Skip to contentSkip to conversation

KVM: s390: add counters for vsie performance

PATCH Superseded Nico Boehr 3 participants 8 messages Sep 4, 2023 – Sep 7, 2023

Subsystems (3)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Nico Boehr

author
#0 lore raw

v3 cover KVM: s390: add counters for vsie performance

v3:
---
* rename te -> entry (David)
* add counters for gmap reuse and gmap create (David)

v2:
---
* also count shadowing of pages (Janosch)
* fix naming of counters (Janosch)
* mention shadowing of multiple levels is counted in each level (Claudio)
* fix inaccuate commit description regarding gmap notifier (Claudio)

When running a guest-3 via VSIE, guest-1 needs to shadow the page table structures of guest-2.

To reflect changes of the guest-2 in the _shadowed_ page table structures, the _shadowing_ sturctures sometimes need to be rebuilt. Since this is a costly operation, it should be avoided whenever possible.

This series adds kvm stat counters to count the number of shadow gmaps created and a tracepoint whenever something is unshadowed. This is a first step to try and improve VSIE performance.

Please note that "KVM: s390: add tracepoint in gmap notifier" has some checkpatch --strict findings. I did not fix these since the tracepoint definition would then look completely different from all the other tracepoints in arch/s390/kvm/trace-s390.h. If you want me to fix that, please let me know.

While developing this, a question regarding the stat counters came up:
there's usually no locking involved when the stat counters are incremented. On s390, GCC accidentally seems to do the right thing(TM) most of the time by generating a agsi instruction (which should be atomic given proper alignment). However, it's not guaranteed, so would we rather want to go with an atomic for the stat counters to avoid losing events? Or do we just accept the fact that we might loose events sometimes? Is there anything that speaks against having an atomic in kvm_stat?

Nico Boehr (2):
KVM: s390: add stat counter for shadow gmap events
KVM: s390: add tracepoint in gmap notifier

 arch/s390/include/asm/kvm_host.h |  7 +++++++
 arch/s390/kvm/gaccess.c          |  7 +++++++
 arch/s390/kvm/kvm-s390.c         | 11 ++++++++++-
 arch/s390/kvm/trace-s390.h       | 23 +++++++++++++++++++++++
 arch/s390/kvm/vsie.c             |  5 ++++-
 5 files changed, 51 insertions(+), 2 deletions(-)
Signature · 1 line
2.41.0
author
#1 lore raw

PATCH v3 1/2 KVM: s390: add stat counter for shadow gmap events

The shadow gmap tracks memory of nested guests (guest-3). In certain scenarios, the shadow gmap needs to be rebuilt, which is a costly operation since it involves a SIE exit into guest-1 for every entry in the respective shadow level.

Add kvm stat counters when new shadow structures are created at various levels. Also add a counter gmap_shadow_create when a completely fresh shadow gmap is created as well as a counter gmap_shadow_reuse when an existing gmap is being reused.

Note that when several levels are shadowed at once, counters on all affected levels will be increased.

Also note that not all page table levels need to be present and a ASCE can directly point to e.g. a segment table. In this case, a new segment table will always be equivalent to a new shadow gmap and hence will be counted as gmap_shadow_create and not as gmap_shadow_segment.

---
 arch/s390/include/asm/kvm_host.h | 7 +++++++
 arch/s390/kvm/gaccess.c          | 7 +++++++
 arch/s390/kvm/kvm-s390.c         | 9 ++++++++-
 arch/s390/kvm/vsie.c             | 5 ++++-
 4 files changed, 26 insertions(+), 2 deletions(-)
arch/s390/include/asm/kvm_host.h +7 -0
index 91bfecb91321..5243cbd57e2f 100644
@@ -777,6 +777,13 @@ struct kvm_vm_stat {
 	u64 inject_service_signal;
 	u64 inject_virtio;
 	u64 aen_forward;
+	u64 gmap_shadow_create;
+	u64 gmap_shadow_reuse;
+	u64 gmap_shadow_r1_entry;
+	u64 gmap_shadow_r2_entry;
+	u64 gmap_shadow_r3_entry;
+	u64 gmap_shadow_sg_entry;
+	u64 gmap_shadow_pg_entry;
 };
 
