Skip to contentSkip to conversation

migration/rdma: Use huge page register VM memory

PATCH Superseded LIZHAOXIN1 [李照鑫] 3 participants 5 messages Jun 7, 2021 – Jun 10, 2021

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 migration/rdma: Use huge page register VM memory

When using libvirt for RDMA live migration, if the VM memory is too large, it will take a lot of time to deregister the VM at the source side, resulting in a long downtime (VM 64G, deregister vm time is about 400ms).

Although the VM's memory uses 2M huge pages, the MLNX driver still uses 4K pages for pin memory, as well as for unpin. So we use huge pages to skip the process of pin memory and unpin memory to reduce downtime.

The test environment:
kernel: linux-5.12
MLNX: ConnectX-4 LX
libvirt command:
virsh migrate --live --p2p --persistent --copy-storage-inc --listen-address \
0.0.0.0 --rdma-pin-all --migrateuri rdma://192.168.0.2 [VM] qemu+tcp://192.168.0.2/system

migration/rdma.c +20 -7
index 1cdb4561f3..9823449297 100644
@@ -1123,13 +1123,26 @@ static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma)
     RDMALocalBlocks *local = &rdma->local_ram_blocks;
 
     for (i = 0; i < local->nb_blocks; i++) {
-        local->block[i].mr =
-            ibv_reg_mr(rdma->pd,
-                    local->block[i].local_host_addr,
-                    local->block[i].length,
-                    IBV_ACCESS_LOCAL_WRITE |
-                    IBV_ACCESS_REMOTE_WRITE
-                    );
+        if (strcmp(local->block[i].block_name,"pc.ram") == 0) {
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE |
+                        IBV_ACCESS_ON_DEMAND |
+                        IBV_ACCESS_HUGETLB
+                        );
+        } else {
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE
+                        );
+        }
+
         if (!local->block[i].mr) {
             perror("Failed to register local dest ram block!\n");
             break;
#1 lore raw
Quoted 23 lines · On Mon, Jun 07, 2021 at 01:57:02PM +0000, LIZHAOXIN1 [李照鑫] wrote:

When using libvirt for RDMA live migration, if the VM memory is too large, it will take a lot of time to deregister the VM at the source side, resulting in a long downtime (VM 64G, deregister vm time is about 400ms).

Although the VM's memory uses 2M huge pages, the MLNX driver still uses 4K pages for pin memory, as well as for unpin. So we use huge pages to skip the process of pin memory and unpin memory to reduce downtime.

The test environment:
kernel: linux-5.12
MLNX: ConnectX-4 LX
libvirt command:
virsh migrate --live --p2p --persistent --copy-storage-inc --listen-address \
0.0.0.0 --rdma-pin-all --migrateuri rdma://192.168.0.2 [VM] qemu+tcp://192.168.0.2/system

migration/rdma.c +0 -0
index 1cdb4561f3..9823449297 100644
@@ -1123,13 +1123,26 @@ static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma)
     RDMALocalBlocks *local = &rdma->local_ram_blocks;

migration/rdma.c static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) · line 1126

Quoted 10 lines
migration/rdma.c +1 -7
 
     for (i = 0; i < local->nb_blocks; i++) {
-        local->block[i].mr =
-            ibv_reg_mr(rdma->pd,
-                    local->block[i].local_host_addr,
-                    local->block[i].length,
-                    IBV_ACCESS_LOCAL_WRITE |
-                    IBV_ACCESS_REMOTE_WRITE
-                    );
+        if (strcmp(local->block[i].block_name,"pc.ram") == 0) {

'pc.ram' is an x86 architecture specific name, so this will still leave a problem on other architectures I assume.

Quoted 12 lines
migration/rdma.c +12 -0
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE |
+                        IBV_ACCESS_ON_DEMAND |
+                        IBV_ACCESS_HUGETLB
+                        );
+        } else {
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,

migration/rdma.c static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) · line 1148

Quoted 10 lines
migration/rdma.c +7 -0
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE
+                        );
+        }
+
         if (!local->block[i].mr) {
             perror("Failed to register local dest ram block!\n");
             break;

Regards,
Daniel

Signature · 3 lines
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
#2 lore raw
Quoted 25 lines · * Daniel P. Berrangé ([email protected]) wrote:
Quoted 24 lines · On Mon, Jun 07, 2021 at 01:57:02PM +0000, LIZHAOXIN1 [李照鑫] wrote:

When using libvirt for RDMA live migration, if the VM memory is too large, it will take a lot of time to deregister the VM at the source side, resulting in a long downtime (VM 64G, deregister vm time is about 400ms).

Although the VM's memory uses 2M huge pages, the MLNX driver still uses 4K pages for pin memory, as well as for unpin. So we use huge pages to skip the process of pin memory and unpin memory to reduce downtime.

