| Message ID | 1369298842-6295-5-git-send-email-gaowanlong@cn.fujitsu.com |
|---|---|
| State | New |
| Headers | show |
On 05/23/2013 04:47 PM, 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. Any comments ? Thanks, Wanlong Gao > > Signed-off-by: Wanlong Gao <gaowanlong@cn.fujitsu.com> > --- > 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); > } > >
Ping............ > On 05/23/2013 04:47 PM, 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. > > Any comments ? > > Thanks, > Wanlong Gao > >> >> Signed-off-by: Wanlong Gao <gaowanlong@cn.fujitsu.com> >> --- >> 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); >> } >> >> > >
Any comments? > 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. > > Signed-off-by: Wanlong Gao <gaowanlong@cn.fujitsu.com> > --- > 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); > } > >
Any comments? > 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. > > Signed-off-by: Wanlong Gao <gaowanlong@cn.fujitsu.com> > --- > 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); > } > >
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 <gaowanlong@cn.fujitsu.com> > > --- > > 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); > > } > > > > > > >
On 05/31/2013 02:22 AM, Eduardo Habkost wrote: > 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) You proposal seems good, but as I said above, when 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(). Then the "numactl" directions through hugetlbfs files will not work, neither does the AutoNUMA. We should set the numa binding directions manually before PCI-passthrough assigned devices. Any external tools can't resolve this problem but pinning memory manually inside QEMU. So, IMO, we should both support manually pinning memory nodes inside QEMU and give interfaces to allow external tools to judge the memory bind policy. Thanks, Wanlong Gao > > >>> >>> Signed-off-by: Wanlong Gao <gaowanlong@cn.fujitsu.com> >>> --- >>> 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); >>> } >>> >>> >> >> >> >
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); }
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. Signed-off-by: Wanlong Gao <gaowanlong@cn.fujitsu.com> --- exec.c | 21 +++++++++++++++++++++ include/sysemu/sysemu.h | 1 + vl.c | 13 +++++++++++++ 3 files changed, 35 insertions(+)