Skip to contentSkip to conversation

virtio-mem: block size and address-assignment optimizations

PATCH Superseded David Hildenbrand 2 participants 9 messages Sep 28, 2020 – Sep 29, 2020

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover virtio-mem: block size and address-assignment optimizations

Let's try to detect the actual THP size and use it as default block size (unless the page size of the backend indicates that THP don't apply). Always allow to set a block size of 1 MiB, but warn if the configured block size is smaller than the default. Handle large block sizes better, avoiding a virtio-spec violation and optimizing address auto-detection.

For existing setups (x86-64), the default block size won't change (was, and will be 2 MiB on anonymous memory). For existing x86-64 setups, the address auto-detection won't change in relevant setups (esp., anonymous memory and hugetlbfs with 2 MiB pages and no manual configuration of the block size). I don't see the need for compatibility handling (especially, as virtio-mem is still not considered production-ready).

Most of this is a preparation for future architectures, using hugetlbfs to full extend, and using manually configured, larger block sizes (relevant for vfio in the future).

v1 -> v2:
- Tweak some patch descriptions
- "virtio-mem: Probe THP size to determine default block size"
-- Beautify THP detection a bit.
-- Assume THP might only get used if the memory backend page size corresponds
to the real hostpage size.
-- Use virtio_mem_default_block_size(RAMBlock *rb) to handle selection
of the default block size for a RAMBlock.
-- Implement virtio_mem_get_block_size() as preparation for patch #5
- "memory-device: Add get_min_alignment() callback"
-- Simplify documentation.
- "virito-mem: Implement get_min_alignment()"
-- Simplify due to changes in patch #1.

David Hildenbrand (5):
virtio-mem: Probe THP size to determine default block size
virtio-mem: Check that "memaddr" is multiples of the block size
memory-device: Support big alignment requirements
memory-device: Add get_min_alignment() callback
virito-mem: Implement get_min_alignment()

 hw/mem/memory-device.c         |  20 ++++--
 hw/virtio/virtio-mem-pci.c     |   7 +++
 hw/virtio/virtio-mem.c         | 110 +++++++++++++++++++++++++++++++--
 include/hw/mem/memory-device.h |  10 +++
 4 files changed, 137 insertions(+), 10 deletions(-)
Signature · 1 line
2.26.2
#1 lore raw

PATCH v2 1/5 virtio-mem: Probe THP size to determine default block size

Let's allow a minimum block size of 1 MiB in all configurations. Select the default block size based on
- The page size of the memory backend.
- The THP size if the memory backend size corresponds to the real hsot page size.
- The global minimum of 1 MiB.
and warn if something smaller is configured by the user.

VIRTIO_MEM only supports Linux (depends on LINUX), so we can probe the THP size unconditionally.

For now we only support virtio-mem on x86-64 - there isn't a user-visiable change (x86-64 only supports 2 MiB THP on the PMD level) - the default was, and will be 2 MiB.

If we ever have THP on the PUD level (e.g., 1 GiB THP on x86-64), we expect to have a trigger to explicitly opt-in for the new THP granularity.

Cc "Michael S. Tsirkin" <[email protected]>
Cc Wei Yang <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
Cc Igor Mammedov <[email protected]>
Cc Pankaj Gupta <[email protected]>
---
 hw/virtio/virtio-mem.c | 105 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 101 insertions(+), 4 deletions(-)
hw/virtio/virtio-mem.c +101 -4
index 8fbec77ccc..9b1461cf9d 100644
@@ -33,10 +33,83 @@
 #include "trace.h"
 