 struct kvm_arch_memory_slot {
arch/s390/kvm/gaccess.c +7 -0
index 6d6bc19b37dc..ff8349d17b33 100644
@@ -1382,6 +1382,7 @@ static int kvm_s390_shadow_tables(struct gmap *sg, unsigned long saddr,
 				  unsigned long *pgt, int *dat_protection,
 				  int *fake)
 {
+	struct kvm *kvm;
 	struct gmap *parent;
 	union asce asce;
 	union vaddress vaddr;
@@ -1390,6 +1391,7 @@ static int kvm_s390_shadow_tables(struct gmap *sg, unsigned long saddr,
 
 	*fake = 0;
 	*dat_protection = 0;
+	kvm = sg->private;
 	parent = sg->parent;
 	vaddr.addr = saddr;
 	asce.val = sg->orig_asce;
@@ -1450,6 +1452,7 @@ static int kvm_s390_shadow_tables(struct gmap *sg, unsigned long saddr,
 		rc = gmap_shadow_r2t(sg, saddr, rfte.val, *fake);
 		if (rc)
 			return rc;
+		kvm->stat.gmap_shadow_r1_entry++;
 	}
 		fallthrough;
 	case ASCE_TYPE_REGION2: {
@@ -1478,6 +1481,7 @@ static int kvm_s390_shadow_tables(struct gmap *sg, unsigned long saddr,
 		rc = gmap_shadow_r3t(sg, saddr, rste.val, *fake);
 		if (rc)
 			return rc;
+		kvm->stat.gmap_shadow_r2_entry++;
 	}
 		fallthrough;
 	case ASCE_TYPE_REGION3: {
@@ -1515,6 +1519,7 @@ static int kvm_s390_shadow_tables(struct gmap *sg, unsigned long saddr,
 		rc = gmap_shadow_sgt(sg, saddr, rtte.val, *fake);
 		if (rc)
 			return rc;
+		kvm->stat.gmap_shadow_r3_entry++;
 	}
 		fallthrough;
 	case ASCE_TYPE_SEGMENT: {
@@ -1548,6 +1553,7 @@ static int kvm_s390_shadow_tables(struct gmap *sg, unsigned long saddr,
 		rc = gmap_shadow_pgt(sg, saddr, ste.val, *fake);
 		if (rc)
 			return rc;
+		kvm->stat.gmap_shadow_sg_entry++;
 	}
 	}
 	/* Return the parent address of the page table */
@@ -1618,6 +1624,7 @@ int kvm_s390_shadow_fault(struct kvm_vcpu *vcpu, struct gmap *sg,
 	pte.p |= dat_protection;
 	if (!rc)
 		rc = gmap_shadow_page(sg, saddr, __pte(pte.val));
+	vcpu->kvm->stat.gmap_shadow_pg_entry++;
 	ipte_unlock(vcpu->kvm);
 	mmap_read_unlock(sg->mm);
 	return rc;
arch/s390/kvm/kvm-s390.c +8 -1
index d1e768bcfe1d..9379471081fa 100644
@@ -66,7 +66,14 @@ const struct _kvm_stats_desc kvm_vm_stats_desc[] = {
 	STATS_DESC_COUNTER(VM, inject_pfault_done),
 	STATS_DESC_COUNTER(VM, inject_service_signal),
 	STATS_DESC_COUNTER(VM, inject_virtio),
-	STATS_DESC_COUNTER(VM, aen_forward)
+	STATS_DESC_COUNTER(VM, aen_forward),
+	STATS_DESC_COUNTER(VM, gmap_shadow_reuse),
+	STATS_DESC_COUNTER(VM, gmap_shadow_create),
+	STATS_DESC_COUNTER(VM, gmap_shadow_r1_entry),
+	STATS_DESC_COUNTER(VM, gmap_shadow_r2_entry),
+	STATS_DESC_COUNTER(VM, gmap_shadow_r3_entry),
+	STATS_DESC_COUNTER(VM, gmap_shadow_sg_entry),
+	STATS_DESC_COUNTER(VM, gmap_shadow_pg_entry),
 };
 
