[PATCH v4 0/6] optimize the downtime for vfio migration

[PATCH v4 0/6] optimize the downtime for vfio migration

[Longpeng(Mike)]

Hi guys,
 
In vfio migration resume phase, the cost would increase if the
vfio device has more unmasked vectors. We try to optimize it in
this series.
 
You can see the commit message in PATCH 6 for details.
 
Patch 1-3 are simple cleanups and fixup.
Patch 4-5 are the preparations for the optimization.
Patch 6 optimizes the vfio msix setup path.

Changes v3->v4:
 - fix several typos and grammatical errors [Alex]
 - remove the patches that fix and clean the MSIX common part
   from this series [Alex]
 - Patch 6:
    - use vector->use directly and fill it with -1 on error
      paths [Alex]
    - add comment before enable deferring to commit [Alex]
    - move the code that do_use/release on vector 0 into an
      "else" branch [Alex]
    - introduce vfio_prepare_kvm_msi_virq_batch() that enables
      the 'defer_kvm_irq_routing' flag [Alex]
    - introduce vfio_commit_kvm_msi_virq_batch() that clears the
      'defer_kvm_irq_routing' flag and does further work [Alex]

Changes v2->v3:
 - fix two errors [Longpeng]

Changes v1->v2:
 - fix several typos and grammatical errors [Alex, Philippe]
 - split fixups and cleanups into separate patches  [Alex, Philippe]
 - introduce kvm_irqchip_add_deferred_msi_route to
   minimize code changes    [Alex]
 - enable the optimization in msi setup path    [Alex]

Longpeng (Mike) (6):
  vfio: simplify the conditional statements in vfio_msi_enable
  vfio: move re-enabling INTX out of the common helper
  vfio: simplify the failure path in vfio_msi_enable
  kvm: irqchip: extract kvm_irqchip_add_deferred_msi_route
  Revert "vfio: Avoid disabling and enabling vectors repeatedly in VFIO
    migration"
  vfio: defer to commit kvm irq routing when enable msi/msix

 accel/kvm/kvm-all.c  |  15 ++++-
 hw/vfio/pci.c        | 182 +++++++++++++++++++++++++++++++++------------------
 hw/vfio/pci.h        |   1 +
 include/sysemu/kvm.h |   6 ++
 4 files changed, 140 insertions(+), 64 deletions(-)

-- 
1.8.3.1

[PATCH v4 1/6] vfio: simplify the conditional statements in vfio_msi_enable

[Longpeng(Mike)]

It's unnecessary to test against the specific return value of
VFIO_DEVICE_SET_IRQS, since any positive return is an error
indicating the number of vectors we should retry with.

Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
---
 hw/vfio/pci.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index 4feaa1c..c1fba40 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -650,7 +650,7 @@ retry:
     if (ret) {
         if (ret < 0) {
             error_report("vfio: Error: Failed to setup MSI fds: %m");
-        } else if (ret != vdev->nr_vectors) {
+        } else {
             error_report("vfio: Error: Failed to enable %d "
                          "MSI vectors, retry with %d", vdev->nr_vectors, ret);
         }
@@ -668,7 +668,7 @@ retry:
         g_free(vdev->msi_vectors);
         vdev->msi_vectors = NULL;
 
-        if (ret > 0 && ret != vdev->nr_vectors) {
+        if (ret > 0) {
             vdev->nr_vectors = ret;
             goto retry;
         }
-- 
1.8.3.1

[PATCH v4 2/6] vfio: move re-enabling INTX out of the common helper

[Longpeng(Mike)]

Move re-enabling INTX out, and the callers should decide to
re-enable it or not.

Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
---
 hw/vfio/pci.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index c1fba40..a4985c0 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -690,7 +690,6 @@ retry:
 
 static void vfio_msi_disable_common(VFIOPCIDevice *vdev)
 {
-    Error *err = NULL;
     int i;
 
     for (i = 0; i < vdev->nr_vectors; i++) {
@@ -709,15 +708,11 @@ static void vfio_msi_disable_common(VFIOPCIDevice *vdev)
     vdev->msi_vectors = NULL;
     vdev->nr_vectors = 0;
     vdev->interrupt = VFIO_INT_NONE;
-
-    vfio_intx_enable(vdev, &err);
-    if (err) {
-        error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
-    }
 }
 
 static void vfio_msix_disable(VFIOPCIDevice *vdev)
 {
+    Error *err = NULL;
     int i;
 
     msix_unset_vector_notifiers(&vdev->pdev);
@@ -738,6 +733,10 @@ static void vfio_msix_disable(VFIOPCIDevice *vdev)
     }
 
     vfio_msi_disable_common(vdev);
+    vfio_intx_enable(vdev, &err);
+    if (err) {
+        error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
+    }
 
     memset(vdev->msix->pending, 0,
            BITS_TO_LONGS(vdev->msix->entries) * sizeof(unsigned long));
@@ -747,8 +746,14 @@ static void vfio_msix_disable(VFIOPCIDevice *vdev)
 
 static void vfio_msi_disable(VFIOPCIDevice *vdev)
 {
+    Error *err = NULL;
+
     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSI_IRQ_INDEX);
     vfio_msi_disable_common(vdev);
+    vfio_intx_enable(vdev, &err);
+    if (err) {
+        error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
+    }
 
     trace_vfio_msi_disable(vdev->vbasedev.name);
 }
-- 
1.8.3.1

[PATCH v4 3/6] vfio: simplify the failure path in vfio_msi_enable

[Longpeng(Mike)]

Use vfio_msi_disable_common to simplify the error handling
in vfio_msi_enable.

Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
---
 hw/vfio/pci.c | 16 ++--------------
 1 file changed, 2 insertions(+), 14 deletions(-)

diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index a4985c0..aeeb6cd 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -47,6 +47,7 @@
 
 static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
+static void vfio_msi_disable_common(VFIOPCIDevice *vdev);
 
 /*
  * Disabling BAR mmaping can be slow, but toggling it around INTx can
@@ -655,24 +656,12 @@ retry:
                          "MSI vectors, retry with %d", vdev->nr_vectors, ret);
         }
 
-        for (i = 0; i < vdev->nr_vectors; i++) {
-            VFIOMSIVector *vector = &vdev->msi_vectors[i];
-            if (vector->virq >= 0) {
-                vfio_remove_kvm_msi_virq(vector);
-            }
-            qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
-                                NULL, NULL, NULL);
-            event_notifier_cleanup(&vector->interrupt);
-        }
-
-        g_free(vdev->msi_vectors);
-        vdev->msi_vectors = NULL;
+        vfio_msi_disable_common(vdev);
 
         if (ret > 0) {
             vdev->nr_vectors = ret;
             goto retry;
         }
-        vdev->nr_vectors = 0;
 
         /*
          * Failing to setup MSI doesn't really fall within any specification.
@@ -680,7 +669,6 @@ retry:
          * out to fall back to INTx for this device.
          */
         error_report("vfio: Error: Failed to enable MSI");
-        vdev->interrupt = VFIO_INT_NONE;
 
         return;
     }
-- 
1.8.3.1

[PATCH v4 4/6] kvm: irqchip: extract kvm_irqchip_add_deferred_msi_route

[Longpeng(Mike)]

Extract a common helper that add MSI route for specific vector
but does not commit immediately.

Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
---
 accel/kvm/kvm-all.c  | 15 +++++++++++++--
 include/sysemu/kvm.h |  6 ++++++
 2 files changed, 19 insertions(+), 2 deletions(-)

diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
index db8d83b..8627f7c 100644
--- a/accel/kvm/kvm-all.c
+++ b/accel/kvm/kvm-all.c
@@ -1953,7 +1953,7 @@ int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
     return kvm_set_irq(s, route->kroute.gsi, 1);
 }
 
-int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
+int kvm_irqchip_add_deferred_msi_route(KVMState *s, int vector, PCIDevice *dev)
 {
     struct kvm_irq_routing_entry kroute = {};
     int virq;
@@ -1996,7 +1996,18 @@ int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
 
     kvm_add_routing_entry(s, &kroute);
     kvm_arch_add_msi_route_post(&kroute, vector, dev);
-    kvm_irqchip_commit_routes(s);
+
+    return virq;
+}
+
+int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
+{
+    int virq;
+
+    virq = kvm_irqchip_add_deferred_msi_route(s, vector, dev);
+    if (virq >= 0) {
+        kvm_irqchip_commit_routes(s);
+    }
 
     return virq;
 }
diff --git a/include/sysemu/kvm.h b/include/sysemu/kvm.h
index a1ab1ee..8de0d9a 100644
--- a/include/sysemu/kvm.h
+++ b/include/sysemu/kvm.h
@@ -476,6 +476,12 @@ void kvm_init_cpu_signals(CPUState *cpu);
  * @return: virq (>=0) when success, errno (<0) when failed.
  */
 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev);
+/**
+ * Add MSI route for specific vector but does not commit to KVM
+ * immediately
+ */
+int kvm_irqchip_add_deferred_msi_route(KVMState *s, int vector,
+                                       PCIDevice *dev);
 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
                                  PCIDevice *dev);
 void kvm_irqchip_commit_routes(KVMState *s);
-- 
1.8.3.1

[PATCH v4 5/6] Revert "vfio: Avoid disabling and enabling vectors repeatedly in VFIO migration"

[Longpeng(Mike)]

Commit ecebe53fe993 ("vfio: Avoid disabling and enabling vectors
repeatedly in VFIO migration") avoids inefficiently disabling and
enabling vectors repeatedly and lets the unmasked vectors be enabled
one by one.

But we want to batch multiple routes and defer the commit, and only
commit once outside the loop of setting vector notifiers, so we
cannot enable the vectors one by one in the loop now.

Revert that commit and we will take another way in the next patch,
it can not only avoid disabling/enabling vectors repeatedly, but
also satisfy our requirement of defer to commit.

Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
---
 hw/vfio/pci.c | 20 +++-----------------
 1 file changed, 3 insertions(+), 17 deletions(-)

diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index aeeb6cd..0bd832b 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -569,9 +569,6 @@ static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
 
 static void vfio_msix_enable(VFIOPCIDevice *vdev)
 {
-    PCIDevice *pdev = &vdev->pdev;
-    unsigned int nr, max_vec = 0;
-
     vfio_disable_interrupts(vdev);
 
     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->msix->entries);
@@ -590,22 +587,11 @@ static void vfio_msix_enable(VFIOPCIDevice *vdev)
      * triggering to userspace, then immediately release the vector, leaving
      * the physical device with no vectors enabled, but MSI-X enabled, just
      * like the guest view.
-     * If there are already unmasked vectors (in migration resume phase and
-     * some guest startups) which will be enabled soon, we can allocate all
-     * of them here to avoid inefficiently disabling and enabling vectors
-     * repeatedly later.
      */
-    if (!pdev->msix_function_masked) {
-        for (nr = 0; nr < msix_nr_vectors_allocated(pdev); nr++) {
-            if (!msix_is_masked(pdev, nr)) {
-                max_vec = nr;
-            }
-        }
-    }
-    vfio_msix_vector_do_use(pdev, max_vec, NULL, NULL);
-    vfio_msix_vector_release(pdev, max_vec);
+    vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
+    vfio_msix_vector_release(&vdev->pdev, 0);
 
-    if (msix_set_vector_notifiers(pdev, vfio_msix_vector_use,
+    if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
                                   vfio_msix_vector_release, NULL)) {
         error_report("vfio: msix_set_vector_notifiers failed");
     }
-- 
1.8.3.1

[PATCH v4 6/6] vfio: defer to commit kvm irq routing when enable msi/msix

[Longpeng(Mike)]

In migration resume phase, all unmasked msix vectors need to be
setup when loading the VF state. However, the setup operation would
take longer if the VM has more VFs and each VF has more unmasked
vectors.

The hot spot is kvm_irqchip_commit_routes, it'll scan and update
all irqfds that are already assigned each invocation, so more
vectors means need more time to process them.

vfio_pci_load_config
  vfio_msix_enable
    msix_set_vector_notifiers
      for (vector = 0; vector < dev->msix_entries_nr; vector++) {
        vfio_msix_vector_do_use
          vfio_add_kvm_msi_virq
            kvm_irqchip_commit_routes <-- expensive
      }

We can reduce the cost by only committing once outside the loop.
The routes are cached in kvm_state, we commit them first and then
bind irqfd for each vector.

The test VM has 128 vcpus and 8 VF (each one has 65 vectors),
we measure the cost of the vfio_msix_enable for each VF, and
we can see 90+% costs can be reduce.

VF      Count of irqfds[*]  Original        With this patch

1st           65            8               2
2nd           130           15              2
3rd           195           22              2
4th           260           24              3
5th           325           36              2
6th           390           44              3
7th           455           51              3
8th           520           58              4
Total                       258ms           21ms

[*] Count of irqfds
How many irqfds that already assigned and need to process in this
round.

The optimization can be applied to msi type too.

Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
---
 hw/vfio/pci.c | 129 ++++++++++++++++++++++++++++++++++++++++++++++------------
 hw/vfio/pci.h |   1 +
 2 files changed, 105 insertions(+), 25 deletions(-)

diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index 0bd832b..dca2d0c 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -413,8 +413,6 @@ static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix)
 static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
                                   int vector_n, bool msix)
 {
-    int virq;
-
     if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
         return;
     }
@@ -423,20 +421,31 @@ static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
         return;
     }
 
-    virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev);
-    if (virq < 0) {
+    vector->virq = kvm_irqchip_add_deferred_msi_route(kvm_state, vector_n,
+                                                      &vdev->pdev);
+    if (vector->virq < 0) {
         event_notifier_cleanup(&vector->kvm_interrupt);
+        vector->virq = -1;
+        return;
+    }
+
+    if (vdev->defer_kvm_irq_routing) {
+        /*
+         * The vector->virq will be reset to -1 if we fail to add the
+         * corresponding irqfd in vfio_commit_kvm_msi_virq_batch().
+         */
         return;
     }
 
+    kvm_irqchip_commit_routes(kvm_state);
+
     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
-                                       NULL, virq) < 0) {
-        kvm_irqchip_release_virq(kvm_state, virq);
+                                           NULL, vector->virq) < 0) {
+        kvm_irqchip_release_virq(kvm_state, vector->virq);
         event_notifier_cleanup(&vector->kvm_interrupt);
+        vector->virq = -1;
         return;
     }
-
-    vector->virq = virq;
 }
 
 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
@@ -501,11 +510,13 @@ static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
      * increase them as needed.
      */
     if (vdev->nr_vectors < nr + 1) {
-        vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
         vdev->nr_vectors = nr + 1;
-        ret = vfio_enable_vectors(vdev, true);
-        if (ret) {
-            error_report("vfio: failed to enable vectors, %d", ret);
+        if (!vdev->defer_kvm_irq_routing) {
+            vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
+            ret = vfio_enable_vectors(vdev, true);
+            if (ret) {
+                error_report("vfio: failed to enable vectors, %d", ret);
+            }
         }
     } else {
         Error *err = NULL;
@@ -567,6 +578,46 @@ static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
     }
 }
 
