Currently, we configure QEMU to prealloc memory almost by default. Well, by default for NVDIMMs, hugepages and if user asked us to (via memoryBacking <allocation mode="immediate"/>). However, there are two cases where this approach is not the best: 1) in case when guest's NVDIMM is backed by real life NVDIMM. In this case users should put <pmem/> into the <memory/> device <source/>, like this: <memory model='nvdimm' access='shared'> <source> <path>/dev/pmem0</path> <pmem/> </source> </memory> Instructing QEMU to do prealloc in this case means that each page of the NVDIMM is "touched" (the first byte is read and written back - see QEMU commit v2.9.0-rc1~26^2) which cripples device wear. 2) if free-page-reporting is turned on. While the free-page-reporting feature might not have a catchy or obvious name, when enabled it instructs KVM and subsequently QEMU to free pages no longer used by guest resulting in smaller memory footprint. And preallocating whole memory goes against this. The BZ comment 11 mentions another, third case 'virtio-mem' but that is not implemented in libvirt, yet. Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1894053 Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com&gt; --- src/qemu/qemu_command.c | 11 +++++++++-- .../memory-hotplug-nvdimm-pmem.x86_64-latest.args | 2 +- 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/src/qemu/qemu_command.c b/src/qemu/qemu_command.c index 479bcc0b0c..3df8b5ac76 100644 --- a/src/qemu/qemu_command.c +++ b/src/qemu/qemu_command.c @@ -2977,7 +2977,11 @@ qemuBuildMemoryBackendProps(virJSONValuePtr *backendProps, if (discard == VIR_TRISTATE_BOOL_ABSENT) discard = def->mem.discard; - if (def->mem.allocation == VIR_DOMAIN_MEMORY_ALLOCATION_IMMEDIATE) + /* The whole point of free_page_reporting is that as soon as guest frees + * any memory it is freed in the host too. Prealloc doesn't make much sense + * then. */ + if (def->mem.allocation == VIR_DOMAIN_MEMORY_ALLOCATION_IMMEDIATE && + def->memballoon->free_page_reporting != VIR_TRISTATE_SWITCH_ON) prealloc = true;

if (virDomainNumatuneGetMode(def->numa, mem->targetNode, &mode) < 0 && @@ -3064,7 +3068,10 @@ qemuBuildMemoryBackendProps(virJSONValuePtr *backendProps, if (mem->nvdimmPath) { memPath = g_strdup(mem->nvdimmPath); - prealloc = true;

+ /* If the NVDIMM is a real device then there's nothing to prealloc. + * If anyhing, we would be only wearing off the device. */ + if (!mem->nvdimmPmem) + prealloc = true;

} else if (useHugepage) { if (qemuGetDomainHupageMemPath(priv->driver, def, pagesize, &memPath) < 0) return -1; diff --git a/tests/qemuxml2argvdata/memory-hotplug-nvdimm-pmem.x86_64-latest.args b/tests/qemuxml2argvdata/memory-hotplug-nvdimm-pmem.x86_64-latest.args index cac02a6f6d..fb4ae4b518 100644 --- a/tests/qemuxml2argvdata/memory-hotplug-nvdimm-pmem.x86_64-latest.args +++ b/tests/qemuxml2argvdata/memory-hotplug-nvdimm-pmem.x86_64-latest.args @@ -20,7 +20,7 @@ file=/tmp/lib/domain--1-QEMUGuest1/master-key.aes \ -object memory-backend-ram,id=ram-node0,size=224395264 \ -numa node,nodeid=0,cpus=0-1,memdev=ram-node0 \ -object memory-backend-file,id=memnvdimm0,mem-path=/tmp/nvdimm,share=no,\ -prealloc=yes,size=536870912,pmem=yes \ +size=536870912,pmem=yes \ -device nvdimm,node=0,memdev=memnvdimm0,id=nvdimm0,slot=0 \ -uuid c7a5fdbd-edaf-9455-926a-d65c16db1809 \ -display none \ -- 2.26.2