 const struct kvm_stats_header kvm_vm_stats_header = {
arch/s390/kvm/vsie.c +4 -1
index 61499293c2ac..02dcbe82a8e5 100644
@@ -1214,8 +1214,10 @@ static int acquire_gmap_shadow(struct kvm_vcpu *vcpu,
 	 * we're holding has been unshadowed. If the gmap is still valid,
 	 * we can safely reuse it.
 	 */
-	if (vsie_page->gmap && gmap_shadow_valid(vsie_page->gmap, asce, edat))
+	if (vsie_page->gmap && gmap_shadow_valid(vsie_page->gmap, asce, edat)) {
+		vcpu->kvm->stat.gmap_shadow_reuse++;
 		return 0;
+	}
 
 	/* release the old shadow - if any, and mark the prefix as unmapped */
 	release_gmap_shadow(vsie_page);
@@ -1223,6 +1225,7 @@ static int acquire_gmap_shadow(struct kvm_vcpu *vcpu,
 	if (IS_ERR(gmap))
 		return PTR_ERR(gmap);
 	gmap->private = vcpu->kvm;
+	vcpu->kvm->stat.gmap_shadow_create++;
 	WRITE_ONCE(vsie_page->gmap, gmap);
 	return 0;
 }
Signature · 1 line
2.41.0
Reviewed-by on v3 1/2
#2 lore raw

Re: [PATCH v3 1/2] KVM: s390: add stat counter for shadow gmap events

Quoted 13 lines · Nico Boehr

The shadow gmap tracks memory of nested guests (guest-3). In certain scenarios, the shadow gmap needs to be rebuilt, which is a costly operation since it involves a SIE exit into guest-1 for every entry in the respective shadow level.

Add kvm stat counters when new shadow structures are created at various levels. Also add a counter gmap_shadow_create when a completely fresh shadow gmap is created as well as a counter gmap_shadow_reuse when an existing gmap is being reused.

Note that when several levels are shadowed at once, counters on all affected levels will be increased.

Quoted 6 lines

Also note that not all page table levels need to be present and a ASCE can directly point to e.g. a segment table. In this case, a new segment table will always be equivalent to a new shadow gmap and hence will be counted as gmap_shadow_create and not as gmap_shadow_segment.

Signature · 3 lines
Cheers,

David / dhildenb
author
#3 lore raw

PATCH v3 2/2 KVM: s390: add tracepoint in gmap notifier

In reply to Nico Boehr

The gmap notifier is called for changes in table entries with the notifier bit set. To diagnose performance issues, it can be useful to see what causes certain changes in the gmap.

Hence, add a tracepoint in the gmap notifier.

---
 arch/s390/kvm/kvm-s390.c   |  2 ++
 arch/s390/kvm/trace-s390.h | 23 +++++++++++++++++++++++
 2 files changed, 25 insertions(+)
arch/s390/kvm/kvm-s390.c +2 -0
index 9379471081fa..f643c9b9f2f3 100644
@@ -3987,6 +3987,8 @@ static void kvm_gmap_notifier(struct gmap *gmap, unsigned long start,
 	unsigned long prefix;
 	unsigned long i;
 
+	trace_kvm_s390_gmap_notifier(start, end, gmap_is_shadow(gmap));
+
 	if (gmap_is_shadow(gmap))
 		return;
 	if (start >= 1UL << 31)
arch/s390/kvm/trace-s390.h +23 -0
index 6f0209d45164..5dabd0b64d6e 100644
@@ -333,6 +333,29 @@ TRACE_EVENT(kvm_s390_airq_suppressed,
 		      __entry->id, __entry->isc)
 	);
 