 /*
- * Use QEMU_VMALLOC_ALIGN, so no THP will have to be split when unplugging
- * memory (e.g., 2MB on x86_64).
+ * Let's not allow blocks smaller than 1 MiB, for example, to keep the tracking
+ * bitmap small.
  */
-#define VIRTIO_MEM_MIN_BLOCK_SIZE ((uint32_t)QEMU_VMALLOC_ALIGN)
+#define VIRTIO_MEM_MIN_BLOCK_SIZE ((uint32_t)(1 * MiB))
+
+#if defined(__x86_64__) || defined(__arm__) || defined(__aarch64__) || \
+    defined(__powerpc64__)
+#define VIRTIO_MEM_DEFAULT_THP_SIZE ((uint32_t)(2 * MiB))
+#else
+        /* fallback to 1 MiB (e.g., the THP size on s390x) */
+#define VIRTIO_MEM_DEFAULT_THP_SIZE VIRTIO_MEM_MIN_BLOCK_SIZE
+#endif
+
+/*
+ * We want to have a reasonable default block size such that
+ * 1. We avoid splitting THPs when unplugging memory, which degrades
+ *    performance.
+ * 2. We avoid placing THPs for plugged blocks that also cover unplugged
+ *    blocks.
+ *
+ * The actual THP size might differ between Linux kernels, so we try to probe
+ * it. In the future (if we ever run into issues regarding 2.), we might want
+ * to disable THP in case we fail to properly probe the THP size, or if the
+ * block size is configured smaller than the THP size.
+ */
+static uint32_t thp_size;
+
+#define HPAGE_PMD_SIZE_PATH "/sys/kernel/mm/transparent_hugepage/hpage_pmd_size"
+static uint32_t virtio_mem_thp_size(void)
+{
+    gchar *content = NULL;
+    const char *endptr;
+    uint64_t tmp;
+
+    if (thp_size) {
+        return thp_size;
+    }
+
+    /*
+     * Try to probe the actual THP size, fallback to (sane but eventually
+     * incorrect) default sizes.
+     */
+    if (g_file_get_contents(HPAGE_PMD_SIZE_PATH, &content, NULL, NULL) &&
+        !qemu_strtou64(content, &endptr, 0, &tmp) &&
+        (!endptr || *endptr == '\n')) {
+        /*
+         * Sanity-check the value, if it's too big (e.g., aarch64 with 64k base
+         * pages) or weird, fallback to something smaller.
+         */
+        if (!tmp || !is_power_of_2(tmp) || tmp > 16 * MiB) {
+            warn_report("Read unsupported THP size: %" PRIx64, tmp);
+        } else {
+            thp_size = tmp;
+        }
+    }
+
+    if (!thp_size) {
+        thp_size = VIRTIO_MEM_DEFAULT_THP_SIZE;
+        warn_report("Could not detect THP size, falling back to %" PRIx64
+                    "  MiB.", thp_size / MiB);
+    }
+
+    g_free(content);
+    return thp_size;
+}
+
+static uint64_t virtio_mem_default_block_size(RAMBlock *rb)
+{
+    const uint64_t page_size = qemu_ram_pagesize(rb);
+
+    /* We can have hugetlbfs with a page size smaller than the THP size. */
+    if (page_size == qemu_real_host_page_size) {
+        return MAX(page_size, virtio_mem_thp_size());
+    }
+    return MAX(page_size, VIRTIO_MEM_MIN_BLOCK_SIZE);
+}
+
 /*
  * Size the usable region bigger than the requested size if possible. Esp.
  * Linux guests will only add (aligned) memory blocks in case they fully
@@ -437,10 +510,23 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp)
     rb = vmem->memdev->mr.ram_block;
     page_size = qemu_ram_pagesize(rb);
 
+    /*
+     * If the block size wasn't configured by the user, use a sane default. This
+     * allows using hugetlbfs backends of any page size without manual
+     * intervention.
+     */
+    if (!vmem->block_size) {
+        vmem->block_size = virtio_mem_default_block_size(rb);
+    }
+
     if (vmem->block_size < page_size) {
         error_setg(errp, "'%s' property has to be at least the page size (0x%"
                    PRIx64 ")", VIRTIO_MEM_BLOCK_SIZE_PROP, page_size);
         return;
+    } else if (vmem->block_size < virtio_mem_default_block_size(rb)) {
+        warn_report("'%s' property is smaller than the default block size (%"
+                    PRIx64 " MiB)", VIRTIO_MEM_BLOCK_SIZE_PROP,
+                    virtio_mem_default_block_size(rb) / MiB);
     } else if (!QEMU_IS_ALIGNED(vmem->requested_size, vmem->block_size)) {
         error_setg(errp, "'%s' property has to be multiples of '%s' (0x%" PRIx64
                    ")", VIRTIO_MEM_REQUESTED_SIZE_PROP,
@@ -731,6 +817,18 @@ static void virtio_mem_get_block_size(Object *obj, Visitor *v, const char *name,
     const VirtIOMEM *vmem = VIRTIO_MEM(obj);
     uint64_t value = vmem->block_size;
 
+    /*
+     * If not configured by the user (and we're not realized yet), use the
+     * default block size we would use with the current memory backend.
+     */
+    if (!value) {
+        if (vmem->memdev && memory_region_is_ram(&vmem->memdev->mr)) {
+            value = virtio_mem_default_block_size(vmem->memdev->mr.ram_block);
+        } else {
+            value = virtio_mem_thp_size();
+        }
+    }
+
     visit_type_size(v, name, &value, errp);
 }
 
@@ -810,7 +908,6 @@ static void virtio_mem_instance_init(Object *obj)
 {
     VirtIOMEM *vmem = VIRTIO_MEM(obj);
 
-    vmem->block_size = VIRTIO_MEM_MIN_BLOCK_SIZE;
     notifier_list_init(&vmem->size_change_notifiers);
     vmem->precopy_notifier.notify = virtio_mem_precopy_notify;
 
Signature · 1 line
2.26.2
#7 lore raw

PATCH v2 4/5 memory-device: Add get_min_alignment() callback

In reply to David Hildenbrand

Add a callback that can be used to express additional alignment requirements (exceeding the ones from the memory region).

