_to_hartid_mask(cmask, &hmask);
 		sbi_remote_sfence_vma(cpumask_bits(&hmask), start, size);
-- 
2.21.0

Re: [PATCH v4 1/3] RISC-V: Do not invoke SBI call if cpumask is empty

[Christoph Hellwig]

On Thu, Aug 22, 2019 at 12:51:49AM -0700, Atish Patra wrote:
> SBI calls are expensive. If cpumask is empty, there is no need to
> trap via SBI as no remote tlb flushing is required.
> 
> Signed-off-by: Atish Patra <atish.patra@wdc.com>

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [PATCH v4 2/3] RISC-V: Issue a local tlbflush if possible.

[Christoph Hellwig]

On Thu, Aug 22, 2019 at 12:51:50AM -0700, Atish Patra wrote:
> In RISC-V, tlb flush happens via SBI which is expensive. If the local
> cpu is the only cpu in cpumask, there is no need to invoke a SBI call.
> 
> Just do a local flush and return.
> 
> Signed-off-by: Atish Patra <atish.patra@wdc.com>

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [PATCH v4 3/3] RISC-V: Issue a tlb page flush if possible

[Christoph Hellwig]

On Thu, Aug 22, 2019 at 12:51:51AM -0700, Atish Patra wrote:
> If tlbflush request is for page only, there is no need to do a
> complete local tlb shootdown.
> 
> Just do a local tlb flush for the given address.

Looks good, although I suspect in many cases even doing multiple
single-page sfence.vma calls might be cheaper than the global one.

But I think that is worth a ѕeparate discussion, preferably with actual
numbers.

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [PATCH v4 3/3] RISC-V: Issue a tlb page flush if possible

[Atish Patra]

On Thu, 2019-08-22 at 10:11 +0200, Christoph Hellwig wrote:
> On Thu, Aug 22, 2019 at 12:51:51AM -0700, Atish Patra wrote:
> > If tlbflush request is for page only, there is no need to do a
> > complete local tlb shootdown.
> > 
> > Just do a local tlb flush for the given address.
> 
> Looks good, although I suspect in many cases even doing multiple
> single-page sfence.vma calls might be cheaper than the global one.
> 
> But I think that is worth a ѕeparate discussion, preferably with
> actual
> numbers.
> 

Yup. Finding a good threashold is always tricky without real
benchmarks.

> Reviewed-by: Christoph Hellwig <hch@lst.de>

Thanks for the review.

-- 
Regards,
Atish

Re: [PATCH v4 0/3] Optimize tlbflush path

[Paul Walmsley]

Hi Atish,

On Thu, 22 Aug 2019, Atish Patra wrote:

> This series adds few optimizations to reduce the trap cost in the tlb
> flush path. We should only make SBI calls to remote tlb flush only if
> absolutely required. 

The patches look great.  My understanding is that these optimization 
patches may actually be a partial workaround for the TLB flushing bug that 
we've been looking at for the last month or so, which can corrupt memory 
or crash the system.

If that's the case, let's first root-cause the underlying bug.  Otherwise 
we'll just be papering over the actual issue, which probably could still 
occur even with this series, correct?  Since it contains no explicit 
fixes?


- Paul

Re: [PATCH v4 0/3] Optimize tlbflush path

[Atish Patra]

On Fri, 2019-08-30 at 19:50 -0700, Paul Walmsley wrote:
> Hi Atish,
> 
> On Thu, 22 Aug 2019, Atish Patra wrote:
> 
> > This series adds few optimizations to reduce the trap cost in the
> > tlb
> > flush path. We should only make SBI calls to remote tlb flush only
> > if
> > absolutely required. 
> 
> The patches look great.  My understanding is that these optimization 
> patches may actually be a partial workaround for the TLB flushing bug
> that 
> we've been looking at for the last month or so, which can corrupt
> memory 
> or crash the system.
> 
> If that's the case, let's first root-cause the underlying
> bug.  Otherwise 
> we'll just be papering over the actual issue, which probably could
> still 
> occur even with this series, correct?  Since it contains no explicit 
> fixes?
> 
> 
I have verified the glibc locale install issue both in Qemu and
Unleashed. I don't see any issue with OpenSBI master + Linux v5.3
kernel.

As per our investigation, it looks like a hardware errata with
Unleashed board as the memory corruption issue only happens in case of
tlb range flush. In RISC-V, sfence.vma can only be issued at page
boundary. If the range is larger than that, OpenSBI has to issue
multiple sfence.vma calls back to back leading to possible memory
corruption.

Currently, OpenSBI has a platform feature i.e. "tlb_range_flush_limit"
that allows to configure tlb flush threshold per platform. Any tlb
flush range request greater than this threshold, is converted to a full
flush. Currently, it is set to the default value 4K for every
platform[1]. Glibc locale install memory corruption only happens if
this threshold is changed to a higher value i.e. 1G. This doesn't
change anything in OpenSBI code path except the fact that it will issue
many sfence.vma instructions back to back instead of one. 

If the hardware team at SiFive can look into this as well, it would be
great.

To conclude, we think this issue need to be investigated by hardware
team and the kernel patch can be merged to get the performance benefit.

[1] 
https://github.com/riscv/opensbi/blob/master/include/sbi/sbi_platform.h#L40



> - Paul
> 
> _______________________________________________
> linux-riscv mailing list
> linux-riscv@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-riscv

-- 
Regards,
Atish

Re: [PATCH v4 0/3] Optimize tlbflush path

[Paul Walmsley]

On Fri, 30 Aug 2019, Paul Walmsley wrote:

> On Thu, 22 Aug 2019, Atish Patra wrote:
> 
> > This series adds few optimizations to reduce the trap cost in the tlb
> > flush path. We should only make SBI calls to remote tlb flush only if
> > absolutely required. 
> 
> The patches look great.  My understanding is that these optimization 
> patches may actually be a partial workaround for the TLB flushing bug that 
> we've been looking at for the last month or so, which can corrupt memory 
> or crash the system.

I don't think we're any closer to root-causing this issue.  Meanwhile, 
OpenSBI has merged patches to work around it.  So since many of us have 
looked at Atish's TLB optimization patches, and we all think they are 
useful optimizations, let's plan to merge it for v5.5-rc1.


- Paul

Re: [PATCH v4 0/3] Optimize tlbflush path

[Atish Patra]

On Thu, 2019-10-24 at 20:09 -0700, Paul Walmsley wrote:
> On Fri, 30 Aug 2019, Paul Walmsley wrote:
> 
> > On Thu, 22 Aug 2019, Atish Patra wrote:
> > 
> > > This series adds few optimizations to reduce the trap cost in the
> > > tlb
> > > flush path. We should only make SBI calls to remote tlb flush
> > > only if
> > > absolutely required. 
> > 
> > The patches look great.  My understanding is that these
> > optimization 
> > patches may actually be a partial workaround for the TLB flushing
> > bug that 
> > we've been looking at for the last month or so, which can corrupt
> > memory 
> > or crash the system.
> 
> I don't think we're any closer to root-causing this
> issue.  Meanwhile, 
> OpenSBI has merged patches to work around it.  So since many of us
> have 
> looked at Atish's TLB optimization patches, and we all think they
> are 
> useful optimizations, let's plan to merge it for v5.5-rc1.
> 
> 

Thanks. These patches still applies cleanly on 5.4-rc5.
Let me know if you need something from myside.

> - Paul

-- 
Regards,
Atish