[ovs-dev] [PATCH] dpif-netdev: Enter quiescent state after each offloading operation.

Tue Mar 3 14:44:13 UTC 2020

On 2/28/2020 11:03 AM, Ilya Maximets wrote:
> On 2/27/20 10:00 AM, Eli Britstein wrote:
>> On 2/26/2020 2:31 PM, Ilya Maximets wrote:
>>> On 2/26/20 1:05 PM, Eli Britstein wrote:
>>>> On 2/24/2020 1:06 PM, Ilya Maximets wrote:
>>>>> On 2/23/20 3:32 PM, Eli Britstein wrote:
>>>>>> On 2/21/2020 4:54 PM, Ilya Maximets wrote:
>>>>>>> If the offloading queue is big and filled continuously, offloading
>>>>>>> thread may have no chance to quiesce blocking rcu callbacks and
>>>>>>> other threads waiting for synchronization.
>>>>>>>
>>>>>>> Fix that by entering momentary quiescent state after each operation
>>>>>>> since we're not holding any rcu-protected memory here.
>>>>>>>
>>>>>>> Fixes: 02bb2824e51d ("dpif-netdev: do hw flow offload in a thread")
>>>>>>> Reported-by: Eli Britstein <elibr at mellanox.com>
>>>>>>> Reported-at: https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmail.openvswitch.org%2Fpipermail%2Fovs-discuss%2F2020-February%2F049768.html&data=02%7C01%7Celibr%40mellanox.com%7Ceb2c70875cb14a4d4b4108d7bc2d2676%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C637184774411852418&sdata=8YgXyAJ248Mqhiq4tbUcAb357Dqdu1ue%2BBXp%2F6pX6tY%3D&reserved=0
>>>>>>> Signed-off-by: Ilya Maximets <i.maximets at ovn.org>
>>>>>>> ---
>>>>>>>      lib/dpif-netdev.c | 1 +
>>>>>>>      1 file changed, 1 insertion(+)
>>>>>>>
>>>>>>> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
>>>>>>> index d393aab5e..a798db45d 100644
>>>>>>> --- a/lib/dpif-netdev.c
>>>>>>> +++ b/lib/dpif-netdev.c
>>>>>>> @@ -2512,6 +2512,7 @@ dp_netdev_flow_offload_main(void *data OVS_UNUSED)
>>>>>>>              VLOG_DBG("%s to %s netdev flow\n",
>>>>>>>                       ret == 0 ? "succeed" : "failed", op);
>>>>>>>              dp_netdev_free_flow_offload(offload);
>>>>>>> +        ovsrcu_quiesce();
>>>>>> This seems to solve the issue of responsiveness, but I have encountered a crash using it, while there is a lot of flow deletions.
>>>>>>
>>>>>> #0  0x0000000001e462dc in get_unaligned_u32 (p=0x7f57f7813500) at lib/unaligned.h:86
>>>>>> #1  0x0000000001e46388 in hash_bytes (p_=0x7f57f7813500, n=16, basis=0) at lib/hash.c:38
>>>>>> #2  0x0000000001f7f836 in ufid_to_rte_flow_data_find (ufid=0x7f57f7813500) at lib/netdev-offload-dpdk.c:134
>>>>>> #3  0x0000000001f88693 in netdev_offload_dpdk_flow_del (netdev=0x42260da0, ufid=0x7f57f7813500, stats=0x0) at lib/netdev-offload-dpdk.c:3361
>>>>>> #4  0x0000000001e6a00d in netdev_flow_del (netdev=0x42260da0, ufid=0x7f57f7813500, stats=0x0) at lib/netdev-offload.c:296
>>>>> This might mean 2 things:
>>>>> 1. Incorrect usage of RCU protected data inside netdev-offload-dpdk.
>>>>> 2. We're not allowed to quiesce if we have any item in offloading queue,
>>>>>       i.e. bad RCU usage scheme in dpif-netdev.
>>>>>
>>>>> Both cases are bad and I'm not sure if we can easily fix case #2.
>>>>> I'll take a look and try to find a root cause.
>>>> It was my bad. Sorry. This is another issue, that exists in my branch only (CT offloads), so it's not related. I was able to fix it.
>>> OK.  Good to know.
>>>
>>>> However, I wanted to reproduce with master branch, but then I cannot create a lot of flows as I get: "upcall: datapath flow limit reached".
>>>>
>>>> I tried setting large number using "ovs-appctl upcall/set-flow-limit 400000", but it didn't help.
>>> Number of flows in dpif-netdev is artificially limited by MAX_FLOWS (65536)
>>> value.  There was a recent discussion about this limit and I think we could
>>> just remove it.  I have a small patch for this, will send soon.
>>> For now, you can just raise the hardcoded value of MAX_FLOWS in dpif-netdev.c.
>> The issue is not in dpif-netdev.c. It's in ofproto/ofproto-dpif-upcall.c. If I'm not wrong, when it happens, there is not attempt to install the flow, so anyway MAX_FLOWS in dpif-netdev.c is not the blocker.
> There are couple of things that controls the maximum number of datapath flows.
> First is the other_config:flow-limit that defaults to 200000.  Second is the
> datapath internal MAX_FLOWS define.  And also, "ovs-appctl upcall/set-flow-limit"
> should be able to modify this value.  Sounds strange.
>
>> BTW, my setup is the PF and VF rep in a bridge. I run testpmd with macswap on the VF, and send 10k different src MAC packets (OpenFlow configuration is NORMAL). So, I expect total 20k flows, also less than 64k of MAX_FLOWS.
> Are you sure that packets are not corrupted?  Could you dump datapath flows
> and verify that these flows are what you've expected?

This is for example for 4 different MACs 
(e4:10:00:00:00:01-e4:10:00:00:00:04). As expected:

$ ovs-appctl dpctl/dump-flows -m
flow-dump from pmd on cpu core: 8
ufid:65aae19f-06a9-4887-8d2d-60b04187f7cd, 
skb_priority(0/0),skb_mark(0/0),ct_state(0/0),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),recirc_id(0),dp_hash(0/0),in_port(vm1),packet_type(ns=0,id=0),eth(src=e4:11:00:00:00:01,dst=e4:10:00:00:00:01),eth_type(0x0800),ipv4(src=1.1.1.1/0.0.0.0,dst=1.1.1.2/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1025/0,dst=1025/0), 
packets:11316766, bytes:1244844260, used:0.013s, offloaded:yes, dp:dpdk, 
actions:pf, dp-extra-info:miniflow_bits(5,1)
ufid:a0e4b9bc-8973-41d1-b2cf-29f2a14935d8, 
skb_priority(0/0),skb_mark(0/0),ct_state(0/0),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),recirc_id(0),dp_hash(0/0),in_port(pf),packet_type(ns=0,id=0),eth(src=e4:10:00:00:00:03,dst=e4:11:00:00:00:01),eth_type(0x0800),ipv4(src=1.1.1.2/0.0.0.0,dst=1.1.1.1/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1025/0,dst=1025/0), 
packets:109476803, bytes:12042448330, used:0.013s, offloaded:yes, 
dp:dpdk, actions:vm1, dp-extra-info:miniflow_bits(5,1)
ufid:bbb7fe49-422b-434c-bd70-32d6a90c44c4, 
skb_priority(0/0),skb_mark(0/0),ct_state(0/0),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),recirc_id(0),dp_hash(0/0),in_port(vm1),packet_type(ns=0,id=0),eth(src=e4:11:00:00:00:01,dst=e4:10:00:00:00:03),eth_type(0x0800),ipv4(src=1.1.1.1/0.0.0.0,dst=1.1.1.2/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1025/0,dst=1025/0), 
packets:11147160, bytes:1226187600, used:0.013s, offloaded:yes, dp:dpdk, 
actions:pf, dp-extra-info:miniflow_bits(5,1)
ufid:bebe7de1-b2d6-48d6-87aa-3fab025a793d, 
skb_priority(0/0),skb_mark(0/0),ct_state(0/0),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),recirc_id(0),dp_hash(0/0),in_port(pf),packet_type(ns=0,id=0),eth(src=e4:10:00:00:00:02,dst=e4:11:00:00:00:01),eth_type(0x0800),ipv4(src=1.1.1.2/0.0.0.0,dst=1.1.1.1/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1025/0,dst=1025/0), 
packets:110643486, bytes:12170783460, used:0.013s, offloaded:yes, 
dp:dpdk, actions:vm1, dp-extra-info:miniflow_bits(5,1)
ufid:24e09efe-f02c-41d5-a7c1-3d0aea28cbe0, 
skb_priority(0/0),skb_mark(0/0),ct_state(0/0),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),recirc_id(0),dp_hash(0/0),in_port(vm1),packet_type(ns=0,id=0),eth(src=e4:11:00:00:00:01,dst=e4:10:00:00:00:02),eth_type(0x0800),ipv4(src=1.1.1.1/0.0.0.0,dst=1.1.1.2/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1025/0,dst=1025/0), 
packets:11304574, bytes:1243503140, used:0.013s, offloaded:yes, dp:dpdk, 
actions:pf, dp-extra-info:miniflow_bits(5,1)
ufid:4fba2205-37d8-42a1-8899-57bdbbc1c539, 
skb_priority(0/0),skb_mark(0/0),ct_state(0/0),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),recirc_id(0),dp_hash(0/0),in_port(pf),packet_type(ns=0,id=0),eth(src=e4:10:00:00:00:01,dst=e4:11:00:00:00:01),eth_type(0x0800),ipv4(src=1.1.1.2/0.0.0.0,dst=1.1.1.1/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1025/0,dst=1025/0), 
packets:110644871, bytes:12170935810, used:0.013s, offloaded:yes, 
dp:dpdk, actions:vm1, dp-extra-info:miniflow_bits(5,1)
ufid:99848c40-a8cd-4732-b520-30a27010964b, 
skb_priority(0/0),skb_mark(0/0),ct_state(0/0),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),recirc_id(0),dp_hash(0/0),in_port(pf),packet_type(ns=0,id=0),eth(src=e4:10:00:00:00:04,dst=e4:11:00:00:00:01),eth_type(0x0800),ipv4(src=1.1.1.2/0.0.0.0,dst=1.1.1.1/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1025/0,dst=1025/0), 
packets:110644198, bytes:12170861780, used:0.013s, offloaded:yes, 
dp:dpdk, actions:vm1, dp-extra-info:miniflow_bits(5,1)
ufid:293e5fa1-cce6-41d2-9b3c-db01105f3cc2, 
skb_priority(0/0),skb_mark(0/0),ct_state(0/0),ct_zone(0/0),ct_mark(0/0),ct_label(0/0),recirc_id(0),dp_hash(0/0),in_port(vm1),packet_type(ns=0,id=0),eth(src=e4:11:00:00:00:01,dst=e4:10:00:00:00:04),eth_type(0x0800),ipv4(src=1.1.1.1/0.0.0.0,dst=1.1.1.2/0.0.0.0,proto=17/0,tos=0/0,ttl=64/0,frag=no),udp(src=1025/0,dst=1025/0), 
packets:11327727, bytes:1246049970, used:0.013s, offloaded:yes, dp:dpdk, 
actions:pf, dp-extra-info:miniflow_bits(5,1)

>
>>>>>> In other trials, I see the ovs-appctl stuck again, but then I cannot attach gdb (or use pstack). It hangs. In this scenario, I see this in dmesg:
>>>>>>
>>>>>> [  404.450694] ovs-vswitchd[8574]: segfault at 7f8bd4e87280 ip 0000000001e463c3 sp 00007f8cfaf33028 error 4 in ovs-vswitchd[400000+200c000]
>>>>>>
>>>>>> Yanqin Wei <Yanqin.Wei at arm.com> suggested to add ovsrcu_try_quiesce call (in the same place). It seems more stable (haven't see such crash as above), but it has the same stuck symptom as above.
>>>>> ovsrcu_try_quiesce() is the same as ovsrcu_quiesce() except that is fails
>>>>> if mutex can not be taken right now.  There is no real difference.  Your
>>>>> crashes are less frequent just because thread enters quiescent state less
>>>>> frequently.
>>>> If so, I think "ovsrcu_try_quiesce" is more suitable here. Don't you think?
>>> I don't think that taking of one more mutex could make harm in
>>> offloading thread.  If we'll want to care about minor sleeping
>>> spikes, we should start from replacing the dp_flow_offload.mutex
>>> and sleeping on the conditional variable.
>> Agree it's not critical. However, if it's not mandatory, I think it's better to avoid it.
> Since it's not critical, I think it's better to not block RCU callbacks
> and other threads that waits for RCU synchronization.

I am not clear. I thought ovsrcu_try_quiesce is more lightweight than 
ovsrcu_quiesce. Now I see in the code there is indeed a lock there 
(seq_try_lock()).

So, what's the logic behind your previous comment:

Your
crashes are less frequent just because thread enters quiescent state less
frequently.

Is it because of that lock in ovsrcu_try_quiesce the thread enters 
quiescent less frequently? So actually ovsrcu_try_quiesce is heavier 
than ovsrcu_quiesce.

>
>> In addition I think of making it multi-threaded (thread per port) to boost insertion/deletion rates.
>>
> This will not give any performance unless you've found a good way how to
> avoid taking of global dp->port_mutex.  Also, "per port" seems too much.
Of course. Need to find a way not to have a global lock. Regarding 
"per-port", maybe not a thread per-port, but per-port request queue. The 
number of threads can maybe configurable, each handling several queues.