kvm/irqchip: Speed up KVM_SET_GSI_ROUTING
authorChristian Borntraeger <borntraeger@de.ibm.com>
Thu, 16 Jan 2014 12:44:20 +0000 (13:44 +0100)
committerPaolo Bonzini <pbonzini@redhat.com>
Mon, 5 May 2014 14:29:11 +0000 (16:29 +0200)
When starting lots of dataplane devices the bootup takes very long on
Christian's s390 with irqfd patches. With larger setups he is even
able to trigger some timeouts in some components.  Turns out that the
KVM_SET_GSI_ROUTING ioctl takes very long (strace claims up to 0.1 sec)
when having multiple CPUs.  This is caused by the  synchronize_rcu and
the HZ=100 of s390.  By changing the code to use a private srcu we can
speed things up.  This patch reduces the boot time till mounting root
from 8 to 2 seconds on my s390 guest with 100 disks.

Uses of hlist_for_each_entry_rcu, hlist_add_head_rcu, hlist_del_init_rcu
are fine because they do not have lockdep checks (hlist_for_each_entry_rcu
uses rcu_dereference_raw rather than rcu_dereference, and write-sides
do not do rcu lockdep at all).

Note that we're hardly relying on the "sleepable" part of srcu.  We just
want SRCU's faster detection of grace periods.

Testing was done by Andrew Theurer using netperf tests STREAM, MAERTS
and RR.  The difference between results "before" and "after" the patch
has mean -0.2% and standard deviation 0.6%.  Using a paired t-test on the
data points says that there is a 2.5% probability that the patch is the
cause of the performance difference (rather than a random fluctuation).

(Restricting the t-test to RR, which is the most likely to be affected,
changes the numbers to respectively -0.3% mean, 0.7% stdev, and 8%
probability that the numbers actually say something about the patch.
The probability increases mostly because there are fewer data points).

Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Christian Borntraeger <borntraeger@de.ibm.com> # s390
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
include/linux/kvm_host.h
virt/kvm/eventfd.c
virt/kvm/irq_comm.c
virt/kvm/irqchip.c
virt/kvm/kvm_main.c

index 1e125b055327958edb96f02bc916e6c1e9e40845..970c68197c698898df483198995d9a0c558c0770 100644 (file)
@@ -370,6 +370,7 @@ struct kvm {
        struct mm_struct *mm; /* userspace tied to this vm */
        struct kvm_memslots *memslots;
        struct srcu_struct srcu;
+       struct srcu_struct irq_srcu;
 #ifdef CONFIG_KVM_APIC_ARCHITECTURE
        u32 bsp_vcpu_id;
 #endif
index 912ec5a95e2c129fae74ca855b6dccf6f0c99f37..20c3af7692c5d32d6de0c7a88a120a98b14d252e 100644 (file)
@@ -31,6 +31,7 @@
 #include <linux/list.h>
 #include <linux/eventfd.h>
 #include <linux/kernel.h>
+#include <linux/srcu.h>
 #include <linux/slab.h>
 
 #include "iodev.h"
@@ -118,19 +119,22 @@ static void
 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
 {
        struct _irqfd_resampler *resampler;
+       struct kvm *kvm;
        struct _irqfd *irqfd;
+       int idx;
 
        resampler = container_of(kian, struct _irqfd_resampler, notifier);
+       kvm = resampler->kvm;
 
-       kvm_set_irq(resampler->kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
+       kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
                    resampler->notifier.gsi, 0, false);
 
-       rcu_read_lock();
+       idx = srcu_read_lock(&kvm->irq_srcu);
 
        list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
                eventfd_signal(irqfd->resamplefd, 1);
 
-       rcu_read_unlock();
+       srcu_read_unlock(&kvm->irq_srcu, idx);
 }
 
 static void
@@ -142,7 +146,7 @@ irqfd_resampler_shutdown(struct _irqfd *irqfd)
        mutex_lock(&kvm->irqfds.resampler_lock);
 
        list_del_rcu(&irqfd->resampler_link);
-       synchronize_rcu();
+       synchronize_srcu(&kvm->irq_srcu);
 
        if (list_empty(&resampler->list)) {
                list_del(&resampler->link);
@@ -221,17 +225,18 @@ irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
        unsigned long flags = (unsigned long)key;
        struct kvm_kernel_irq_routing_entry *irq;
        struct kvm *kvm = irqfd->kvm;
+       int idx;
 
        if (flags & POLLIN) {
-               rcu_read_lock();
-               irq = rcu_dereference(irqfd->irq_entry);
+               idx = srcu_read_lock(&kvm->irq_srcu);
+               irq = srcu_dereference(irqfd->irq_entry, &kvm->irq_srcu);
                /* An event has been signaled, inject an interrupt */
                if (irq)
                        kvm_set_msi(irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
                                        false);
                else
                        schedule_work(&irqfd->inject);
-               rcu_read_unlock();
+               srcu_read_unlock(&kvm->irq_srcu, idx);
        }
 
        if (flags & POLLHUP) {
@@ -363,7 +368,7 @@ kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
                }
 
                list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
-               synchronize_rcu();
+               synchronize_srcu(&kvm->irq_srcu);
 
                mutex_unlock(&kvm->irqfds.resampler_lock);
        }
@@ -465,7 +470,7 @@ kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
                         * another thread calls kvm_irq_routing_update before
                         * we flush workqueue below (we synchronize with
                         * kvm_irq_routing_update using irqfds.lock).
-                        * It is paired with synchronize_rcu done by caller
+                        * It is paired with synchronize_srcu done by caller
                         * of that function.
                         */
                        rcu_assign_pointer(irqfd->irq_entry, NULL);
@@ -524,7 +529,7 @@ kvm_irqfd_release(struct kvm *kvm)
 
 /*
  * Change irq_routing and irqfd.
- * Caller must invoke synchronize_rcu afterwards.
+ * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
  */
 void kvm_irq_routing_update(struct kvm *kvm,
                            struct kvm_irq_routing_table *irq_rt)
index e2e6b4473a96fafc98dc85aa6c2e8fa49fcd1c09..ced4a542a031313cefd0915bbfa3943d3433c240 100644 (file)
@@ -163,6 +163,7 @@ int kvm_set_irq_inatomic(struct kvm *kvm, int irq_source_id, u32 irq, int level)
        struct kvm_kernel_irq_routing_entry *e;
        int ret = -EINVAL;
        struct kvm_irq_routing_table *irq_rt;
+       int idx;
 
        trace_kvm_set_irq(irq, level, irq_source_id);
 
@@ -174,8 +175,8 @@ int kvm_set_irq_inatomic(struct kvm *kvm, int irq_source_id, u32 irq, int level)
         * Since there's no easy way to do this, we only support injecting MSI
         * which is limited to 1:1 GSI mapping.
         */
-       rcu_read_lock();
-       irq_rt = rcu_dereference(kvm->irq_routing);
+       idx = srcu_read_lock(&kvm->irq_srcu);
+       irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
        if (irq < irq_rt->nr_rt_entries)
                hlist_for_each_entry(e, &irq_rt->map[irq], link) {
                        if (likely(e->type == KVM_IRQ_ROUTING_MSI))
@@ -184,7 +185,7 @@ int kvm_set_irq_inatomic(struct kvm *kvm, int irq_source_id, u32 irq, int level)
                                ret = -EWOULDBLOCK;
                        break;
                }
-       rcu_read_unlock();
+       srcu_read_unlock(&kvm->irq_srcu, idx);
        return ret;
 }
 
@@ -253,22 +254,22 @@ void kvm_unregister_irq_mask_notifier(struct kvm *kvm, int irq,
        mutex_lock(&kvm->irq_lock);
        hlist_del_rcu(&kimn->link);
        mutex_unlock(&kvm->irq_lock);
-       synchronize_rcu();
+       synchronize_srcu(&kvm->irq_srcu);
 }
 
 void kvm_fire_mask_notifiers(struct kvm *kvm, unsigned irqchip, unsigned pin,
                             bool mask)
 {
        struct kvm_irq_mask_notifier *kimn;
-       int gsi;
+       int idx, gsi;
 
-       rcu_read_lock();
-       gsi = rcu_dereference(kvm->irq_routing)->chip[irqchip][pin];
+       idx = srcu_read_lock(&kvm->irq_srcu);
+       gsi = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu)->chip[irqchip][pin];
        if (gsi != -1)
                hlist_for_each_entry_rcu(kimn, &kvm->mask_notifier_list, link)
                        if (kimn->irq == gsi)
                                kimn->func(kimn, mask);
-       rcu_read_unlock();
+       srcu_read_unlock(&kvm->irq_srcu, idx);
 }
 
 int kvm_set_routing_entry(struct kvm_irq_routing_table *rt,
index 20dc9e4a8f6cee7b6d86a6decc92410715c9264c..b43c275775cd5a1d9e5bbc7f842ceaa81f2a0c69 100644 (file)
@@ -26,6 +26,7 @@
 
 #include <linux/kvm_host.h>
 #include <linux/slab.h>
+#include <linux/srcu.h>
 #include <linux/export.h>
 #include <trace/events/kvm.h>
 #include "irq.h"
 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 {
        struct kvm_irq_ack_notifier *kian;
-       int gsi;
+       int gsi, idx;
 
-       rcu_read_lock();
-       gsi = rcu_dereference(kvm->irq_routing)->chip[irqchip][pin];
+       idx = srcu_read_lock(&kvm->irq_srcu);
+       gsi = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu)->chip[irqchip][pin];
        if (gsi != -1)
                hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
                                         link)
                        if (kian->gsi == gsi) {
-                               rcu_read_unlock();
+                               srcu_read_unlock(&kvm->irq_srcu, idx);
                                return true;
                        }
 
