Skip to contentSkip to conversation

kvm: Implement atomic memory region resizes

RFC PATCH Superseded David Hildenbrand 3 participants 14 messages Mar 3, 2020 – Mar 6, 2020 KVM

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover kvm: Implement atomic memory region resizes

Currently, when doing a

    memory_region_ram_resize() -> memory_region_set_size()

the old KVM slot will first get removed and the new, resized one, will be re-added. This is fine as long as no IOCTL is currently using any data from such a memory slot (e.g., when building ACPI tables). However, if e.g., a VCPU is in KVM_RUN and tries to access any data on such a slot while we're growing it, we will get wrong faults while the slot is temporarily removed.

Let's allow to resize memory regions while the guest is running and might be using the regions. Inhibit any KVM ioctl while we are replacing the memory slot(s).

This is a preparation for virtio-mem (initially, x86-64 only), which wants to resize (esp. grow) ram memory regions while the guest is running via memory_region_ram_resize().

Using pause_all_vcpus()/resume_all_vcpus() is not possible, as it will temporarily drop the BQL - something most callers can't handle (esp. when called from vcpu context e.g., in virtio code).

Once we can handle resizes in the kernel (e.g., via KVM_SET_USER_MEMORY_REGION), we can make inhibiting optional at runtime.

Instead of inhibiting during the region_resize(), we could inhibit for the hole memory transaction (from begin() to commit()). This could be nice, because also splitting of memory regions would be atomic (I remember there was a BUG report regarding that), however, I am not sure if that might impact any RT users.

Tested so far with x86-64 KVM only. Thoughts? Anything important I am missing? Any alternatives that don't require kernel changes?

David Hildenbrand (4):
openpic_kvm: Use kvm_device_ioctl() instead of ioctl() intc/s390_flic_kvm.c: Use kvm_device_ioctl() instead of ioctl()
memory: Add region_resize() callback to memory notifier
kvm: Implement atomic memory region resizes via region_resize()

 accel/kvm/kvm-all.c     | 121 +++++++++++++++++++++++++++++++++++++---
 hw/intc/openpic_kvm.c   |   8 +--
 hw/intc/s390_flic_kvm.c |  22 ++++----
 include/exec/memory.h   |  18 ++++++
 include/hw/core/cpu.h   |   3 +
 memory.c                |  72 ++++++++++++++++++++++--
 6 files changed, 217 insertions(+), 27 deletions(-)
Signature · 1 line
2.24.1
#1 lore raw

RFC PATCH 1/4 openpic_kvm: Use kvm_device_ioctl() instead of ioctl()

Let's use the official variant, which will e.g., trace the call.

Cc Markus Armbruster <[email protected]>
Cc "Philippe Mathieu-Daudé" <[email protected]>
Cc "Marc-André Lureau" <[email protected]>
Cc Paolo Bonzini <[email protected]>
---
 hw/intc/openpic_kvm.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
hw/intc/openpic_kvm.c +4 -4
index e4bf47d885..0d8abc1604 100644
@@ -70,7 +70,7 @@ static void kvm_openpic_write(void *opaque, hwaddr addr, uint64_t val,
     attr.attr = addr;
     attr.addr = (uint64_t)(unsigned long)&val32;
 
-    ret = ioctl(opp->fd, KVM_SET_DEVICE_ATTR, &attr);
+    ret = kvm_device_ioctl(opp->fd, KVM_SET_DEVICE_ATTR, &attr);
     if (ret < 0) {
         qemu_log_mask(LOG_UNIMP, "%s: %s %" PRIx64 "\n", __func__,
                       strerror(errno), attr.attr);
@@ -96,7 +96,7 @@ static uint64_t kvm_openpic_read(void *opaque, hwaddr addr, unsigned size)
     attr.attr = addr;
     attr.addr = (uint64_t)(unsigned long)&val;
 
-    ret = ioctl(opp->fd, KVM_GET_DEVICE_ATTR, &attr);
+    ret = kvm_device_ioctl(opp->fd, KVM_GET_DEVICE_ATTR, &attr);
     if (ret < 0) {
         qemu_log_mask(LOG_UNIMP, "%s: %s %" PRIx64 "\n", __func__,
                       strerror(errno), attr.attr);
@@ -145,7 +145,7 @@ static void kvm_openpic_region_add(MemoryListener *listener,
     attr.attr = KVM_DEV_MPIC_BASE_ADDR;
     attr.addr = (uint64_t)(unsigned long)&reg_base;
 
-    ret = ioctl(opp->fd, KVM_SET_DEVICE_ATTR, &attr);
+    ret = kvm_device_ioctl(opp->fd, KVM_SET_DEVICE_ATTR, &attr);
     if (ret < 0) {
         fprintf(stderr, "%s: %s %" PRIx64 "\n", __func__,
                 strerror(errno), reg_base);
@@ -179,7 +179,7 @@ static void kvm_openpic_region_del(MemoryListener *listener,
     attr.attr = KVM_DEV_MPIC_BASE_ADDR;
     attr.addr = (uint64_t)(unsigned long)&reg_base;
 
-    ret = ioctl(opp->fd, KVM_SET_DEVICE_ATTR, &attr);
+    ret = kvm_device_ioctl(opp->fd, KVM_SET_DEVICE_ATTR, &attr);
     if (ret < 0) {
         fprintf(stderr, "%s: %s %" PRIx64 "\n", __func__,
                 strerror(errno), reg_base);
Signature · 1 line
2.24.1
#2 lore raw

RFC PATCH 2/4 intc/s390_flic_kvm.c: Use kvm_device_ioctl() instead of ioctl()

In reply to David Hildenbrand

Let's use the official variant, which will e.g., trace the call.

