Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
Dear Nitin, if you read Chetan’s email bellow, you will see that this one is already excluded… Also, it will not be easy to explain how this patch blows tx function in dpdk mlx5 pmd… — Damjan > On 17 Feb 2020, at 11:12, Nitin Saxena wrote: > > Hi Prashant/Chetan, > > I would try following change first to solve the problem in 1908 > > commit b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897b > Author: Damjan Marion > Date: Tue Mar 12 18:14:15 2019 +0100 > > vlib: don't use vector for keeping buffer indices in > > Type: refactor > > Change-Id: I72221b97d7e0bf5c93e20bbda4473ca67bfcdeb4 > Signed-off-by: Damjan Marion damar...@cisco.com > > You can also try copying src/plugins/dpdk/buffer.c from stable/2001 branch to > stable/1908 > > Thanks, > Nitin > > From: vpp-dev@lists.fd.io On Behalf Of Damjan Marion > via Lists.Fd.Io > Sent: Monday, February 17, 2020 1:52 PM > To: chetan bhasin > Cc: vpp-dev@lists.fd.io > Subject: [EXT] Re: [vpp-dev] Regarding buffers-per-numa parameter > > External Email > > On 17 Feb 2020, at 07:37, chetan bhasin wrote: > > Bottom line is stable/vpp 908 does not work with higher number of buffers but > stable/vpp2001 does. Could you please advise which area we can look at ,as it > would be difficult for us to move to vpp2001 at this time. > > I really don’t have idea what caused this problem to disappear. > You may try to use “git bisect” to find out which commit fixed it…. > > — > Damjan > > > > On Mon, Feb 17, 2020 at 11:01 AM chetan bhasin via Lists.Fd.Io > wrote: > Thanks Damjan for the reply! > > Following are my observations on Intel X710/XL710 pci- > 1) I took latest code base from stable/vpp19.08 : Seeing error as " > ethernet-input l3 mac mismatch" > With Buffers 537600 > vpp# show buffers > | > Pool NameIndex NUMA Size Data Size Total Avail Cached Used > default-numa-0 0 0 2496 2048 537600 510464 131925817 > default-numa-1 1 1 2496 2048 537600 5288963908314 > > vpp# show hardware-interfaces > NameIdx Link Hardware > BondEthernet0 3 up BondEthernet0 > Link speed: unknown > Ethernet address 3c:fd:fe:b5:5e:40 > FortyGigabitEthernet12/0/0 1 up FortyGigabitEthernet12/0/0 > Link speed: 40 Gbps > Ethernet address 3c:fd:fe:b5:5e:40 > Intel X710/XL710 Family > carrier up full duplex mtu 9206 > flags: admin-up pmd rx-ip4-cksum > rx: queues 16 (max 320), desc 1024 (min 64 max 4096 align 32) > tx: queues 16 (max 320), desc 4096 (min 64 max 4096 align 32) > pci: device 8086:1583 subsystem 8086:0001 address :12:00.00 numa 0 > max rx packet len: 9728 > promiscuous: unicast off all-multicast on > vlan offload: strip off filter off qinq off > rx offload avail: vlan-strip ipv4-cksum udp-cksum tcp-cksum qinq-strip >outer-ipv4-cksum vlan-filter vlan-extend jumbo-frame >scatter keep-crc > rx offload active: ipv4-cksum > tx offload avail: vlan-insert ipv4-cksum udp-cksum tcp-cksum sctp-cksum >tcp-tso outer-ipv4-cksum qinq-insert vxlan-tnl-tso >gre-tnl-tso ipip-tnl-tso geneve-tnl-tso multi-segs >mbuf-fast-free > tx offload active: none > rss avail: ipv4-frag ipv4-tcp ipv4-udp ipv4-sctp ipv4-other > ipv6-frag >ipv6-tcp ipv6-udp ipv6-sctp ipv6-other l2-payload > rss active:ipv4-frag ipv4-tcp ipv4-udp ipv4-other ipv6-frag > ipv6-tcp >ipv6-udp ipv6-other > tx burst function: i40e_xmit_pkts_vec_avx2 > rx burst function: i40e_recv_pkts_vec_avx2 > tx errors 17 > rx frames ok4585 > rx bytes ok 391078 > extended stats: > rx good packets 4585 > rx good bytes 391078 > tx errors 17 > rx multicast packets 4345 > rx broadcast packets 243 > rx unknown protocol packets 4588 > rx size 65 to 127 packets 4529 > rx size 128 to 255 packets 32 > rx size 256 to 511 packets 26 > rx size 1024 to 1522 packets 1 > tx size 65 to 127 packets 33 > FortyGigabitEthernet12/0/1 2 up FortyGigabitEthernet12/0/1 > Link speed: 40 Gbps > Ethernet address 3c:fd:fe:b5:5e:40 > Intel X710/XL710 Family > carrier up full duplex
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
Thanks Damjan and Nikhil for your time. I also find below logs via dmesg (Intel X710/XL710 ) [root@bfs-dl360g10-25 vpp]# uname -a Linux bfs-dl360g10-25 3.10.0-957.5.1.el7.x86_64 #1 SMP Wed Dec 19 10:46:58 EST 2018 x86_64 x86_64 x86_64 GNU/Linux [root@bfs-dl360g10-25 vpp]# uname -r 3.10.0-957.5.1.el7.x86_64 Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: DRHD: handling fault status reg 400 Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: DRHD: handling fault status reg 402 Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: [DMA Read] Request device [12:00.0] fault addr 5ec7f31000 [fault reason 06] PTE Read access is not set Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: DRHD: handling fault status reg 502 Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: [DMA Read] Request device [12:00.0] fault addr 5ec804 [fault reason 06] PTE Read access is not set Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: [DMA Read] Request device [12:00.0] fault addr 5ec53be000 [fault reason 06] PTE Read access is not set Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: DRHD: handling fault status reg 700 Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: DRHD: handling fault status reg 702 Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: [DMA Read] Request device [12:00.0] fault addr 5ec6f24000 [fault reason 06] PTE Read access is not set Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: [DMA Write] Request device [12:00.0] fault addr 5ec60eb000 [fault reason 05] PTE Write access is not set Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: [DMA Write] Request device [12:00.0] fault addr 5ec6684000 [fault reason 05] PTE Write access is not set Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: [DMA Write] Request device [12:00.0] fault addr 5ec607d000 [fault reason 05] PTE Write access is not set Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: DRHD: handling fault status reg 300 Feb 17 10:38:46 bfs-dl360g10-05 kernel: DMAR: DRHD: handling fault status reg 302 Thanks, Chetan On Mon, Feb 17, 2020 at 3:47 PM Damjan Marion wrote: > > Dear Nitin, > > if you read Chetan’s email bellow, you will see that this one is already > excluded… > > Also, it will not be easy to explain how this patch blows tx function in > dpdk mlx5 pmd… > > — > Damjan > > > On 17 Feb 2020, at 11:12, Nitin Saxena wrote: > > > > Hi Prashant/Chetan, > > > > I would try following change first to solve the problem in 1908 > > > > commit b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897b > > Author: Damjan Marion > > Date: Tue Mar 12 18:14:15 2019 +0100 > > > > vlib: don't use vector for keeping buffer indices in > > > > Type: refactor > > > > Change-Id: I72221b97d7e0bf5c93e20bbda4473ca67bfcdeb4 > > Signed-off-by: Damjan Marion damar...@cisco.com > > > > You can also try copying src/plugins/dpdk/buffer.c from stable/2001 > branch to stable/1908 > > > > Thanks, > > Nitin > > > > From: vpp-dev@lists.fd.io On Behalf Of Damjan > Marion via Lists.Fd.Io > > Sent: Monday, February 17, 2020 1:52 PM > > To: chetan bhasin > > Cc: vpp-dev@lists.fd.io > > Subject: [EXT] Re: [vpp-dev] Regarding buffers-per-numa parameter > > > > External Email > > > > On 17 Feb 2020, at 07:37, chetan bhasin > wrote: > > > > Bottom line is stable/vpp 908 does not work with higher number of > buffers but stable/vpp2001 does. Could you please advise which area we can > look at ,as it would be difficult for us to move to vpp2001 at this time. > > > > I really don’t have idea what caused this problem to disappear. > > You may try to use “git bisect” to find out which commit fixed it…. > > > > — > > Damjan > > > > > > > > On Mon, Feb 17, 2020 at 11:01 AM chetan bhasin via Lists.Fd.Io > wrote: > > Thanks Damjan for the reply! > > > > Following are my observations on Intel X710/XL710 pci- > > 1) I took latest code base from stable/vpp19.08 : Seeing error as " > ethernet-input l3 mac mismatch" > > With Buffers 537600 > > vpp# show buffers >| > > Pool NameIndex NUMA Size Data Size Total Avail Cached > Used > > default-numa-0 0 0 2496 2048 537600 510464 1319 > 25817 > > default-numa-1 1 1 2496 2048 537600 528896390 > 8314 > > > > vpp# show hardware-interfaces > > NameIdx Link Hardware > > BondEthernet0 3 up BondEthernet0 > > Link speed: unknown > > Ethernet address 3c:fd:fe:b5:5e:40 > > FortyGigabitEthernet12/0/0 1 up FortyGigabitEthernet12/0/0 > > Link speed: 40 Gbps > > Ethernet address 3c:fd:fe:b5:5e:40 > > Intel X710/XL710 Family > > carrier up full duplex mtu 9206 > > flags: admin-up pmd rx-ip4-cksum > > rx: queues 16 (max 320), desc 1024 (min 64 max 4096 align 32) > > tx: queues 16 (max 320), desc 4096 (min 64 max 4096 align 32) > > pci: device 8086:1583 subsystem 8086:0001 address :12:00.00 numa > 0 > > max rx packet len: 9728 > > promiscuous: unicas
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
Hi Damjan, >> if you read Chetan’s email bellow, you will see that this one is already >> excluded… Sorry I missed that part. After seeing diffs between stable/1908 and stable/2001, commit: b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897 is the only visible git commit in dpdk plugin which is playing with mempool buffers. If it does not solve the problem then I suspect problem lies outside dpdk plugin. I am guessing DPDK-19.08 is being used here with VPP-19.08 Hi Chetan, > > 3) I took previous commit of "vlib: don't use vector for keeping buffer > indices in the pool " ie "df0191ead2cf39611714b6603cdc5bdddc445b57" : > Everything looks fine with Buffers 537600. In which branch, Commit: df0191ead2cf39611714b6603cdc5bdddc445b57 is previous commit of b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? Thanks, Nitin > -Original Message- > From: Damjan Marion > Sent: Monday, February 17, 2020 3:47 PM > To: Nitin Saxena > Cc: chetan bhasin ; vpp-dev@lists.fd.io > Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter > > > Dear Nitin, > > if you read Chetan’s email bellow, you will see that this one is already > excluded… > > Also, it will not be easy to explain how this patch blows tx function in dpdk > mlx5 pmd… > > — > Damjan > > > On 17 Feb 2020, at 11:12, Nitin Saxena wrote: > > > > Hi Prashant/Chetan, > > > > I would try following change first to solve the problem in 1908 > > > > commit b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897b > > Author: Damjan Marion > > Date: Tue Mar 12 18:14:15 2019 +0100 > > > > vlib: don't use vector for keeping buffer indices in > > > > Type: refactor > > > > Change-Id: I72221b97d7e0bf5c93e20bbda4473ca67bfcdeb4 > > Signed-off-by: Damjan Marion damar...@cisco.com > > > > You can also try copying src/plugins/dpdk/buffer.c from stable/2001 > branch to stable/1908 > > > > Thanks, > > Nitin > > > > From: vpp-dev@lists.fd.io On Behalf Of Damjan > Marion via Lists.Fd.Io > > Sent: Monday, February 17, 2020 1:52 PM > > To: chetan bhasin > > Cc: vpp-dev@lists.fd.io > > Subject: [EXT] Re: [vpp-dev] Regarding buffers-per-numa parameter > > > > External Email > > > > On 17 Feb 2020, at 07:37, chetan bhasin > wrote: > > > > Bottom line is stable/vpp 908 does not work with higher number of buffers > but stable/vpp2001 does. Could you please advise which area we can look at > ,as it would be difficult for us to move to vpp2001 at this time. > > > > I really don’t have idea what caused this problem to disappear. > > You may try to use “git bisect” to find out which commit fixed it…. > > > > — > > Damjan > > > > > > > > On Mon, Feb 17, 2020 at 11:01 AM chetan bhasin via Lists.Fd.Io > wrote: > > Thanks Damjan for the reply! > > > > Following are my observations on Intel X710/XL710 pci- > > 1) I took latest code base from stable/vpp19.08 : Seeing error as " > ethernet-input l3 mac mismatch" > > With Buffers 537600 > > vpp# show buffers > | > > Pool NameIndex NUMA Size Data Size Total Avail Cached > > Used > > default-numa-0 0 0 2496 2048 537600 510464 1319 > > 25817 > > default-numa-1 1 1 2496 2048 537600 528896390 > > 8314 > > > > vpp# show hardware-interfaces > > NameIdx Link Hardware > > BondEthernet0 3 up BondEthernet0 > > Link speed: unknown > > Ethernet address 3c:fd:fe:b5:5e:40 > > FortyGigabitEthernet12/0/0 1 up FortyGigabitEthernet12/0/0 > > Link speed: 40 Gbps > > Ethernet address 3c:fd:fe:b5:5e:40 > > Intel X710/XL710 Family > > carrier up full duplex mtu 9206 > > flags: admin-up pmd rx-ip4-cksum > > rx: queues 16 (max 320), desc 1024 (min 64 max 4096 align 32) > > tx: queues 16 (max 320), desc 4096 (min 64 max 4096 align 32) > > pci: device 8086:1583 subsystem 8086:0001 address :12:00.00 numa > 0 > > max rx packet len: 9728 > > promiscuous: unicast off all-multicast on > > vlan offload: strip off filter off qinq off > > rx offload avail: vlan-strip ipv4-cksum udp-cksum tcp-cksum qinq-strip > >outer-ipv4-cksum vlan-filter vlan-extend jumbo-frame > >scatter keep-crc > > rx offload active: ipv4-cksum > > tx offload avail: vlan-insert ipv4-cksum udp-cksum tcp-cks
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
>> I am guessing DPDK-19.08 is being used here with VPP-19.08 Typo, dpdk-19.05 and not dpdk-19.08 > -Original Message- > From: vpp-dev@lists.fd.io On Behalf Of Nitin Saxena > Sent: Monday, February 17, 2020 5:34 PM > To: Damjan Marion > Cc: chetan bhasin ; vpp-dev@lists.fd.io > Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter > > Hi Damjan, > > >> if you read Chetan’s email bellow, you will see that this one is > >> already excluded… > Sorry I missed that part. After seeing diffs between stable/1908 and > stable/2001, commit: b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897 is the > only visible git commit in dpdk plugin which is playing with mempool buffers. > If it does not solve the problem then I suspect problem lies outside dpdk > plugin. I am guessing DPDK-19.08 is being used here with VPP-19.08 > > Hi Chetan, > > > 3) I took previous commit of "vlib: don't use vector for keeping > > > buffer > > indices in the pool " ie "df0191ead2cf39611714b6603cdc5bdddc445b57" : > > Everything looks fine with Buffers 537600. > In which branch, Commit: df0191ead2cf39611714b6603cdc5bdddc445b57 is > previous commit of b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? > > Thanks, > Nitin > > -Original Message- > > From: Damjan Marion > > Sent: Monday, February 17, 2020 3:47 PM > > To: Nitin Saxena > > Cc: chetan bhasin ; vpp-dev@lists.fd.io > > Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter > > > > > > Dear Nitin, > > > > if you read Chetan’s email bellow, you will see that this one is > > already excluded… > > > > Also, it will not be easy to explain how this patch blows tx function > > in dpdk > > mlx5 pmd… > > > > — > > Damjan > > > > > On 17 Feb 2020, at 11:12, Nitin Saxena wrote: > > > > > > Hi Prashant/Chetan, > > > > > > I would try following change first to solve the problem in 1908 > > > > > > commit b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897b > > > Author: Damjan Marion > > > Date: Tue Mar 12 18:14:15 2019 +0100 > > > > > > vlib: don't use vector for keeping buffer indices in > > > > > > Type: refactor > > > > > > Change-Id: I72221b97d7e0bf5c93e20bbda4473ca67bfcdeb4 > > > Signed-off-by: Damjan Marion damar...@cisco.com > > > > > > You can also try copying src/plugins/dpdk/buffer.c from stable/2001 > > branch to stable/1908 > > > > > > Thanks, > > > Nitin > > > > > > From: vpp-dev@lists.fd.io On Behalf Of Damjan > > Marion via Lists.Fd.Io > > > Sent: Monday, February 17, 2020 1:52 PM > > > To: chetan bhasin > > > Cc: vpp-dev@lists.fd.io > > > Subject: [EXT] Re: [vpp-dev] Regarding buffers-per-numa parameter > > > > > > External Email > > > > > > On 17 Feb 2020, at 07:37, chetan bhasin > > wrote: > > > > > > Bottom line is stable/vpp 908 does not work with higher number of > > > buffers > > but stable/vpp2001 does. Could you please advise which area we can > > look at ,as it would be difficult for us to move to vpp2001 at this time. > > > > > > I really don’t have idea what caused this problem to disappear. > > > You may try to use “git bisect” to find out which commit fixed it…. > > > > > > — > > > Damjan > > > > > > > > > > > > On Mon, Feb 17, 2020 at 11:01 AM chetan bhasin via Lists.Fd.Io > > wrote: > > > Thanks Damjan for the reply! > > > > > > Following are my observations on Intel X710/XL710 pci- > > > 1) I took latest code base from stable/vpp19.08 : Seeing error as " > > ethernet-input l3 mac mismatch" > > > With Buffers 537600 vpp# show buffers > > | > > > Pool NameIndex NUMA Size Data Size Total Avail Cached > > > Used > > > default-numa-0 0 0 2496 2048 537600 510464 1319 > > > 25817 > > > default-numa-1 1 1 2496 2048 537600 528896390 > > > 8314 > > > > > > vpp# show hardware-interfaces > > > NameIdx Link Hardware > > > BondEthernet0 3 up BondEthernet0 > > > Link speed: unknown > > > Ethernet address 3c:fd:fe:b5:5e:40 > > > FortyGigabitEthernet12/0/0 1 up FortyGigabitEthernet12/0/0
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