Will be used by virtio-mem to express special alignment requirements due to manually configured, big block sizes (e.g., 1GB with an ordinary memory-backend-ram). This avoids failing later when realizing, because auto-detection wasn't able to assign a properly aligned address.

Cc "Michael S. Tsirkin" <[email protected]>
Cc Wei Yang <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
Cc Igor Mammedov <[email protected]>
Cc Pankaj Gupta <[email protected]>
---
 hw/mem/memory-device.c         | 11 +++++++++--
 include/hw/mem/memory-device.h | 10 ++++++++++
 2 files changed, 19 insertions(+), 2 deletions(-)
hw/mem/memory-device.c +9 -2
index 8a736f1a26..cf0627fd01 100644
@@ -259,7 +259,7 @@ void memory_device_pre_plug(MemoryDeviceState *md, MachineState *ms,
 {
     const MemoryDeviceClass *mdc = MEMORY_DEVICE_GET_CLASS(md);
     Error *local_err = NULL;
-    uint64_t addr, align;
+    uint64_t addr, align = 0;
     MemoryRegion *mr;
 
     mr = mdc->get_memory_region(md, &local_err);
@@ -267,7 +267,14 @@ void memory_device_pre_plug(MemoryDeviceState *md, MachineState *ms,
         goto out;
     }
 
-    align = legacy_align ? *legacy_align : memory_region_get_alignment(mr);
+    if (legacy_align) {
+        align = *legacy_align;
+    } else {
+        if (mdc->get_min_alignment) {
+            align = mdc->get_min_alignment(md);
+        }
+        align = MAX(align, memory_region_get_alignment(mr));
+    }
     addr = mdc->get_addr(md);
     addr = memory_device_get_free_addr(ms, !addr ? NULL : &addr, align,
                                        memory_region_size(mr), &local_err);
include/hw/mem/memory-device.h +10 -0
index cde52e83c9..dcef8d65f7 100644
@@ -88,6 +88,16 @@ struct MemoryDeviceClass {
      */
     MemoryRegion *(*get_memory_region)(MemoryDeviceState *md, Error **errp);
 
+    /*
+     * Optional: Return the desired minimum alignment of the device in guest
+     * physical address space. The final alignment is computed based on this
+     * alignment and the alignment requirements of the memory region.
+     *
+     * Called when plugging the memory device to detect the required alignment
+     * during address assignment.
+     */
+    uint64_t (*get_min_alignment)(const MemoryDeviceState *md);
+
     /*
      * Translate the memory device into #MemoryDeviceInfo.
      */
Signature · 1 line
2.26.2
#5 lore raw

PATCH v2 2/5 virtio-mem: Check that "memaddr" is multiples of the block size

In reply to David Hildenbrand

The spec requires us to set the "addr" in guest physical address space to multiples of the block size. In some cases, this is not the case right
now: For example, when starting a VM with 4 GiB boot memory and a virtio-mem device with a block size of 2 GiB, "memaddr" will be auto-assigned to 0x140000000 / 5 GiB.

We'll try to improve auto-assignment for memory devices next, to avoid bailing out in case memory device code selects a bad address.

Note: The Linux driver doesn't support such big block sizes yet.

Cc "Michael S. Tsirkin" <[email protected]>
Cc Wei Yang <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
Cc Igor Mammedov <[email protected]>
Cc Pankaj Gupta <[email protected]>
---
 hw/virtio/virtio-mem.c | 5 +++++
 1 file changed, 5 insertions(+)
hw/virtio/virtio-mem.c +5 -0
index 9b1461cf9d..878c0b4f21 100644
@@ -532,6 +532,11 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp)
                    ")", VIRTIO_MEM_REQUESTED_SIZE_PROP,
                    VIRTIO_MEM_BLOCK_SIZE_PROP, vmem->block_size);
         return;
+    } else if (!QEMU_IS_ALIGNED(vmem->addr, vmem->block_size)) {
+        error_setg(errp, "'%s' property has to be multiples of '%s' (0x%" PRIx64
+                   ")", VIRTIO_MEM_ADDR_PROP, VIRTIO_MEM_BLOCK_SIZE_PROP,
+                   vmem->block_size);
+        return;
     } else if (!QEMU_IS_ALIGNED(memory_region_size(&vmem->memdev->mr),
                                 vmem->block_size)) {
         error_setg(errp, "'%s' property memdev size has to be multiples of"
Signature · 1 line
2.26.2
#6 lore raw

PATCH v2 3/5 memory-device: Support big alignment requirements

In reply to David Hildenbrand

Let's warn instead of bailing out - the worst thing that can happen is that we'll fail hot/coldplug later. The user got warned, and this should be rare.