Cc Cornelia Huck <[email protected]>
Cc Halil Pasic <[email protected]>
Cc Christian Borntraeger <[email protected]>
---
 hw/intc/s390_flic_kvm.c | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)
hw/intc/s390_flic_kvm.c +11 -11
index a306b26faa..5151582ba0 100644
@@ -68,7 +68,7 @@ static int flic_get_all_irqs(KVMS390FLICState *flic,
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_GET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_GET_DEVICE_ATTR, &attr);
 
     return rc == -1 ? -errno : rc;
 }
@@ -80,7 +80,7 @@ static void flic_enable_pfault(KVMS390FLICState *flic)
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     if (rc) {
         fprintf(stderr, "flic: couldn't enable pfault\n");
@@ -94,7 +94,7 @@ static void flic_disable_wait_pfault(KVMS390FLICState *flic)
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     if (rc) {
         fprintf(stderr, "flic: couldn't disable pfault\n");
@@ -118,7 +118,7 @@ static int flic_enqueue_irqs(void *buf, uint64_t len,
         .attr = len,
     };
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     return rc ? -errno : 0;
 }
@@ -197,7 +197,7 @@ static int kvm_s390_clear_io_flic(S390FLICState *fs, uint16_t subchannel_id,
     if (unlikely(!flic->clear_io_supported)) {
         return -ENOSYS;
     }
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
     return rc ? -errno : 0;
 }
 
@@ -218,7 +218,7 @@ static int kvm_s390_modify_ais_mode(S390FLICState *fs, uint8_t isc,
         return -ENOSYS;
     }
 
-    return ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr) ? -errno : 0;
+    return kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr) ? -errno : 0;
 }
 
 static int kvm_s390_inject_airq(S390FLICState *fs, uint8_t type,
@@ -235,7 +235,7 @@ static int kvm_s390_inject_airq(S390FLICState *fs, uint8_t type,
         return -ENOSYS;
     }
 
-    return ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr) ? -errno : 0;
+    return kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr) ? -errno : 0;
 }
 
 /**
@@ -296,7 +296,7 @@ static int kvm_s390_register_io_adapter(S390FLICState *fs, uint32_t id,
         return 0;
     }
 
-    r = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    r = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     return r ? -errno : 0;
 }
@@ -321,7 +321,7 @@ static int kvm_s390_io_adapter_map(S390FLICState *fs, uint32_t id,
         return 0;
     }
 
-    r = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    r = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
     return r ? -errno : 0;
 }
 
@@ -519,7 +519,7 @@ static int kvm_flic_ais_post_load(void *opaque, int version_id)
         return -ENOSYS;
     }
 
-    return ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr) ? -errno : 0;
+    return kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr) ? -errno : 0;
 }
 
 static const VMStateDescription kvm_s390_flic_ais_tmp = {
@@ -636,7 +636,7 @@ static void kvm_s390_flic_reset(DeviceState *dev)
         }
     }
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
     if (rc) {
         trace_flic_reset_failed(errno);
     }
Signature · 1 line
2.24.1
#3 lore raw

Re: [PATCH RFC 2/4] intc/s390_flic_kvm.c: Use kvm_device_ioctl() instead of ioctl()

Quoted 43 lines · David Hildenbrand

Let's use the official variant, which will e.g., trace the call.

Cc Cornelia Huck <[email protected]>
Cc Halil Pasic <[email protected]>
Cc Christian Borntraeger <[email protected]>
---
 hw/intc/s390_flic_kvm.c | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)
hw/intc/s390_flic_kvm.c +3 -3
index a306b26faa..5151582ba0 100644
@@ -68,7 +68,7 @@ static int flic_get_all_irqs(KVMS390FLICState *flic,
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_GET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_GET_DEVICE_ATTR, &attr);
 
     return rc == -1 ? -errno : rc;
 }
@@ -80,7 +80,7 @@ static void flic_enable_pfault(KVMS390FLICState *flic)
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     if (rc) {
         fprintf(stderr, "flic: couldn't enable pfault\n");
@@ -94,7 +94,7 @@ static void flic_disable_wait_pfault(KVMS390FLICState *flic)
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     if (rc) {
         fprintf(stderr, "flic: couldn't disable pfault\n");

hw/intc/s390_flic_kvm.c static int flic_enqueue_irqs(void *buf, uint64_t len, · line 123

Quoted 7 lines
hw/intc/s390_flic_kvm.c +1 -1
@@ -118,7 +118,7 @@ static int flic_enqueue_irqs(void *buf, uint64_t len,
         .attr = len,
     };
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     return rc ? -errno : 0;

This is no longer necessary as this is done by kvm_device_ioctl, no?

Same for other location.s

#4 lore raw

Re: [PATCH RFC 2/4] intc/s390_flic_kvm.c: Use kvm_device_ioctl() instead of ioctl()

Quoted 49 lines · Christian Borntraeger
Quoted 46 lines · David Hildenbrand

Let's use the official variant, which will e.g., trace the call.

