Skip to contentSkip to conversation

optimize the downtime for vfio migration

PATCH In review Longpeng(Mike) 2 participants 7 messages Mar 26, 2022 – Mar 28, 2022 KVM

lore
Share & tools
lore
#0 lore raw

v6 cover optimize the downtime for vfio migration

From: Longpeng <[email protected]>

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 are the preparations for the optimization. Patch 5 optimizes the vfio msix setup path.

v5: https://lore.kernel.org/all/[email protected]/T/

Change v5->v6:
- remove the Patch 4("kvm: irqchip: extract kvm_irqchip_add_deferred_msi_route") of v5, and use KVMRouteChange API instead. [Paolo, Longpeng]

Changes v4->v5:
- setup the notifier and irqfd in the same function to makes
the code neater. [Alex]

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) (5):
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
Revert "vfio: Avoid disabling and enabling vectors repeatedly in VFIO
migration"
vfio: defer to commit kvm irq routing when enable msi/msix

 hw/vfio/pci.c | 183 +++++++++++++++++++++++++++++++-------------------
 hw/vfio/pci.h |   2 +
 2 files changed, 115 insertions(+), 70 deletions(-)
Signature · 1 line
2.23.0
#1 lore raw

PATCH v6 1/5 vfio: simplify the conditional statements in vfio_msi_enable

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.

---
 hw/vfio/pci.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
hw/vfio/pci.c +2 -2
index d07a4e99b1..1433989aeb 100644
@@ -653,7 +653,7 @@ 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);
         }
@@ -671,7 +671,7 @@ 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;
         }
Signature · 1 line
2.23.0
#2 lore raw

PATCH v6 2/5 vfio: move re-enabling INTX out of the common helper

In reply to Longpeng(Mike)

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

---
 hw/vfio/pci.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)
hw/vfio/pci.c +11 -6
index 1433989aeb..a1916e2e36 100644
@@ -693,7 +693,6 @@ retry:
 
 static void vfio_msi_disable_common(VFIOPCIDevice *vdev)
 {
-    Error *err = NULL;
     int i;
 
     for (i = 0; i < vdev->nr_vectors; i++) {
@@ -712,15 +711,11 @@ 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);
@@ -741,6 +736,10 @@ 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));
@@ -750,8 +749,14 @@ 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);
 }
Signature · 1 line
2.23.0
#3 lore raw

PATCH v6 3/5 vfio: simplify the failure path in vfio_msi_enable

In reply to Longpeng(Mike)

Use vfio_msi_disable_common to simplify the error handling in vfio_msi_enable.

---
 hw/vfio/pci.c | 16 ++--------------
 1 file changed, 2 insertions(+), 14 deletions(-)
hw/vfio/pci.c +2 -14
index a1916e2e36..6f49e71cd4 100644
@@ -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
@@ -658,24 +659,12 @@ 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.
@@ -683,7 +672,6 @@ retry:
          * out to fall back to INTx for this device.
          */
         error_report("vfio: Error: Failed to enable MSI");
-        vdev->interrupt = VFIO_INT_NONE;
 
         return;
     }
Signature · 1 line
2.23.0
#4 lore raw

PATCH v6 4/5 Revert "vfio: Avoid disabling and enabling vectors repeatedly in VFIO migration"

In reply to 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.

---
 hw/vfio/pci.c | 20 +++-----------------
 1 file changed, 3 insertions(+), 17 deletions(-)
hw/vfio/pci.c +3 -17
index 6f49e71cd4..6801391cf6 100644
@@ -572,9 +572,6 @@ 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);
@@ -593,22 +590,11 @@ 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");
     }
Signature · 1 line
2.23.0
#5 lore raw

PATCH v6 5/5 vfio: defer to commit kvm irq routing when enable msi/msix

In reply to 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.

---
 hw/vfio/pci.c | 130 +++++++++++++++++++++++++++++++++++++-------------
 hw/vfio/pci.h |   2 +
 2 files changed, 99 insertions(+), 33 deletions(-)
hw/vfio/pci.c +97 -33
index 6801391cf6..c58d7fa00f 100644
@@ -45,6 +45,9 @@
 
 #define TYPE_VFIO_PCI_NOHOTPLUG "vfio-pci-nohotplug"
 
+/* Protected by BQL */
+static KVMRouteChange vfio_route_change;
+
 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);
@@ -413,33 +416,36 @@ 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)
 {
-    KVMRouteChange c;
-    int virq;
-
     if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
         return;
     }
 
