ovs: do not allocate memory from offline numa node

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO, 0) It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

---

<4>[   24.368805] ------------[ cut here ]------------
<2>[   24.368846] kernel BUG at include/linux/gfp.h:325!
<4>[   24.368868] invalid opcode: 0000 [#1] SMP
<4>[   24.368892] Modules linked in: openvswitch vxlan udp_tunnel ip6_udp_tunnel gre libcrc32c kvm_amd kvm crc32_pclmul ghash_clmulni_intel aesni_intel ablk_helper cryptd lrw mgag200 ttm drm_kms_helper drm gf128mul glue_helper serio_raw aes_x86_64 sysimgblt sysfillrect syscopyarea sp5100_tco amd64_edac_mod edac_core edac_mce_amd i2c_piix4 k10temp fam15h_power microcode raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq raid1 raid0 igb multipath i2c_algo_bit i2c_core linear dca psmouse ptp ahci pata_atiixp pps_core libahci
<4>[   24.369225] CPU: 22 PID: 987 Comm: ovs-vswitchd Not tainted 3.18.19-24 #1
<4>[   24.369255] Hardware name: Supermicro H8DGU/H8DGU, BIOS 3.0b       05/07/2013
<4>[   24.369286] task: ffff8807f2433240 ti: ffff8807ec9a0000 task.ti: ffff8807ec9a0000
<4>[   24.369317] RIP: 0010:[<ffffffff8119da34>]  [<ffffffff8119da34>] new_slab+0x2d4/0x380
<4>[   24.369359] RSP: 0018:ffff8807ec9a35d8  EFLAGS: 00010246
<4>[   24.369383] RAX: 0000000000000000 RBX: ffff8807ff403c00 RCX: 0000000000000000
<4>[   24.369412] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 00000000002012d0
<4>[   24.369441] RBP: ffff8807ec9a3608 R08: ffff8807f193cfe0 R09: 000000010080000a
<4>[   24.369471] R10: 00000000f193cf01 R11: 0000000000015f38 R12: 0000000000000000
<4>[   24.369501] R13: 0000000000000080 R14: 0000000000000000 R15: 00000000000000d0
<4>[   24.369531] FS:  00007febb0cbe980(0000) GS:ffff8807ffd80000(0000) knlGS:0000000000000000
<4>[   24.369563] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
<4>[   24.369588] CR2: 00007efc53abc1b8 CR3: 00000007f213f000 CR4: 00000000000407e0
<4>[   24.369618] Stack:
<4>[   24.369630]  ffff8807ec9a3618 0000000000000000 0000000000000000 ffff8807ffd958c0
<4>[   24.369669]  ffff8807ff403c00 00000000000080d0 ffff8807ec9a36f8 ffffffff816cc548
<4>[   24.370755]  ffff8807ec9a3708 0000000000000296 0000000000000004 0000000000000000
<4>[   24.371777] Call Trace:
<4>[   24.372929]  [<ffffffff816cc548>] __slab_alloc+0x33b/0x459
<4>[   24.374179]  [<ffffffffa0192a09>] ? ovs_flow_alloc+0x59/0x110 [openvswitch]