Hi Nitin, https://github.com/FDio/vpp/commits/stable/2001/src/vlib As per stable/2001 branch , the given change is checked-in around Oct 28 2019. df0191ead2cf39611714b6603cdc5bdddc445b57 is previous commit of b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? Yes (branch vpp 20.01) Thanks, Chetan Bhasin On Mon, Feb 17, 2020 at 5:33 PM Nitin Saxena wrote: > Hi Damjan, > > >> if you read Chetan’s email bellow, you will see that this one is > already excluded… > Sorry I missed that part. After seeing diffs between stable/1908 and > stable/2001, commit: b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897 is the only > visible git commit in dpdk plugin which is playing with mempool buffers. If > it does not solve the problem then I suspect problem lies outside dpdk > plugin. I am guessing DPDK-19.08 is being used here with VPP-19.08 > > Hi Chetan, > > > 3) I took previous commit of "vlib: don't use vector for keeping > buffer > > indices in the pool " ie "df0191ead2cf39611714b6603cdc5bdddc445b57" : > > Everything looks fine with Buffers 537600. > In which branch, Commit: df0191ead2cf39611714b6603cdc5bdddc445b57 is > previous commit of b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? > > Thanks, > Nitin > > -Original Message- > > From: Damjan Marion > > Sent: Monday, February 17, 2020 3:47 PM > > To: Nitin Saxena > > Cc: chetan bhasin ; vpp-dev@lists.fd.io > > Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter > > > > > > Dear Nitin, > > > > if you read Chetan’s email bellow, you will see that this one is already > > excluded… > > > > Also, it will not be easy to explain how this patch blows tx function in > dpdk > > mlx5 pmd… > > > > — > > Damjan > > > > > On 17 Feb 2020, at 11:12, Nitin Saxena wrote: > > > > > > Hi Prashant/Chetan, > > > > > > I would try following change first to solve the problem in 1908 > > > > > > commit b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897b > > > Author: Damjan Marion > > > Date: Tue Mar 12 18:14:15 2019 +0100 > > > > > > vlib: don't use vector for keeping buffer indices in > > > > > > Type: refactor > > > > > > Change-Id: I72221b97d7e0bf5c93e20bbda4473ca67bfcdeb4 > > > Signed-off-by: Damjan Marion damar...@cisco.com > > > > > > You can also try copying src/plugins/dpdk/buffer.c from stable/2001 > > branch to stable/1908 > > > > > > Thanks, > > > Nitin > > > > > > From: vpp-dev@lists.fd.io On Behalf Of Damjan > > Marion via Lists.Fd.Io > > > Sent: Monday, February 17, 2020 1:52 PM > > > To: chetan bhasin > > > Cc: vpp-dev@lists.fd.io > > > Subject: [EXT] Re: [vpp-dev] Regarding buffers-per-numa parameter > > > > > > External Email > > > > > > On 17 Feb 2020, at 07:37, chetan bhasin > > wrote: > > > > > > Bottom line is stable/vpp 908 does not work with higher number of > buffers > > but stable/vpp2001 does. Could you please advise which area we can look > at > > ,as it would be difficult for us to move to vpp2001 at this time. > > > > > > I really don’t have idea what caused this problem to disappear. > > > You may try to use “git bisect” to find out which commit fixed it…. > > > > > > — > > > Damjan > > > > > > > > > > > > On Mon, Feb 17, 2020 at 11:01 AM chetan bhasin via Lists.Fd.Io > > wrote: > > > Thanks Damjan for the reply! > > > > > > Following are my observations on Intel X710/XL710 pci- > > > 1) I took latest code base from stable/vpp19.08 : Seeing error as " > > ethernet-input l3 mac mismatch" > > > With Buffers 537600 > > > vpp# show buffers > > | > > > Pool NameIndex NUMA Size Data Size Total Avail > Cached Used > > > default-numa-0 0 0 2496 2048 537600 510464 1319 > 25817 > > > default-numa-1 1 1 2496 2048 537600 528896390 > 8314 > > > > > > vpp# show hardware-interfaces > > > NameIdx Link Hardware > > > BondEthernet0 3 up BondEthernet0 > > > Link speed: unknown > > > Ethernet address 3c:fd:fe:b5:5e:40 > > > FortyGigabitEthernet12/0/0 1 up > FortyGigabitEthernet12/0/0 > > > Link speed: 40 Gbps > > > Ethernet ad
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
Hi, One more finding related to intel nic and number of buffers (537600) vpp branch driver card buffers Traffic Err stable/1908 uio_pci_genric X722(10G) 537600 Working *stable/1908* *vfio-pci* *XL710(40G)* *537600 * *Not Working* *l3 mac mismatch* stable/2001 uio_pci_genric X722(10G) 537600 Working stable/2001 vfio-pci XL710(40G) 537600 Working Thanks, Chetan On Mon, Feb 17, 2020 at 7:17 PM chetan bhasin via Lists.Fd.Io wrote: > Hi Nitin, > > https://github.com/FDio/vpp/commits/stable/2001/src/vlib > As per stable/2001 branch , the given change is checked-in around Oct 28 > 2019. > > df0191ead2cf39611714b6603cdc5bdddc445b57 is previous commit of > b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? > Yes (branch vpp 20.01) > > Thanks, > Chetan Bhasin > > On Mon, Feb 17, 2020 at 5:33 PM Nitin Saxena wrote: > >> Hi Damjan, >> >> >> if you read Chetan’s email bellow, you will see that this one is >> already excluded… >> Sorry I missed that part. After seeing diffs between stable/1908 and >> stable/2001, commit: b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897 is the only >> visible git commit in dpdk plugin which is playing with mempool buffers. If >> it does not solve the problem then I suspect problem lies outside dpdk >> plugin. I am guessing DPDK-19.08 is being used here with VPP-19.08 >> >> Hi Chetan, >> > > 3) I took previous commit of "vlib: don't use vector for keeping >> buffer >> > indices in the pool " ie "df0191ead2cf39611714b6603cdc5bdddc445b57" : >> > Everything looks fine with Buffers 537600. >> In which branch, Commit: df0191ead2cf39611714b6603cdc5bdddc445b57 is >> previous commit of b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? >> >> Thanks, >> Nitin >> > -----Original Message- >> > From: Damjan Marion >> > Sent: Monday, February 17, 2020 3:47 PM >> > To: Nitin Saxena >> > Cc: chetan bhasin ; vpp-dev@lists.fd.io >> > Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter >> > >> > >> > Dear Nitin, >> > >> > if you read Chetan’s email bellow, you will see that this one is already >> > excluded… >> > >> > Also, it will not be easy to explain how this patch blows tx function >> in dpdk >> > mlx5 pmd… >> > >> > — >> > Damjan >> > >> > > On 17 Feb 2020, at 11:12, Nitin Saxena wrote: >> > > >> > > Hi Prashant/Chetan, >> > > >> > > I would try following change first to solve the problem in 1908 >> > > >> > > commit b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897b >> > > Author: Damjan Marion >> > > Date: Tue Mar 12 18:14:15 2019 +0100 >> > > >> > > vlib: don't use vector for keeping buffer indices in >> > > >> > > Type: refactor >> > > >> > > Change-Id: I72221b97d7e0bf5c93e20bbda4473ca67bfcdeb4 >> > > Signed-off-by: Damjan Marion damar...