Cc Cornelia Huck <[email protected]>
Cc Halil Pasic <[email protected]>
Cc Christian Borntraeger <[email protected]>
---
 hw/intc/s390_flic_kvm.c | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)
hw/intc/s390_flic_kvm.c +3 -3
index a306b26faa..5151582ba0 100644
@@ -68,7 +68,7 @@ static int flic_get_all_irqs(KVMS390FLICState *flic,
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_GET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_GET_DEVICE_ATTR, &attr);
 
     return rc == -1 ? -errno : rc;
 }
@@ -80,7 +80,7 @@ static void flic_enable_pfault(KVMS390FLICState *flic)
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     if (rc) {
         fprintf(stderr, "flic: couldn't enable pfault\n");
@@ -94,7 +94,7 @@ static void flic_disable_wait_pfault(KVMS390FLICState *flic)
     };
     int rc;
 
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     if (rc) {
         fprintf(stderr, "flic: couldn't disable pfault\n");
@@ -118,7 +118,7 @@ static int flic_enqueue_irqs(void *buf, uint64_t len,
         .attr = len,
     };
 
Quoted 9 lines

hw/intc/s390_flic_kvm.c static int flic_enqueue_irqs(void *buf, uint64_t len, · line 123

Quoted 4 lines
hw/intc/s390_flic_kvm.c +1 -1
-    rc = ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
+    rc = kvm_device_ioctl(flic->fd, KVM_SET_DEVICE_ATTR, &attr);
 
     return rc ? -errno : 0;

This is no longer necessary as this is done by kvm_device_ioctl, no?

Same for other location.s

Right, missed that. Thanks for pointing that out!

Signature · 3 lines
Thanks,

David / dhildenb
#5 lore raw

RFC PATCH 3/4 memory: Add region_resize() callback to memory notifier

In reply to David Hildenbrand

Let's provide a way for memory notifiers to get notified about a resize. If the region_resize() callback is not implemented by a notifier, we mimic the old behavior by removing the old section and adding the new, resized section.

This callback is helpful when backends (like KVM) want to implement atomic resizes of memory sections (e.g., resize while VCPUs are running and using the section).

For now, we won't bother about a resize() callback for coalesced MMIO regions. The main future use case is for virtio-mem to resize ram memory regions while the guest is running.

Cc Richard Henderson <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc "Dr. David Alan Gilbert" <[email protected]>
Cc Eduardo Habkost <[email protected]>
Cc Marcel Apfelbaum <[email protected]>
Cc Igor Mammedov <[email protected]>
---
 include/exec/memory.h | 18 +++++++++++
 memory.c              | 72 ++++++++++++++++++++++++++++++++++++++++---
 2 files changed, 85 insertions(+), 5 deletions(-)
include/exec/memory.h +18 -0
index 74805dd448..704eb64d75 100644
@@ -492,6 +492,24 @@ struct MemoryListener {
      */
     void (*region_nop)(MemoryListener *listener, MemoryRegionSection *section);
 
+    /**
+     * @region_resize:
+     *
+     * Called during an address space update transaction,
+     * for a section of the address space that is in the same place in the
+     * address space as in the last transaction, however, the size changed.
+     * Dirty memory logging can change as well.
+     *
+     * Note: If this callback is not implemented, the resize is communicated
+     *       via a region_del(), followed by a region_add() instead.
+     *
+     * @listener: The #MemoryListener.
+     * @section: The old #MemoryRegionSection.
+     * @new: The new size.
+     */
+    void (*region_resize)(MemoryListener *listener,
+                          MemoryRegionSection *section, Int128 new);
+
     /**
      * @log_start:
      *
memory.c +67 -5

90 lines not on this page · raw message

Signature · 1 line
2.24.1
#6 lore raw

RFC PATCH 4/4 kvm: Implement atomic memory region resizes via region_resize()

In reply to David Hildenbrand

virtio-mem wants to resize (esp. grow) ram memory regions while the guest is already aware of them and makes use of them. Resizing a KVM slot can only currently be done by removing it and re-adding it. While the kvm slot is temporarily removed, VCPUs that try to read from these slots will fault. But also, other ioctls might depend on all slots being in place.

Let's inhibit most KVM ioctls while performing the resize. Once we have an ioctl that can perform atomic resizes (e.g., KVM_SET_USER_MEMORY_REGION extensions), we can make inhibiting optional at runtime.

Also, make sure to hold the kvm_slots_lock while performing both actions (removing+re-adding).

Note: Resizes of memory regions currently seems to happen during bootup only, so I don't think any existing RT users should be affected.

Cc Richard Henderson <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc "Dr. David Alan Gilbert" <[email protected]>
Cc Eduardo Habkost <[email protected]>
Cc Marcel Apfelbaum <[email protected]>
Cc Igor Mammedov <[email protected]>
---
 accel/kvm/kvm-all.c   | 121 +++++++++++++++++++++++++++++++++++++++---
 include/hw/core/cpu.h |   3 ++
 2 files changed, 117 insertions(+), 7 deletions(-)
accel/kvm/kvm-all.c +114 -7
index 439a4efe52..bba58db098 100644
@@ -149,6 +149,21 @@ bool kvm_msi_use_devid;
 static bool kvm_immediate_exit;
 static hwaddr kvm_max_slot_size = ~0;
 
