Skip to contentSkip to conversation

Provide vcpu dispatch statistics

RFC PATCH New Naveen N. Rao 1 participant 7 messages May 6, 2019 PowerPC

lore
Share & tools
lore
#0 lore raw

cover Provide vcpu dispatch statistics

This series adds a new procfs file /proc/powerpc/vcpudispatch_stats for providing statistics around how the LPAR processors are dispatched by the POWER Hypervisor, in a shared LPAR environment. Patch 6/6 has more details on how the statistics are gathered.

An example output:

    $ sudo cat /proc/powerpc/vcpudispatch_stats
    cpu0 6839 4126 0 2683 30 0 6821 18 0
    cpu1 2515 1274 0 1229 12 0 2509 6 0
    cpu2 2317 1198 0 1109 10 0 2312 5 0
    cpu3 2259 1165 0 1088 6 0 2256 3 0
    cpu4 2205 1143 0 1056 6 0 2202 3 0
    cpu5 2165 1121 0 1038 6 0 2162 3 0
    cpu6 2183 1127 0 1050 6 0 2180 3 0
    cpu7 2193 1133 0 1052 8 0 2187 6 0
    cpu8 2165 1115 0 1032 18 0 2156 9 0
    cpu9 2301 1252 0 1033 16 0 2293 8 0
    cpu10 2197 1138 0 1041 18 0 2187 10 0
    cpu11 2273 1185 0 1062 26 0 2260 13 0
    cpu12 2186 1125 0 1043 18 0 2177 9 0
    cpu13 2161 1115 0 1030 16 0 2153 8 0
    cpu14 2206 1153 0 1033 20 0 2196 10 0
    cpu15 2163 1115 0 1032 16 0 2155 8 0

In the output above, for cpu0, we see that there have been 6839 vcpu dispatches since statistics were enabled. The next 5 numbers represent dispatch dispersions without taking the vcpu home nodes into account.
- 4126 dispatches were on the same physical cpu as the previous time
- 2683 were on a different core, but within the same chip,
- while 30 times, a vcpu was dispatched on a different chip compared to its last dispatch.

The next 3 numbers represent statistics in relation to the vcpu home node. We see that there have been 6821 dispatches in the vcpu home node, while 18 dispatches were in a different chip.

TODO:
- Consider need for adding cond_resched() in some places.
- More testing, especially on larger machines.

- Naveen

Naveen N. Rao (6):
powerpc/pseries: Use macros for referring to the DTL enable mask powerpc/pseries: Do not save the previous DTL mask value powerpc/pseries: Factor out DTL buffer allocation and registration
routines
powerpc/pseries: Generalize hcall_vphn() powerpc/pseries: Introduce helpers to gatekeep DTLB usage powerpc/pseries: Provide vcpu dispatch statistics

 arch/powerpc/include/asm/lppaca.h         |  11 +
 arch/powerpc/include/asm/plpar_wrappers.h |   4 +
 arch/powerpc/include/asm/topology.h       |   4 +
 arch/powerpc/mm/book3s64/vphn.h           |   8 +
 arch/powerpc/mm/numa.c                    | 139 +++++-
 arch/powerpc/platforms/pseries/dtl.c      |  22 +-
 arch/powerpc/platforms/pseries/lpar.c     | 543 +++++++++++++++++++++-
 arch/powerpc/platforms/pseries/setup.c    |  34 +-
 8 files changed, 690 insertions(+), 75 deletions(-)
Signature · 1 line
2.21.0
#1 lore raw

RFC PATCH 1/6 powerpc/pseries: Use macros for referring to the DTL enable mask

Introduce macros to encode the DTL enable mask fields and use those instead of hardcoding numbers.

---
 arch/powerpc/include/asm/lppaca.h      | 11 +++++++++++
 arch/powerpc/platforms/pseries/dtl.c   |  8 +-------
 arch/powerpc/platforms/pseries/lpar.c  |  2 +-
 arch/powerpc/platforms/pseries/setup.c |  2 +-
 4 files changed, 14 insertions(+), 9 deletions(-)
arch/powerpc/include/asm/lppaca.h +11 -0
index 7c23ce8a5a4c..2c7e31187726 100644
@@ -154,6 +154,17 @@ struct dtl_entry {
 #define DISPATCH_LOG_BYTES	4096	/* bytes per cpu */
 #define N_DISPATCH_LOG		(DISPATCH_LOG_BYTES / sizeof(struct dtl_entry))
 