This will be necessary for memory devices with rather big (user-defined) alignment requirements - say a virtio-mem device with a 2G block size - which will become important, for example, when supporting vfio in the future.

Cc "Michael S. Tsirkin" <[email protected]>
Cc Wei Yang <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
Cc Igor Mammedov <[email protected]>
Cc Pankaj Gupta <[email protected]>
---
 hw/mem/memory-device.c | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)
hw/mem/memory-device.c +5 -4
index 4bc9cf0917..8a736f1a26 100644
@@ -119,9 +119,10 @@ static uint64_t memory_device_get_free_addr(MachineState *ms,
 
     /* start of address space indicates the maximum alignment we expect */
     if (!QEMU_IS_ALIGNED(range_lob(&as), align)) {
-        error_setg(errp, "the alignment (0x%" PRIx64 ") is not supported",
-                   align);
-        return 0;
+        warn_report("the alignment (0x%" PRIx64 ") exceeds the expected"
+                    " maximum alignment, memory will get fragmented and not"
+                    " all 'maxmem' might be usable for memory devices.",
+                    align);
     }
 
     memory_device_check_addable(ms, size, &err);
@@ -151,7 +152,7 @@ static uint64_t memory_device_get_free_addr(MachineState *ms,
             return 0;
         }
     } else {
-        if (range_init(&new, range_lob(&as), size)) {
+        if (range_init(&new, QEMU_ALIGN_UP(range_lob(&as), align), size)) {
             error_setg(errp, "can't add memory device, device too big");
             return 0;
         }
Signature · 1 line
2.26.2
#8 lore raw

PATCH v2 5/5 virito-mem: Implement get_min_alignment()

In reply to David Hildenbrand

The block size determines the alignment requirements. Implement get_min_alignment() of the TYPE_MEMORY_DEVICE interface.

This allows auto-assignment of a properly aligned address in guest physical address space. For example, when specifying a 2GB block size for a virtio-mem device with 10GB with a memory setup "-m 4G, 20G", we'll no longer fail when realizing.

Cc "Michael S. Tsirkin" <[email protected]>
Cc Wei Yang <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
Cc Igor Mammedov <[email protected]>
Cc Pankaj Gupta <[email protected]>
---
 hw/virtio/virtio-mem-pci.c | 7 +++++++
 1 file changed, 7 insertions(+)
hw/virtio/virtio-mem-pci.c +7 -0
index 590cec041b..aac975a17c 100644
@@ -75,6 +75,12 @@ static void virtio_mem_pci_fill_device_info(const MemoryDeviceState *md,
     info->type = MEMORY_DEVICE_INFO_KIND_VIRTIO_MEM;
 }
 
+static uint64_t virtio_mem_pci_get_min_alignment(const MemoryDeviceState *md)
+{
+    return object_property_get_uint(OBJECT(md), VIRTIO_MEM_BLOCK_SIZE_PROP,
+                                    &error_abort);
+}
+
 static void virtio_mem_pci_size_change_notify(Notifier *notifier, void *data)
 {
     VirtIOMEMPCI *pci_mem = container_of(notifier, VirtIOMEMPCI,
@@ -109,6 +115,7 @@ static void virtio_mem_pci_class_init(ObjectClass *klass, void *data)
     mdc->get_plugged_size = virtio_mem_pci_get_plugged_size;
     mdc->get_memory_region = virtio_mem_pci_get_memory_region;
     mdc->fill_device_info = virtio_mem_pci_fill_device_info;
+    mdc->get_min_alignment = virtio_mem_pci_get_min_alignment;
 }
 
 static void virtio_mem_pci_instance_init(Object *obj)
Signature · 1 line
2.26.2
#2 lore raw

Re: [PATCH v2 1/5] virtio-mem: Probe THP size to determine default block size

In reply to David Hildenbrand
Quoted 4 lines

Let's allow a minimum block size of 1 MiB in all configurations. Select the default block size based on
- The page size of the memory backend.
- The THP size if the memory backend size corresponds to the real hsot

s/hsot/host

Quoted 8 lines

page size.
- The global minimum of 1 MiB.
and warn if something smaller is configured by the user.

VIRTIO_MEM only supports Linux (depends on LINUX), so we can probe the THP size unconditionally.

For now we only support virtio-mem on x86-64 - there isn't a user-visiable

s/visiable/visible

Quoted 91 lines

change (x86-64 only supports 2 MiB THP on the PMD level) - the default was, and will be 2 MiB.