+/*
+ * While holding this lock in write, no new KVM ioctls can be started, but
+ * kvm ioctl inhibitors will have to wait for existing ones to finish
+ * (indicated by cpu->in_ioctl and kvm_in_ioctl, both updated with this lock
+ * held in read when entering the ioctl).
+ */
+pthread_rwlock_t kvm_ioctl_lock;
+/*
+ * Atomic counter of active KVM ioctls except
+ * - The KVM ioctl inhibitor is doing an ioctl
+ * - kvm_ioctl(): Harmless and not interesting for inhibitors.
+ * - kvm_vcpu_ioctl(): Tracked via cpu->in_ioctl.
+ */
+static int kvm_in_ioctl;
+
 static const KVMCapabilityInfo kvm_required_capabilites[] = {
     KVM_CAP_INFO(USER_MEMORY),
     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
@@ -1023,6 +1038,7 @@ void kvm_set_max_memslot_size(hwaddr max_slot_size)
     kvm_max_slot_size = max_slot_size;
 }
 
+/* Called with KVMMemoryListener.slots_lock held */
 static void kvm_set_phys_mem(KVMMemoryListener *kml,
                              MemoryRegionSection *section, bool add)
 {
@@ -1052,14 +1068,12 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml,
     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
           (start_addr - section->offset_within_address_space);
 
-    kvm_slots_lock(kml);
-
     if (!add) {
         do {
             slot_size = MIN(kvm_max_slot_size, size);
             mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
             if (!mem) {
-                goto out;
+                return;
             }
             if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
                 kvm_physical_sync_dirty_bitmap(kml, section);
@@ -1079,7 +1093,7 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml,
             start_addr += slot_size;
             size -= slot_size;
         } while (size);
-        goto out;
+        return;
     }
 
     /* register the new slot */
@@ -1108,9 +1122,6 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml,
         ram += slot_size;
         size -= slot_size;
     } while (size);
-
-out:
-    kvm_slots_unlock(kml);
 }
 
 static void kvm_region_add(MemoryListener *listener,
@@ -1119,7 +1130,9 @@ static void kvm_region_add(MemoryListener *listener,
     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 
     memory_region_ref(section->mr);
+    kvm_slots_lock(kml);
     kvm_set_phys_mem(kml, section, true);
+    kvm_slots_unlock(kml);
 }
 
 static void kvm_region_del(MemoryListener *listener,
@@ -1127,10 +1140,68 @@ static void kvm_region_del(MemoryListener *listener,
 {
     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 
+    kvm_slots_lock(kml);
     kvm_set_phys_mem(kml, section, false);
+    kvm_slots_unlock(kml);
     memory_region_unref(section->mr);
 }
 
+/*
+ * Certain updates (e.g., resizing memory regions) require temporarily removing
+ * kvm memory slots. Make sure any ioctl sees a consistent memory slot state.
+ */
+static void kvm_ioctl_inhibit_begin(void)
+{
+    CPUState *cpu;
+
+    /*
+     * We allow to inhibit only when holding the BQL, so we can identify
+     * when an inhibitor wants to issue an ioctl easily.
+     */
+    g_assert(qemu_mutex_iothread_locked());
+
+    pthread_rwlock_wrlock(&kvm_ioctl_lock);
+
+    /* Inhibiting happens rarely, we can keep things simple and spin here. */
+    while (true) {
+        bool any_cpu_in_ioctl = false;
+
+        CPU_FOREACH(cpu) {
+            if (atomic_read(&cpu->in_ioctl)) {
+                any_cpu_in_ioctl = true;
+                qemu_cpu_kick(cpu);
+            }
+        }
+        if (!any_cpu_in_ioctl && !atomic_read(&kvm_in_ioctl)) {
+            break;
+        }
+        g_usleep(100);
+    }
+}
+
+static void kvm_ioctl_inhibit_end(void)
+{
+    pthread_rwlock_unlock(&kvm_ioctl_lock);
+}
+
+static void kvm_region_resize(MemoryListener *listener,
+                              MemoryRegionSection *section, Int128 new)
+{
+    KVMMemoryListener *kml = container_of(listener, KVMMemoryListener,
+                                          listener);
+    MemoryRegionSection new_section = *section;
+
+    new_section.size = new;
+
+    kvm_slots_lock(kml);
+    /* Inhibit KVM ioctls while temporarily removing slots. */
+    kvm_ioctl_inhibit_begin();
+    kvm_set_phys_mem(kml, section, false);
+    kvm_set_phys_mem(kml, &new_section, true);
+    kvm_ioctl_inhibit_end();
+    kvm_slots_unlock(kml);
+}
+
 static void kvm_log_sync(MemoryListener *listener,
                          MemoryRegionSection *section)
 {
@@ -1249,6 +1320,7 @@ void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
 
     kml->listener.region_add = kvm_region_add;
     kml->listener.region_del = kvm_region_del;
+    kml->listener.region_resize = kvm_region_resize;
     kml->listener.log_start = kvm_log_start;
     kml->listener.log_stop = kvm_log_stop;
     kml->listener.log_sync = kvm_log_sync;
@@ -1894,6 +1966,7 @@ static int kvm_init(MachineState *ms)
     assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size);
 
     s->sigmask_len = 8;
+    pthread_rwlock_init(&kvm_ioctl_lock, NULL);
 
 #ifdef KVM_CAP_SET_GUEST_DEBUG
     QTAILQ_INIT(&s->kvm_sw_breakpoints);
@@ -2304,6 +2377,34 @@ static void kvm_eat_signals(CPUState *cpu)
     } while (sigismember(&chkset, SIG_IPI));
 }
 