@cisco.com >> > > >> > > You can also try copying src/plugins/dpdk/buffer.c from stable/2001 >> > branch to stable/1908 >> > > >> > > Thanks, >> > > Nitin >> > > >> > > From: vpp-dev@lists.fd.io On Behalf Of Damjan >> > Marion via Lists.Fd.Io >> > > Sent: Monday, February 17, 2020 1:52 PM >> > > To: chetan bhasin >> > > Cc: vpp-dev@lists.fd.io >> > > Subject: [EXT] Re: [vpp-dev] Regarding buffers-per-numa parameter >> > > >> > > External Email >> > > >> > > On 17 Feb 2020, at 07:37, chetan bhasin >> > wrote: >> > > >> > > Bottom line is stable/vpp 908 does not work with higher number of >> buffers >> > but stable/vpp2001 does. Could you please advise which area we can look >> at >> > ,as it would be difficult for us to move to vpp2001 at this time. >> > > >> > > I really don’t have idea what caused this problem to disappear. >> > > You may try to use “git bisect” to find out which commit fixed it…. >> > > >> > > — >> > > Damjan >> > > >> > > >> > > >> > > On Mon, Feb 17, 2020 at 11:01 AM chetan bhasin via Lists.Fd.Io >> > wrote: >> > > Thanks Damjan for the reply! >> > > >> > > Following are my observations on Intel X710/XL710 pci- >> > > 1) I took latest code base from stable/vpp19.08 : Seeing error as " >> > ethernet-input
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
HI Chethan, Your packet trace shows that packet data is all 0 and that’s why you are running into l3 mac mismatch. I am guessing something messed with IOMMU due to which translation is not happening. Although packet length is correct. You can try out AVF plugin to iron out where problem exists, in dpdk plugin or vlib Thanks, Nitin From: chetan bhasin Sent: Tuesday, February 18, 2020 12:50 PM To: me Cc: Nitin Saxena ; vpp-dev Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter Hi, One more finding related to intel nic and number of buffers (537600) vpp branch driver card buffers Traffic Err stable/1908 uio_pci_genric X722(10G) 537600 Working stable/1908 vfio-pci XL710(40G) 537600 Not Working l3 mac mismatch stable/2001 uio_pci_genric X722(10G) 537600 Working stable/2001 vfio-pci XL710(40G) 537600 Working Thanks, Chetan On Mon, Feb 17, 2020 at 7:17 PM chetan bhasin via Lists.Fd.Io<https://urldefense.proofpoint.com/v2/url?u=http-3A__Lists.Fd.Io&d=DwMFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=S4H7jibYAtA5YOvfL3IkGduCfk9LbZMPOAecQGDzWV0&m=qxJrqbz5sNlCrzJTOZjaJ0jHeaW077bX6ZxmV308jfg&s=ffS1Y8GHllzjueMUVW31gwrVEIK1HVSNTKk2yA-VjG8&e=> mailto:gmail@lists.fd.io>> wrote: Hi Nitin, https://github.com/FDio/vpp/commits/stable/2001/src/vlib<https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_FDio_vpp_commits_stable_2001_src_vlib&d=DwMFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=S4H7jibYAtA5YOvfL3IkGduCfk9LbZMPOAecQGDzWV0&m=qxJrqbz5sNlCrzJTOZjaJ0jHeaW077bX6ZxmV308jfg&s=LljqKCmXwjl4uzuLM_oB-jhjYV5xVGFpHPDomTZwKAU&e=> As per stable/2001 branch , the given change is checked-in around Oct 28 2019. df0191ead2cf39611714b6603cdc5bdddc445b57 is previous commit of b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? Yes (branch vpp 20.01) Thanks, Chetan Bhasin On Mon, Feb 17, 2020 at 5:33 PM Nitin Saxena mailto:nsax...@marvell.com>> wrote: Hi Damjan, >> if you read Chetan’s email bellow, you will see that this one is already >> excluded… Sorry I missed that part. After seeing diffs between stable/1908 and stable/2001, commit: b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897 is the only visible git commit in dpdk plugin which is playing with mempool buffers. If it does not solve the problem then I suspect problem lies outside dpdk plugin. I am guessing DPDK-19.08 is being used here with VPP-19.08 Hi Chetan, > > 3) I took previous commit of "vlib: don't use vector for keeping buffer > indices in the pool " ie "df0191ead2cf39611714b6603cdc5bdddc445b57" : > Everything looks fine with Buffers 537600. In which branch, Commit: df0191ead2cf39611714b6603cdc5bdddc445b57 is previous commit of b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? Thanks, Nitin > -Original Message- > From: Damjan Marion mailto:dmar...@me.com>> > Sent: Monday, February 17, 2020 3:47 PM > To: Nitin Saxena mailto:nsax...@marvell.com>> > Cc: chetan bhasin > mailto:chetan.bhasin...@gmail.com>>; > vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io> > Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter > > > Dear Nitin, > > if you read Chetan’s email bellow, you will see that this one is already > excluded… > > Also, it will not be easy to explain how this patch blows tx function in dpdk > mlx5 pmd… > > — > Damjan > > > On 17 Feb 2020, at 11:12, Nitin Saxena > > mailto:nsax...@marvell.com>> wrote: > > > > Hi Prashant/Chetan, > > > > I would try following change first to solve the problem in 1908 > > > > commit b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897b > > Author: Damjan Marion mailto:damar...@cisco.com>> > > Date: Tue Mar 12 18:14:15 2019 +0100 > > > > vlib: don't use vector for keeping buffer indices in > > > > Type: refactor > > > > Change-Id: I72221b97d7e0bf5c93e20bbda4473ca67bfcdeb4 > > Signed-off-by: Damjan Marion > > damar...@cisco.com<mailto:damar...@cisco.com> > > > > You can also try copying src/plugins/dpdk/buffer.c from stable/2001 > branch to stable/1908 > > > > Thanks, > > Nitin > > > > From: vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io> > > mailto:vpp-dev@lists.fd.io>> On Behalf Of Damjan > Marion via > Lists.Fd.Io<https://urldefense.proofpoint.com/v2/url?u=http-3A__Lists.Fd.Io&d=DwMFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=S4H7jibYAtA5YOvfL3IkGduCfk9LbZMPOAecQGDzWV0&m=qxJrqbz5sNlCrzJTOZjaJ0jHeaW077bX6ZxmV308jfg&s=ffS1Y8GHllzjueMUVW31gwrVEIK1HVSNTKk2yA-VjG8&e=> > > Sent: Monday, February 17, 2020 1:52 PM > > To: chetan bhasin > > mailto:chetan.bhasin...@gmail.com>> > > Cc: vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io> > > Subject: [EXT] Re
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
Hi Nitin,Damjan, For 40G *XL710* buffers : 537600 (500K+) 1) vpp 19.08 (sept 2019) : it worked with vpp 19.08 (sept release) after removing intel_iommu=on from Grub params. 2) stable/vpp2001(latest) : It worked even we have "intel_iommu=on" in Grub params On stable/vpp2001 , I found a check-in before which it did not work with " intel_iommu=on " as grub params, but after the below change-list it work even with grub params. commit 45495480c8165090722389b08075df06ccfcd7ef Author: Yulong Pei Date: Thu Oct 17 18:41:52 2019 +0800 vlib: linux: fix wrong iommu_group value issue when using dpdk-plugin Before above change in vpp 20.01 , when we bring up vpp with vfio-pci, vpp change /sys/module/vfio/parameters/enable_unsafe_noiommu_mode to "Y" , and we face issue with traffic but after the change sys file value remain as "N" in /sys/module/vfio/parameters/enable_unsafe_noiommu_mode and traffic works fine. As it is bare metal so we can remove intel_iommu=on from grub to make it work without any patches . Any suggestions? Regards, Chetan On Tue, Feb 18, 2020 at 1:07 PM Nitin Saxena wrote: > HI Chethan, > > > > Your packet trace shows that packet data is all 0 and that’s why you are > running into l3 mac mismatch. > > I am guessing something messed with IOMMU due to which translation is not > happening. Although packet length is correct. > > You can try out AVF plugin to iron out where problem exists, in dpdk > plugin or vlib > > > > Thanks, > > Nitin > > > > *From:* chetan bhasin > *Sent:* Tuesday, February 18, 2020 12:50 PM > *To:* me > *Cc:* Nitin Saxena ; vpp-dev > *Subject:* Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter > > > > Hi, > > One more finding related to intel nic and number of buffers (537600) > > > > vpp branch > > driver > > card > > buffers > > Traffic > > Err > > stable/1908 > > uio_pci_genric > > X722(10G) > > 537600 > > Working > > > > *stable/1908* > > *vfio-pci* > > *XL710(40G)* > > *537600 * > > *Not Working* > > *l3 mac mismatch* > > stable/2001 > > uio_pci_genric > > X722(10G) > > 537600 > > Working > > > > stable/2001 > > vfio-pci > > XL710(40G) > > 537600 > > Working > > > > > > > > Thanks, > > Chetan > > > > On Mon, Feb 17, 2020 at 7:17 PM chetan bhasin via Lists.Fd.Io > <https://urldefense.proofpoint.com/v2/url?u=http-3A__Lists.Fd.Io&d=DwMFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=S4H7jibYAtA5YOvfL3IkGduCfk9LbZMPOAecQGDzWV0&m=qxJrqbz5sNlCrzJTOZjaJ0jHeaW077bX6ZxmV308jfg&s=ffS1Y8GHllzjueMUVW31gwrVEIK1HVSNTKk2yA-VjG8&e=> > wrote: > > Hi Nitin, > > > > https://github.com/FDio/vpp/commits/stable/2001/src/vlib > <https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_FDio_vpp_commits_stable_2001_src_vlib&d=DwMFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=S4H7jibYAtA5YOvfL3IkGduCfk9LbZMPOAecQGDzWV0&m=qxJrqbz5sNlCrzJTOZjaJ0jHeaW077bX6ZxmV308jfg&s=LljqKCmXwjl4uzuLM_oB-jhjYV5xVGFpHPDomTZwKAU&e=> > > As per stable/2001 branch , the given change is checked-in around Oct 28 > 2019. > > > > df0191ead2cf39611714b6603cdc5bdddc445b57 is previous commit of > b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? > Yes (branch vpp 20.01) > > > > Thanks, > > Chetan Bhasin > > > > On Mon, Feb 17, 2020 at 5:33 PM Nitin Saxena wrote: > > Hi Damjan, > > >> if you read Chetan’s email bellow, you will see that this one is > already excluded… > Sorry I missed that part. After seeing diffs between stable/1908 and > stable/2001, commit: b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897 is the only > visible git commit in dpdk plugin which is playing with mempool buffers. If > it does not solve the problem then I suspect problem lies outside dpdk > plugin. I am guessing DPDK-19.08 is being used here with VPP-19.08 > > Hi Chetan, > > > 3) I took previous commit of "vlib: don't use vector for keeping > buffer > > indices in the pool " ie "df0191ead2cf39611714b6603cdc5bdddc445b57" : > > Everything looks fine with Buffers 537600. > In which branch, Commit: df0191ead2cf39611714b6603cdc5bdddc445b57 is > previous commit of b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? > > Thanks, > Nitin > > -Original Message- > > From: Damjan Marion > > Sent: Monday, February 17, 2020 3:47 PM > > To: Nitin Saxena > > Cc: chetan bhasin ; vpp-dev@lists.fd.io > > Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter > > > > > > Dear Nitin, > > > > if
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
> On 21 Feb 2020, at 10:31, chetan bhasin wrote: > > Hi Nitin,Damjan, > > For 40G XL710 buffers : 537600 (500K+) > 1) vpp 19.08 (sept 2019) : it worked with vpp 19.08 (sept release) after > removing intel_iommu=on from Grub params. > 2) stable/vpp2001(latest) : It worked even we have "intel_iommu=on" in Grub > params > > > On stable/vpp2001 , I found a check-in before which it did not work with " > intel_iommu=on " as grub params, but after the below change-list it work even > with grub params. > commit 45495480c8165090722389b08075df06ccfcd7ef > Author: Yulong Pei mailto:yulong@intel.com>> > Date: Thu Oct 17 18:41:52 2019 +0800 > vlib: linux: fix wrong iommu_group value issue when using dpdk-plugin > > Before above change in vpp 20.01 , when we bring up vpp with vfio-pci, vpp > change /sys/module/vfio/parameters/enable_unsafe_noiommu_mode to "Y" , and > we face issue with traffic but after the change sys file value remain as > "N" in /sys/module/vfio/parameters/enable_unsafe_noiommu_mode and traffic > works fine. > > As it is bare metal so we can remove intel_iommu=on from grub to make it work > without any patches . Any suggestions? IOMMU gives you following: - protection and security - it prevents misbehaving NIC to read/write intentionally or unintentionally memory it is not supposed to access - VA -> PA translation If you are running bare-metal, single tenant security is probably not concern, but still it can protect NIC from doing something bad eventually because of driver issues. VA -> PA translation helps with performance, as driver doesn’t need to lookup for PA when submitting descriptors but this is not critical perf issue. So it is up to you to decide, work without IOMMU or patch your old VPP version…. > > Regards, > Chetan > > On Tue, Feb 18, 2020 at 1:07 PM Nitin Saxena <mailto:nsax...@marvell.com>> wrote: > HI Chethan, > > > > Your packet trace shows that packet data is all 0 and that’s why you are > running into l3 mac mismatch. > > I am guessing something messed with IOMMU due to which translation is not > happening. Although packet length is correct. > > You can try out AVF plugin to iron out where problem exists, in dpdk plugin > or vlib > > > > Thanks, > > Nitin > > > > From: chetan bhasin <mailto:chetan.bhasin...@gmail.com>> > Sent: Tuesday, February 18, 2020 12:50 PM > To: me mailto:chetan.bhasin...@gmail.com>> > Cc: Nitin Saxena mailto:nsax...@marvell.com>>; vpp-dev > mailto:vpp-dev@lists.fd.io>> > Subject: Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter > > > > Hi, > > One more finding related to intel nic and number of buffers (537600) > > > > vpp branch > > driver > > card > > buffers > > Traffic > > Err > > stable/1908 > > uio_pci_genric > > X722(10G) > > 537600 > > Working > > > > stable/1908 > > vfio-pci > > XL710(40G) > > 537600 > > Not Working > > l3 mac mismatch > > stable/2001 > > uio_pci_genric > > X722(10G) > > 537600 > > Working > > > > stable/2001 > > vfio-pci > > XL710(40G) > > 537600 > > Working > > > > > > > > Thanks, > > Chetan > > > > On Mon, Feb 17, 2020 at 7:17 PM chetan bhasin via Lists.Fd.Io > <https://urldefense.proofpoint.com/v2/url?u=http-3A__Lists.Fd.Io&d=DwMFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=S4H7jibYAtA5YOvfL3IkGduCfk9LbZMPOAecQGDzWV0&m=qxJrqbz5sNlCrzJTOZjaJ0jHeaW077bX6ZxmV308jfg&s=ffS1Y8GHllzjueMUVW31gwrVEIK1HVSNTKk2yA-VjG8&e=> > mailto:gmail@lists.fd.io>> > wrote: > > Hi Nitin, > > > > https://github.com/FDio/vpp/commits/stable/2001/src/vlib > <https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_FDio_vpp_commits_stable_2001_src_vlib&d=DwMFaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=S4H7jibYAtA5YOvfL3IkGduCfk9LbZMPOAecQGDzWV0&m=qxJrqbz5sNlCrzJTOZjaJ0jHeaW077bX6ZxmV308jfg&s=LljqKCmXwjl4uzuLM_oB-jhjYV5xVGFpHPDomTZwKAU&e=> > As per stable/2001 branch , the given change is checked-in around Oct 28 2019. > > > > df0191ead2cf39611714b6603cdc5bdddc445b57 is previous commit of > b6e8b1a7c8bf9f9fbd05cdc3c90111d9e7a6897? > Yes (branch vpp 20.01) > > > > Thanks, > > Chetan Bhasin > > > > On Mon, Feb 17, 2020 at 5:33 PM Nitin Saxena <mailto:nsax...@marvell.com>> wrote: > > Hi Damjan, > > >> if you read Chetan’s email bellow, you will see that this one is already &
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
Thanks a lot Damjan for quick response ! We will try latest stable/1908 that has the given patch. *With Mellanox Technologies MT27710 Family [ConnectX-4 Lx] :* 1) stable/vpp1908 : If we configure buffers (250k) and have 2048 huge pages of 2MB (total 4GB), we see issue with traffic. "l3 mac mismatch" 2) stable/vpp1908 :If we configure 4 huge pages of 1GB via grub parameters , vpp works even with 400K buffers. Could you please guide us what's the best approach here ? For point 1) we see following logs in one of the vpp thread - #5 0x7f3375afbae2 in rte_vlog (level=, logtype=77, format=0x7f3376768df8 *"net_mlx5: port %u unable to find virtually contiguous chunk for address (%p). rte_memseg_contig_walk() failed.\n%.0s",* ap=ap@entry=0x7f3379c4fac8) at /nfs-bfs/workspace/gkeown/integra/mainline/ngp/mainline/third-party/vpp/vanilla_1908/build-root/build-vpp-native/external/dpdk-19.05/lib/librte_eal/common/eal_common_log.c:427 #6 0x7f3375ab2c12 in rte_log (level=level@entry=5, logtype=, format=format@entry=0x7f3376768df8 "net_mlx5: port %u unable to find virtually contiguous chunk for address (%p). rte_memseg_contig_walk() failed.\n%.0s") at /nfs-bfs/workspace/gkeown/integra/mainline/ngp/mainline/third-party/vpp/vanilla_1908/build-root/build-vpp-native/external/dpdk-19.05/lib/librte_eal/common/eal_common_log.c:443 #7 0x7f3375dc47fa in mlx5_mr_create_primary (dev=dev@entry=0x7f3376e9d940 , entry=entry@entry=0x7ef5c00d02ca, addr=addr@entry=69384463936) at /nfs-bfs/workspace/gkeown/integra/mainline/ngp/mainline/third-party/vpp/vanilla_1908/build-root/build-vpp-native/external/dpdk-19.05/drivers/net/mlx5/mlx5_mr.c:627 Thanks, Chetan On Fri, Feb 21, 2020 at 3:13 PM Damjan Marion wrote: > > > On 21 Feb 2020, at 10:31, chetan bhasin > wrote: > > Hi Nitin,Damjan, > > For 40G *XL710* buffers : 537600 (500K+) > 1) vpp 19.08 (sept 2019) : it worked with vpp 19.08 (sept release) after > removing intel_iommu=on from Grub params. > 2) stable/vpp2001(latest) : It worked even we have "intel_iommu=on" in > Grub params > > > On stable/vpp2001 , I found a check-in before which it did not work with " > intel_iommu=on " as grub params, but after the below change-list it work > even with grub params. > commit 45495480c8165090722389b08075df06ccfcd7ef > Author: Yulong Pei > Date: Thu Oct 17 18:41:52 2019 +0800 > vlib: linux: fix wrong iommu_group value issue when using dpdk-plugin > > Before above change in vpp 20.01 , when we bring up vpp with vfio-pci, vpp > change /sys/module/vfio/parameters/enable_unsafe_noiommu_mode to "Y" , > and we face issue with traffic but after the change sys file value remain > as "N" in /sys/module/vfio/parameters/enable_unsafe_noiommu_mode and > traffic works fine. > > As it is bare metal so we can remove intel_iommu=on from grub to make it > work without any patches . Any suggestions? > > > IOMMU gives you following: > - protection and security - it prevents misbehaving NIC to read/write > intentionally or unintentionally memory it is not supposed to access > - VA -> PA translation > > If you are running bare-metal, single tenant security is probably not > concern, but still it can protect NIC from doing something bad eventually > because of driver issues. > VA -> PA translation helps with performance, as driver doesn’t need to > lookup for PA when submitting descriptors but this is not critical perf > issue. > > So it is up to you to decide, work without IOMMU or patch your old VPP > version…. > > > Regards, > Chetan > > On Tue, Feb 18, 2020 at 1:07 PM Nitin Saxena wrote: > >> HI Chethan, >> >> >> >> Your packet trace shows that packet data is all 0 and that’s why you are >> running into l3 mac mismatch. >> >> I am guessing something messed with IOMMU due to which translation is not >> happening. Although packet length is correct. >> >> You can try out AVF plugin to iron out where problem exists, in dpdk >> plugin or vlib >> >> >> >> Thanks, >> >> Nitin >> >> >> >> *From:* chetan bhasin >> *Sent:* Tuesday, February 18, 2020 12:50 PM >> *To:* me >> *Cc:* Nitin Saxena ; vpp-dev >> *Subject:* Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter >> >> >> >> Hi, >> >> One more finding related to intel nic and number of buffers (537600) >> >> >> >> vpp branch >> >> driver >> >> card >> >> buffers >> >> Traffic >> >> Err >> >> stable/1908 >> >> uio_pci_genric >> >&g
Re: [EXT] [vpp-dev] Regarding buffers-per-numa parameter
> On 21 Feb 2020, at 11:48, chetan bhasin wrote: > > Thanks a lot Damjan for quick response ! > > We will try latest stable/1908 that has the given patch. > > With Mellanox Technologies MT27710 Family [ConnectX-4 Lx] : > 1) stable/vpp1908 : If we configure buffers (250k) and have 2048 huge pages > of 2MB (total 4GB), we see issue with traffic. "l3 mac mismatch" > 2) stable/vpp1908 :If we configure 4 huge pages of 1GB via grub parameters , > vpp works even with 400K buffers. > > Could you please guide us what's the best approach here ? > > For point 1) we see following logs in one of the vpp thread - > > #5 0x7f3375afbae2 in rte_vlog (level=, logtype=77, > format=0x7f3376768df8 "net_mlx5: port %u unable to find virtually > contiguous chunk for address (%p). rte_memseg_contig_walk() failed.\n%.0s", > ap=ap@entry=0x7f3379c4fac8) > at > /nfs-bfs/workspace/gkeown/integra/mainline/ngp/mainline/third-party/vpp/vanilla_1908/build-root/build-vpp-native/external/dpdk-19.05/lib/librte_eal/common/eal_common_log.c:427 > #6 0x7f3375ab2c12 in rte_log (level=level@entry=5, logtype= out>, > format=format@entry=0x7f3376768df8 "net_mlx5: port %u unable to find > virtually contiguous chunk for address (%p). rte_memseg_contig_walk() > failed.\n%.0s") > at > /nfs-bfs/workspace/gkeown/integra/mainline/ngp/mainline/third-party/vpp/vanilla_1908/build-root/build-vpp-native/external/dpdk-19.05/lib/librte_eal/common/eal_common_log.c:443 > #7 0x7f3375dc47fa in mlx5_mr_create_primary > (dev=dev@entry=0x7f3376e9d940 , > entry=entry@entry=0x7ef5c00d02ca, addr=addr@entry=69384463936) > at > /nfs-bfs/workspace/gkeown/integra/mainline/ngp/mainline/third-party/vpp/vanilla_1908/build-root/build-vpp-native/external/dpdk-19.05/drivers/net/mlx5/mlx5_mr.c:627 No idea about mlx5 PMD, it is a bit special and we encourage people to use rdma-core plugin instead currently performance is lower but we will have DirectVerbs code merged soon... — Damjan > > > Thanks, > Chetan > > > On Fri, Feb 21, 2020 at 3:13 PM Damjan Marion <mailto:dmar...@me.com>> wrote: > > >> On 21 Feb 2020, at 10:31, chetan bhasin > <mailto:chetan.bhasin...@gmail.com>> wrote: >> >> Hi Nitin,Damjan, >> >> For 40G XL710 buffers : 537600 (500K+) >> 1) vpp 19.08 (sept 2019) : it worked with vpp 19.08 (sept release) after >> removing intel_iommu=on from Grub params. >> 2) stable/vpp2001(latest) : It worked even we have "intel_iommu=on" in Grub >> params >> >> >> On stable/vpp2001 , I found a check-in before which it did not work with " >> intel_iommu=on " as grub params, but after the below change-list it work >> even with grub params. >> commit 45495480c8165090722389b08075df06ccfcd7ef >> Author: Yulong Pei mailto:yulong@intel.com>> >> Date: Thu Oct 17 18:41:52 2019 +0800 >> vlib: linux: fix wrong iommu_group value issue when using dpdk-plugin >> >> Before above change in vpp 20.01 , when we bring up vpp with vfio-pci, vpp >> change /sys/module/vfio/parameters/enable_unsafe_noiommu_mode to "Y" , and >> we face issue with traffic but after the change sys file value remain as >> "N" in /sys/module/vfio/parameters/enable_unsafe_noiommu_mode and traffic >> works fine. >> >> As it is bare metal so we can remove intel_iommu=on from grub to make it >> work without any patches . Any suggestions? > > IOMMU gives you following: > - protection and security - it prevents misbehaving NIC to read/write > intentionally or unintentionally memory it is not supposed to access > - VA -> PA translation > > If you are running bare-metal, single tenant security is probably not > concern, but still it can protect NIC from doing something bad eventually > because of driver issues. > VA -> PA translation helps with performance, as driver doesn’t need to lookup > for PA when submitting descriptors but this is not critical perf issue. > > So it is up to you to decide, work without IOMMU or patch your old VPP > version…. > >> >> Regards, >> Chetan >> >> On Tue, Feb 18, 2020 at 1:07 PM Nitin Saxena > <mailto:nsax...@marvell.com>> wrote: >> HI Chethan, >> >> >> >> Your packet trace shows that packet data is all 0 and that’s why you are >> running into l3 mac mismatch. >> >> I am guessing something messed with IOMMU due to which translation is not >> happening. Although packet length is correct. >> >> You can try out AVF plugin to iron out w