On 11/30/20 11:16 AM, Daniel P. Berrangé wrote: > On Mon, Nov 30, 2020 at 11:06:14AM +0100, Michal Privoznik wrote: > > Currently, we configure QEMU to prealloc memory almost by > > default. Well, by default for NVDIMMs, hugepages and if user > > asked us to (via memoryBacking <allocation mode="immediate"/>). > > > > However, there are two cases where this approach is not the best: > > > > 1) in case when guest's NVDIMM is backed by real life NVDIMM. In > > this case users should put <pmem/> into the <memory/> device > > <source/>, like this: > > > > <memory model='nvdimm' access='shared'> > > <source> > > <path>/dev/pmem0</path> > > <pmem/> > > </source> > > </memory> > > > > Instructing QEMU to do prealloc in this case means that each > > page of the NVDIMM is "touched" (the first byte is read and > > written back - see QEMU commit v2.9.0-rc1~26^2) which cripples > > device wear. > > > > 2) if free-page-reporting is turned on. While the > > free-page-reporting feature might not have a catchy or obvious > > name, when enabled it instructs KVM and subsequently QEMU to > > free pages no longer used by guest resulting in smaller memory > > footprint. And preallocating whole memory goes against this. > > > > The BZ comment 11 mentions another, third case 'virtio-mem' but > > that is not implemented in libvirt, yet. > > > > Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1894053 > > Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com&gt; > > --- > > src/qemu/qemu_command.c | 11 +++++++++-- > > .../memory-hotplug-nvdimm-pmem.x86_64-latest.args | 2 +- > > 2 files changed, 10 insertions(+), 3 deletions(-) > > > > diff --git a/src/qemu/qemu_command.c b/src/qemu/qemu_command.c > > index 479bcc0b0c..3df8b5ac76 100644 > > --- a/src/qemu/qemu_command.c > > +++ b/src/qemu/qemu_command.c > > @@ -2977,7 +2977,11 @@ qemuBuildMemoryBackendProps(virJSONValuePtr *backendProps, > > if (discard == VIR_TRISTATE_BOOL_ABSENT) > > discard = def->mem.discard; > > - if (def->mem.allocation == VIR_DOMAIN_MEMORY_ALLOCATION_IMMEDIATE) > > + /* The whole point of free_page_reporting is that as soon as guest frees > > + * any memory it is freed in the host too. Prealloc doesn't make much sense > > + * then. */ > > + if (def->mem.allocation == VIR_DOMAIN_MEMORY_ALLOCATION_IMMEDIATE && > > + def->memballoon->free_page_reporting != VIR_TRISTATE_SWITCH_ON) > > prealloc = true; > > If the user asked for allocation == immediate, we should not be > silently ignoring that request. Isn't the scenario described simply > a wierd user configuration scenario and if they don't want that, then > then they can set <allocation mode="ondemand"/> instead. Okay. > > > if (virDomainNumatuneGetMode(def->numa, mem->targetNode, &mode) < 0 && > > @@ -3064,7 +3068,10 @@ qemuBuildMemoryBackendProps(virJSONValuePtr *backendProps, > > if (mem->nvdimmPath) { > > memPath = g_strdup(mem->nvdimmPath); > > - prealloc = true; > > > > > + /* If the NVDIMM is a real device then there's nothing to prealloc. > > + * If anyhing, we would be only wearing off the device. */ > > + if (!mem->nvdimmPmem) > > + prealloc = true; > > I wonder if QEMU itself should take this optimization to skip its > allocation logic ? Also would make sense. This is that kind of bug which lies in between libvirt and qemu. Although, since we are worried in silently ignoring user requests, then wouldn't this be exactly what QEMU would be doing? I mean, if an user/libvirt put both .prealloc=yes and .pmem=yes onto cmd line then these would cancel out, wouldn't they?

On Mon, Nov 30, 2020 at 11:48:28AM +0100, Michal Privoznik wrote: > On 11/30/20 11:16 AM, Daniel P. Berrangé wrote: > > On Mon, Nov 30, 2020 at 11:06:14AM +0100, Michal Privoznik wrote: > > > Currently, we configure QEMU to prealloc memory almost by > > > default. Well, by default for NVDIMMs, hugepages and if user > > > asked us to (via memoryBacking <allocation mode="immediate"/>). > > > > > > However, there are two cases where this approach is not the best: > > > > > > 1) in case when guest's NVDIMM is backed by real life NVDIMM. In > > > this case users should put <pmem/> into the <memory/> device > > > <source/>, like this: > > > > > > <memory model='nvdimm' access='shared'> > > > <source> > > > <path>/dev/pmem0</path> > > > <pmem/> > > > </source> > > > </memory> > > > > > > Instructing QEMU to do prealloc in this case means that each > > > page of the NVDIMM is "touched" (the first byte is read and > > > written back - see QEMU commit v2.9.0-rc1~26^2) which cripples > > > device wear. > > > > > > 2) if free-page-reporting is turned on. While the > > > free-page-reporting feature might not have a catchy or obvious > > > name, when enabled it instructs KVM and subsequently QEMU to > > > free pages no longer used by guest resulting in smaller memory > > > footprint. And preallocating whole memory goes against this. > > > > > > The BZ comment 11 mentions another, third case 'virtio-mem' but > > > that is not implemented in libvirt, yet. > > > > > > Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1894053 > > > Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com&gt; > > > --- > > > src/qemu/qemu_command.c | 11 +++++++++-- > > > .../memory-hotplug-nvdimm-pmem.x86_64-latest.args | 2 +- > > > 2 files changed, 10 insertions(+), 3 deletions(-) > > > > > > diff --git a/src/qemu/qemu_command.c b/src/qemu/qemu_command.c > > > index 479bcc0b0c..3df8b5ac76 100644 > > > --- a/src/qemu/qemu_command.c > > > +++ b/src/qemu/qemu_command.c > > > @@ -2977,7 +2977,11 @@ qemuBuildMemoryBackendProps(virJSONValuePtr *backendProps, > > > if (discard == VIR_TRISTATE_BOOL_ABSENT) > > > discard = def->mem.discard; > > > - if (def->mem.allocation == VIR_DOMAIN_MEMORY_ALLOCATION_IMMEDIATE) > > > + /* The whole point of free_page_reporting is that as soon as guest frees > > > + * any memory it is freed in the host too. Prealloc doesn't make much sense > > > + * then. */ > > > + if (def->mem.allocation == VIR_DOMAIN_MEMORY_ALLOCATION_IMMEDIATE && > > > + def->memballoon->free_page_reporting != VIR_TRISTATE_SWITCH_ON) > > > prealloc = true; > > > > If the user asked for allocation == immediate, we should not be > > silently ignoring that request. Isn't the scenario described simply > > a wierd user configuration scenario and if they don't want that, then > > then they can set <allocation mode="ondemand"/> instead. > > Okay. > > > > > > if (virDomainNumatuneGetMode(def->numa, mem->targetNode, &mode) < 0 && > > > @@ -3064,7 +3068,10 @@ qemuBuildMemoryBackendProps(virJSONValuePtr *backendProps, > > > if (mem->nvdimmPath) { > > > memPath = g_strdup(mem->nvdimmPath); > > > - prealloc = true; > > > > > > > > > + /* If the NVDIMM is a real device then there's nothing to prealloc. > > > + * If anyhing, we would be only wearing off the device. */ > > > + if (!mem->nvdimmPmem) > > > + prealloc = true; > > > > I wonder if QEMU itself should take this optimization to skip its > > allocation logic ?