+static void vfio_prepare_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
+{
+    assert(!vdev->defer_kvm_irq_routing);
+    vdev->defer_kvm_irq_routing = true;
+}
+
+static void vfio_commit_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
+{
+    VFIOMSIVector *vector;
+    int i;
+
+    if (!vdev->defer_kvm_irq_routing) {
+        return;
+    }
+
+    vdev->defer_kvm_irq_routing = false;
+
+    if (!vdev->nr_vectors) {
+        return;
+    }
+
+    kvm_irqchip_commit_routes(kvm_state);
+
+    for (i = 0; i < vdev->nr_vectors; i++) {
+        vector = &vdev->msi_vectors[i];
+
+        if (!vector->use || vector->virq < 0) {
+            continue;
+        }
+
+        if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state,
+                                               &vector->kvm_interrupt,
+                                               NULL, vector->virq) < 0) {
+            kvm_irqchip_release_virq(kvm_state, vector->virq);
+            event_notifier_cleanup(&vector->kvm_interrupt);
+            vector->virq = -1;
+        }
+    }
+}
+
 static void vfio_msix_enable(VFIOPCIDevice *vdev)
 {
     vfio_disable_interrupts(vdev);
@@ -576,26 +627,45 @@ static void vfio_msix_enable(VFIOPCIDevice *vdev)
     vdev->interrupt = VFIO_INT_MSIX;
 
     /*
-     * Some communication channels between VF & PF or PF & fw rely on the
-     * physical state of the device and expect that enabling MSI-X from the
-     * guest enables the same on the host.  When our guest is Linux, the
-     * guest driver call to pci_enable_msix() sets the enabling bit in the
-     * MSI-X capability, but leaves the vector table masked.  We therefore
-     * can't rely on a vector_use callback (from request_irq() in the guest)
-     * to switch the physical device into MSI-X mode because that may come a
-     * long time after pci_enable_msix().  This code enables vector 0 with
-     * triggering to userspace, then immediately release the vector, leaving
-     * the physical device with no vectors enabled, but MSI-X enabled, just
-     * like the guest view.
+     * Setting vector notifiers triggers synchronous vector-use
+     * callbacks for each active vector.  Deferring to commit the KVM
+     * routes once rather than per vector provides a substantial
+     * performance improvement.
      */
-    vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
-    vfio_msix_vector_release(&vdev->pdev, 0);
+    vfio_prepare_kvm_msi_virq_batch(vdev);
 
     if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
                                   vfio_msix_vector_release, NULL)) {
         error_report("vfio: msix_set_vector_notifiers failed");
     }
 
+    vfio_commit_kvm_msi_virq_batch(vdev);
+
+    if (vdev->nr_vectors) {
+        int ret;
+
+        ret = vfio_enable_vectors(vdev, true);
+        if (ret) {
+            error_report("vfio: failed to enable vectors, %d", ret);
+        }
+    } else {
+        /*
+         * Some communication channels between VF & PF or PF & fw rely on the
+         * physical state of the device and expect that enabling MSI-X from the
+         * guest enables the same on the host.  When our guest is Linux, the
+         * guest driver call to pci_enable_msix() sets the enabling bit in the
+         * MSI-X capability, but leaves the vector table masked.  We therefore
+         * can't rely on a vector_use callback (from request_irq() in the guest)
+         * to switch the physical device into MSI-X mode because that may come a
+         * long time after pci_enable_msix().  This code enables vector 0 with
+         * triggering to userspace, then immediately release the vector, leaving
+         * the physical device with no vectors enabled, but MSI-X enabled, just
+         * like the guest view.
+         */
+        vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
+        vfio_msix_vector_release(&vdev->pdev, 0);
+    }
+
     trace_vfio_msix_enable(vdev->vbasedev.name);
 }
 
@@ -605,6 +675,13 @@ static void vfio_msi_enable(VFIOPCIDevice *vdev)
 
     vfio_disable_interrupts(vdev);
 
+    /*
+     * Setting vector notifiers needs to enable route for each vector.
+     * Deferring to commit the KVM routes once rather than per vector
+     * provides a substantial performance improvement.
+     */
+    vfio_prepare_kvm_msi_virq_batch(vdev);
+
     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
 retry:
     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->nr_vectors);
@@ -630,6 +707,8 @@ retry:
         vfio_add_kvm_msi_virq(vdev, vector, i, false);
     }
 
+    vfio_commit_kvm_msi_virq_batch(vdev);
+
     /* Set interrupt type prior to possible interrupts */
     vdev->interrupt = VFIO_INT_MSI;
 
diff --git a/hw/vfio/pci.h b/hw/vfio/pci.h
index 6477751..d3c5177 100644
--- a/hw/vfio/pci.h
+++ b/hw/vfio/pci.h
@@ -171,6 +171,7 @@ struct VFIOPCIDevice {
     bool no_kvm_ioeventfd;
     bool no_vfio_ioeventfd;
     bool enable_ramfb;
+    bool defer_kvm_irq_routing;
     VFIODisplay *dpy;
     Notifier irqchip_change_notifier;
 };
-- 
1.8.3.1

Re: [PATCH v4 6/6] vfio: defer to commit kvm irq routing when enable msi/msix

[Alex Williamson]

On Thu, 14 Oct 2021 08:48:52 +0800
"Longpeng(Mike)" <longpeng2@huawei.com> wrote:

> In migration resume phase, all unmasked msix vectors need to be
> setup when loading the VF state. However, the setup operation would
> take longer if the VM has more VFs and each VF has more unmasked
> vectors.
> 
> The hot spot is kvm_irqchip_commit_routes, it'll scan and update
> all irqfds that are already assigned each invocation, so more
> vectors means need more time to process them.
> 
> vfio_pci_load_config
>   vfio_msix_enable
>     msix_set_vector_notifiers
>       for (vector = 0; vector < dev->msix_entries_nr; vector++) {
>         vfio_msix_vector_do_use
>           vfio_add_kvm_msi_virq
>             kvm_irqchip_commit_routes <-- expensive
>       }
> 
> We can reduce the cost by only committing once outside the loop.
> The routes are cached in kvm_state, we commit them first and then
> bind irqfd for each vector.
> 
> The test VM has 128 vcpus and 8 VF (each one has 65 vectors),
> we measure the cost of the vfio_msix_enable for each VF, and
> we can see 90+% costs can be reduce.
> 
> VF      Count of irqfds[*]  Original        With this patch
> 
> 1st           65            8               2
> 2nd           130           15              2
> 3rd           195           22              2
> 4th           260           24              3
> 5th           325           36              2
> 6th           390           44              3
> 7th           455           51              3
> 8th           520           58              4
> Total                       258ms           21ms
> 
> [*] Count of irqfds
> How many irqfds that already assigned and need to process in this
> round.
> 
> The optimization can be applied to msi type too.
> 
> Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
> ---
>  hw/vfio/pci.c | 129 ++++++++++++++++++++++++++++++++++++++++++++++------------
>  hw/vfio/pci.h |   1 +
>  2 files changed, 105 insertions(+), 25 deletions(-)
> 
> diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
> index 0bd832b..dca2d0c 100644
> --- a/hw/vfio/pci.c
> +++ b/hw/vfio/pci.c
> @@ -413,8 +413,6 @@ static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix)
>  static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
>                                    int vector_n, bool msix)
>  {
> -    int virq;
> -
>      if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
>          return;
>      }
> @@ -423,20 +421,31 @@ static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
>          return;
>      }
>  
> -    virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev);
> -    if (virq < 0) {
> +    vector->virq = kvm_irqchip_add_deferred_msi_route(kvm_state, vector_n,
> +                                                      &vdev->pdev);
> +    if (vector->virq < 0) {
>          event_notifier_cleanup(&vector->kvm_interrupt);
> +        vector->virq = -1;

Nit, it seems like all negative values are equivalent here, I don't
think we need to explicitly set virq to -1 given that it's already < 0.

> +        return;
> +    }
> +
> +    if (vdev->defer_kvm_irq_routing) {
> +        /*
> +         * The vector->virq will be reset to -1 if we fail to add the
> +         * corresponding irqfd in vfio_commit_kvm_msi_virq_batch().
> +         */
>          return;
>      }
>  
> +    kvm_irqchip_commit_routes(kvm_state);
> +
>      if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
> -                                       NULL, virq) < 0) {
> -        kvm_irqchip_release_virq(kvm_state, virq);
> +                                           NULL, vector->virq) < 0) {
> +        kvm_irqchip_release_virq(kvm_state, vector->virq);
>          event_notifier_cleanup(&vector->kvm_interrupt);
> +        vector->virq = -1;
>          return;
>      }
> -
> -    vector->virq = virq;
>  }
>  
>  static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
> @@ -501,11 +510,13 @@ static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
>       * increase them as needed.
>       */
>      if (vdev->nr_vectors < nr + 1) {
> -        vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
>          vdev->nr_vectors = nr + 1;
> -        ret = vfio_enable_vectors(vdev, true);
> -        if (ret) {
> -            error_report("vfio: failed to enable vectors, %d", ret);
> +        if (!vdev->defer_kvm_irq_routing) {
> +            vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
> +            ret = vfio_enable_vectors(vdev, true);
> +            if (ret) {
> +                error_report("vfio: failed to enable vectors, %d", ret);
> +            }
>          }
>      } else {
>          Error *err = NULL;
> @@ -567,6 +578,46 @@ static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
>      }
>  }
>  
> +static void vfio_prepare_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
> +{
> +    assert(!vdev->defer_kvm_irq_routing);
> +    vdev->defer_kvm_irq_routing = true;
> +}
> +
> +static void vfio_commit_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
> +{
> +    VFIOMSIVector *vector;
> +    int i;
> +
> +    if (!vdev->defer_kvm_irq_routing) {
> +        return;
> +    }
> +
> +    vdev->defer_kvm_irq_routing = false;
> +
> +    if (!vdev->nr_vectors) {
> +        return;
> +    }
> +
> +    kvm_irqchip_commit_routes(kvm_state);
> +
> +    for (i = 0; i < vdev->nr_vectors; i++) {
> +        vector = &vdev->msi_vectors[i];
> +
> +        if (!vector->use || vector->virq < 0) {
> +            continue;
> +        }
> +
> +        if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state,
> +                                               &vector->kvm_interrupt,
> +                                               NULL, vector->virq) < 0) {
> +            kvm_irqchip_release_virq(kvm_state, vector->virq);
> +            event_notifier_cleanup(&vector->kvm_interrupt);
> +            vector->virq = -1;
> +        }

I started trying to get rid of this code that largely duplicates the
error case of vfio_add_kvm_msi_virq() and questioned why we setup the
notifier separate from connecting it to the irqfd.  If we setup the
notifier and irqfd in the same function we can clean things up a bit
more and confine the deferred state tests in the vector-use code.  I
think we can also assert if we have an unmatched batch commit call and
we probably don't need to test both that a vector is in use and has a
virq, one should not be true without the other.