If we ever have THP on the PUD level (e.g., 1 GiB THP on x86-64), we expect to have a trigger to explicitly opt-in for the new THP granularity.

Cc "Michael S. Tsirkin" <[email protected]>
Cc Wei Yang <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
Cc Igor Mammedov <[email protected]>
Cc Pankaj Gupta <[email protected]>
---
 hw/virtio/virtio-mem.c | 105 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 101 insertions(+), 4 deletions(-)
hw/virtio/virtio-mem.c +63 -3
index 8fbec77ccc..9b1461cf9d 100644
@@ -33,10 +33,83 @@
 #include "trace.h"
 
 /*
- * Use QEMU_VMALLOC_ALIGN, so no THP will have to be split when unplugging
- * memory (e.g., 2MB on x86_64).
+ * Let's not allow blocks smaller than 1 MiB, for example, to keep the tracking
+ * bitmap small.
  */
-#define VIRTIO_MEM_MIN_BLOCK_SIZE ((uint32_t)QEMU_VMALLOC_ALIGN)
+#define VIRTIO_MEM_MIN_BLOCK_SIZE ((uint32_t)(1 * MiB))
+
+#if defined(__x86_64__) || defined(__arm__) || defined(__aarch64__) || \
+    defined(__powerpc64__)
+#define VIRTIO_MEM_DEFAULT_THP_SIZE ((uint32_t)(2 * MiB))
+#else
+        /* fallback to 1 MiB (e.g., the THP size on s390x) */
+#define VIRTIO_MEM_DEFAULT_THP_SIZE VIRTIO_MEM_MIN_BLOCK_SIZE
+#endif
+
+/*
+ * We want to have a reasonable default block size such that
+ * 1. We avoid splitting THPs when unplugging memory, which degrades
+ *    performance.
+ * 2. We avoid placing THPs for plugged blocks that also cover unplugged
+ *    blocks.
+ *
+ * The actual THP size might differ between Linux kernels, so we try to probe
+ * it. In the future (if we ever run into issues regarding 2.), we might want
+ * to disable THP in case we fail to properly probe the THP size, or if the
+ * block size is configured smaller than the THP size.
+ */
+static uint32_t thp_size;
+
+#define HPAGE_PMD_SIZE_PATH "/sys/kernel/mm/transparent_hugepage/hpage_pmd_size"
+static uint32_t virtio_mem_thp_size(void)
+{
+    gchar *content = NULL;
+    const char *endptr;
+    uint64_t tmp;
+
+    if (thp_size) {
+        return thp_size;
+    }
+
+    /*
+     * Try to probe the actual THP size, fallback to (sane but eventually
+     * incorrect) default sizes.
+     */
+    if (g_file_get_contents(HPAGE_PMD_SIZE_PATH, &content, NULL, NULL) &&
+        !qemu_strtou64(content, &endptr, 0, &tmp) &&
+        (!endptr || *endptr == '\n')) {
+        /*
+         * Sanity-check the value, if it's too big (e.g., aarch64 with 64k base
+         * pages) or weird, fallback to something smaller.
+         */
+        if (!tmp || !is_power_of_2(tmp) || tmp > 16 * MiB) {
+            warn_report("Read unsupported THP size: %" PRIx64, tmp);
+        } else {
+            thp_size = tmp;
+        }
+    }
+
+    if (!thp_size) {
+        thp_size = VIRTIO_MEM_DEFAULT_THP_SIZE;
+        warn_report("Could not detect THP size, falling back to %" PRIx64
+                    "  MiB.", thp_size / MiB);
+    }
+
+    g_free(content);
+    return thp_size;

hw/virtio/virtio-mem.c line 109

Quoted 10 lines
hw/virtio/virtio-mem.c +10 -0
+}
+
+static uint64_t virtio_mem_default_block_size(RAMBlock *rb)
+{
+    const uint64_t page_size = qemu_ram_pagesize(rb);
+
+    /* We can have hugetlbfs with a page size smaller than the THP size. */
+    if (page_size == qemu_real_host_page_size) {
+        return MAX(page_size, virtio_mem_thp_size());
+    }

This condition is special, can think of hugetlbfs smaller in size than THP size configured.

hw/virtio/virtio-mem.c line 110

Quoted 1 line
hw/virtio/virtio-mem.c +1 -0
+    return MAX(page_size, VIRTIO_MEM_MIN_BLOCK_SIZE);

Do we still need this? Or we can have only one return for both the cases? Probably, I am missing something here.

