Saeed Mahameed says:
====================
mlx5-updates-2021-09-30
1) From Yevgeny Kliteynik:
This patch series deals with vport handling in SW steering.
For every vport, SW steering queries FW for this vport's properties,
such as RX/TX ICM addresses to be able to add this vport as dest action.
The following patches rework vport capabilities managements and add support
for Scalable Functions (SFs).
- Patch 1 fixes the vport number data type all over the DR code to 16 bits
in accordance with HW spec.
- Patch 2 replaces local SW steering WIRE_PORT macro with the existing
mlx5 define.
- Patch 3 adds missing query for vport 0 and and handles eswitch manager
capabilities for ECPF (BlueField in embedded CPU mode).
- Patch 4 fixes error messages for failure to obtain vport caps from
different locations in the code to have the same verbosity level and
similar wording.
- Patch 5 adds support for csum recalculation flow tables on SFs: it
implements these FTs management in XArray instead of the fixed size array,
thus adding support for csum recalculation table for any valid vport.
- Patch 6 is the main patch of this whole series: it refactors vports
capabilities handling and adds SFs support.
2) Minor and trivial updates and cleanups
* tag 'mlx5-updates-2021-09-30' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux:
net/mlx5e: Use array_size() helper
net/mlx5: Use struct_size() helper in kvzalloc()
net/mlx5: Use kvcalloc() instead of kvzalloc()
net/mlx5: Tolerate failures in debug features while driver load
net/mlx5: Warn for devlink reload when there are VFs alive
net/mlx5: DR, Add missing string for action type SAMPLER
net/mlx5: DR, init_next_match only if needed
net/mlx5: DR, Fix typo 'offeset' to 'offset'
net/mlx5: DR, Increase supported num of actions to 32
net/mlx5: DR, Add support for SF vports
net/mlx5: DR, Support csum recalculation flow table on SFs
net/mlx5: DR, Align error messages for failure to obtain vport caps
net/mlx5: DR, Add missing query for vport 0
net/mlx5: DR, Replace local WIRE_PORT macro with the existing MLX5_VPORT_UPLINK
net/mlx5: DR, Fix vport number data type to u16
====================
Link: https://lore.kernel.org/r/20210930232050.41779-1-saeed@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some architectures do not include uapi/asm/socket.h
Fixes: 2bb2f5fb21 ("net: add new socket option SO_RESERVE_MEM")
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts commit 4f42ad2011, reversing
changes made to ea2dd331bf.
These chanfges break the build when mctp is modular.
Signed-off-by: David S. Miller <davem@davemloft.net>
Each region has an independently configurable number of maximum
snapshots. This information is not reported to userspace, making it not
very discoverable. Fix this by adding a new
DEVLINK_ATTR_REGION_MAX_SNAPSHOST attribute which is used to report this
maximum.
Ex:
$devlink region
pci/0000:af:00.0/nvm-flash: size 10485760 snapshot [] max 1
pci/0000:af:00.0/device-caps: size 4096 snapshot [] max 10
pci/0000:af:00.1/nvm-flash: size 10485760 snapshot [] max 1
pci/0000:af:00.1/device-caps: size 4096 snapshot [] max 10
This information enables users to understand why a new region command
may fail due to having too many existing snapshots.
Reported-by: Gurucharan G <gurucharanx.g@intel.com> (A Contingent worker at Intel)
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jeremy Kerr says:
====================
MCTP kunit tests
This change adds some initial kunit tests for the MCTP core. We'll
expand the coverage in a future series, and augment with a few
selftests, but this establishes a baseline set of tests for now.
Thanks to the kunit folks for the framework!
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Add multi-packet route input tests, for message reassembly. These will
feed packets to be received by a bound socket, or dropped.
Signed-off-by: Jeremy Kerr <jk@codeconstruct.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a few tests for single-packet route inputs, testing the
mctp_route_input function.