+static void kvm_cpu_set_in_ioctl(CPUState *cpu, bool in_ioctl)
+{
+    if (unlikely(qemu_mutex_iothread_locked())) {
+        return;
+    }
+    if (in_ioctl) {
+        pthread_rwlock_rdlock(&kvm_ioctl_lock);
+        atomic_set(&cpu->in_ioctl, true);
+        pthread_rwlock_unlock(&kvm_ioctl_lock);
+    } else {
+        atomic_set(&cpu->in_ioctl, false);
+    }
+}
+
+static void kvm_set_in_ioctl(bool in_ioctl)
+{
+    if (likely(qemu_mutex_iothread_locked())) {
+        return;
+    }
+    if (in_ioctl) {
+        pthread_rwlock_rdlock(&kvm_ioctl_lock);
+        atomic_inc(&kvm_in_ioctl);
+        pthread_rwlock_unlock(&kvm_ioctl_lock);
+    } else {
+        atomic_dec(&kvm_in_ioctl);
+    }
+}
+
 int kvm_cpu_exec(CPUState *cpu)
 {
     struct kvm_run *run = cpu->kvm_run;
@@ -2488,7 +2589,9 @@ int kvm_vm_ioctl(KVMState *s, int type, ...)
     va_end(ap);
 
     trace_kvm_vm_ioctl(type, arg);
+    kvm_set_in_ioctl(true);
     ret = ioctl(s->vmfd, type, arg);
+    kvm_set_in_ioctl(false);
     if (ret == -1) {
         ret = -errno;
     }
@@ -2506,7 +2609,9 @@ int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
     va_end(ap);
 
     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
+    kvm_cpu_set_in_ioctl(cpu, true);
     ret = ioctl(cpu->kvm_fd, type, arg);
+    kvm_cpu_set_in_ioctl(cpu, false);
     if (ret == -1) {
         ret = -errno;
     }
@@ -2524,7 +2629,9 @@ int kvm_device_ioctl(int fd, int type, ...)
     va_end(ap);
 
     trace_kvm_device_ioctl(fd, type, arg);
+    kvm_set_in_ioctl(true);
     ret = ioctl(fd, type, arg);
+    kvm_set_in_ioctl(false);
     if (ret == -1) {
         ret = -errno;
     }
include/hw/core/cpu.h +3 -0
index 73e9a869a4..4fbff6f3d7 100644
@@ -431,6 +431,9 @@ struct CPUState {
     /* shared by kvm, hax and hvf */
     bool vcpu_dirty;
 
+    /* kvm only for now: CPU is in kvm_vcpu_ioctl() (esp. KVM_RUN) */
+    bool in_ioctl;
+
     /* Used to keep track of an outstanding cpu throttle thread for migration
      * autoconverge
      */
Signature · 1 line
2.24.1
maintainer
#7 lore raw

Re: [PATCH RFC 4/4] kvm: Implement atomic memory region resizes via region_resize()

Quoted 4 lines · David Hildenbrand

virtio-mem wants to resize (esp. grow) ram memory regions while the guest is already aware of them and makes use of them. Resizing a KVM slot can only currently be done by removing it and re-adding it. While the kvm slot is temporarily removed, VCPUs that try to read from these slots will fault.

Only fetches I think? Data reads and write would be treated as MMIO accesses and they should just work (using either the old or new FlatView).

Quoted 11 lines

But also, other ioctls might depend on all slots being in place.

Let's inhibit most KVM ioctls while performing the resize. Once we have an ioctl that can perform atomic resizes (e.g., KVM_SET_USER_MEMORY_REGION extensions), we can make inhibiting optional at runtime.

Also, make sure to hold the kvm_slots_lock while performing both actions (removing+re-adding).

Note: Resizes of memory regions currently seems to happen during bootup only, so I don't think any existing RT users should be affected.

rwlocks are not efficient, they cause cache line contention. For MMIO-heavy workloads the impact will be very large (well, not that large because right now they all take the BQL, but one can always hope).

I would very much prefer to add a KVM_SET_USER_MEMORY_REGION extension right away.

Paolo

Quoted 257 lines
Cc Richard Henderson <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc "Dr. David Alan Gilbert" <[email protected]>
Cc Eduardo Habkost <[email protected]>
Cc Marcel Apfelbaum <[email protected]>
Cc Igor Mammedov <[email protected]>
---
 accel/kvm/kvm-all.c   | 121 +++++++++++++++++++++++++++++++++++++++---
 include/hw/core/cpu.h |   3 ++
 2 files changed, 117 insertions(+), 7 deletions(-)
accel/kvm/kvm-all.c +114 -7
index 439a4efe52..bba58db098 100644
@@ -149,6 +149,21 @@ bool kvm_msi_use_devid;
 static bool kvm_immediate_exit;
 static hwaddr kvm_max_slot_size = ~0;
 
+/*
+ * While holding this lock in write, no new KVM ioctls can be started, but
+ * kvm ioctl inhibitors will have to wait for existing ones to finish
+ * (indicated by cpu->in_ioctl and kvm_in_ioctl, both updated with this lock
+ * held in read when entering the ioctl).
+ */
+pthread_rwlock_t kvm_ioctl_lock;
+/*
+ * Atomic counter of active KVM ioctls except
+ * - The KVM ioctl inhibitor is doing an ioctl
+ * - kvm_ioctl(): Harmless and not interesting for inhibitors.
+ * - kvm_vcpu_ioctl(): Tracked via cpu->in_ioctl.
+ */
+static int kvm_in_ioctl;
+
 static const KVMCapabilityInfo kvm_required_capabilites[] = {
     KVM_CAP_INFO(USER_MEMORY),
     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
@@ -1023,6 +1038,7 @@ void kvm_set_max_memslot_size(hwaddr max_slot_size)
     kvm_max_slot_size = max_slot_size;
 }
 