Does this look like an improvement to you and would you like to roll it
into this patch?  Thanks,

Alex

 pci.c |   58 ++++++++++++++++++++++------------------------------------
 1 file changed, 22 insertions(+), 36 deletions(-)

commit 6fb9336e3fe9e3775b0a0e7aadaff781fb52c0e7
Author: Alex Williamson <alex.williamson@redhat.com>
Date:   Thu Oct 21 13:35:12 2021 -0600

    cleanup

diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index 1792c30049da..5b3a86dd5292 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -417,35 +417,33 @@ static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
         return;
     }
 
-    if (event_notifier_init(&vector->kvm_interrupt, 0)) {
-        return;
-    }
-
     vector->virq = kvm_irqchip_add_deferred_msi_route(kvm_state, vector_n,
                                                       &vdev->pdev);
+}
+
+static void vfio_connect_kvm_msi_virq(VFIOMSIVector *vector)
+{
     if (vector->virq < 0) {
-        event_notifier_cleanup(&vector->kvm_interrupt);
-        vector->virq = -1;
         return;
     }
 
-    if (vdev->defer_kvm_irq_routing) {
-        /*
-         * The vector->virq will be reset to -1 if we fail to add the
-         * corresponding irqfd in vfio_commit_kvm_msi_virq_batch().
-         */
-        return;
+    if (event_notifier_init(&vector->kvm_interrupt, 0)) {
+        goto fail_notifier;
     }
 
-    kvm_irqchip_commit_routes(kvm_state);
-
     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
                                            NULL, vector->virq) < 0) {
-        kvm_irqchip_release_virq(kvm_state, vector->virq);
-        event_notifier_cleanup(&vector->kvm_interrupt);
-        vector->virq = -1;
-        return;
+        goto fail_kvm;
     }
+
+    return;
+
+fail_kvm:
+    event_notifier_cleanup(&vector->kvm_interrupt);
+fail_notifier:
+    kvm_irqchip_release_virq(kvm_state, vector->virq);
+    vector->virq = -1;
+    return;
 }
 
 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
@@ -501,6 +499,10 @@ static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
     } else {
         if (msg) {
             vfio_add_kvm_msi_virq(vdev, vector, nr, true);
+            if (!vdev->defer_kvm_irq_routing) {
+                kvm_irqchip_commit_routes(kvm_state);
+                vfio_connect_kvm_msi_virq(vector);
+            }
         }
     }
 
@@ -586,13 +588,9 @@ static void vfio_prepare_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
 
 static void vfio_commit_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
 {
-    VFIOMSIVector *vector;
     int i;
 
-    if (!vdev->defer_kvm_irq_routing) {
-        return;
-    }
-
+    assert(vdev->defer_kvm_irq_routing);
     vdev->defer_kvm_irq_routing = false;
 
     if (!vdev->nr_vectors) {
@@ -602,19 +600,7 @@ static void vfio_commit_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
     kvm_irqchip_commit_routes(kvm_state);
 
     for (i = 0; i < vdev->nr_vectors; i++) {
-        vector = &vdev->msi_vectors[i];
-
-        if (!vector->use || vector->virq < 0) {
-            continue;
-        }
-
-        if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state,
-                                               &vector->kvm_interrupt,
-                                               NULL, vector->virq) < 0) {
-            kvm_irqchip_release_virq(kvm_state, vector->virq);
-            event_notifier_cleanup(&vector->kvm_interrupt);
-            vector->virq = -1;
-        }
+        vfio_connect_kvm_msi_virq(&vdev->msi_vectors[i]);
     }
 }
 

RE: [PATCH v4 6/6] vfio: defer to commit kvm irq routing when enable msi/msix

[Longpeng (Mike, Cloud Infrastructure Service Product Dept.)]