Signed-off-by: Jeremy Kerr <jk@codeconstruct.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a few tests for the initial packet ingress through
mctp_pkttype_receive function; mainly packet header sanity checks. Full
input routing checks will be added as a separate change.
Signed-off-by: Jeremy Kerr <jk@codeconstruct.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This change adds the first kunit test for the mctp subsystem, and an
initial test for the fragmentation path.
We're adding tests under a new net/mctp/test/ directory.
Signed-off-by: Jeremy Kerr <jk@codeconstruct.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Saeed Mahameed says:
====================
mlx5 fixes 2021-09-30
This series introduces some fixes to mlx5 driver.
Please pull and let me know if there is any problem.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Make use of the struct_size() helper instead of an open-coded version,
in order to avoid any potential type mistakes or integer overflows that,
in the worse scenario, could lead to heap overflows.
Link: https://github.com/KSPP/linux/issues/160
Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
FW tracer and resource dump are debug features. Although failing to
initialize them may indicate an error, don't let this stop device
loading.
Signed-off-by: Aya Levin <ayal@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
When performing PF reload, VF can't communicate with FW until
it recovers and reloads as well.
Add a warning message when performing devlink reload while
VFs are still present. Thus, giving a notice of an unfavorable
behavior that might occur as a result of a consequential reloads
and cause interruption of VF recovery.
Signed-off-by: Lama Kayal <lkayal@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Allocate next steering table entry only if the remaining space requires to.
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Increase max supported number of actions in the same rule.
Signed-off-by: Hamdan Igbaria <hamdani@nvidia.com>
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Move all the vport capabilities to a separate struct and store vport caps
in XArray: SFs vport numbers will not come in the same range as VF vports,
so the existing implementation of vport capabilities as a fixed size array
is not suitable here.
XArray is a perfect fit: it is efficient when the indices used are densely
clustered. In addition to being a perfect fit as a dynamic data structure,
XArray also provides locking - it uses RCU and an internal spinlock to
synchronise access, so no additional protection needed.
Now except for the eswitch manager vport, all other vports (including the
uplink vport) are handled in the same way: when a new go-to-vport action
is added, this vport's caps are loaded from the xarray. If it is the first
time for this particular vport number, then its capabilities are queried
from FW and filled in into the appropriate entry.
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Reviewed-by: Muhammad Sammar <muhammads@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Implement csum recalculation flow tables in XAarray instead of a fixed
array, thus adding support for csum recalc table on any valid vport
number, which enables this support for SFs.
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Reviewed-by: Muhammad Sammar <muhammads@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Print similar error messages when an invalid vport number is
provided during action creation and during STEv0/1 creation.
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Reviewed-by: Muhammad Sammar <muhammads@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Currently, vport 0 capabilities are not set.
To fix this, we now querying both eswitch manager and vport 0.
Eswitch manager has an access to all the vports - for eswitch manager PF, all
vports can be referred as other vports. The exception is embedded CPU mode,
where there is vport 0 of ECPF and the PF vport 0.
Here is how vport are queried:
For Connect-X5/6:
PF vport (0) and vports 1..n: vport number, other = true
esw_manager is vport 0 (PF)
For BlueField (in embedded CPU mode):
ECPF vport: vport = 0, other = false
PF vport (0) and 1..n: vport number, other = true
esw_manager = vport 0 (ECPF)
Also, note that there's no need for other_vport function parameter
in dr_domain_query_vport - this value is now deduced locally in the
function.
Signed-off-by: Yuval Avnery <yuvalav@mellanox.com>
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Reviewed-by: Muhammad Sammar <muhammads@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
SW steering defines its own macro for uplink vport number.
Replace this macro with an already existing mlx5 macro.
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
According to the HW spec, vport number is a 16-bit value.
Fix vport usage all over the code to u16 data type.
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Reviewed-by: Muhammad Sammar <muhammads@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Pull networking fixes from Jakub Kicinski:
"Networking fixes, including fixes from mac80211, netfilter and bpf.