+/* Called with KVMMemoryListener.slots_lock held */
 static void kvm_set_phys_mem(KVMMemoryListener *kml,
                              MemoryRegionSection *section, bool add)
 {
@@ -1052,14 +1068,12 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml,
     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
           (start_addr - section->offset_within_address_space);
 
-    kvm_slots_lock(kml);
-
     if (!add) {
         do {
             slot_size = MIN(kvm_max_slot_size, size);
             mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
             if (!mem) {
-                goto out;
+                return;
             }
             if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
                 kvm_physical_sync_dirty_bitmap(kml, section);
@@ -1079,7 +1093,7 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml,
             start_addr += slot_size;
             size -= slot_size;
         } while (size);
-        goto out;
+        return;
     }
 
     /* register the new slot */
@@ -1108,9 +1122,6 @@ static void kvm_set_phys_mem(KVMMemoryListener *kml,
         ram += slot_size;
         size -= slot_size;
     } while (size);
-
-out:
-    kvm_slots_unlock(kml);
 }
 
 static void kvm_region_add(MemoryListener *listener,
@@ -1119,7 +1130,9 @@ static void kvm_region_add(MemoryListener *listener,
     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 
     memory_region_ref(section->mr);
+    kvm_slots_lock(kml);
     kvm_set_phys_mem(kml, section, true);
+    kvm_slots_unlock(kml);
 }
 
 static void kvm_region_del(MemoryListener *listener,
@@ -1127,10 +1140,68 @@ static void kvm_region_del(MemoryListener *listener,
 {
     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 
+    kvm_slots_lock(kml);
     kvm_set_phys_mem(kml, section, false);
+    kvm_slots_unlock(kml);
     memory_region_unref(section->mr);
 }
 
+/*
+ * Certain updates (e.g., resizing memory regions) require temporarily removing
+ * kvm memory slots. Make sure any ioctl sees a consistent memory slot state.
+ */
+static void kvm_ioctl_inhibit_begin(void)
+{
+    CPUState *cpu;
+
+    /*
+     * We allow to inhibit only when holding the BQL, so we can identify
+     * when an inhibitor wants to issue an ioctl easily.
+     */
+    g_assert(qemu_mutex_iothread_locked());
+
+    pthread_rwlock_wrlock(&kvm_ioctl_lock);
+
+    /* Inhibiting happens rarely, we can keep things simple and spin here. */
+    while (true) {
+        bool any_cpu_in_ioctl = false;
+
+        CPU_FOREACH(cpu) {
+            if (atomic_read(&cpu->in_ioctl)) {
+                any_cpu_in_ioctl = true;
+                qemu_cpu_kick(cpu);
+            }
+        }
+        if (!any_cpu_in_ioctl && !atomic_read(&kvm_in_ioctl)) {
+            break;
+        }
+        g_usleep(100);
+    }
+}
+
+static void kvm_ioctl_inhibit_end(void)
+{
+    pthread_rwlock_unlock(&kvm_ioctl_lock);
+}
+
+static void kvm_region_resize(MemoryListener *listener,
+                              MemoryRegionSection *section, Int128 new)
+{
+    KVMMemoryListener *kml = container_of(listener, KVMMemoryListener,
+                                          listener);
+    MemoryRegionSection new_section = *section;
+
+    new_section.size = new;
+
+    kvm_slots_lock(kml);
+    /* Inhibit KVM ioctls while temporarily removing slots. */
+    kvm_ioctl_inhibit_begin();
+    kvm_set_phys_mem(kml, section, false);
+    kvm_set_phys_mem(kml, &new_section, true);
+    kvm_ioctl_inhibit_end();
+    kvm_slots_unlock(kml);
+}
+
 static void kvm_log_sync(MemoryListener *listener,
                          MemoryRegionSection *section)
 {
@@ -1249,6 +1320,7 @@ void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
 
     kml->listener.region_add = kvm_region_add;
     kml->listener.region_del = kvm_region_del;
+    kml->listener.region_resize = kvm_region_resize;
     kml->listener.log_start = kvm_log_start;
     kml->listener.log_stop = kvm_log_stop;
     kml->listener.log_sync = kvm_log_sync;
@@ -1894,6 +1966,7 @@ static int kvm_init(MachineState *ms)
     assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size);
 
     s->sigmask_len = 8;
+    pthread_rwlock_init(&kvm_ioctl_lock, NULL);
 
 #ifdef KVM_CAP_SET_GUEST_DEBUG
     QTAILQ_INIT(&s->kvm_sw_breakpoints);
@@ -2304,6 +2377,34 @@ static void kvm_eat_signals(CPUState *cpu)
     } while (sigismember(&chkset, SIG_IPI));
 }
 