-    if (event_notifier_init(&vector->kvm_interrupt, 0)) {
+    vector->virq = kvm_irqchip_add_msi_route(&vfio_route_change,
+                                             vector_n, &vdev->pdev);
+}
+
+static void vfio_connect_kvm_msi_virq(VFIOMSIVector *vector)
+{
+    if (vector->virq < 0) {
         return;
     }
 
-    c = kvm_irqchip_begin_route_changes(kvm_state);
-    virq = kvm_irqchip_add_msi_route(&c, vector_n, &vdev->pdev);
-    if (virq < 0) {
-        event_notifier_cleanup(&vector->kvm_interrupt);
-        return;
+    if (event_notifier_init(&vector->kvm_interrupt, 0)) {
+        goto fail_notifier;
     }
-    kvm_irqchip_commit_route_changes(&c);
 
     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
-                                       NULL, virq) < 0) {
-        kvm_irqchip_release_virq(kvm_state, virq);
-        event_notifier_cleanup(&vector->kvm_interrupt);
-        return;
+                                           NULL, vector->virq) < 0) {
+        goto fail_kvm;
     }
 
-    vector->virq = virq;
+    return;
+
+fail_kvm:
+    event_notifier_cleanup(&vector->kvm_interrupt);
+fail_notifier:
+    kvm_irqchip_release_virq(kvm_state, vector->virq);
+    vector->virq = -1;
 }
 
 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
@@ -494,7 +500,14 @@ 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) {
+                vfio_add_kvm_msi_virq(vdev, vector, nr, true);
+            } else {
+                vfio_route_change = kvm_irqchip_begin_route_changes(kvm_state);
+                vfio_add_kvm_msi_virq(vdev, vector, nr, true);
+                kvm_irqchip_commit_route_changes(&vfio_route_change);
+                vfio_connect_kvm_msi_virq(vector);
+            }
         }
     }
 
@@ -504,11 +517,13 @@ 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;
@@ -570,6 +585,27 @@ 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;
+    vfio_route_change = kvm_irqchip_begin_route_changes(kvm_state);
+}
+
+static void vfio_commit_kvm_msi_virq_batch(VFIOPCIDevice *vdev)
+{
+    int i;
+
+    assert(vdev->defer_kvm_irq_routing);
+    vdev->defer_kvm_irq_routing = false;
+
+    kvm_irqchip_commit_route_changes(&vfio_route_change);
+
+    for (i = 0; i < vdev->nr_vectors; i++) {
+        vfio_connect_kvm_msi_virq(&vdev->msi_vectors[i]);
+    }
+}
+
 static void vfio_msix_enable(VFIOPCIDevice *vdev)
 {
     vfio_disable_interrupts(vdev);
@@ -579,26 +615,45 @@ 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);
 }
 
@@ -608,6 +663,13 @@ 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);
@@ -633,6 +695,8 @@ 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;
 
hw/vfio/pci.h +2 -0
index 64777516d1..7c236a52f4 100644
@@ -19,6 +19,7 @@
 #include "qemu/queue.h"
 #include "qemu/timer.h"
 #include "qom/object.h"
+#include "sysemu/kvm.h"
 
 #define PCI_ANY_ID (~0)
 
@@ -171,6 +172,7 @@ struct VFIOPCIDevice {
     bool no_kvm_ioeventfd;
     bool no_vfio_ioeventfd;
     bool enable_ramfb;
+    bool defer_kvm_irq_routing;
     VFIODisplay *dpy;
     Notifier irqchip_change_notifier;
 };
Signature · 1 line
2.23.0
#6 lore raw
In reply to Longpeng(Mike)
Quoted 57 lines · On Sat, 26 Mar 2022 14:02:21 +0800 "Longpeng(Mike)" <[email protected]>…

From: Longpeng <[email protected]>

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 are the preparations for the optimization. Patch 5 optimizes the vfio msix setup path.

v5: https://lore.kernel.org/all/[email protected]/T/

Change v5->v6:
- remove the Patch 4("kvm: irqchip: extract kvm_irqchip_add_deferred_msi_route") of v5, and use KVMRouteChange API instead. [Paolo, Longpeng]

Changes v4->v5:
- setup the notifier and irqfd in the same function to makes
the code neater. [Alex]

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) (5):
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
Revert "vfio: Avoid disabling and enabling vectors repeatedly in VFIO
migration"
vfio: defer to commit kvm irq routing when enable msi/msix

Quoted 4 lines
 hw/vfio/pci.c | 183 +++++++++++++++++++++++++++++++-------------------
 hw/vfio/pci.h |   2 +
 2 files changed, 115 insertions(+), 70 deletions(-)

Nice to see you found a solution with Paolo's suggestion for begin/commit batching. Looks ok to me; I'll queue this for after the v7.0 QEMU release and look for further reviews and comments in the interim. Thanks,

Alex