Current release - regressions:
- bpf, cgroup: assign cgroup in cgroup_sk_alloc when called from
interrupt
- mdio: revert mechanical patches which broke handling of optional
resources
- dev_addr_list: prevent address duplication
Previous releases - regressions:
- sctp: break out if skb_header_pointer returns NULL in sctp_rcv_ootb
(NULL deref)
- Revert "mac80211: do not use low data rates for data frames with no
ack flag", fixing broadcast transmissions
- mac80211: fix use-after-free in CCMP/GCMP RX
- netfilter: include zone id in tuple hash again, minimize collisions
- netfilter: nf_tables: unlink table before deleting it (race -> UAF)
- netfilter: log: work around missing softdep backend module
- mptcp: don't return sockets in foreign netns
- sched: flower: protect fl_walk() with rcu (race -> UAF)
- ixgbe: fix NULL pointer dereference in ixgbe_xdp_setup
- smsc95xx: fix stalled rx after link change
- enetc: fix the incorrect clearing of IF_MODE bits
- ipv4: fix rtnexthop len when RTA_FLOW is present
- dsa: mv88e6xxx: 6161: use correct MAX MTU config method for this
SKU
- e100: fix length calculation & buffer overrun in ethtool::get_regs
Previous releases - always broken:
- mac80211: fix using stale frag_tail skb pointer in A-MSDU tx
- mac80211: drop frames from invalid MAC address in ad-hoc mode
- af_unix: fix races in sk_peer_pid and sk_peer_cred accesses (race
-> UAF)
- bpf, x86: Fix bpf mapping of atomic fetch implementation
- bpf: handle return value of BPF_PROG_TYPE_STRUCT_OPS prog
- netfilter: ip6_tables: zero-initialize fragment offset
- mhi: fix error path in mhi_net_newlink
- af_unix: return errno instead of NULL in unix_create1() when over
the fs.file-max limit
Misc:
- bpf: exempt CAP_BPF from checks against bpf_jit_limit
- netfilter: conntrack: make max chain length random, prevent
guessing buckets by attackers
- netfilter: nf_nat_masquerade: make async masq_inet6_event handling
generic, defer conntrack walk to work queue (prevent hogging RTNL
lock)"
* tag 'net-5.15-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (77 commits)
af_unix: fix races in sk_peer_pid and sk_peer_cred accesses
net: stmmac: fix EEE init issue when paired with EEE capable PHYs
net: dev_addr_list: handle first address in __hw_addr_add_ex
net: sched: flower: protect fl_walk() with rcu
net: introduce and use lock_sock_fast_nested()
net: phy: bcm7xxx: Fixed indirect MMD operations
net: hns3: disable firmware compatible features when uninstall PF
net: hns3: fix always enable rx vlan filter problem after selftest
net: hns3: PF enable promisc for VF when mac table is overflow
net: hns3: fix show wrong state when add existing uc mac address
net: hns3: fix mixed flag HCLGE_FLAG_MQPRIO_ENABLE and HCLGE_FLAG_DCB_ENABLE
net: hns3: don't rollback when destroy mqprio fail
net: hns3: remove tc enable checking
net: hns3: do not allow call hns3_nic_net_open repeatedly
ixgbe: Fix NULL pointer dereference in ixgbe_xdp_setup
net: bridge: mcast: Associate the seqcount with its protecting lock.
net: mdio-ipq4019: Fix the error for an optional regs resource
net: hns3: fix hclge_dbg_dump_tm_pg() stack usage
net: mdio: mscc-miim: Fix the mdio controller
af_unix: Return errno instead of NULL in unix_create1().
...
TX-port-TS hijacks the PTP traffic to a specific HW TX-queue. This
conflicts with MQPRIO in channel mode, which specifies explicitly which
TC accepts the packet. This patch mutually excludes the above
configuration.
Fixes: ec60c4581b ("net/mlx5e: Support MQPRIO channel mode")
Signed-off-by: Aya Levin <ayal@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
When setting number of completion EQs of the SF, consider number of
online CPUs.