+/*
+ * Dispatch trace log event enable mask:
+ *   0x1: voluntary virtual processor waits
+ *   0x2: time-slice preempts
+ *   0x4: virtual partition memory page faults
+ */
+#define DTL_LOG_CEDE		0x1
+#define DTL_LOG_PREEMPT		0x2
+#define DTL_LOG_FAULT		0x4
+#define DTL_LOG_ALL		(DTL_LOG_CEDE | DTL_LOG_PREEMPT | DTL_LOG_FAULT)
+
 extern struct kmem_cache *dtl_cache;
 
 /*
arch/powerpc/platforms/pseries/dtl.c +1 -7
index ef6595153642..051ea2de1e1a 100644
@@ -40,13 +40,7 @@ struct dtl {
 };
 static DEFINE_PER_CPU(struct dtl, cpu_dtl);
 
-/*
- * Dispatch trace log event mask:
- * 0x7: 0x1: voluntary virtual processor waits
- *      0x2: time-slice preempts
- *      0x4: virtual partition memory page faults
- */
-static u8 dtl_event_mask = 0x7;
+static u8 dtl_event_mask = DTL_LOG_ALL;
 
 
 /*
arch/powerpc/platforms/pseries/lpar.c +1 -1
index 1034ef1fe2b4..23f2ac6793b7 100644
@@ -126,7 +126,7 @@ void vpa_init(int cpu)
 			pr_err("WARNING: DTL registration of cpu %d (hw %d) "
 			       "failed with %ld\n", smp_processor_id(),
 			       hwcpu, ret);
-		lppaca_of(cpu).dtl_enable_mask = 2;
+		lppaca_of(cpu).dtl_enable_mask = DTL_LOG_PREEMPT;
 	}
 }
 
arch/powerpc/platforms/pseries/setup.c +1 -1
index e4f0dfd4ae33..fabaefff8399 100644
@@ -316,7 +316,7 @@ static int alloc_dispatch_logs(void)
 		pr_err("WARNING: DTL registration of cpu %d (hw %d) failed "
 		       "with %d\n", smp_processor_id(),
 		       hard_smp_processor_id(), ret);
-	get_paca()->lppaca_ptr->dtl_enable_mask = 2;
+	get_paca()->lppaca_ptr->dtl_enable_mask = DTL_LOG_PREEMPT;
 
 	return 0;
 }
Signature · 1 line
2.21.0
#2 lore raw

RFC PATCH 2/6 powerpc/pseries: Do not save the previous DTL mask value

In reply to Naveen N. Rao

When CONFIG_VIRT_CPU_ACCOUNTING_NATIVE is enabled, we always initialize DTL enable mask to DTL_LOG_PREEMPT (0x2). There are no other places where the mask is changed. As such, when reading the DTL log buffer through debugfs, there is no need to save and restore the previous mask value.

We don't need to save and restore the earlier mask value if CONFIG_VIRT_CPU_ACCOUNTING_NATIVE is not enabled. So, remove the field from the structure as well.

---
 arch/powerpc/platforms/pseries/dtl.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
arch/powerpc/platforms/pseries/dtl.c +1 -3
index 051ea2de1e1a..fb05804adb2f 100644
@@ -55,7 +55,6 @@ struct dtl_ring {
 	struct dtl_entry *write_ptr;
 	struct dtl_entry *buf;
 	struct dtl_entry *buf_end;
-	u8	saved_dtl_mask;
 };
 
 static DEFINE_PER_CPU(struct dtl_ring, dtl_rings);
@@ -105,7 +104,6 @@ static int dtl_start(struct dtl *dtl)
 	dtlr->write_ptr = dtl->buf;
 
 	/* enable event logging */
-	dtlr->saved_dtl_mask = lppaca_of(dtl->cpu).dtl_enable_mask;
 	lppaca_of(dtl->cpu).dtl_enable_mask |= dtl_event_mask;
 
 	dtl_consumer = consume_dtle;
@@ -123,7 +121,7 @@ static void dtl_stop(struct dtl *dtl)
 	dtlr->buf = NULL;
 
 	/* restore dtl_enable_mask */
-	lppaca_of(dtl->cpu).dtl_enable_mask = dtlr->saved_dtl_mask;
+	lppaca_of(dtl->cpu).dtl_enable_mask = DTL_LOG_PREEMPT;
 
 	if (atomic_dec_and_test(&dtl_count))
 		dtl_consumer = NULL;
Signature · 1 line
2.21.0
#3 lore raw

RFC PATCH 3/6 powerpc/pseries: Factor out DTL buffer allocation and registration routines

In reply to Naveen N. Rao

Introduce new helpers for DTL buffer allocation and registration and have the existing code use those.

---
 arch/powerpc/include/asm/plpar_wrappers.h |  2 +
 arch/powerpc/platforms/pseries/lpar.c     | 66 ++++++++++++++++-------
 arch/powerpc/platforms/pseries/setup.c    | 34 +-----------
 3 files changed, 52 insertions(+), 50 deletions(-)
arch/powerpc/include/asm/plpar_wrappers.h +2 -0
index cff5a411e595..d08feb1bc2bd 100644
@@ -88,6 +88,8 @@ static inline long register_dtl(unsigned long cpu, unsigned long vpa)
 	return vpa_call(H_VPA_REG_DTL, cpu, vpa);
 }
 
+extern void register_dtl_buffer(int cpu);
+extern void alloc_dtl_buffers(void);
 extern void vpa_init(int cpu);
 
 static inline long plpar_pte_enter(unsigned long flags,
arch/powerpc/platforms/pseries/lpar.c +48 -18

83 lines not on this page · raw message

arch/powerpc/platforms/pseries/setup.c +2 -32
index fabaefff8399..b6995e5cc5c9 100644
@@ -277,46 +277,16 @@ struct kmem_cache *dtl_cache;
  */
 static int alloc_dispatch_logs(void)
 {
-	int cpu, ret;
-	struct paca_struct *pp;
-	struct dtl_entry *dtl;
-
 	if (!firmware_has_feature(FW_FEATURE_SPLPAR))
 		return 0;
 
 	if (!dtl_cache)
 		return 0;
 
-	for_each_possible_cpu(cpu) {
-		pp = paca_ptrs[cpu];
-		dtl = kmem_cache_alloc(dtl_cache, GFP_KERNEL);
-		if (!dtl) {
-			pr_warn("Failed to allocate dispatch trace log for cpu %d\n",
-				cpu);
-			pr_warn("Stolen time statistics will be unreliable\n");
-			break;
-		}
-
-		pp->dtl_ridx = 0;
-		pp->dispatch_log = dtl;
-		pp->dispatch_log_end = dtl + N_DISPATCH_LOG;
-		pp->dtl_curr = dtl;
-	}
+	alloc_dtl_buffers();
 
 	/* Register the DTL for the current (boot) cpu */
-	dtl = get_paca()->dispatch_log;
-	get_paca()->dtl_ridx = 0;
-	get_paca()->dtl_curr = dtl;
-	get_paca()->lppaca_ptr->dtl_idx = 0;
-
-	/* hypervisor reads buffer length from this field */
-	dtl->enqueue_to_dispatch_time = cpu_to_be32(DISPATCH_LOG_BYTES);
-	ret = register_dtl(hard_smp_processor_id(), __pa(dtl));
-	if (ret)
-		pr_err("WARNING: DTL registration of cpu %d (hw %d) failed "
-		       "with %d\n", smp_processor_id(),
-		       hard_smp_processor_id(), ret);
-	get_paca()->lppaca_ptr->dtl_enable_mask = DTL_LOG_PREEMPT;
+	register_dtl_buffer(smp_processor_id());
 
 	return 0;
 }
Signature · 1 line
2.21.0
#4 lore raw

RFC PATCH 4/6 powerpc/pseries: Generalize hcall_vphn()

In reply to Naveen N. Rao

H_HOME_NODE_ASSOCIATIVITY hcall can take two different flags and return different associativity information in each case. Generalize the existing hcall_vphn() function to take flags as an argument and to return the result. Update the only existing user to pass the proper arguments.

---
 arch/powerpc/mm/book3s64/vphn.h |  8 ++++++++
 arch/powerpc/mm/numa.c          | 27 +++++++++++++--------------
 2 files changed, 21 insertions(+), 14 deletions(-)
arch/powerpc/mm/book3s64/vphn.h +8 -0
index f0b93c2dd578..f7ff1e0c3801 100644
@@ -11,6 +11,14 @@
  */
 #define VPHN_ASSOC_BUFSIZE (VPHN_REGISTER_COUNT*sizeof(u64)/sizeof(u16) + 1)
 