The test environment:
kernel: linux-5.12
MLNX: ConnectX-4 LX
libvirt command:
virsh migrate --live --p2p --persistent --copy-storage-inc --listen-address \
0.0.0.0 --rdma-pin-all --migrateuri rdma://192.168.0.2 [VM] qemu+tcp://192.168.0.2/system

migration/rdma.c +0 -0
index 1cdb4561f3..9823449297 100644
@@ -1123,13 +1123,26 @@ static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma)
     RDMALocalBlocks *local = &rdma->local_ram_blocks;
 
Quoted 12 lines

migration/rdma.c static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) · line 1126

Quoted 9 lines
migration/rdma.c +1 -7
     for (i = 0; i < local->nb_blocks; i++) {
-        local->block[i].mr =
-            ibv_reg_mr(rdma->pd,
-                    local->block[i].local_host_addr,
-                    local->block[i].length,
-                    IBV_ACCESS_LOCAL_WRITE |
-                    IBV_ACCESS_REMOTE_WRITE
-                    );
+        if (strcmp(local->block[i].block_name,"pc.ram") == 0) {

'pc.ram' is an x86 architecture specific name, so this will still leave a problem on other architectures I assume.

Yes, and also break even on PC when using NUMA.
I think the thing to do here is to call qemu_ram_pagesize on the RAMBlock;

if (qemu_ram_pagesize(RAMBlock....) != qemu_real_host_page_size)
it's a huge page

I guess it's probably best to do that in qemu_rdma_init_one_block or something?

I wonder how that all works when there's a mix of different huge page sizes?

Dave

Quoted 29 lines
Quoted 12 lines
migration/rdma.c +12 -0
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE |
+                        IBV_ACCESS_ON_DEMAND |
+                        IBV_ACCESS_HUGETLB
+                        );
+        } else {
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,

migration/rdma.c static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) · line 1148

Quoted 10 lines
migration/rdma.c +7 -0
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE
+                        );
+        }
+
         if (!local->block[i].mr) {
             perror("Failed to register local dest ram block!\n");
             break;

Regards,
Daniel

Signature · 3 lines
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
Signature · 1 line
Dr. David Alan Gilbert / [email protected] / Manchester, UK
#4 lore raw
In reply to Daniel P. Berrangé

Yes, 'pc.ram' is the specific name for x86. I have read that memory_region_allocate_system_memory assigns different names to other architectures.
Thanks for your reminding.

Regards,
lizhaoxin.

-----邮件原件-----
发件人: Daniel P. Berrangé <[email protected]>
发送时间: 2021年6月7日 22:18
收件人: LIZHAOXIN1 [李照鑫] <[email protected]>
抄送: [email protected]; [email protected]; [email protected]; sunhao2 [孙昊] <[email protected]>; DENGLINWEN [邓林文] <[email protected]>; YANGFENG1 [杨峰] <[email protected]> 主题: Re: [PATCH] migration/rdma: Use huge page register VM memory

Quoted 25 lines · On Mon, Jun 07, 2021 at 01:57:02PM +0000, LIZHAOXIN1 [李照鑫] wrote:

When using libvirt for RDMA live migration, if the VM memory is too large, it will take a lot of time to deregister the VM at the source side, resulting in a long downtime (VM 64G, deregister vm time is about 400ms).

Although the VM's memory uses 2M huge pages, the MLNX driver still uses 4K pages for pin memory, as well as for unpin. So we use huge pages to skip the process of pin memory and unpin memory to reduce downtime.

The test environment:
kernel: linux-5.12
MLNX: ConnectX-4 LX
libvirt command:
virsh migrate --live --p2p --persistent --copy-storage-inc --listen-address \
0.0.0.0 --rdma-pin-all --migrateuri rdma://192.168.0.2 [VM] qemu+tcp://192.168.0.2/system

migration/rdma.c index renamed from migration/rdma.c +0 -0

1cdb4561f3..9823449297 100644

migration/rdma.c +0 -0
@@ -1123,13 +1123,26 @@ static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma)
     RDMALocalBlocks *local = &rdma->local_ram_blocks;

migration/rdma.c static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) · line 1126

Quoted 10 lines
migration/rdma.c +1 -7
 
     for (i = 0; i < local->nb_blocks; i++) {
-        local->block[i].mr =
-            ibv_reg_mr(rdma->pd,
-                    local->block[i].local_host_addr,
-                    local->block[i].length,
-                    IBV_ACCESS_LOCAL_WRITE |
-                    IBV_ACCESS_REMOTE_WRITE
-                    );
+        if (strcmp(local->block[i].block_name,"pc.ram") == 0) {

'pc.ram' is an x86 architecture specific name, so this will still leave a problem on other architectures I assume.