Without this consideration, when number of online cpus are less than 8,
unnecessary 8 completion EQs are allocated.
Fixes: c36326d38d ("net/mlx5: Round-Robin EQs over IRQs")
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
The maximum irq_index can be 2047, This means irq_name should have 4
characters reserve for the irq_index. Hence, increase it to 4.
Fixes: 3af26495a2 ("net/mlx5: Enlarge interrupt field in CREATE_EQ")
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
When in Real-time mode, HW clock is synced with the PTP daemon. Hence
driver should not re-calibrate the next pulse (via MTPPSE repetitive
events mechanism).
This patch arms repetitive events only in free-running mode.
Fixes: 432119de33 ("net/mlx5: Add cyc2time HW translation mode support")
Signed-off-by: Aya Levin <ayal@nvidia.com>
Reviewed-by: Eran Ben Elisha <eranbe@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Allow configuration of 1PPS start time only with time-stamp representing
a round second. Prior to this patch driver allowed setting of a
non-round-second which is not supported by the device. Avoid unexpected
behavior by restricting start-time configuration to a round-second.
Fixes: 4272f9b88d ("net/mlx5e: Change 1PPS out scheme")
Signed-off-by: Aya Levin <ayal@nvidia.com>
Reviewed-by: Eran Ben Elisha <eranbe@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
* Add netdev->tc_to_txq rollback in case of failure in
mlx5e_update_netdev_queues().
* Fix broken transition between the two modes:
MQPRIO DCB mode with tc==8, and MQPRIO channel mode.
* Disable MQPRIO channel mode if re-attaching with a different number
of channels.
* Improve code sharing.
Fixes: ec60c4581b ("net/mlx5e: Support MQPRIO channel mode")
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
The value for maximum number of channels is first calculated based
on the netdev's profile and current function resources (specifically,
number of MSIX vectors, which depends among other things on the number
of online cores in the system).
This value is then used to calculate the netdev's number of rxqs/txqs.
Once created (by alloc_etherdev_mqs), the number of netdev's rxqs/txqs
is constant and we must not exceed it.
To achieve this, keep the maximum number of channels in sync upon any
netdevice re-attach.
Use mlx5e_get_max_num_channels() for calculating the number of netdev's
rxqs/txqs. After netdev is created, use mlx5e_calc_max_nch() (which
coinsiders core device resources, profile, and netdev) to init or
update priv->max_nch.
Before this patch, the value of priv->max_nch might get out of sync,
mistakenly allowing accesses to out-of-bounds objects, which would
crash the system.
Track the number of channels stats structures used in a separate
field, as they are persistent to suspend/resume operations. All the
collected stats of every channel index that ever existed should be
preserved. They are reset only when struct mlx5e_priv is,
in mlx5e_priv_cleanup(), which is part of the profile changing flow.
There is no point anymore in blocking a profile change due to max_nch
mismatch in mlx5e_netdev_change_profile(). Remove the limitation.
Fixes: a1f240f180 ("net/mlx5e: Adjust to max number of channles when re-attaching")
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Aya Levin <ayal@nvidia.com>
Reviewed-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Currently in Rx data path IPsec crypto offloaded packets uses
csum_none flag, so checksum is handled by the stack, this naturally
have some performance/cpu utilization impact on such flows. As Nvidia
NIC starting from ConnectX6DX provides checksum complete value out of
the box also for such flows there is no sense in taking csum_none path,
furthermore the stack (xfrm) have the method to handle checksum complete
corrections for such flows i.e. IPsec trailer removal and consequently
checksum value adjustment.
Because of the above and in addition the ConnectX6DX is the first HW
which supports IPsec crypto offload then it is safe to report csum
complete for IPsec offloaded traffic.
Fixes: b2ac7541e3 ("net/mlx5e: IPsec: Add Connect-X IPsec Rx data path offload")
Signed-off-by: Raed Salem <raeds@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Pull gpio fixes from Bartosz Golaszewski:
"A single fix for the gpio-pca953x driver and two commits updating the
MAINTAINERS entries for Mun Yew Tham (GPIO specific) and myself
(treewide after a change in professional situation).
Summary:
- don't ignore I2C errors in gpio-pca953x
- update MAINTAINERS entries for Mun Yew Tham and myself"
* tag 'gpio-fixes-for-v5.15-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/brgl/linux:
MAINTAINERS: Update Mun Yew Tham as Altera Pio Driver maintainer
MAINTAINERS: update my email address
gpio: pca953x: do not ignore i2c errors
Pull rdma fixes from Jason Gunthorpe:
"Not much too exciting here, although two syzkaller bugs that seem to
have 9 lives may have finally been squashed.
Several core bugs and a batch of driver bug fixes:
- Fix compilation problems in qib and hfi1
- Do not corrupt the joined multicast group state when using
SEND_ONLY
- Several CMA bugs, a reference leak for listening and two syzkaller
crashers
- Various bug fixes for irdma
- Fix a Sleeping while atomic bug in usnic
- Properly sanitize kernel pointers in dmesg
- Two bugs in the 64b CQE support for hns"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
RDMA/hns: Add the check of the CQE size of the user space
RDMA/hns: Fix the size setting error when copying CQE in clean_cq()
RDMA/hfi1: Fix kernel pointer leak
RDMA/usnic: Lock VF with mutex instead of spinlock
RDMA/hns: Work around broken constant propagation in gcc 8
RDMA/cma: Ensure rdma_addr_cancel() happens before issuing more requests
RDMA/cma: Do not change route.addr.src_addr.ss_family
RDMA/irdma: Report correct WC error when there are MW bind errors
RDMA/irdma: Report correct WC error when transport retry counter is exceeded
RDMA/irdma: Validate number of CQ entries on create CQ
RDMA/irdma: Skip CQP ring during a reset
MAINTAINERS: Update Broadcom RDMA maintainers
RDMA/cma: Fix listener leak in rdma_cma_listen_on_all() failure
IB/cma: Do not send IGMP leaves for sendonly Multicast groups
IB/qib: Fix clang confusion of NULL pointer comparison
Jann Horn reported that SO_PEERCRED and SO_PEERGROUPS implementations
are racy, as af_unix can concurrently change sk_peer_pid and sk_peer_cred.
In order to fix this issue, this patch adds a new spinlock that needs
to be used whenever these fields are read or written.
Jann also pointed out that l2cap_sock_get_peer_pid_cb() is currently
reading sk->sk_peer_pid which makes no sense, as this field
is only possibly set by AF_UNIX sockets.
We will have to clean this in a separate patch.
This could be done by reverting b48596d1dc "Bluetooth: L2CAP: Add get_peer_pid callback"
or implementing what was truly expected.
Fixes: 109f6e39fa ("af_unix: Allow SO_PEERCRED to work across namespaces.")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Jann Horn <jannh@google.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Luiz Augusto von Dentz <luiz.von.dentz@intel.com>
Cc: Marcel Holtmann <marcel@holtmann.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Eric Dumazet says:
====================
net: snmp: minor optimizations
Fetching many SNMP counters on hosts with large number of cpus
takes a lot of time. mptcp still uses the old non-batched
fashion which is not cache friendly.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Using snmp_get_cpu_field_batch() allows for better cpu cache
utilization, especially on hosts with large number of cpus.
Also remove special handling when mptcp mibs where not yet
allocated.
I chose to use temporary storage on the stack to keep this patch simple.
We might in the future use the storage allocated in netstat_seq_show().
Combined with prior patch (inlining snmp_get_cpu_field)
time to fetch and output mptcp counters on a 256 cpu host [1]
goes from 75 usec to 16 usec.
[1] L1 cache size is 32KB, it is not big enough to hold all dataset.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This trivial function is called ~90,000 times on 256 cpus hosts,
when reading /proc/net/netstat. And this number keeps inflating.
Inlining it saves many cycles.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add counters for XDP REDIRECT success and failure. This brings the
redirect path in line with metrics gathered via the other XDP paths.
Signed-off-by: Joshua Roys <roysjosh@gmail.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When STMMAC is paired with Energy-Efficient Ethernet(EEE) capable PHY,
and the PHY is advertising EEE by default, we need to enable EEE on the
xPCS side too, instead of having user to manually trigger the enabling
config via ethtool.
Fixed this by adding xpcs_config_eee() call in stmmac_eee_init().
Fixes: 7617af3d1a ("net: pcs: Introducing support for DWC xpcs Energy Efficient Ethernet")
Cc: Michael Sit Wei Hong <michael.wei.hong.sit@intel.com>
Signed-off-by: Wong Vee Khee <vee.khee.wong@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Originally, ixgbe driver doesn't allow the mounting of xdpdrv if the
server is equipped with more than 64 cpus online. So it turns out that
the loading of xdpdrv causes the "NOMEM" failure.
Actually, we can adjust the algorithm and then make it work through
mapping the current cpu to some xdp ring with the protect of @tx_lock.
Here are some numbers before/after applying this patch with xdp-example
loaded on the eth0X:
As client (tx path):
Before After
TCP_STREAM send-64 734.14 714.20
TCP_STREAM send-128 1401.91 1395.05
TCP_STREAM send-512 5311.67 5292.84
TCP_STREAM send-1k 9277.40 9356.22 (not stable)
TCP_RR send-1 22559.75 21844.22
TCP_RR send-128 23169.54 22725.13
TCP_RR send-512 21670.91 21412.56
As server (rx path):
Before After
TCP_STREAM send-64 1416.49 1383.12
TCP_STREAM send-128 3141.49 3055.50
TCP_STREAM send-512 9488.73 9487.44
TCP_STREAM send-1k 9491.17 9356.22 (not stable)
TCP_RR send-1 23617.74 23601.60
...
Notice: the TCP_RR mode is unstable as the official document explains.
I tested many times with different parameters combined through netperf.
Though the result is not that accurate, I cannot see much influence on
this patch. The static key is places on the hot path, but it actually
shouldn't cause a huge regression theoretically.
Co-developed-by: Shujin Li <lishujin@kuaishou.com>
Signed-off-by: Shujin Li <lishujin@kuaishou.com>
Signed-off-by: Jason Xing <xingwanli@kuaishou.com>
Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Wei Wang says:
====================
net: add new socket option SO_RESERVE_MEM
This patch series introduces a new socket option SO_RESERVE_MEM.
This socket option provides a mechanism for users to reserve a certain
amount of memory for the socket to use. When this option is set, kernel
charges the user specified amount of memory to memcg, as well as
sk_forward_alloc. This amount of memory is not reclaimable and is
available in sk_forward_alloc for this socket.
With this socket option set, the networking stack spends less cycles
doing forward alloc and reclaim, which should lead to better system
performance, with the cost of an amount of pre-allocated and
unreclaimable memory, even under memory pressure.
With a tcp_stream test with 10 flows running on a simulated 100ms RTT
link, I can see the cycles spent in __sk_mem_raise_allocated() dropping
by ~0.02%. Not a whole lot, since we already have logic in
sk_mem_uncharge() to only reclaim 1MB when sk_forward_alloc has more
than 2MB free space. But on a system suffering memory pressure
constently, the savings should be more.
The first patch is the implementation of this socket option. The
following 2 patches change the tcp stack to make use of this reserved
memory when under memory pressure. This makes the tcp stack behavior
more flexible when under memory pressure, and provides a way for user to
control the distribution of the memory among its sockets.
With a TCP connection on a simulated 100ms RTT link, the default
throughput under memory pressure is ~500Kbps. With SO_RESERVE_MEM set to
100KB, the throughput under memory pressure goes up to ~3.5Mbps.
Change since v2:
- Added description for new field added in struct sock in patch 1
Change since v1:
- Added performance stats in cover letter and rebased
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