+static void kvm_cpu_set_in_ioctl(CPUState *cpu, bool in_ioctl)
+{
+    if (unlikely(qemu_mutex_iothread_locked())) {
+        return;
+    }
+    if (in_ioctl) {
+        pthread_rwlock_rdlock(&kvm_ioctl_lock);
+        atomic_set(&cpu->in_ioctl, true);
+        pthread_rwlock_unlock(&kvm_ioctl_lock);
+    } else {
+        atomic_set(&cpu->in_ioctl, false);
+    }
+}
+
+static void kvm_set_in_ioctl(bool in_ioctl)
+{
+    if (likely(qemu_mutex_iothread_locked())) {
+        return;
+    }
+    if (in_ioctl) {
+        pthread_rwlock_rdlock(&kvm_ioctl_lock);
+        atomic_inc(&kvm_in_ioctl);
+        pthread_rwlock_unlock(&kvm_ioctl_lock);
+    } else {
+        atomic_dec(&kvm_in_ioctl);
+    }
+}
+
 int kvm_cpu_exec(CPUState *cpu)
 {
     struct kvm_run *run = cpu->kvm_run;
@@ -2488,7 +2589,9 @@ int kvm_vm_ioctl(KVMState *s, int type, ...)
     va_end(ap);
 
     trace_kvm_vm_ioctl(type, arg);
+    kvm_set_in_ioctl(true);
     ret = ioctl(s->vmfd, type, arg);
+    kvm_set_in_ioctl(false);
     if (ret == -1) {
         ret = -errno;
     }
@@ -2506,7 +2609,9 @@ int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
     va_end(ap);
 
     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
+    kvm_cpu_set_in_ioctl(cpu, true);
     ret = ioctl(cpu->kvm_fd, type, arg);
+    kvm_cpu_set_in_ioctl(cpu, false);
     if (ret == -1) {
         ret = -errno;
     }
@@ -2524,7 +2629,9 @@ int kvm_device_ioctl(int fd, int type, ...)
     va_end(ap);
 
     trace_kvm_device_ioctl(fd, type, arg);
+    kvm_set_in_ioctl(true);
     ret = ioctl(fd, type, arg);
+    kvm_set_in_ioctl(false);
     if (ret == -1) {
         ret = -errno;
     }
include/hw/core/cpu.h +3 -0
index 73e9a869a4..4fbff6f3d7 100644
@@ -431,6 +431,9 @@ struct CPUState {
     /* shared by kvm, hax and hvf */
     bool vcpu_dirty;
 
+    /* kvm only for now: CPU is in kvm_vcpu_ioctl() (esp. KVM_RUN) */
+    bool in_ioctl;
+
     /* Used to keep track of an outstanding cpu throttle thread for migration
      * autoconverge
      */
#8 lore raw

Re: [PATCH RFC 4/4] kvm: Implement atomic memory region resizes via region_resize()

Quoted 6 lines · Paolo Bonzini
Quoted 4 lines · David Hildenbrand

virtio-mem wants to resize (esp. grow) ram memory regions while the guest is already aware of them and makes use of them. Resizing a KVM slot can only currently be done by removing it and re-adding it. While the kvm slot is temporarily removed, VCPUs that try to read from these slots will fault.

s/try to read/try to access/

Quoted 2 lines

Only fetches I think? Data reads and write would be treated as MMIO accesses and they should just work (using either the old or new FlatView).

On x86-64, I saw KVM fault printks getting printed (it was about 1-2 years ago, though, when I realized this was a problem). Could be that these were fetches. At least the guest eventually crashed :)

On other archs (esp. s390x) guests will directly receive a PGM_ADDRESSING from KVM if they stumble over memory that is not covered by a kvm slot.

Quoted 10 lines
Quoted 10 lines

But also, other ioctls might depend on all slots being in place.

Let's inhibit most KVM ioctls while performing the resize. Once we have an ioctl that can perform atomic resizes (e.g., KVM_SET_USER_MEMORY_REGION extensions), we can make inhibiting optional at runtime.

Also, make sure to hold the kvm_slots_lock while performing both actions (removing+re-adding).

Quoted 6 lines
Quoted 2 lines

Note: Resizes of memory regions currently seems to happen during bootup only, so I don't think any existing RT users should be affected.

rwlocks are not efficient, they cause cache line contention. For MMIO-heavy workloads the impact will be very large (well, not that large because right now they all take the BQL, but one can always hope).

Yeah, rwlocks are not optimal and I am still looking for better alternatives (suggestions welcome :) ). Using RCU might not work, because the rcu_read region might be too big (esp. while in KVM_RUN).

I had a prototype which used a bunch of atomics + qemu_cond_wait. But it was quite elaborate and buggy.

(I assume only going into KVM_RUN is really affected, and I do wonder if it will be noticeable at all. Doing an ioctl is always already an expensive operation.)

I can look into per-cpu locks instead of the rwlock.

Quoted 4 lines

I would very much prefer to add a KVM_SET_USER_MEMORY_REGION extension right away.

I really want to avoid dependencies on kernel features to at least make it work for now. Especially, resizing memory slots in KVM (especially while dirty bitmaps, rmaps, etc. are active) is non-trivial.

Thanks Paolo!

Signature · 3 lines
Thanks,

David / dhildenb
maintainer
#9 lore raw

Re: [PATCH RFC 4/4] kvm: Implement atomic memory region resizes via region_resize()

Quoted 4 lines · David Hildenbrand

Yeah, rwlocks are not optimal and I am still looking for better alternatives (suggestions welcome :) ). Using RCU might not work, because the rcu_read region might be too big (esp. while in KVM_RUN).

Quoted 8 lines

I had a prototype which used a bunch of atomics + qemu_cond_wait. But it was quite elaborate and buggy.