+/*
+ * The H_HOME_NODE_ASSOCIATIVITY hcall takes two values for flags:
+ * 1 for retrieving associativity information for a guest cpu
+ * 2 for retrieving associativity information for a host/hypervisor cpu
+ */
+#define VPHN_FLAG_VCPU	1
+#define VPHN_FLAG_PCPU	2
+
 extern int vphn_unpack_associativity(const long *packed, __be32 *unpacked);
 
 #endif
arch/powerpc/mm/numa.c +13 -14
index 57e64273cb33..57f006b6214b 100644
@@ -1087,6 +1087,17 @@ static void reset_topology_timer(void);
 static int topology_timer_secs = 1;
 static int topology_inited;
 
+static long hcall_vphn(unsigned long cpu, u64 flags, __be32 *associativity)
+{
+	long rc;
+	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
+
+	rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, cpu);
+	vphn_unpack_associativity(retbuf, associativity);
+
+	return rc;
+}
+
 /*
  * Change polling interval for associativity changes.
  */
@@ -1165,25 +1176,13 @@ static int update_cpu_associativity_changes_mask(void)
  * Retrieve the new associativity information for a virtual processor's
  * home node.
  */
-static long hcall_vphn(unsigned long cpu, __be32 *associativity)
-{
-	long rc;
-	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
-	u64 flags = 1;
-	int hwcpu = get_hard_smp_processor_id(cpu);
-
-	rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
-	vphn_unpack_associativity(retbuf, associativity);
-
-	return rc;
-}
-
 static long vphn_get_associativity(unsigned long cpu,
 					__be32 *associativity)
 {
 	long rc;
 
-	rc = hcall_vphn(cpu, associativity);
+	rc = hcall_vphn(get_hard_smp_processor_id(cpu),
+				VPHN_FLAG_VCPU, associativity);
 
 	switch (rc) {
 	case H_FUNCTION:
Signature · 1 line
2.21.0
#5 lore raw

RFC PATCH 5/6 powerpc/pseries: Introduce helpers to gatekeep DTLB usage

In reply to Naveen N. Rao

Since we would be introducing a new user of the DTL buffer in a subsequent patch, add helpers to gatekeep use of the DTL buffer. The current usage of the DTL buffer from debugfs is at a per-cpu level (corresponding to the cpu debugfs file that is opened). Subsequently, we will have users enabling/accessing DTLB for all online cpus. These helpers allow any number of per-cpu users, or a single global user exclusively.

---
 arch/powerpc/include/asm/plpar_wrappers.h |  2 ++
 arch/powerpc/platforms/pseries/dtl.c      | 10 ++++++-
 arch/powerpc/platforms/pseries/lpar.c     | 36 +++++++++++++++++++++++
 3 files changed, 47 insertions(+), 1 deletion(-)
arch/powerpc/include/asm/plpar_wrappers.h +2 -0
index d08feb1bc2bd..ab7dd454b6eb 100644
@@ -88,6 +88,8 @@ static inline long register_dtl(unsigned long cpu, unsigned long vpa)
 	return vpa_call(H_VPA_REG_DTL, cpu, vpa);
 }
 
+extern bool register_dtl_buffer_access(bool global);
+extern void unregister_dtl_buffer_access(bool global);
 extern void register_dtl_buffer(int cpu);
 extern void alloc_dtl_buffers(void);
 extern void vpa_init(int cpu);
arch/powerpc/platforms/pseries/dtl.c +9 -1
index fb05804adb2f..dd28296c9903 100644
@@ -193,11 +193,15 @@ static int dtl_enable(struct dtl *dtl)
 	if (dtl->buf)
 		return -EBUSY;
 
+	if (register_dtl_buffer_access(false))
+		return -EBUSY;
+
 	n_entries = dtl_buf_entries;
 	buf = kmem_cache_alloc_node(dtl_cache, GFP_KERNEL, cpu_to_node(dtl->cpu));
 	if (!buf) {
 		printk(KERN_WARNING "%s: buffer alloc failed for cpu %d\n",
 				__func__, dtl->cpu);
+		unregister_dtl_buffer_access(false);
 		return -ENOMEM;
 	}
 
@@ -214,8 +218,11 @@ static int dtl_enable(struct dtl *dtl)
 	}
 	spin_unlock(&dtl->lock);
 
-	if (rc)
+	if (rc) {
+		unregister_dtl_buffer_access(false);
 		kmem_cache_free(dtl_cache, buf);
+	}
+
 	return rc;
 }
 
@@ -227,6 +234,7 @@ static void dtl_disable(struct dtl *dtl)
 	dtl->buf = NULL;
 	dtl->buf_entries = 0;
 	spin_unlock(&dtl->lock);
+	unregister_dtl_buffer_access(false);
 }
 
 /* file interface */
arch/powerpc/platforms/pseries/lpar.c +36 -0
index 3375ca8cefb5..6af5a2a11deb 100644
@@ -65,6 +65,42 @@ EXPORT_SYMBOL(plpar_hcall);
 EXPORT_SYMBOL(plpar_hcall9);
 EXPORT_SYMBOL(plpar_hcall_norets);
 
+static DEFINE_SPINLOCK(dtl_buffer_refctr_lock);
+static unsigned int dtl_buffer_global_refctr, dtl_buffer_percpu_refctr;
+
+bool register_dtl_buffer_access(bool global)
+{
+	int rc = 0;
+
+	spin_lock(&dtl_buffer_refctr_lock);
+
+	if ((global && (dtl_buffer_global_refctr || dtl_buffer_percpu_refctr))
+			|| (!global && dtl_buffer_global_refctr)) {
+		rc = -1;
+	} else {
+		if (global)
+			dtl_buffer_global_refctr++;
+		else
+			dtl_buffer_percpu_refctr++;
+	}
+
+	spin_unlock(&dtl_buffer_refctr_lock);
+
+	return rc;
+}
+
+void unregister_dtl_buffer_access(bool global)
+{
+	spin_lock(&dtl_buffer_refctr_lock);
+
+	if (global)
+		dtl_buffer_global_refctr--;
+	else
+		dtl_buffer_percpu_refctr--;
+
+	spin_unlock(&dtl_buffer_refctr_lock);
+}
+
 void alloc_dtl_buffers(void)
 {
 	int cpu;
Signature · 1 line
2.21.0
#6 lore raw

RFC PATCH 6/6 powerpc/pseries: Provide vcpu dispatch statistics

In reply to Naveen N. Rao

For Shared Processor LPARs, the POWER Hypervisor maintains a relatively static mapping of the LPAR processors (vcpus) to physical processor chips (representing the "home" node) and tries to always dispatch vcpus on their associated physical processor chip. However, under certain scenarios, vcpus may be dispatched on a different processor chip (away from its home node). The actual physical processor number on which a certain vcpu is dispatched is available to the guest in the 'processor_id' field of each DTL entry.

The guest can discover the home node of each vcpu through the H_HOME_NODE_ASSOCIATIVITY(flags=1) hcall. The guest can also discover the associativity of physical processors, as represented in the DTL entry, through the H_HOME_NODE_ASSOCIATIVITY(flags=2) hcall.

These can then be compared to determine if the vcpu was dispatched on its home node or not. If the vcpu was not dispatched on the home node, it is possible to determine if the vcpu was dispatched in a different chip, socket or drawer.

Introduce a procfs file /proc/powerpc/vcpudispatch_stats that can be used to obtain these statistics. Writing '1' to this file enables collecting the statistics, while writing '0' disables the statistics. The statistics themselves are available by reading the procfs file. By default, the DTLB log for each vcpu is processed 50 times a second so as not to miss any entries. This processing frequency can be changed through /proc/powerpc/vcpudispatch_stats_freq.

---
 arch/powerpc/include/asm/topology.h   |   4 +
 arch/powerpc/mm/numa.c                | 112 +++++++
 arch/powerpc/platforms/pseries/lpar.c | 445 +++++++++++++++++++++++++-
 3 files changed, 559 insertions(+), 2 deletions(-)
arch/powerpc/include/asm/topology.h +4 -0
index f85e2b01c3df..7c064731a0f2 100644
@@ -93,6 +93,10 @@ extern int prrn_is_enabled(void);
 extern int find_and_online_cpu_nid(int cpu);
 extern int timed_topology_update(int nsecs);
 extern void __init shared_proc_topology_init(void);
+extern int init_cpu_associativity(void);
+extern void destroy_cpu_associativity(void);
+extern int cpu_relative_dispatch_distance(int last_disp_cpu, int cur_disp_cpu);
+extern int cpu_home_node_dispatch_distance(int disp_cpu);
 #else
 static inline int start_topology_update(void)
 {
arch/powerpc/mm/numa.c +112 -0

124 lines not on this page · raw message

arch/powerpc/platforms/pseries/lpar.c +443 -2

499 lines not on this page · raw message

Signature · 1 line
2.21.0