Quoted 12 lines
migration/rdma.c +12 -0
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE |
+                        IBV_ACCESS_ON_DEMAND |
+                        IBV_ACCESS_HUGETLB
+                        );
+        } else {
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,

migration/rdma.c static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) · line 1148

Quoted 10 lines
migration/rdma.c +7 -0
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE
+                        );
+        }
+
         if (!local->block[i].mr) {
             perror("Failed to register local dest ram block!\n");
             break;

Regards,
Daniel

Signature · 3 lines
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
#3 lore raw
In reply to Dr. David Alan Gilbert

Yes, When I configured two NUMAs for the VM, the name of the memory is 'ram-node*', and other architectures had different names. As you suggested, I use qemu_ram_pagesize() and qemu_real_host_page_size to determine which Ramblocks use huge page.
I will send the patch second version later.

when there's a mix of different huge page sizes, there is no difference in their behavior, register or pin are just to prevent the memory from being swapped out. Huge page itself will not be swapped out,so huge page no need deregister or unpin.

The libvirt xml of my VM is
...
<memoryBacking>

    <hugepages>
      <page size='2048' unit='KiB' nodeset='0'/>
      <page size='1048576' unit='KiB' nodeset='1'/>
    </hugepages>
  </memoryBacking>
...
<numa>
      <cell id='0' cpus='0-7' memory='31457280' unit='KiB' memAccess='shared'/>
      <cell id='1' cpus='8-15' memory='2097152' unit='KiB' memAccess='shared'/>

</numa>
...

After testing, the RDMA live migration is normal, and the downtime is significantly reduced.

-----邮件原件-----
发件人: Dr. David Alan Gilbert <[email protected]>
发送时间: 2021年6月7日 23:00
收件人: Daniel P. Berrangé <[email protected]>
抄送: LIZHAOXIN1 [李照鑫] <[email protected]>; [email protected]; [email protected]; sunhao2 [孙昊] <[email protected]>; DENGLINWEN [邓林文] <[email protected]>; YANGFENG1 [杨峰] <[email protected]> 主题: Re: [PATCH] migration/rdma: Use huge page register VM memory

Quoted 27 lines · * Daniel P. Berrangé ([email protected]) wrote:
Quoted 26 lines · On Mon, Jun 07, 2021 at 01:57:02PM +0000, LIZHAOXIN1 [李照鑫] wrote:

When using libvirt for RDMA live migration, if the VM memory is too large, it will take a lot of time to deregister the VM at the source side, resulting in a long downtime (VM 64G, deregister vm time is about 400ms).

Although the VM's memory uses 2M huge pages, the MLNX driver still uses 4K pages for pin memory, as well as for unpin. So we use huge pages to skip the process of pin memory and unpin memory to reduce downtime.

The test environment:
kernel: linux-5.12
MLNX: ConnectX-4 LX
libvirt command:
virsh migrate --live --p2p --persistent --copy-storage-inc --listen-address \
0.0.0.0 --rdma-pin-all --migrateuri rdma://192.168.0.2 [VM] qemu+tcp://192.168.0.2/system

migration/rdma.c index renamed from migration/rdma.c +0 -0

1cdb4561f3..9823449297 100644

migration/rdma.c +0 -0
@@ -1123,13 +1123,26 @@ static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma)
     RDMALocalBlocks *local = &rdma->local_ram_blocks;
 
Quoted 12 lines

migration/rdma.c static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) · line 1126

Quoted 9 lines
migration/rdma.c +1 -7
     for (i = 0; i < local->nb_blocks; i++) {
-        local->block[i].mr =
-            ibv_reg_mr(rdma->pd,
-                    local->block[i].local_host_addr,
-                    local->block[i].length,
-                    IBV_ACCESS_LOCAL_WRITE |
-                    IBV_ACCESS_REMOTE_WRITE
-                    );
+        if (strcmp(local->block[i].block_name,"pc.ram") == 0) {

'pc.ram' is an x86 architecture specific name, so this will still leave a problem on other architectures I assume.

Yes, and also break even on PC when using NUMA.
I think the thing to do here is to call qemu_ram_pagesize on the RAMBlock;

if (qemu_ram_pagesize(RAMBlock....) != qemu_real_host_page_size)
it's a huge page

I guess it's probably best to do that in qemu_rdma_init_one_block or something?

I wonder how that all works when there's a mix of different huge page sizes?

Dave

Quoted 23 lines
Quoted 12 lines
migration/rdma.c +12 -0
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE |
+                        IBV_ACCESS_ON_DEMAND |
+                        IBV_ACCESS_HUGETLB
+                        );
+        } else {
+            local->block[i].mr =
+                ibv_reg_mr(rdma->pd,

migration/rdma.c static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) · line 1148

Quoted 10 lines
migration/rdma.c +7 -0
+                        local->block[i].local_host_addr,
+                        local->block[i].length,
+                        IBV_ACCESS_LOCAL_WRITE |
+                        IBV_ACCESS_REMOTE_WRITE
+                        );
+        }
+
         if (!local->block[i].mr) {
             perror("Failed to register local dest ram block!\n");
             break;
Quoted 6 lines

Regards,
Daniel
--

|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK