[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt] [RFC PATCH v2 1/4] numa: describe siblings distances within cells



On Wed, 28 Jun 2017 15:21:29 +0100
"Daniel P. Berrange" <berrange redhat com> wrote:

    Hi Daniel,

  Appreciate your input and comments. Let me work on package and return
  asap.

  One quick question; For future i am looking into way to somehow auto-
  tune node (cell/socket/cpu) assignment under <numa> directive. Such
  in a way brings intelligence into libvirt and therefor not wanted.

  Perhaps this is up to s/w stack utilizing guests per libvirt. If
  there are thoughts on such matter then i am very interested to hear
  about them.

> On Wed, Jun 28, 2017 at 03:56:36PM +0200, Wim Ten Have wrote:
> > From: Wim ten Have <wim ten have oracle com>
> > 
> > Add libvirtd NUMA cell domain administration functionality to
> > describe underlying cell id sibling distances in full fashion
> > when configuring HVM guests.
> > 
> > [below is an example of a 4 node setup]
> > 
> >   <cpu>
> >     <numa>
> >       <cell id='0' cpus='0' memory='2097152' unit='KiB'>
> >         <distances>
> >           <sibling id='0' value='10'/>
> >           <sibling id='1' value='21'/>
> >           <sibling id='2' value='31'/>
> >           <sibling id='3' value='41'/>
> >         </distances>
> >       </cell>
> >       <cell id='1' cpus='1' memory='2097152' unit='KiB'>
> >         <distances>
> >           <sibling id='0' value='21'/>
> >           <sibling id='1' value='10'/>
> >           <sibling id='2' value='31'/>
> >           <sibling id='3' value='41'/>
> >         </distances>
> >       </cell>
> >       <cell id='2' cpus='2' memory='2097152' unit='KiB'>
> >         <distances>
> >           <sibling id='0' value='31'/>
> >           <sibling id='1' value='21'/>
> >           <sibling id='2' value='10'/>
> >           <sibling id='3' value='21'/>
> >         </distances>
> >       <cell id='3' cpus='3' memory='2097152' unit='KiB'>
> >         <distances>
> >           <sibling id='0' value='41'/>
> >           <sibling id='1' value='31'/>
> >           <sibling id='2' value='21'/>
> >           <sibling id='3' value='10'/>
> >         </distances>
> >       </cell>
> >     </numa>
> >   </cpu>
> > 
> > Changes under this commit concern all those that require adding
> > the valid data structures, virDomainNuma* functional routines and the
> > XML doc schema enhancements to enforce appropriate administration.
> > 
> > These changes are accompanied with topic related documentation under
> > docs/formatdomain.html within the "CPU model and topology" extending the
> > "Guest NUMA topology" paragraph.
> > 
> > For terminology we refer to sockets as "nodes" where access to each
> > others' distinct resources such as memory make them "siblings" with a
> > designated "distance" between them.  A specific design is described under
> > the ACPI (Advanced Configuration and Power Interface Specification)
> > within the chapter explaining the system's SLIT (System Locality Distance
> > Information Table).
> > 
> > These patches extend core libvirt's XML description of a virtual machine's
> > hardware to include NUMA distance information for sibling nodes, which
> > is then passed to Xen guests via libxl. Recently qemu landed support for
> > constructing the SLIT since commit 0f203430dd ("numa: Allow setting NUMA
> > distance for different NUMA nodes"), hence these core libvirt extensions
> > can also help other drivers in supporting this feature.
> > 
> > These changes alter the docs/schemas/cputypes.rng enforcing
> > domain administration to follow the syntax below per numa cell id.
> > 
> > These changes also alter docs/schemas/basictypes.rng to add
> > "numaDistanceValue" which is an "unsignedInt" with a minimum value
> > of 10 as 0-9 are reserved values and can not be used as System
> > Locality Distance Information Table data.
> > 
> > Signed-off-by: Wim ten Have <wim ten have oracle com>
> > ---
> >  docs/formatdomain.html.in   |  64 ++++++++++-
> >  docs/schemas/basictypes.rng |   8 ++
> >  docs/schemas/cputypes.rng   |  18 +++
> >  src/conf/cpu_conf.c         |   2 +-
> >  src/conf/numa_conf.c        | 260 +++++++++++++++++++++++++++++++++++++++++++-
> >  src/conf/numa_conf.h        |  25 ++++-
> >  src/libvirt_private.syms    |   6 +
> >  7 files changed, 376 insertions(+), 7 deletions(-)
> > 
> > diff --git a/docs/formatdomain.html.in b/docs/formatdomain.html.in
> > index 36bea67..b9736e3 100644
> > --- a/docs/formatdomain.html.in
> > +++ b/docs/formatdomain.html.in
> > @@ -1510,7 +1510,69 @@
> >      </p>
> >  
> >      <p>
> > -      This guest NUMA specification is currently available only for QEMU/KVM.
> > +      This guest NUMA specification is currently available only for
> > +      QEMU/KVM and Xen.  Whereas Xen driver also allows for a distinct
> > +      description of NUMA arranged <code>sibling</code> <code>cell</code>
> > +      <code>distances</code> <span class="since">Since 3.6.0</span>.
> > +    </p>
> > +
> > +    <p>
> > +      Under NUMA h/w architecture, distinct resources such as memory
> > +      create a designated distance between <code>cell</code> and
> > +      <code>siblings</code> that now can be described with the help of
> > +      <code>distances</code>. A detailed describtion can be found within
> > +      the ACPI (Advanced Configuration and Power Interface Specification)
> > +      within the chapter explaining the system's SLIT (System Locality
> > +      Distance Information Table).
> > +    </p>
> > +
> > +<pre>
> > +...
> > +&lt;cpu&gt;
> > +  ...
> > +  &lt;numa&gt;
> > +    &lt;cell id='0' cpus='0,4-7' memory='512000' unit='KiB'&gt;
> > +      &lt;distances&gt;
> > +        &lt;sibling id='0' value='10'/&gt;
> > +        &lt;sibling id='1' value='21'/&gt;
> > +        &lt;sibling id='2' value='31'/&gt;
> > +        &lt;sibling id='3' value='41'/&gt;
> > +      &lt;/distances&gt;
> > +    &lt;/cell&gt;
> > +    &lt;cell id='1' cpus='1,8-10,12-15' memory='512000' unit='KiB' memAccess='shared'&gt;
> > +      &lt;distances&gt;
> > +        &lt;sibling id='0' value='21'/&gt;
> > +        &lt;sibling id='1' value='10'/&gt;
> > +        &lt;sibling id='2' value='21'/&gt;
> > +        &lt;sibling id='3' value='31'/&gt;
> > +      &lt;/distances&gt;
> > +    &lt;/cell&gt;
> > +    &lt;cell id='2' cpus='2,11' memory='512000' unit='KiB' memAccess='shared'&gt;
> > +      &lt;distances&gt;
> > +        &lt;sibling id='0' value='31'/&gt;
> > +        &lt;sibling id='1' value='21'/&gt;
> > +        &lt;sibling id='2' value='10'/&gt;
> > +        &lt;sibling id='3' value='21'/&gt;
> > +      &lt;/distances&gt;
> > +    &lt;/cell&gt;
> > +    &lt;cell id='3' cpus='3' memory='512000' unit='KiB'&gt;
> > +      &lt;distances&gt;
> > +        &lt;sibling id='0' value='41'/&gt;
> > +        &lt;sibling id='1' value='31'/&gt;
> > +        &lt;sibling id='2' value='21'/&gt;
> > +        &lt;sibling id='3' value='10'/&gt;
> > +      &lt;/distances&gt;
> > +    &lt;/cell&gt;
> > +  &lt;/numa&gt;
> > +  ...
> > +&lt;/cpu&gt;
> > +...</pre>
> > +
> > +    <p>
> > +      Under Xen driver, if no <code>distances</code> are given to describe
> > +      the SLIT data between different cells, it will default to a scheme
> > +      using 10 for local and 21 for other cells.  Which is the assumption
> > +      of guest OS when no SLIT is specified.
> >      </p>  
> 
> We need to specify what happens if partial data is given too. In particular
> if you give a distance for  A -> B, and not B -> A, then we should assume
> symmetry.

  Indeed.

> Rather than having the defaults be duplicated in drivers, we should have
> a helper API in domain_conf.c that can be used to query the distance
> between 2 nodes that can be called by drivers. This can then apply
> defaults in a consistent manner.

  This comment in a way helps partially towards my questions for auto-tune as
  a prerequisite.  On my question, do you think it is wise to explore or do you
  think libvirt is not the place to have this done?

> > @@ -66,6 +68,12 @@ struct _virDomainNuma {
> >          virBitmapPtr nodeset;   /* host memory nodes where this guest node resides */
> >          virDomainNumatuneMemMode mode;  /* memory mode selection */
> >          virDomainMemoryAccess memAccess; /* shared memory access configuration */
> > +
> > +        struct _virDomainNumaDistance {
> > +          unsigned int value;    /* locality value for node i*j */  
> 
> Strictly speaking distances have an orderng, ie this is 'node i -> j'
> in that specific direction. 'node j -> i' can be different.
> 
> > +          unsigned int cellid;  
> 
> Needs extra indent - 4 spaces per indent level
> 
> > +        } *distances;           /* remote node distances */
> > +        size_t ndistances;
> >      } *mem_nodes;           /* guest node configuration */
> >      size_t nmem_nodes;
> >  
> > @@ -686,6 +694,95 @@ virDomainNumatuneNodesetIsAvailable(virDomainNumaPtr numatune,
> >  }  
> 
> > +    for (i = 0; i < ndistances; i++) {
> > +        unsigned int sibling_id = i, sibling_value;
> > +
> > +        /* siblings are in order of parsing or explicitly numbered */
> > +        if ((tmp = virXMLPropString(nodes[i], "id"))) {
> > +            if (virStrToLong_uip(tmp, NULL, 10, &sibling_id) < 0) {
> > +                virReportError(VIR_ERR_XML_ERROR,
> > +                               _("Invalid 'id' attribute in NUMA distances for sibling: '%s'"),
> > +                               tmp);
> > +                goto cleanup;
> > +            }
> > +
> > +            if (sibling_id >= ndistances) {
> > +                virReportError(VIR_ERR_XML_ERROR, "%s",
> > +                               _("Exactly one 'sibling' element per NUMA distance "
> > +                                 "is allowed, non-contiguous ranges or ranges not "
> > +                                 "starting from 0 are not allowed"));  
> 
> We should allow non-contiguous ranges, starting from any node id. ie users should
> be free to provide sparse information, relying on defaults for the rest.
> 
> > +                goto cleanup;
> > +            }
> > +        }
> > +        VIR_FREE(tmp);
> > +
> > +        /* We need a locality value */
> > +        if (!(tmp = virXMLPropString(nodes[i], "value"))) {
> > +            virReportError(VIR_ERR_XML_ERROR,
> > +                           _("Missing 'value' attribute in NUMA distances for sibling id: '%d'"),
> > +                           sibling_id);
> > +            goto cleanup;
> > +        }
> > +
> > +        /* It needs to be applicable */
> > +        if (virStrToLong_uip(tmp, NULL, 10, &sibling_value) < 0) {
> > +            virReportError(VIR_ERR_XML_ERROR,
> > +                           _("Invalid 'value' attribute in NUMA distances for sibling id: '%d'"),
> > +                           sibling_id);
> > +            goto cleanup;
> > +        }
> > +        VIR_FREE(tmp);
> > +
> > +        distances[sibling_id].cellid = sibling_id;
> > +        distances[sibling_id].value = sibling_value;
> > +    }
> > +
> > +    def->mem_nodes[cur_cell].distances = distances;
> > +    def->mem_nodes[cur_cell].ndistances = ndistances;
> > +
> > +    ret = 0;
> > +
> > + cleanup:
> > +    if (ret)
> > +        VIR_FREE(distances);
> > +    VIR_FREE(nodes);
> > +    VIR_FREE(tmp);
> > +
> > +    return ret;
> > +}
> > +
> >  int
> >  virDomainNumaDefCPUParseXML(virDomainNumaPtr def,
> >                              xmlXPathContextPtr ctxt)
> > @@ -788,6 +885,14 @@ virDomainNumaDefCPUParseXML(virDomainNumaPtr def,
> >              def->mem_nodes[cur_cell].memAccess = rc;
> >              VIR_FREE(tmp);
> >          }
> > +
> > +        /* Parse NUMA distances info */
> > +        if (virDomainNumaDefNodeDistanceParseXML(def, ctxt, cur_cell) < 0) {
> > +                virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
> > +                          _("NUMA cell %d has incorrect 'distances' configured"),
> > +                          cur_cell);
> > +                goto cleanup;
> > +        }  
> 
> This method already reported an detailed error message, which is being
> overwritten here with a useless generic message.
> 
> 
> > +size_t
> > +virDomainNumaSetNodeCount(virDomainNumaPtr numa, size_t nmem_nodes)
> > +{
> > +    if (!numa || !nmem_nodes)
> > +        return 0;  
> 
> This error condition doesn't report any error message...
> 
> > +
> > +    if (numa->mem_nodes) {
> > +        virReportError(VIR_ERR_INTERNAL_ERROR, "%s",
> > +                       _("Cannot alter an existing nmem_nodes set"));
> > +        return 0;
> > +    }  
> 
> ...but this does...
> 
> > +
> > +    if (VIR_ALLOC_N(numa->mem_nodes, nmem_nodes) < 0)
> > +        return 0;  
> 
> ...and so does this.

  I maybe missing your thought on specific one.  There's other use
  on memory allocation under this layer but if there is no memory
  left how could it possibly report other then have specific function
  return error?

> Callers need to have consistent error reporting for all scenarios

  Let me carefully go over whole again.

Regards,
- Wim.

> > +
> > +    numa->nmem_nodes = nmem_nodes;
> > +
> > +    return numa->nmem_nodes;
> > +}  
> 
> 
> 
> > +
> > +
> > +size_t
> > +virDomainNumaGetNodeDistance(virDomainNumaPtr numa,
> > +                             size_t node,
> > +                             size_t cellid)
> > +{
> > +    virDomainNumaDistancePtr distances;
> > +
> > +    if (!numa)
> > +        return 0;  
> 
> Just mark 'numa' parameters as ATTRIBUTE_NONNULL and make
> sure callers don't run this method if numa is not
> configured.
> 
> > +
> > +    distances = numa->mem_nodes[node].distances;  
> 
> You've not bounds-checked 'node' vs mem_nodes size.
> 
> > +    if (!numa->mem_nodes[node].ndistances || !distances)
> > +        return 0;  
> 
> This is where we should plug in defaults for partially
> specified data.
> 
> > +
> > +    return distances[cellid].value;
> > +}  
> 
> What are calling semantics here ?  I think we should
> 
> > +
> > +
> > +size_t
> > +virDomainNumaSetNodeDistance(virDomainNumaPtr numa,
> > +                             size_t node,
> > +                             size_t cellid,
> > +                             unsigned int value)
> > +{
> > +    virDomainNumaDistancePtr distances;
> > +
> > +    /*
> > +     * Advanced Configuration and Power Interface
> > +     * Specification version 6.1. Chapter 5.2.17
> > +     * System Locality Distance Information Table
> > +     * ... Distance values of 0-9 are reserved.
> > +     */
> > +    if (!numa || value < 10)
> > +        return 0;  
> 
> ATTRIBUTE_NONNULL instead, and report an error for
> values < 10
> 
> 
> > +
> > +    distances = numa->mem_nodes[node].distances;
> > +
> > +    if (numa->mem_nodes[node].ndistances > 0 &&
> > +        distances[cellid].value)
> > +        return 0;  
> 
> I'm not sure what this check is doing - it seems to be
> refusing to let you change the value if one is already
> set. I don't see the point of that, but if it is needed
> then we should report errors
> 
> > +
> > +    distances[cellid].cellid = cellid;
> > +    distances[cellid].value = value;
> > +
> > +    return distances[cellid].value;
> > +}
> > +
> > +
> > +size_t
> > +virDomainNumaGetNodeDistanceCount(virDomainNumaPtr numa,
> > +                                  size_t node)
> > +{
> > +    if (!numa)
> > +        return 0;  
> 
> ATTRIBUTE_NONNULL instead
> 
> > +
> > +    return numa->mem_nodes[node].ndistances;
> > +}
> > +
> > +
> > +size_t
> > +virDomainNumaSetNodeDistanceCount(virDomainNumaPtr numa,
> > +                                  size_t node,
> > +                                  size_t ndistances)
> > +{
> > +    virDomainNumaDistancePtr distances;
> > +
> > +    if (!numa || !ndistances)
> > +        return 0;  
> 
> ATTRIBUTE_NONNULL for first.
> 
> For the second we should presumably free any existing
> distance data
> 
> > +
> > +    distances = numa->mem_nodes[node].distances;
> > +    if (distances) {
> > +        virReportError(VIR_ERR_INTERNAL_ERROR,
> > +                       _("Cannot alter an existing nmem_nodes distances set for node: %zu"),
> > +                       node);
> > +        return 0;
> > +    }
> > +
> > +    if (VIR_ALLOC_N(distances, ndistances) < 0)
> > +        return 0;
> > +
> > +    numa->mem_nodes[node].distances = distances;
> > +    numa->mem_nodes[node].ndistances = ndistances;
> > +
> > +    return numa->mem_nodes[node].ndistances;
> > +}
> > +
> > +
> >  virBitmapPtr
> >  virDomainNumaGetNodeCpumask(virDomainNumaPtr numa,
> >                              size_t node)
> > @@ -937,6 +1175,20 @@ virDomainNumaGetNodeCpumask(virDomainNumaPtr numa,
> >  }
> >  
> >  
> > +virBitmapPtr
> > +virDomainNumaSetNodeCpumask(virDomainNumaPtr numa,
> > +                            size_t node,
> > +                            virBitmapPtr cpumask)
> > +{
> > +    if (!numa || !cpumask)
> > +        return NULL;
> > +
> > +    numa->mem_nodes[node].cpumask = cpumask;
> > +
> > +    return numa->mem_nodes[node].cpumask;
> > +}
> > +  
> Regards,
> Daniel


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]