<4>[   24.375390]  [<ffffffff8114da93>] ? get_page_from_freelist+0x483/0x9f0
<4>[   24.376623]  [<ffffffff8136b15e>] ? memzero_explicit+0xe/0x10
<4>[   24.377767]  [<ffffffffa0192a09>] ? ovs_flow_alloc+0x59/0x110 [openvswitch]
<4>[   24.378951]  [<ffffffff8119e12c>] kmem_cache_alloc_node+0x9c/0x1b0
<4>[   24.379916]  [<ffffffff8119f08b>] ? kmem_cache_alloc+0x18b/0x1a0
<4>[   24.390806]  [<ffffffffa01929cd>] ? ovs_flow_alloc+0x1d/0x110 [openvswitch]
<4>[   24.391779]  [<ffffffffa0192a09>] ovs_flow_alloc+0x59/0x110 [openvswitch]
<4>[   24.392875]  [<ffffffffa018b18b>] ovs_flow_cmd_new+0x5b/0x360 [openvswitch]
<4>[   24.394004]  [<ffffffff8114e16c>] ? __alloc_pages_nodemask+0x16c/0xaf0
<4>[   24.394973]  [<ffffffff815bba77>] ? __alloc_skb+0x87/0x2a0
<4>[   24.395926]  [<ffffffff8138b240>] ? nla_parse+0x90/0x110
<4>[   24.476276]  [<ffffffff815fe453>] genl_family_rcv_msg+0x373/0x3d0
<4>[   24.477704]  [<ffffffff811a09dc>] ? __kmalloc_node_track_caller+0x6c/0x220
<4>[   24.478859]  [<ffffffff815fe4f4>] genl_rcv_msg+0x44/0x80
<4>[   24.479987]  [<ffffffff815fe4b0>] ? genl_family_rcv_msg+0x3d0/0x3d0
<4>[   24.481325]  [<ffffffff815fda49>] netlink_rcv_skb+0xb9/0xe0
<4>[   24.482466]  [<ffffffff815fdd6c>] genl_rcv+0x2c/0x40
<4>[   24.483554]  [<ffffffff815fd04b>] netlink_unicast+0x12b/0x1c0
<4>[   24.484739]  [<ffffffff815fd472>] netlink_sendmsg+0x392/0x6d0
<4>[   24.485942]  [<ffffffff815b2f9f>] sock_sendmsg+0xaf/0xc0
<4>[   24.486953]  [<ffffffff815fd2e2>] ? netlink_sendmsg+0x202/0x6d0
<4>[   24.487969]  [<ffffffff815b3622>] ___sys_sendmsg.part.19+0x322/0x330
<4>[   24.489167]  [<ffffffff815b3839>] ? SYSC_sendto+0xf9/0x130
<4>[   24.490217]  [<ffffffff815b367a>] ___sys_sendmsg+0x4a/0x70
<4>[   24.491162]  [<ffffffff815b40c9>] __sys_sendmsg+0x49/0x90
<4>[   24.492082]  [<ffffffff815b4129>] SyS_sendmsg+0x19/0x20
<4>[   24.493181]  [<ffffffff816d6c09>] system_call_fastpath+0x12/0x17
<4>[   24.494124] Code: 40 e9 ea fe ff ff 90 e8 6b 69 ff ff 49 89 c4 e9 07 fe ff ff 4c 89 f7 ff d0 e9 26 ff ff ff 49 c7 04 06 00 00 00 00 e9 3c ff ff ff <0f> 0b ba 00 10 00 00 be 5a 00 00 00 4c 89 ef 48 d3 e2 e8 65 2a
<1>[   24.496071] RIP  [<ffffffff8119da34>] new_slab+0x2d4/0x380
<4>[   24.497152]  RSP <ffff8807ec9a35d8>
<4>[   24.498945] ---[ end trace 6f97360ff4a9ee45 ]---
---
 net/openvswitch/flow_table.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

net/openvswitch/flow_table.c +0 -0

index f2ea83ba4763..c7f74aab34b9 100644

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO, 0) It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO, 0) It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

net/openvswitch/flow_table.c +0 -0

  	if (!stats)
  		goto err;
 

Quoted 7 lines · Konstantin Khlebnikov

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO, 0) It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

...

Quoted 6 lines

net/openvswitch/flow_table.c +0 -0

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 

net/openvswitch/flow_table.c struct sw_flow *ovs_flow_alloc(void) · line 97

Quoted 5 lines

net/openvswitch/flow_table.c +2 -1

      /* Initialize the default stat node. */
      stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Quoted 21 lines · Vlastimil Babka

Quoted 9 lines · Konstantin Khlebnikov

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
0)
It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

...

Quoted 6 lines

net/openvswitch/flow_table.c +0 -0

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 

Quoted 9 lines

net/openvswitch/flow_table.c struct sw_flow *ovs_flow_alloc(void) · line 97

Quoted 5 lines

net/openvswitch/flow_table.c +2 -1

      /* Initialize the default stat node. */
      stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Another question to ask would be is it possible for node 0 to be online, but be a memoryless node?

I would say you are better off just making this call kmem_cache_alloc. I don't see anything that indicates the memory has to come from node 0, so adding the extra overhead doesn't provide any value.

Quoted 33 lines · Alexander Duyck

Quoted 21 lines · Vlastimil Babka

Quoted 9 lines · Konstantin Khlebnikov

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
0)
It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

...

Quoted 6 lines

net/openvswitch/flow_table.c +0 -0

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 

Quoted 9 lines

net/openvswitch/flow_table.c struct sw_flow *ovs_flow_alloc(void) · line 97

Quoted 5 lines

net/openvswitch/flow_table.c +2 -1

     /* Initialize the default stat node. */
     stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Quoted 6 lines

Another question to ask would be is it possible for node 0 to be online, but be a memoryless node?

I would say you are better off just making this call kmem_cache_alloc. I don't see anything that indicates the memory has to come from node 0, so adding the extra overhead doesn't provide any value.

I agree that this at least makes me wonder, though I actually have concerns in the opposite direction - I see assumptions about this being on node 0 in net/openvswitch/flow.c.

Jarno, since you original wrote this code, can you take a look to see if everything still makes sense?

Quoted 49 lines · Jesse Gross

Quoted 33 lines · Alexander Duyck

Quoted 21 lines · Vlastimil Babka

Quoted 9 lines · Konstantin Khlebnikov

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
0)
It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

...

Quoted 6 lines

net/openvswitch/flow_table.c +0 -0

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 

Quoted 9 lines

net/openvswitch/flow_table.c struct sw_flow *ovs_flow_alloc(void) · line 97

Quoted 5 lines

net/openvswitch/flow_table.c +2 -1

     /* Initialize the default stat node. */
     stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Quoted 6 lines

Another question to ask would be is it possible for node 0 to be online, but be a memoryless node?

I would say you are better off just making this call kmem_cache_alloc. I don't see anything that indicates the memory has to come from node 0, so adding the extra overhead doesn't provide any value.

I agree that this at least makes me wonder, though I actually have concerns in the opposite direction - I see assumptions about this being on node 0 in net/openvswitch/flow.c.

Quoted 2 lines

Jarno, since you original wrote this code, can you take a look to see if everything still makes sense?

We keep the pre-allocated stats node at array index 0, which is initially used by all CPUs, but if CPUs from multiple numa nodes start updating the stats, we allocate additional stats nodes (up to one per numa node), and the CPUs on node 0 keep using the preallocated entry. If stats cannot be allocated from CPUs local node, then those CPUs keep using the entry at index 0. Currently the code in net/openvswitch/flow.c will try to allocate the local memory repeatedly, which may not be optimal when there is no memory at the local node.

Allocating the memory for the index 0 from other than node 0, as discussed here, just means that the CPUs on node 0 will keep on using non-local memory for stats. In a scenario where there are CPUs on two nodes (0, 1), but only the node 1 has memory, a shared flow entry will still end up having separate memory allocated for both nodes, but both of the nodes would be at node 1. However, there is still a high likelihood that the memory allocations would not share a cache line, which should prevent the nodes from invalidating each other’s caches. Based on this I do not see a problem relaxing the memory allocation for the default stats node. If node 0 has memory, however, it would be better to allocate the memory from node 0.

Quoted 49 lines · On Wed, Oct 7, 2015 at 10:47 AM, Jarno Rajahalme <[email protected]…

Quoted 49 lines · Jesse Gross

Quoted 33 lines · Alexander Duyck

Quoted 21 lines · Vlastimil Babka

Quoted 9 lines · Konstantin Khlebnikov

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
0)
It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

...

Quoted 6 lines

net/openvswitch/flow_table.c +0 -0

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 

Quoted 9 lines

net/openvswitch/flow_table.c struct sw_flow *ovs_flow_alloc(void) · line 97

Quoted 5 lines

net/openvswitch/flow_table.c +2 -1

    /* Initialize the default stat node. */
    stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Quoted 6 lines

Another question to ask would be is it possible for node 0 to be online, but be a memoryless node?

I would say you are better off just making this call kmem_cache_alloc. I don't see anything that indicates the memory has to come from node 0, so adding the extra overhead doesn't provide any value.

I agree that this at least makes me wonder, though I actually have concerns in the opposite direction - I see assumptions about this being on node 0 in net/openvswitch/flow.c.

Quoted 6 lines

Quoted 2 lines

Jarno, since you original wrote this code, can you take a look to see if everything still makes sense?

We keep the pre-allocated stats node at array index 0, which is initially used by all CPUs, but if CPUs from multiple numa nodes start updating the stats, we allocate additional stats nodes (up to one per numa node), and the CPUs on node 0 keep using the preallocated entry. If stats cannot be allocated from CPUs local node, then those CPUs keep using the entry at index 0. Currently the code in net/openvswitch/flow.c will try to allocate the local memory repeatedly, which may not be optimal when there is no memory at the local node.

Allocating the memory for the index 0 from other than node 0, as discussed here, just means that the CPUs on node 0 will keep on using non-local memory for stats. In a scenario where there are CPUs on two nodes (0, 1), but only the node 1 has memory, a shared flow entry will still end up having separate memory allocated for both nodes, but both of the nodes would be at node 1. However, there is still a high likelihood that the memory allocations would not share a cache line, which should prevent the nodes from invalidating each other’s caches. Based on this I do not see a problem relaxing the memory allocation for the default stats node. If node 0 has memory, however, it would be better to allocate the memory from node 0.

Thanks for going through all of that.

It seems like the question that is being raised is whether it actually makes sense to try to get the initial memory on node 0, especially since it seems to introduce some corner cases? Is there any reason why the flow is more likely to hit node 0 than a randomly chosen one? (Assuming that this is a multinode system, otherwise it's kind of a moot point.) We could have a separate pointer to the default allocated memory, so it wouldn't conflict with memory that was intentionally allocated for node 0.

On Oct 8, 2015, at 4:03 PM, Jesse Gross <[email protected]> wrote:

On Wed, Oct 7, 2015 at 10:47 AM, Jarno Rajahalme <[email protected]> wrote:

On Oct 6, 2015, at 6:01 PM, Jesse Gross <[email protected]> wrote:

On Mon, Oct 5, 2015 at 1:25 PM, Alexander Duyck
<[email protected]> wrote:

On 10/05/2015 06:59 AM, Vlastimil Babka wrote:

On 10/02/2015 12:18 PM, Konstantin Khlebnikov wrote:

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
0)
It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

...

net/openvswitch/flow_table.c +2 -1

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 
    /* Initialize the default stat node. */
    stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);
 
 
 

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Another question to ask would be is it possible for node 0 to be online, but be a memoryless node?

I would say you are better off just making this call kmem_cache_alloc. I don't see anything that indicates the memory has to come from node 0, so adding the extra overhead doesn't provide any value.

I agree that this at least makes me wonder, though I actually have concerns in the opposite direction - I see assumptions about this being on node 0 in net/openvswitch/flow.c.

Jarno, since you original wrote this code, can you take a look to see if everything still makes sense?

We keep the pre-allocated stats node at array index 0, which is initially used by all CPUs, but if CPUs from multiple numa nodes start updating the stats, we allocate additional stats nodes (up to one per numa node), and the CPUs on node 0 keep using the preallocated entry. If stats cannot be allocated from CPUs local node, then those CPUs keep using the entry at index 0. Currently the code in net/openvswitch/flow.c will try to allocate the local memory repeatedly, which may not be optimal when there is no memory at the local node.

Allocating the memory for the index 0 from other than node 0, as discussed here, just means that the CPUs on node 0 will keep on using non-local memory for stats. In a scenario where there are CPUs on two nodes (0, 1), but only the node 1 has memory, a shared flow entry will still end up having separate memory allocated for both nodes, but both of the nodes would be at node 1. However, there is still a high likelihood that the memory allocations would not share a cache line, which should prevent the nodes from invalidating each other’s caches. Based on this I do not see a problem relaxing the memory allocation for the default stats node. If node 0 has memory, however, it would be better to allocate the memory from node 0.

Thanks for going through all of that.

It seems like the question that is being raised is whether it actually makes sense to try to get the initial memory on node 0, especially since it seems to introduce some corner cases? Is there any reason why the flow is more likely to hit node 0 than a randomly chosen one? (Assuming that this is a multinode system, otherwise it's kind of a moot point.) We could have a separate pointer to the default allocated memory, so it wouldn't conflict with memory that was intentionally allocated for node 0.

It would still be preferable to know from which node the default stats node was allocated, and store it in the appropriate pointer in the array. We could then add a new “default stats node index” that would be used to locate the node in the array of pointers we already have. That way we would avoid extra allocation and processing of the default stats node.

Quoted 99 lines · On Fri, Oct 9, 2015 at 8:54 AM, Jarno Rajahalme <[email protected]…

On Oct 8, 2015, at 4:03 PM, Jesse Gross <[email protected]> wrote:

On Wed, Oct 7, 2015 at 10:47 AM, Jarno Rajahalme <[email protected]> wrote:

On Oct 6, 2015, at 6:01 PM, Jesse Gross <[email protected]> wrote:

On Mon, Oct 5, 2015 at 1:25 PM, Alexander Duyck
<[email protected]> wrote:

On 10/05/2015 06:59 AM, Vlastimil Babka wrote:

On 10/02/2015 12:18 PM, Konstantin Khlebnikov wrote:

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
0)
It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]>

...

net/openvswitch/flow_table.c +2 -1

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 
   /* Initialize the default stat node. */
   stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);
 
 
 

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Another question to ask would be is it possible for node 0 to be online, but be a memoryless node?

I would say you are better off just making this call kmem_cache_alloc. I don't see anything that indicates the memory has to come from node 0, so adding the extra overhead doesn't provide any value.

I agree that this at least makes me wonder, though I actually have concerns in the opposite direction - I see assumptions about this being on node 0 in net/openvswitch/flow.c.

Jarno, since you original wrote this code, can you take a look to see if everything still makes sense?

We keep the pre-allocated stats node at array index 0, which is initially used by all CPUs, but if CPUs from multiple numa nodes start updating the stats, we allocate additional stats nodes (up to one per numa node), and the CPUs on node 0 keep using the preallocated entry. If stats cannot be allocated from CPUs local node, then those CPUs keep using the entry at index 0. Currently the code in net/openvswitch/flow.c will try to allocate the local memory repeatedly, which may not be optimal when there is no memory at the local node.

Allocating the memory for the index 0 from other than node 0, as discussed here, just means that the CPUs on node 0 will keep on using non-local memory for stats. In a scenario where there are CPUs on two nodes (0, 1), but only the node 1 has memory, a shared flow entry will still end up having separate memory allocated for both nodes, but both of the nodes would be at node 1. However, there is still a high likelihood that the memory allocations would not share a cache line, which should prevent the nodes from invalidating each other’s caches. Based on this I do not see a problem relaxing the memory allocation for the default stats node. If node 0 has memory, however, it would be better to allocate the memory from node 0.

Thanks for going through all of that.

It seems like the question that is being raised is whether it actually makes sense to try to get the initial memory on node 0, especially since it seems to introduce some corner cases? Is there any reason why the flow is more likely to hit node 0 than a randomly chosen one? (Assuming that this is a multinode system, otherwise it's kind of a moot point.) We could have a separate pointer to the default allocated memory, so it wouldn't conflict with memory that was intentionally allocated for node 0.

Quoted 5 lines

It would still be preferable to know from which node the default stats node was allocated, and store it in the appropriate pointer in the array. We could then add a new “default stats node index” that would be used to locate the node in the array of pointers we already have. That way we would avoid extra allocation and processing of the default stats node.

I agree, that sounds reasonable to me. Will you make that change?

Besides eliminating corner cases, it might help performance in some cases too by avoiding stressing memory bandwidth on node 0.

Quoted 112 lines · On Oct 9, 2015, at 3:11 PM, Jesse Gross <[email protected]…

Quoted 99 lines · On Fri, Oct 9, 2015 at 8:54 AM, Jarno Rajahalme <[email protected]…

On Oct 8, 2015, at 4:03 PM, Jesse Gross <[email protected] <mailto:[email protected]>> wrote:

On Wed, Oct 7, 2015 at 10:47 AM, Jarno Rajahalme <[email protected] <mailto:[email protected]>> wrote:

On Oct 6, 2015, at 6:01 PM, Jesse Gross <[email protected] <mailto:[email protected]>> wrote:

On Mon, Oct 5, 2015 at 1:25 PM, Alexander Duyck <[email protected] <mailto:[email protected]>> wrote:

On 10/05/2015 06:59 AM, Vlastimil Babka wrote:

On 10/02/2015 12:18 PM, Konstantin Khlebnikov wrote:

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
0)
It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid)) [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected] <mailto:[email protected]>>

...

net/openvswitch/flow_table.c +2 -1

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 
   /* Initialize the default stat node. */
   stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);
 
 
 

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Another question to ask would be is it possible for node 0 to be online, but be a memoryless node?

I would say you are better off just making this call kmem_cache_alloc. I don't see anything that indicates the memory has to come from node 0, so adding the extra overhead doesn't provide any value.

I agree that this at least makes me wonder, though I actually have concerns in the opposite direction - I see assumptions about this being on node 0 in net/openvswitch/flow.c.

Jarno, since you original wrote this code, can you take a look to see if everything still makes sense?

We keep the pre-allocated stats node at array index 0, which is initially used by all CPUs, but if CPUs from multiple numa nodes start updating the stats, we allocate additional stats nodes (up to one per numa node), and the CPUs on node 0 keep using the preallocated entry. If stats cannot be allocated from CPUs local node, then those CPUs keep using the entry at index 0. Currently the code in net/openvswitch/flow.c will try to allocate the local memory repeatedly, which may not be optimal when there is no memory at the local node.

Allocating the memory for the index 0 from other than node 0, as discussed here, just means that the CPUs on node 0 will keep on using non-local memory for stats. In a scenario where there are CPUs on two nodes (0, 1), but only the node 1 has memory, a shared flow entry will still end up having separate memory allocated for both nodes, but both of the nodes would be at node 1. However, there is still a high likelihood that the memory allocations would not share a cache line, which should prevent the nodes from invalidating each other’s caches. Based on this I do not see a problem relaxing the memory allocation for the default stats node. If node 0 has memory, however, it would be better to allocate the memory from node 0.

Thanks for going through all of that.

It seems like the question that is being raised is whether it actually makes sense to try to get the initial memory on node 0, especially since it seems to introduce some corner cases? Is there any reason why the flow is more likely to hit node 0 than a randomly chosen one? (Assuming that this is a multinode system, otherwise it's kind of a moot point.) We could have a separate pointer to the default allocated memory, so it wouldn't conflict with memory that was intentionally allocated for node 0.

Quoted 5 lines

It would still be preferable to know from which node the default stats node was allocated, and store it in the appropriate pointer in the array. We could then add a new “default stats node index” that would be used to locate the node in the array of pointers we already have. That way we would avoid extra allocation and processing of the default stats node.

I agree, that sounds reasonable to me. Will you make that change?