Quoted 55 lines
hw/virtio/virtio-mem.c +27 -1
+}
+
 /*
  * Size the usable region bigger than the requested size if possible. Esp.
  * Linux guests will only add (aligned) memory blocks in case they fully
@@ -437,10 +510,23 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp)
     rb = vmem->memdev->mr.ram_block;
     page_size = qemu_ram_pagesize(rb);
 
+    /*
+     * If the block size wasn't configured by the user, use a sane default. This
+     * allows using hugetlbfs backends of any page size without manual
+     * intervention.
+     */
+    if (!vmem->block_size) {
+        vmem->block_size = virtio_mem_default_block_size(rb);
+    }
+
     if (vmem->block_size < page_size) {
         error_setg(errp, "'%s' property has to be at least the page size (0x%"
                    PRIx64 ")", VIRTIO_MEM_BLOCK_SIZE_PROP, page_size);
         return;
+    } else if (vmem->block_size < virtio_mem_default_block_size(rb)) {
+        warn_report("'%s' property is smaller than the default block size (%"
+                    PRIx64 " MiB)", VIRTIO_MEM_BLOCK_SIZE_PROP,
+                    virtio_mem_default_block_size(rb) / MiB);
     } else if (!QEMU_IS_ALIGNED(vmem->requested_size, vmem->block_size)) {
         error_setg(errp, "'%s' property has to be multiples of '%s' (0x%" PRIx64
                    ")", VIRTIO_MEM_REQUESTED_SIZE_PROP,
@@ -731,6 +817,18 @@ static void virtio_mem_get_block_size(Object *obj, Visitor *v, const char *name,
     const VirtIOMEM *vmem = VIRTIO_MEM(obj);
     uint64_t value = vmem->block_size;
 
+    /*
+     * If not configured by the user (and we're not realized yet), use the
+     * default block size we would use with the current memory backend.
+     */
+    if (!value) {
+        if (vmem->memdev && memory_region_is_ram(&vmem->memdev->mr)) {
+            value = virtio_mem_default_block_size(vmem->memdev->mr.ram_block);
+        } else {
+            value = virtio_mem_thp_size();
+        }
+    }
+
     visit_type_size(v, name, &value, errp);
 }
 
@@ -810,7 +908,6 @@ static void virtio_mem_instance_init(Object *obj)
 {
     VirtIOMEM *vmem = VIRTIO_MEM(obj);
 
-    vmem->block_size = VIRTIO_MEM_MIN_BLOCK_SIZE;
     notifier_list_init(&vmem->size_change_notifiers);
     vmem->precopy_notifier.notify = virtio_mem_precopy_notify;
#3 lore raw

Re: [PATCH v2 1/5] virtio-mem: Probe THP size to determine default block size

Quoted 6 lines · Pankaj Gupta
Quoted 4 lines

Let's allow a minimum block size of 1 MiB in all configurations. Select the default block size based on
- The page size of the memory backend.
- The THP size if the memory backend size corresponds to the real hsot

s/hsot/host

thanks!

Quoted 10 lines
Quoted 8 lines

page size.
- The global minimum of 1 MiB.
and warn if something smaller is configured by the user.

VIRTIO_MEM only supports Linux (depends on LINUX), so we can probe the THP size unconditionally.

For now we only support virtio-mem on x86-64 - there isn't a user-visiable

s/visiable/visible

thanks!

Quoted 96 lines
Quoted 96 lines

change (x86-64 only supports 2 MiB THP on the PMD level) - the default was, and will be 2 MiB.

If we ever have THP on the PUD level (e.g., 1 GiB THP on x86-64), we expect to have a trigger to explicitly opt-in for the new THP granularity.

Cc "Michael S. Tsirkin" <[email protected]>
Cc Wei Yang <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
Cc Igor Mammedov <[email protected]>
Cc Pankaj Gupta <[email protected]>
---
 hw/virtio/virtio-mem.c | 105 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 101 insertions(+), 4 deletions(-)
hw/virtio/virtio-mem.c +68 -3
index 8fbec77ccc..9b1461cf9d 100644
@@ -33,10 +33,83 @@
 #include "trace.h"
 