> -----Original Message-----
> From: Alex Williamson [mailto:alex.williamson@redhat.com]
> Sent: Friday, October 22, 2021 4:51 AM
> To: Longpeng (Mike, Cloud Infrastructure Service Product Dept.)
> <longpeng2@huawei.com>
> Cc: pbonzini@redhat.com; qemu-devel@nongnu.org; kvm@vger.kernel.org; Gonglei
> (Arei) <arei.gonglei@huawei.com>
> Subject: Re: [PATCH v4 6/6] vfio: defer to commit kvm irq routing when enable
> msi/msix
> 
> On Thu, 14 Oct 2021 08:48:52 +0800
> "Longpeng(Mike)" <longpeng2@huawei.com> wrote:
> 
> > In migration resume phase, all unmasked msix vectors need to be
> > setup when loading the VF state. However, the setup operation would
> > take longer if the VM has more VFs and each VF has more unmasked
> > vectors.
> >
> > The hot spot is kvm_irqchip_commit_routes, it'll scan and update
> > all irqfds that are already assigned each invocation, so more
> > vectors means need more time to process them.
> >
> > vfio_pci_load_config
> >   vfio_msix_enable
> >     msix_set_vector_notifiers
> >       for (vector = 0; vector < dev->msix_entries_nr; vector++) {
> >         vfio_msix_vector_do_use
> >           vfio_add_kvm_msi_virq
> >             kvm_irqchip_commit_routes <-- expensive
> >       }
> >
> > We can reduce the cost by only committing once outside the loop.
> > The routes are cached in kvm_state, we commit them first and then
> > bind irqfd for each vector.
> >
> > The test VM has 128 vcpus and 8 VF (each one has 65 vectors),
> > we measure the cost of the vfio_msix_enable for each VF, and
> > we can see 90+% costs can be reduce.
> >
> > VF      Count of irqfds[*]  Original        With this patch
> >
> > 1st           65            8               2
> > 2nd           130           15              2
> > 3rd           195           22              2
> > 4th           260           24              3
> > 5th           325           36              2
> > 6th           390           44              3
> > 7th           455           51              3
> > 8th           520           58              4
> > Total                       258ms           21ms
> >
> > [*] Count of irqfds
> > How many irqfds that already assigned and need to process in this
> > round.
> >
> > The optimization can be applied to msi type too.
> >
> > Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
> > ---
> >  hw/vfio/pci.c | 129
> ++++++++++++++++++++++++++++++++++++++++++++++------------
> >  hw/vfio/pci.h |   1 +
> >  2 files changed, 105 insertions(+), 25 deletions(-)
> >
> > diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
> > index 0bd832b..dca2d0c 100644
> > --- a/hw/vfio/pci.c
> > +++ b/hw/vfio/pci.c
> > @@ -413,8 +413,6 @@ static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool
> msix)
> >  static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector
> *vector,
> >                                    int vector_n, bool msix)
> >  {
> > -    int virq;
> > -
> >      if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
> >          return;
> >      }
> > @@ -423,20 +421,31 @@ static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev,
> VFIOMSIVector *vector,
> >          return;
> >      }
> >
> > -    virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev);
> > -    if (virq < 0) {
> > +    vector->virq = kvm_irqchip_add_deferred_msi_route(kvm_state, vector_n,
> > +                                                      &vdev->pdev);
> > +    if (vector->virq < 0) {
> >          event_notifier_cleanup(&vector->kvm_interrupt);
> > +        vector->virq = -1;
> 
> Nit, it seems like all negative values are equivalent here, I don't
> think we need to explicitly set virq to -1 given that it's already < 0.
> 

OK.

> > +        return;
> > +    }
> > +
> > +    if (vdev->defer_kvm_irq_routing) {
> > +        /*
> > +         * The vector->virq will be reset to -1 if we fail to add the
> > +         * corresponding irqfd in vfio_commit_kvm_msi_virq_batch().
> > +         */
> >          return;
> >      }
> >
> > +    kvm_irqchip_commit_routes(kvm_state);
> > +
> >      if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state,
> &vector->kvm_interrupt,
> > -                                       NULL, virq) < 0) {
> > -        kvm_irqchip_release_virq(kvm_state, virq);
> > +                                           NULL, vector->virq) < 0) {
> > +        kvm_irqchip_release_virq(kvm_state, vector->virq);
> >          event_notifier_cleanup(&vector->kvm_interrupt);
> > +        vector->virq = -1;
> >          return;
> >      }
> > -
> > -    vector->virq = virq;
> >  }
> >
> >  static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
> > @@ -501,11 +510,13 @@ static int vfio_msix_vector_do_use(PCIDevice *pdev,
> unsigned int nr,
> >       * increase them as needed.
> >       */
> >      if (vdev->nr_vectors < nr + 1) {
> > -        vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
> >          vdev->nr_vectors = nr + 1;
> > -        ret = vfio_enable_vectors(vdev, true);
> > -        if (ret) {
> > -            error_report("vfio: failed to enable vectors, %d", ret);
> > +        if (!vdev->defer_kvm_irq_routing) {
> > +            vfio_disable_irqindex(&vdev->vbasedev,
> VFIO_PCI_MSIX_IRQ_INDEX);
> > +            ret = vfio_enable_vectors(vdev, true);
> > +            if (ret) {
> > +                error_report("vfio: failed to enable vectors, %d", ret);
> > +            }
> >          }
> >      } else {
> >          Error *err = NULL;
> > @@ -567,6 +578,46 @@ static void vfio_msix_vector_release(PCIDevice *pdev,
> unsigned int nr)
> >      }
> >  }
> >
> > +static void vfio_prepare_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
> > +{
> > +    assert(!vdev->defer_kvm_irq_routing);
> > +    vdev->defer_kvm_irq_routing = true;
> > +}
> > +
> > +static void vfio_commit_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
> > +{
> > +    VFIOMSIVector *vector;
> > +    int i;
> > +
> > +    if (!vdev->defer_kvm_irq_routing) {
> > +        return;
> > +    }
> > +
> > +    vdev->defer_kvm_irq_routing = false;
> > +
> > +    if (!vdev->nr_vectors) {
> > +        return;
> > +    }
> > +
> > +    kvm_irqchip_commit_routes(kvm_state);
> > +
> > +    for (i = 0; i < vdev->nr_vectors; i++) {
> > +        vector = &vdev->msi_vectors[i];
> > +
> > +        if (!vector->use || vector->virq < 0) {
> > +            continue;
> > +        }
> > +
> > +        if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state,
> > +                                               &vector->kvm_interrupt,
> > +                                               NULL, vector->virq) < 0) {
> > +            kvm_irqchip_release_virq(kvm_state, vector->virq);
> > +            event_notifier_cleanup(&vector->kvm_interrupt);
> > +            vector->virq = -1;
> > +        }
> 
> I started trying to get rid of this code that largely duplicates the
> error case of vfio_add_kvm_msi_virq() and questioned why we setup the
> notifier separate from connecting it to the irqfd.  If we setup the
> notifier and irqfd in the same function we can clean things up a bit
> more and confine the deferred state tests in the vector-use code.  I
> think we can also assert if we have an unmatched batch commit call and
> we probably don't need to test both that a vector is in use and has a
> virq, one should not be true without the other.
> 
> Does this look like an improvement to you and would you like to roll it
> into this patch?  Thanks,
> 

Sure, it looks neater! I've tested it for about two days in local and it
works well. I'll send them (v5) out later. Thanks.

> Alex
> 
>  pci.c |   58 ++++++++++++++++++++++------------------------------------
>  1 file changed, 22 insertions(+), 36 deletions(-)
> 
> commit 6fb9336e3fe9e3775b0a0e7aadaff781fb52c0e7
> Author: Alex Williamson <alex.williamson@redhat.com>
> Date:   Thu Oct 21 13:35:12 2021 -0600
> 
>     cleanup
> 
> diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
> index 1792c30049da..5b3a86dd5292 100644
> --- a/hw/vfio/pci.c
> +++ b/hw/vfio/pci.c
> @@ -417,35 +417,33 @@ static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev,
> VFIOMSIVector *vector,
>          return;
>      }
> 
> -    if (event_notifier_init(&vector->kvm_interrupt, 0)) {
> -        return;
> -    }
> -
>      vector->virq = kvm_irqchip_add_deferred_msi_route(kvm_state, vector_n,
>                                                        &vdev->pdev);
> +}
> +
> +static void vfio_connect_kvm_msi_virq(VFIOMSIVector *vector)
> +{
>      if (vector->virq < 0) {
> -        event_notifier_cleanup(&vector->kvm_interrupt);
> -        vector->virq = -1;
>          return;
>      }
> 
> -    if (vdev->defer_kvm_irq_routing) {
> -        /*
> -         * The vector->virq will be reset to -1 if we fail to add the
> -         * corresponding irqfd in vfio_commit_kvm_msi_virq_batch().
> -         */
> -        return;
> +    if (event_notifier_init(&vector->kvm_interrupt, 0)) {
> +        goto fail_notifier;
>      }
> 
> -    kvm_irqchip_commit_routes(kvm_state);
> -
>      if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state,
> &vector->kvm_interrupt,
>                                             NULL, vector->virq) < 0) {
> -        kvm_irqchip_release_virq(kvm_state, vector->virq);
> -        event_notifier_cleanup(&vector->kvm_interrupt);
> -        vector->virq = -1;
> -        return;
> +        goto fail_kvm;
>      }
> +
> +    return;
> +
> +fail_kvm:
> +    event_notifier_cleanup(&vector->kvm_interrupt);
> +fail_notifier:
> +    kvm_irqchip_release_virq(kvm_state, vector->virq);
> +    vector->virq = -1;
> +    return;
>  }
> 
>  static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
> @@ -501,6 +499,10 @@ static int vfio_msix_vector_do_use(PCIDevice *pdev,
> unsigned int nr,
>      } else {
>          if (msg) {
>              vfio_add_kvm_msi_virq(vdev, vector, nr, true);
> +            if (!vdev->defer_kvm_irq_routing) {
> +                kvm_irqchip_commit_routes(kvm_state);
> +                vfio_connect_kvm_msi_virq(vector);
> +            }
>          }
>      }
> 
> @@ -586,13 +588,9 @@ static void vfio_prepare_kvm_msi_virq_batch(VFIOPCIDevice
> *vdev)
> 
>  static void vfio_commit_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
>  {
> -    VFIOMSIVector *vector;
>      int i;
> 
> -    if (!vdev->defer_kvm_irq_routing) {
> -        return;
> -    }
> -
> +    assert(vdev->defer_kvm_irq_routing);
>      vdev->defer_kvm_irq_routing = false;
> 
>      if (!vdev->nr_vectors) {
> @@ -602,19 +600,7 @@ static void vfio_commit_kvm_msi_virq_batch(VFIOPCIDevice
> *vdev)
>      kvm_irqchip_commit_routes(kvm_state);
> 
>      for (i = 0; i < vdev->nr_vectors; i++) {
> -        vector = &vdev->msi_vectors[i];
> -
> -        if (!vector->use || vector->virq < 0) {
> -            continue;
> -        }
> -
> -        if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state,
> -                                               &vector->kvm_interrupt,
> -                                               NULL, vector->virq) < 0) {
> -            kvm_irqchip_release_virq(kvm_state, vector->virq);
> -            event_notifier_cleanup(&vector->kvm_interrupt);
> -            vector->virq = -1;
> -        }
> +        vfio_connect_kvm_msi_virq(&vdev->msi_vectors[i]);
>      }
>  }
>