Quoted 2 lines

Besides eliminating corner cases, it might help performance in some cases too by avoiding stressing memory bandwidth on node 0.

/**
* kmem_cache_alloc_node - Allocate an object on the specified node
* @cachep: The cache to allocate from.
* @flags: See kmalloc().
* @nodeid: node number of the target node.
*
* Identical to kmem_cache_alloc but it will allocate memory on the given
* node, which can improve the performance for cpu bound structures. *
* Fallback to other node is possible if __GFP_THISNODE is not set. */

/**
* cpuset_mem_spread_node() - On which node to begin search for a file page
* cpuset_slab_spread_node() - On which node to begin search for a slab page *
* If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
* tasks in a cpuset with is_spread_page or is_spread_slab set),
* and if the memory allocation used cpuset_mem_spread_node()
* to determine on which node to start looking, as it will for
* certain page cache or slab cache pages such as used for file
* system buffers and inode caches, then instead of starting on the
* local node to look for a free page, rather spread the starting
* node around the tasks mems_allowed nodes.
*
* We don't have to worry about the returned node being offline
* because "it can't happen", and even if it did, it would be ok. *
* The routines calling guarantee_online_mems() are careful to
* only set nodes in task->mems_allowed that are online. So it

* should not be possible for the following code to return an
* offline node. But if it did, that would be ok, as this routine
* is not returning the node where the allocation must be, only
* the node where the search should start. The zonelist passed to
* __alloc_pages() will include all nodes. If the slab allocator
* is passed an offline node, it will fall back to the local node.
* See kmem_cache_alloc_node().
*/

Quoted 138 lines

Quoted 133 lines · On Oct 9, 2015, at 5:02 PM, Jarno Rajahalme <[email protected]…

On Oct 9, 2015, at 3:11 PM, Jesse Gross <[email protected] <mailto:[email protected]>> wrote:

Quoted 117 lines · On Fri, Oct 9, 2015 at 8:54 AM, Jarno Rajahalme <[email protected]…

On Oct 8, 2015, at 4:03 PM, Jesse Gross <[email protected] <mailto:[email protected]>> wrote:

On Wed, Oct 7, 2015 at 10:47 AM, Jarno Rajahalme <[email protected] <mailto:[email protected]>> wrote:

On Oct 6, 2015, at 6:01 PM, Jesse Gross <[email protected] <mailto:[email protected]>> wrote:

On Mon, Oct 5, 2015 at 1:25 PM, Alexander Duyck <[email protected] <mailto:[email protected]>> wrote:

On 10/05/2015 06:59 AM, Vlastimil Babka wrote:

On 10/02/2015 12:18 PM, Konstantin Khlebnikov wrote:

When openvswitch tries allocate memory from offline numa node 0:
stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
0)
It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid))
[ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h This patch disables numa affinity in this case.

Signed-off-by Konstantin Khlebnikov <[email protected]

<mailto:[email protected]>>

...

net/openvswitch/flow_table.c +2 -1

index f2ea83ba4763..c7f74aab34b9 100644
@@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
 
   /* Initialize the default stat node. */
   stats = kmem_cache_alloc_node(flow_stats_cache,
-                      GFP_KERNEL | __GFP_ZERO, 0);
+                      GFP_KERNEL | __GFP_ZERO,
+                      node_online(0) ? 0 : NUMA_NO_NODE);
 
 
 

Stupid question: can node 0 become offline between this check, and the VM_WARN_ON? :) BTW what kind of system has node 0 offline?

Another question to ask would be is it possible for node 0 to be online, but
be a memoryless node?

I would say you are better off just making this call kmem_cache_alloc. I
don't see anything that indicates the memory has to come from node 0, so adding the extra overhead doesn't provide any value.

I agree that this at least makes me wonder, though I actually have concerns in the opposite direction - I see assumptions about this being on node 0 in net/openvswitch/flow.c.

Jarno, since you original wrote this code, can you take a look to see if everything still makes sense?