+/*
+ * Trace point for gmap notifier calls.
+ */
+TRACE_EVENT(kvm_s390_gmap_notifier,
+		TP_PROTO(unsigned long start, unsigned long end, unsigned int shadow),
+		TP_ARGS(start, end, shadow),
+
+		TP_STRUCT__entry(
+			__field(unsigned long, start)
+			__field(unsigned long, end)
+			__field(unsigned int, shadow)
+			),
+
+		TP_fast_assign(
+			__entry->start = start;
+			__entry->end = end;
+			__entry->shadow = shadow;
+			),
+
+		TP_printk("gmap notified (start:0x%lx end:0x%lx shadow:%d)",
+			__entry->start, __entry->end, __entry->shadow)
+	);
+
 
 #endif /* _TRACE_KVMS390_H */
 
Signature · 1 line
2.41.0
#4 lore raw
In reply to Nico Boehr
Quoted 29 lines · Nico Boehr

v3:
---
* rename te -> entry (David)
* add counters for gmap reuse and gmap create (David)

v2:
---
* also count shadowing of pages (Janosch)
* fix naming of counters (Janosch)
* mention shadowing of multiple levels is counted in each level (Claudio)
* fix inaccuate commit description regarding gmap notifier (Claudio)

When running a guest-3 via VSIE, guest-1 needs to shadow the page table structures of guest-2.

To reflect changes of the guest-2 in the _shadowed_ page table structures, the _shadowing_ sturctures sometimes need to be rebuilt. Since this is a costly operation, it should be avoided whenever possible.

This series adds kvm stat counters to count the number of shadow gmaps created and a tracepoint whenever something is unshadowed. This is a first step to try and improve VSIE performance.

Please note that "KVM: s390: add tracepoint in gmap notifier" has some checkpatch --strict findings. I did not fix these since the tracepoint definition would then look completely different from all the other tracepoints in arch/s390/kvm/trace-s390.h. If you want me to fix that, please let me know.

Quoted 9 lines

While developing this, a question regarding the stat counters came up:
there's usually no locking involved when the stat counters are incremented. On s390, GCC accidentally seems to do the right thing(TM) most of the time by generating a agsi instruction (which should be atomic given proper alignment). However, it's not guaranteed, so would we rather want to go with an atomic for the stat counters to avoid losing events? Or do we just accept the fact that we might loose events sometimes? Is there anything that speaks against having an atomic in kvm_stat?

FWIW the PCI counters (/sys/kernel/debug/pci/<dev>/statistics) use atomic64_add(). Also, s390's memory model is strong enough that these are actually just normal loads/stores/adds (see arch/s390/include/asm/atomic_ops.h) with the generic atomic64_xx() adding debug instrumentation.

author
#5 lore raw

Quoting Niklas Schnelle (2023-09-05 09:53:40)

Quoted 40 lines · Nico Boehr
Quoted 29 lines

v3:
---
* rename te -> entry (David)
* add counters for gmap reuse and gmap create (David)

v2:
---
* also count shadowing of pages (Janosch)
* fix naming of counters (Janosch)
* mention shadowing of multiple levels is counted in each level (Claudio)
* fix inaccuate commit description regarding gmap notifier (Claudio)

When running a guest-3 via VSIE, guest-1 needs to shadow the page table structures of guest-2.

To reflect changes of the guest-2 in the _shadowed_ page table structures, the _shadowing_ sturctures sometimes need to be rebuilt. Since this is a costly operation, it should be avoided whenever possible.

This series adds kvm stat counters to count the number of shadow gmaps created and a tracepoint whenever something is unshadowed. This is a first step to try and improve VSIE performance.