(I assume only going into KVM_RUN is really affected, and I do wonder if it will be noticeable at all. Doing an ioctl is always already an expensive operation.)

I can look into per-cpu locks instead of the rwlock.

Assuming we're only talking about CPU ioctls (seems like a good approximation) maybe you could use start_exclusive/end_exclusive? The current_cpu->in_exclusive_context assignments can be made conditional on "if (current_cpu)".

However that means you have to drop the BQL, see
process_queued_cpu_work. It may be a problem.

Paolo

#10 lore raw

Re: [PATCH RFC 4/4] kvm: Implement atomic memory region resizes via region_resize()

Quoted 8 lines · Paolo Bonzini
Quoted 7 lines · David Hildenbrand

Yeah, rwlocks are not optimal and I am still looking for better alternatives (suggestions welcome :) ). Using RCU might not work, because the rcu_read region might be too big (esp. while in KVM_RUN).

I had a prototype which used a bunch of atomics + qemu_cond_wait. But it was quite elaborate and buggy.

Quoted 7 lines
Quoted 5 lines

(I assume only going into KVM_RUN is really affected, and I do wonder if it will be noticeable at all. Doing an ioctl is always already an expensive operation.)

I can look into per-cpu locks instead of the rwlock.

Assuming we're only talking about CPU ioctls (seems like a good

Yeah, I guess most !CPU ioctls are done under the BQL.

Quoted 6 lines

approximation) maybe you could use start_exclusive/end_exclusive? The current_cpu->in_exclusive_context assignments can be made conditional on "if (current_cpu)".

However that means you have to drop the BQL, see
process_queued_cpu_work. It may be a problem.

Thanks, I'll look into that. I currently have a simple cpu->ioctl_mutex.

Cheers!

Signature · 3 lines
Thanks,

David / dhildenb
#11 lore raw

Re: [PATCH RFC 4/4] kvm: Implement atomic memory region resizes via region_resize()

In reply to Paolo Bonzini
Quoted 14 lines · Paolo Bonzini
Quoted 4 lines · David Hildenbrand

Yeah, rwlocks are not optimal and I am still looking for better alternatives (suggestions welcome :) ). Using RCU might not work, because the rcu_read region might be too big (esp. while in KVM_RUN).

Quoted 8 lines

I had a prototype which used a bunch of atomics + qemu_cond_wait. But it was quite elaborate and buggy.

(I assume only going into KVM_RUN is really affected, and I do wonder if it will be noticeable at all. Doing an ioctl is always already an expensive operation.)

I can look into per-cpu locks instead of the rwlock.

Quoted 8 lines

Assuming we're only talking about CPU ioctls (seems like a good approximation) maybe you could use start_exclusive/end_exclusive? The current_cpu->in_exclusive_context assignments can be made conditional on "if (current_cpu)".

However that means you have to drop the BQL, see
process_queued_cpu_work. It may be a problem.

Yeah, start_exclusive() is expected to be called without the BQL, otherwise the other CPUs would not be able to make progress and can eventually be "caught".

It's essentially the same reason why I can't use high-level pause_all_vcpus()/resume_all_vcpus(). Will drop the BQL which is very bad for resizing code.

Signature · 3 lines
Thanks,

David / dhildenb
maintainer
#12 lore raw

Re: [PATCH RFC 4/4] kvm: Implement atomic memory region resizes via region_resize()

Quoted 8 lines · David Hildenbrand
Quoted 8 lines

Assuming we're only talking about CPU ioctls (seems like a good approximation) maybe you could use start_exclusive/end_exclusive? The current_cpu->in_exclusive_context assignments can be made conditional on "if (current_cpu)".

However that means you have to drop the BQL, see
process_queued_cpu_work. It may be a problem.

Quoted 7 lines

Yeah, start_exclusive() is expected to be called without the BQL, otherwise the other CPUs would not be able to make progress and can eventually be "caught".

It's essentially the same reason why I can't use high-level pause_all_vcpus()/resume_all_vcpus(). Will drop the BQL which is very bad for resizing code.

But any other synchronization primitive that you do which blocks all vCPUs will have the same issue, otherwise you get a deadlock.

Paolo

#13 lore raw

Re: [PATCH RFC 4/4] kvm: Implement atomic memory region resizes via region_resize()

Quoted 13 lines · Paolo Bonzini
Quoted 12 lines · David Hildenbrand
Quoted 8 lines

Assuming we're only talking about CPU ioctls (seems like a good approximation) maybe you could use start_exclusive/end_exclusive? The current_cpu->in_exclusive_context assignments can be made conditional on "if (current_cpu)".

However that means you have to drop the BQL, see
process_queued_cpu_work. It may be a problem.

Yeah, start_exclusive() is expected to be called without the BQL, otherwise the other CPUs would not be able to make progress and can eventually be "caught".

Quoted 6 lines
Quoted 3 lines

It's essentially the same reason why I can't use high-level pause_all_vcpus()/resume_all_vcpus(). Will drop the BQL which is very bad for resizing code.

But any other synchronization primitive that you do which blocks all vCPUs will have the same issue, otherwise you get a deadlock.

This is essentially what this patch solves.

The lock essentially blocks anybody from entering, but not leaving a KVM ioctl. An inhibitor only waits for all IOCTLs to be left. No other lock prohibits that, so I don't think there can ever be a deadlock.

Signature · 3 lines
Thanks,

David / dhildenb