> Also would make sense. This is that kind of bug which lies in between > libvirt and qemu. Although, since we are worried in silently ignoring user > requests, then wouldn't this be exactly what QEMU would be doing? I mean, if > an user/libvirt put both .prealloc=yes and .pmem=yes onto cmd line then > these would cancel out, wouldn't they? The difference is that an real NVDIMM is inherantly preallocated. QEMU

would not be ignoring the prealloc=yes arg - its implementation would merely be a no-op.

Regards, Daniel

On Mon, Nov 30, 2020 at 11:06:14 +0100, Michal Privoznik wrote: > Currently, we configure QEMU to prealloc memory almost by > default. Well, by default for NVDIMMs, hugepages and if user > asked us to (via memoryBacking <allocation mode="immediate"/>). > > However, there are two cases where this approach is not the best: > > 1) in case when guest's NVDIMM is backed by real life NVDIMM. In > this case users should put <pmem/> into the <memory/> device > <source/>, like this: > > <memory model='nvdimm' access='shared'> > <source> > <path>/dev/pmem0</path> > <pmem/> > </source> > </memory> > > Instructing QEMU to do prealloc in this case means that each > page of the NVDIMM is "touched" (the first byte is read and > written back - see QEMU commit v2.9.0-rc1~26^2) which cripples > device wear. > > 2) if free-page-reporting is turned on. While the > free-page-reporting feature might not have a catchy or obvious > name, when enabled it instructs KVM and subsequently QEMU to > free pages no longer used by guest resulting in smaller memory > footprint. And preallocating whole memory goes against this. > > The BZ comment 11 mentions another, third case 'virtio-mem' but > that is not implemented in libvirt, yet. > > Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1894053 > Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com&gt; > --- > src/qemu/qemu_command.c | 11 +++++++++-- > .../memory-hotplug-nvdimm-pmem.x86_64-latest.args | 2 +- > 2 files changed, 10 insertions(+), 3 deletions(-) > > diff --git a/src/qemu/qemu_command.c b/src/qemu/qemu_command.c > index 479bcc0b0c..3df8b5ac76 100644 > --- a/src/qemu/qemu_command.c > +++ b/src/qemu/qemu_command.c > @@ -2977,7 +2977,11 @@ qemuBuildMemoryBackendProps(virJSONValuePtr *backendProps, > if (discard == VIR_TRISTATE_BOOL_ABSENT) > discard = def->mem.discard; > > - if (def->mem.allocation == VIR_DOMAIN_MEMORY_ALLOCATION_IMMEDIATE) > + /* The whole point of free_page_reporting is that as soon as guest frees > + * any memory it is freed in the host too. Prealloc doesn't make much sense > + * then. */ > + if (def->mem.allocation == VIR_DOMAIN_MEMORY_ALLOCATION_IMMEDIATE && > + def->memballoon->free_page_reporting != VIR_TRISTATE_SWITCH_ON) > prealloc = true; IIUC def->mem.allocation is an explicit user-specified configuration, in such case we should not override the user wish based on a different configuration. If they don't make sense together technically, we should reject the config rather than silently changing it. If it's just semantically wrong and pre-existing we may leave it be. Additionally the patch is missing a test case for this scenario.