 /*
- * Use QEMU_VMALLOC_ALIGN, so no THP will have to be split when unplugging
- * memory (e.g., 2MB on x86_64).
+ * Let's not allow blocks smaller than 1 MiB, for example, to keep the tracking
+ * bitmap small.
  */
-#define VIRTIO_MEM_MIN_BLOCK_SIZE ((uint32_t)QEMU_VMALLOC_ALIGN)
+#define VIRTIO_MEM_MIN_BLOCK_SIZE ((uint32_t)(1 * MiB))
+
+#if defined(__x86_64__) || defined(__arm__) || defined(__aarch64__) || \
+    defined(__powerpc64__)
+#define VIRTIO_MEM_DEFAULT_THP_SIZE ((uint32_t)(2 * MiB))
+#else
+        /* fallback to 1 MiB (e.g., the THP size on s390x) */
+#define VIRTIO_MEM_DEFAULT_THP_SIZE VIRTIO_MEM_MIN_BLOCK_SIZE
+#endif
+
+/*
+ * We want to have a reasonable default block size such that
+ * 1. We avoid splitting THPs when unplugging memory, which degrades
+ *    performance.
+ * 2. We avoid placing THPs for plugged blocks that also cover unplugged
+ *    blocks.
+ *
+ * The actual THP size might differ between Linux kernels, so we try to probe
+ * it. In the future (if we ever run into issues regarding 2.), we might want
+ * to disable THP in case we fail to properly probe the THP size, or if the
+ * block size is configured smaller than the THP size.
+ */
+static uint32_t thp_size;
+
+#define HPAGE_PMD_SIZE_PATH "/sys/kernel/mm/transparent_hugepage/hpage_pmd_size"
+static uint32_t virtio_mem_thp_size(void)
+{
+    gchar *content = NULL;
+    const char *endptr;
+    uint64_t tmp;
+
+    if (thp_size) {
+        return thp_size;
+    }
+
+    /*
+     * Try to probe the actual THP size, fallback to (sane but eventually
+     * incorrect) default sizes.
+     */
+    if (g_file_get_contents(HPAGE_PMD_SIZE_PATH, &content, NULL, NULL) &&
+        !qemu_strtou64(content, &endptr, 0, &tmp) &&
+        (!endptr || *endptr == '\n')) {
+        /*
+         * Sanity-check the value, if it's too big (e.g., aarch64 with 64k base
+         * pages) or weird, fallback to something smaller.
+         */
+        if (!tmp || !is_power_of_2(tmp) || tmp > 16 * MiB) {
+            warn_report("Read unsupported THP size: %" PRIx64, tmp);
+        } else {
+            thp_size = tmp;
+        }
+    }
+
+    if (!thp_size) {
+        thp_size = VIRTIO_MEM_DEFAULT_THP_SIZE;
+        warn_report("Could not detect THP size, falling back to %" PRIx64
+                    "  MiB.", thp_size / MiB);
+    }
+
+    g_free(content);
+    return thp_size;
+}
+
+static uint64_t virtio_mem_default_block_size(RAMBlock *rb)
+{
+    const uint64_t page_size = qemu_ram_pagesize(rb);
Quoted 8 lines

hw/virtio/virtio-mem.c line 109

Quoted 5 lines
hw/virtio/virtio-mem.c +5 -0
+
+    /* We can have hugetlbfs with a page size smaller than the THP size. */
+    if (page_size == qemu_real_host_page_size) {
+        return MAX(page_size, virtio_mem_thp_size());
+    }

This condition is special, can think of hugetlbfs smaller in size than THP size configured.

Yeah, there are weird architectures, most prominently arm64:

https://www.kernel.org/doc/html/latest/arm64/hugetlbpage.html

Assume you're on 64K base pages with a probed 512MB THP size (currently). You could have hugetlbfs with 2MB page size via "CONT PTE" bits.

Quoted 4 lines

hw/virtio/virtio-mem.c line 110

Quoted 1 line
hw/virtio/virtio-mem.c +1 -0
+    return MAX(page_size, VIRTIO_MEM_MIN_BLOCK_SIZE);

Do we still need this? Or we can have only one return for both the cases? Probably, I am missing something here.

We still need it. Assume somebody would have 64K hugetlbfs on arm64 (with 4k base pages), we want to make sure we use at least 1MB blocks.

Thanks!

Signature · 3 lines
Thanks,

David / dhildenb
Reviewed-by
#4 lore raw

Re: [PATCH v2 1/5] virtio-mem: Probe THP size to determine default block size

Quoted 127 lines
Quoted 6 lines
Quoted 4 lines

Let's allow a minimum block size of 1 MiB in all configurations. Select the default block size based on
- The page size of the memory backend.
- The THP size if the memory backend size corresponds to the real hsot

s/hsot/host

thanks!

Quoted 10 lines
Quoted 8 lines

page size.
- The global minimum of 1 MiB.
and warn if something smaller is configured by the user.

VIRTIO_MEM only supports Linux (depends on LINUX), so we can probe the THP size unconditionally.

For now we only support virtio-mem on x86-64 - there isn't a user-visiable

s/visiable/visible

thanks!

Quoted 96 lines
Quoted 96 lines

change (x86-64 only supports 2 MiB THP on the PMD level) - the default was, and will be 2 MiB.

If we ever have THP on the PUD level (e.g., 1 GiB THP on x86-64), we expect to have a trigger to explicitly opt-in for the new THP granularity.