Please note that "KVM: s390: add tracepoint in gmap notifier" has some checkpatch --strict findings. I did not fix these since the tracepoint definition would then look completely different from all the other tracepoints in arch/s390/kvm/trace-s390.h. If you want me to fix that, please let me know.

Quoted 9 lines

While developing this, a question regarding the stat counters came up:
there's usually no locking involved when the stat counters are incremented. On s390, GCC accidentally seems to do the right thing(TM) most of the time by generating a agsi instruction (which should be atomic given proper alignment). However, it's not guaranteed, so would we rather want to go with an atomic for the stat counters to avoid losing events? Or do we just accept the fact that we might loose events sometimes? Is there anything that speaks against having an atomic in kvm_stat?

Quoted 5 lines

FWIW the PCI counters (/sys/kernel/debug/pci/<dev>/statistics) use atomic64_add(). Also, s390's memory model is strong enough that these are actually just normal loads/stores/adds (see arch/s390/include/asm/atomic_ops.h) with the generic atomic64_xx() adding debug instrumentation.

In KVM I am mostly concerned about the compiler since we just do counter++
- right now this always seems to result in an agsi instruction, but that's of course not guaranteed.

#6 lore raw
Quoted 47 lines · Nico Boehr

Quoting Niklas Schnelle (2023-09-05 09:53:40)

Quoted 40 lines · Nico Boehr
Quoted 29 lines

v3:
---
* rename te -> entry (David)
* add counters for gmap reuse and gmap create (David)

v2:
---
* also count shadowing of pages (Janosch)
* fix naming of counters (Janosch)
* mention shadowing of multiple levels is counted in each level (Claudio)
* fix inaccuate commit description regarding gmap notifier (Claudio)

When running a guest-3 via VSIE, guest-1 needs to shadow the page table structures of guest-2.

To reflect changes of the guest-2 in the _shadowed_ page table structures, the _shadowing_ sturctures sometimes need to be rebuilt. Since this is a costly operation, it should be avoided whenever possible.

This series adds kvm stat counters to count the number of shadow gmaps created and a tracepoint whenever something is unshadowed. This is a first step to try and improve VSIE performance.

Please note that "KVM: s390: add tracepoint in gmap notifier" has some checkpatch --strict findings. I did not fix these since the tracepoint definition would then look completely different from all the other tracepoints in arch/s390/kvm/trace-s390.h. If you want me to fix that, please let me know.

Quoted 9 lines

While developing this, a question regarding the stat counters came up:
there's usually no locking involved when the stat counters are incremented. On s390, GCC accidentally seems to do the right thing(TM) most of the time by generating a agsi instruction (which should be atomic given proper alignment). However, it's not guaranteed, so would we rather want to go with an atomic for the stat counters to avoid losing events? Or do we just accept the fact that we might loose events sometimes? Is there anything that speaks against having an atomic in kvm_stat?

Quoted 5 lines

FWIW the PCI counters (/sys/kernel/debug/pci/<dev>/statistics) use atomic64_add(). Also, s390's memory model is strong enough that these are actually just normal loads/stores/adds (see arch/s390/include/asm/atomic_ops.h) with the generic atomic64_xx() adding debug instrumentation.

Quoted 3 lines

In KVM I am mostly concerned about the compiler since we just do counter++
- right now this always seems to result in an agsi instruction, but that's of course not guaranteed.

Right, the compiler can do what it wants with that. The question is if we care about a slight imprecision, though.

Probably not worth the trouble for something that never happens and is only used for debugging purposes.

Using atomics would be cleaner, though.

Signature · 3 lines
Cheers,

David / dhildenb
author
#7 lore raw

Quoting David Hildenbrand (2023-09-06 09:37:28)
[...]

Quoted 5 lines

Right, the compiler can do what it wants with that. The question is if we care about a slight imprecision, though.

Probably not worth the trouble for something that never happens and is only used for debugging purposes.

Yep, probably true. Thanks!