We keep the pre-allocated stats node at array index 0, which is initially
used by all CPUs, but if CPUs from multiple numa nodes start updating the
stats, we allocate additional stats nodes (up to one per numa node), and the
CPUs on node 0 keep using the preallocated entry. If stats cannot be allocated from CPUs local node, then those CPUs keep using the entry at index 0. Currently the code in net/openvswitch/flow.c will try to allocate
the local memory repeatedly, which may not be optimal when there is no memory at the local node.

Allocating the memory for the index 0 from other than node 0, as discussed
here, just means that the CPUs on node 0 will keep on using non-local memory
for stats. In a scenario where there are CPUs on two nodes (0, 1), but only
the node 1 has memory, a shared flow entry will still end up having separate
memory allocated for both nodes, but both of the nodes would be at node 1.
However, there is still a high likelihood that the memory allocations would
not share a cache line, which should prevent the nodes from invalidating each other’s caches. Based on this I do not see a problem relaxing the memory allocation for the default stats node. If node 0 has memory, however,
it would be better to allocate the memory from node 0.

Thanks for going through all of that.

It seems like the question that is being raised is whether it actually makes sense to try to get the initial memory on node 0, especially since it seems to introduce some corner cases? Is there any reason why the flow is more likely to hit node 0 than a randomly chosen one? (Assuming that this is a multinode system, otherwise it's kind of a moot point.) We could have a separate pointer to the default allocated memory, so it wouldn't conflict with memory that was intentionally allocated for node 0.

Quoted 8 lines

It would still be preferable to know from which node the default stats node
was allocated, and store it in the appropriate pointer in the array. We could then add a new “default stats node index” that would be used to locate
the node in the array of pointers we already have. That way we would avoid
extra allocation and processing of the default stats node.

I agree, that sounds reasonable to me. Will you make that change?

Quoted 3 lines

Quoted 2 lines

Besides eliminating corner cases, it might help performance in some cases too by avoiding stressing memory bandwidth on node 0.

According to the comment above kmem_cache_alloc_node(), kmem_cache_alloc_node() should not BUG_ON/WARN_ON in this case:

Quoted 4 lines

*//**/*
*/* kmem_cache_alloc_node - Allocate an object on the specified node/* */* @cachep: The cache to allocate from./*
*/* @flags: See kmalloc()./*

Quoted 8 lines

*/* @nodeid: node number of the target node./*
*/*/*
*/* Identical to kmem_cache_alloc but it will allocate memory on the given/*
*/* node, which can improve the performance for cpu bound structures./* */*/*
*/* Fallback to other node is possible if __GFP_THISNODE is not set./* */*//*

See also this from cpuset.c:

Quoted 26 lines

/**
* cpuset_mem_spread_node() - On which node to begin search for a file
page
* cpuset_slab_spread_node() - On which node to begin search for a
slab page
*
* If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
* tasks in a cpuset with is_spread_page or is_spread_slab set),
* and if the memory allocation used cpuset_mem_spread_node()
* to determine on which node to start looking, as it will for
* certain page cache or slab cache pages such as used for file
* system buffers and inode caches, then instead of starting on the
* local node to look for a free page, rather spread the starting
* node around the tasks mems_allowed nodes.
*
* We don't have to worry about the returned node being offline
* because "it can't happen", and even if it did, it would be ok. *
* The routines calling guarantee_online_mems() are careful to
* only set nodes in task->mems_allowed that are online. So it
* should not be possible for the following code to return an
* offline node. But if it did, that would be ok, as this routine
* is not returning the node where the allocation must be, only
* the node where the search should start. The zonelist passed to
* __alloc_pages() will include all nodes. If the slab allocator
* is passed an offline node, it will fall back to the local node.

Quoted 2 lines

* See kmem_cache_alloc_node().
*/

Based on this it seems this is a bug in the memory allocator, it probably should not be calling alloc_pages_exact_node()

when __GFP_THISNODE is not set?

Jarno