Cc "Michael S. Tsirkin" <[email protected]>
Cc Wei Yang <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
Cc Igor Mammedov <[email protected]>
Cc Pankaj Gupta <[email protected]>
---
 hw/virtio/virtio-mem.c | 105 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 101 insertions(+), 4 deletions(-)
hw/virtio/virtio-mem.c +68 -3
index 8fbec77ccc..9b1461cf9d 100644
@@ -33,10 +33,83 @@
 #include "trace.h"
 
 /*
- * Use QEMU_VMALLOC_ALIGN, so no THP will have to be split when unplugging
- * memory (e.g., 2MB on x86_64).
+ * Let's not allow blocks smaller than 1 MiB, for example, to keep the tracking
+ * bitmap small.
  */
-#define VIRTIO_MEM_MIN_BLOCK_SIZE ((uint32_t)QEMU_VMALLOC_ALIGN)
+#define VIRTIO_MEM_MIN_BLOCK_SIZE ((uint32_t)(1 * MiB))
+
+#if defined(__x86_64__) || defined(__arm__) || defined(__aarch64__) || \
+    defined(__powerpc64__)
+#define VIRTIO_MEM_DEFAULT_THP_SIZE ((uint32_t)(2 * MiB))
+#else
+        /* fallback to 1 MiB (e.g., the THP size on s390x) */
+#define VIRTIO_MEM_DEFAULT_THP_SIZE VIRTIO_MEM_MIN_BLOCK_SIZE
+#endif
+
+/*
+ * We want to have a reasonable default block size such that
+ * 1. We avoid splitting THPs when unplugging memory, which degrades
+ *    performance.
+ * 2. We avoid placing THPs for plugged blocks that also cover unplugged
+ *    blocks.
+ *
+ * The actual THP size might differ between Linux kernels, so we try to probe
+ * it. In the future (if we ever run into issues regarding 2.), we might want
+ * to disable THP in case we fail to properly probe the THP size, or if the
+ * block size is configured smaller than the THP size.
+ */
+static uint32_t thp_size;
+
+#define HPAGE_PMD_SIZE_PATH "/sys/kernel/mm/transparent_hugepage/hpage_pmd_size"
+static uint32_t virtio_mem_thp_size(void)
+{
+    gchar *content = NULL;
+    const char *endptr;
+    uint64_t tmp;
+
+    if (thp_size) {
+        return thp_size;
+    }
+
+    /*
+     * Try to probe the actual THP size, fallback to (sane but eventually
+     * incorrect) default sizes.
+     */
+    if (g_file_get_contents(HPAGE_PMD_SIZE_PATH, &content, NULL, NULL) &&
+        !qemu_strtou64(content, &endptr, 0, &tmp) &&
+        (!endptr || *endptr == '\n')) {
+        /*
+         * Sanity-check the value, if it's too big (e.g., aarch64 with 64k base
+         * pages) or weird, fallback to something smaller.
+         */
+        if (!tmp || !is_power_of_2(tmp) || tmp > 16 * MiB) {
+            warn_report("Read unsupported THP size: %" PRIx64, tmp);
+        } else {
+            thp_size = tmp;
+        }
+    }
+
+    if (!thp_size) {
+        thp_size = VIRTIO_MEM_DEFAULT_THP_SIZE;
+        warn_report("Could not detect THP size, falling back to %" PRIx64
+                    "  MiB.", thp_size / MiB);
+    }
+
+    g_free(content);
+    return thp_size;
+}
+
+static uint64_t virtio_mem_default_block_size(RAMBlock *rb)
+{
+    const uint64_t page_size = qemu_ram_pagesize(rb);
Quoted 8 lines

hw/virtio/virtio-mem.c line 109

Quoted 5 lines
hw/virtio/virtio-mem.c +5 -0
+
+    /* We can have hugetlbfs with a page size smaller than the THP size. */
+    if (page_size == qemu_real_host_page_size) {
+        return MAX(page_size, virtio_mem_thp_size());
+    }

This condition is special, can think of hugetlbfs smaller in size than THP size configured.

Quoted 7 lines

Yeah, there are weird architectures, most prominently arm64:

https://www.kernel.org/doc/html/latest/arm64/hugetlbpage.html

Assume you're on 64K base pages with a probed 512MB THP size (currently). You could have hugetlbfs with 2MB page size via "CONT PTE" bits.

Ok. I understand now. Thanks for explaining!

Quoted 8 lines
Quoted 5 lines

hw/virtio/virtio-mem.c line 110

Quoted 2 lines
hw/virtio/virtio-mem.c +1 -0
+    return MAX(page_size, VIRTIO_MEM_MIN_BLOCK_SIZE);

Do we still need this? Or we can have only one return for both the cases? Probably, I am missing something here.

We still need it. Assume somebody would have 64K hugetlbfs on arm64 (with 4k base pages), we want to make sure we use at least 1MB blocks.

ok. got it.

Looks good to me.