Re: [Qemu-devel] [PATCH 5/5] memory: able to pin guest node memory to host node manually

[Eduardo Habkost]

On Thu, May 30, 2013 at 05:57:21PM +0800, Wanlong Gao wrote:
> > Use mbind to pin guest numa node memory to host nodes manually.
> > 
> > If we are not able to pin memory to host node, we may meet the
> > cross node memory access performance regression.
> > 
> > With this patch, we can add manual pinning host node like this:
> > -m 1024 -numa node,cpus=0,nodeid=0,mem=512,pin=0 -numa 
> > node,nodeid=1,cpus=1,mem=512,pin=1
> > 
> > And, if PCI-passthrough is used, direct-attached-device uses DMA transfer
> > between device and qemu process. All pages of the guest will be pinned by 
> > get_user_pages().
> > 
> > KVM_ASSIGN_PCI_DEVICE ioctl
> >   kvm_vm_ioctl_assign_device()
> >     =>kvm_assign_device()
> >       => kvm_iommu_map_memslots()
> >         => kvm_iommu_map_pages()
> >            => kvm_pin_pages()
> > 
> > So, with direct-attached-device, all guest page's page count will be +1 and
> > any page migration will not work. AutoNUMA won't too. And direction by 
> > libvirt is *ignored*.
> > 
> > Above all, we need manual pinning memory to host node to avoid
> > such cross nodes memmory access performance regression.

I believe a similar approach (letting QEMU do the pinning itself) was
already proposed and rejected. See:
http://article.gmane.org/gmane.comp.emulators.kvm.devel/58835
http://article.gmane.org/gmane.comp.emulators.kvm.devel/57684

An alternative approach was proposed at;
http://article.gmane.org/gmane.comp.emulators.qemu/123001
(exporting virtual address information directly)

and another one at:
http://article.gmane.org/gmane.comp.emulators.qemu/157741
(keeping the files inside -mem-path-dir so they could be pinned manually
by other programs)

The approach I was planning to implement was the one proposed at:
http://article.gmane.org/gmane.comp.emulators.kvm.devel/93476
(exporting memory backing information through QMP, instead of depending
on predictable filenames on -mem-path-dir)


> > 
> > Signed-off-by: Wanlong Gao <address@hidden>
> > ---
> >  exec.c                  | 21 +++++++++++++++++++++
> >  include/sysemu/sysemu.h |  1 +
> >  vl.c                    | 13 +++++++++++++
> >  3 files changed, 35 insertions(+)
> > 
> > diff --git a/exec.c b/exec.c
> > index aec65c5..fe929ef 100644
> > --- a/exec.c
> > +++ b/exec.c
> > @@ -36,6 +36,8 @@
> >  #include "qemu/config-file.h"
> >  #include "exec/memory.h"
> >  #include "sysemu/dma.h"
> > +#include "sysemu/sysemu.h"
> > +#include "qemu/bitops.h"
> >  #include "exec/address-spaces.h"
> >  #if defined(CONFIG_USER_ONLY)
> >  #include <qemu.h>
> > @@ -1081,6 +1083,25 @@ ram_addr_t qemu_ram_alloc_from_ptr(ram_addr_t size, 
> > void *host,
> >              memory_try_enable_merging(new_block->host, size);
> >          }
> >      }
> > +
> > +    if (nb_numa_nodes > 0 && !strcmp(mr->name, "pc.ram")) {
> > +        int i;
> > +        uint64_t nodes_mem = 0;
> > +        unsigned long *maskp = g_malloc0(sizeof(*maskp));
> > +        for (i = 0; i < nb_numa_nodes; i++) {
> > +            *maskp = 0;
> > +            if (node_pin[i] != -1) {
> > +                set_bit(node_pin[i], maskp);
> > +                if (qemu_mbind(new_block->host + nodes_mem, node_mem[i],
> > +                               QEMU_MPOL_BIND, maskp, MAX_NODES, 0)) {
> > +                    perror("qemu_mbind");
> > +                    exit(1);
> > +                }
> > +            }
> > +            nodes_mem += node_mem[i];
> > +        }
> > +    }
> > +
> >      new_block->length = size;
> >  
> >      /* Keep the list sorted from biggest to smallest block.  */
> > diff --git a/include/sysemu/sysemu.h b/include/sysemu/sysemu.h
> > index 2fb71af..ebf6580 100644
> > --- a/include/sysemu/sysemu.h
> > +++ b/include/sysemu/sysemu.h
> > @@ -131,6 +131,7 @@ extern QEMUClock *rtc_clock;
> >  #define MAX_CPUMASK_BITS 255
> >  extern int nb_numa_nodes;
> >  extern uint64_t node_mem[MAX_NODES];
> > +extern int node_pin[MAX_NODES];
> >  extern unsigned long *node_cpumask[MAX_NODES];
> >  
> >  #define MAX_OPTION_ROMS 16
> > diff --git a/vl.c b/vl.c
> > index 5555b1d..3768002 100644
> > --- a/vl.c
> > +++ b/vl.c
> > @@ -253,6 +253,7 @@ static QTAILQ_HEAD(, FWBootEntry) fw_boot_order =
> >  
> >  int nb_numa_nodes;
> >  uint64_t node_mem[MAX_NODES];
> > +int node_pin[MAX_NODES];
> >  unsigned long *node_cpumask[MAX_NODES];
> >  
> >  uint8_t qemu_uuid[16];
> > @@ -1390,6 +1391,17 @@ static void numa_add(const char *optarg)
> >              }
> >              node_mem[nodenr] = sval;
> >          }
> > +
> > +        if (get_param_value(option, 128, "pin", optarg) != 0) {
> > +            int unsigned long long pin_node;
> > +            if (parse_uint_full(option, &pin_node, 10) < 0) {
> > +                fprintf(stderr, "qemu: Invalid pinning nodeid: %s\n", 
> > optarg);
> > +                exit(1);
> > +            } else {
> > +                node_pin[nodenr] = pin_node;
> > +            }
> > +        }
> > +
> >          if (get_param_value(option, 128, "cpus", optarg) != 0) {
> >              numa_node_parse_cpus(nodenr, option);
> >          }
> > @@ -2921,6 +2933,7 @@ int main(int argc, char **argv, char **envp)
> >  
> >      for (i = 0; i < MAX_NODES; i++) {
> >          node_mem[i] = 0;
> > +        node_pin[i] = -1;
> >          node_cpumask[i] = bitmap_new(MAX_CPUMASK_BITS);
> >      }
> >  
> > 
> 
> 
> 

-- 
Eduardo