-       rcu_read_unlock();
+       srcu_read_unlock(&kvm->irq_srcu, idx);
 
        return false;
 }
@@ -54,18 +55,18 @@ EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 {
        struct kvm_irq_ack_notifier *kian;
-       int gsi;
+       int gsi, idx;
 
        trace_kvm_ack_irq(irqchip, pin);
 
-       rcu_read_lock();
-       gsi = rcu_dereference(kvm->irq_routing)->chip[irqchip][pin];
+       idx = srcu_read_lock(&kvm->irq_srcu);
+       gsi = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu)->chip[irqchip][pin];
        if (gsi != -1)
                hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
                                         link)
                        if (kian->gsi == gsi)
                                kian->irq_acked(kian);
-       rcu_read_unlock();
+       srcu_read_unlock(&kvm->irq_srcu, idx);
 }
 
 void kvm_register_irq_ack_notifier(struct kvm *kvm,
@@ -85,7 +86,7 @@ void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
        mutex_lock(&kvm->irq_lock);
        hlist_del_init_rcu(&kian->link);
        mutex_unlock(&kvm->irq_lock);
-       synchronize_rcu();
+       synchronize_srcu(&kvm->irq_srcu);
 #ifdef __KVM_HAVE_IOAPIC
        kvm_vcpu_request_scan_ioapic(kvm);
 #endif
@@ -115,7 +116,7 @@ int kvm_set_irq(struct kvm *kvm, int irq_source_id, u32 irq, int level,
                bool line_status)
 {
        struct kvm_kernel_irq_routing_entry *e, irq_set[KVM_NR_IRQCHIPS];
-       int ret = -1, i = 0;
+       int ret = -1, i = 0, idx;
        struct kvm_irq_routing_table *irq_rt;
 
        trace_kvm_set_irq(irq, level, irq_source_id);
@@ -124,12 +125,12 @@ int kvm_set_irq(struct kvm *kvm, int irq_source_id, u32 irq, int level,
         * IOAPIC.  So set the bit in both. The guest will ignore
         * writes to the unused one.
         */
-       rcu_read_lock();
-       irq_rt = rcu_dereference(kvm->irq_routing);
+       idx = srcu_read_lock(&kvm->irq_srcu);
+       irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
        if (irq < irq_rt->nr_rt_entries)
                hlist_for_each_entry(e, &irq_rt->map[irq], link)
                        irq_set[i++] = *e;
-       rcu_read_unlock();
+       srcu_read_unlock(&kvm->irq_srcu, idx);
 
        while(i--) {
                int r;
@@ -226,7 +227,7 @@ int kvm_set_irq_routing(struct kvm *kvm,
        kvm_irq_routing_update(kvm, new);
        mutex_unlock(&kvm->irq_lock);
 
-       synchronize_rcu();
+       synchronize_srcu_expedited(&kvm->irq_srcu);
 
        new = old;
        r = 0;
index fa70c6e642b49107a7c725dec6b8bd5df10a7b58..95b4c2b3906ac0f8c474f86d4819994a4d67f3b2 100644 (file)
@@ -457,11 +457,11 @@ static struct kvm *kvm_create_vm(unsigned long type)
 
        r = kvm_arch_init_vm(kvm, type);
        if (r)
-               goto out_err_nodisable;
+               goto out_err_no_disable;
 
        r = hardware_enable_all();
        if (r)
-               goto out_err_nodisable;
+               goto out_err_no_disable;
 
 #ifdef CONFIG_HAVE_KVM_IRQCHIP
        INIT_HLIST_HEAD(&kvm->mask_notifier_list);
@@ -473,10 +473,12 @@ static struct kvm *kvm_create_vm(unsigned long type)
        r = -ENOMEM;
        kvm->memslots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
        if (!kvm->memslots)
-               goto out_err_nosrcu;
+               goto out_err_no_srcu;
        kvm_init_memslots_id(kvm);
        if (init_srcu_struct(&kvm->srcu))
-               goto out_err_nosrcu;
+               goto out_err_no_srcu;
+       if (init_srcu_struct(&kvm->irq_srcu))
+               goto out_err_no_irq_srcu;
        for (i = 0; i < KVM_NR_BUSES; i++) {
                kvm->buses[i] = kzalloc(sizeof(struct kvm_io_bus),
                                        GFP_KERNEL);
@@ -505,10 +507,12 @@ static struct kvm *kvm_create_vm(unsigned long type)
        return kvm;
 
 out_err:
+       cleanup_srcu_struct(&kvm->irq_srcu);
+out_err_no_irq_srcu:
        cleanup_srcu_struct(&kvm->srcu);
-out_err_nosrcu:
+out_err_no_srcu:
        hardware_disable_all();
-out_err_nodisable:
+out_err_no_disable:
        for (i = 0; i < KVM_NR_BUSES; i++)
                kfree(kvm->buses[i]);
        kfree(kvm->memslots);