Even though every switch model has a different way to access the VTU
Data bits, the base implementation of the VTU GetNext operation remains
the same: wait, write the first VID to iterate from, start the
operation, and read the next VID.
Move this generic implementation into global1_vtu.c and abstract the
handling of the start VID (similarly to the ATU GetNext implementation),
before introducing a new chip operation for specific chips.
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add helpers to access the VTU VID register in the global1_vtu.c file.
At the same time, move mv88e6xxx_g1_vtu_vid_write at the beginning of
_mv88e6xxx_vtu_loadpurge, which adds no functional changes but makes
future patches simpler.
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add helpers to access the VTU SID register in the global1_vtu.c file.
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add helpers to access the VTU FID register in the global1_vtu.c file.
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move the VTU flush operation to global1_vtu.c and call it from a
mv88e6xxx_vtu_setup helper, similarly to the ATU and PVT setup.
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move the helper functions to access the Global 1 VTU Operation register
to a new global1_vtu.c file, and get rid of the old underscore prefix
naming convention. This file will be extended will all VTU/STU related
code.
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
VLAN aware Marvell chips can program 802.1Q VLAN membership as well as
802.1s per VLAN Spanning Tree state using the same 3 VTU Data registers.
Some chips such as 88E6185 use different Data registers offsets for
ports state and membership, and program them in a single operation.
Other chips such as 88E6352 use the same register layout but program
them in distinct operations (an indirect table is used for 802.1s.)
Newer chips such as 88E6390 use the same offsets for both state and
membership in distinct operations, thus require multiple data accesses.
To correctly abstract this, split the "data" structure member of
mv88e6xxx_vtu_entry in two "state" and "member" members, before adding
VTU support for newer chips.
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some chips don't have a VLAN Table Unit, most of them do have a 4K
table, some others as the 88E6390 family has a 13th bit for the VID.
Add a new max_vid member to the info structure, used to check the
presence of a VTU as well as the value used to iterate from in VTU
GetNext operations.
This makes the MV88E6XXX_FLAG_VTU obsolete, thus remove it.
Signed-off-by: Vivien Didelot <vivien.didelot@savoirfairelinux.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a netdev is enslaved to a VRF master, its router interface (RIF)
needs to be destroyed (if exists) and a new one created using the
corresponding virtual router (VR).
>From the driver's perspective, the above is equivalent to an inetaddr
event sent for this netdev. Therefore, when a port netdev (or its
uppers) are enslaved to a VRF master, call the same function that
would've been called had a NETDEV_UP was sent for this netdev in the
inetaddr notification chain.
This patch also fixes a bug when a LAG netdev with an existing RIF is
enslaved to a VRF. Before this patch, each LAG port would drop the
reference on the RIF, but would re-join the same one (in the wrong VR)
soon after. With this patch, the corresponding RIF is first destroyed
and a new one is created using the correct VR.
Fixes: 7179eb5acd ("mlxsw: spectrum_router: Add support for VRFs")
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Reviewed-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Or says:
================
mlx5 neigh update
This series (whose code name is 'neigh update') from Hadar, enhances the
mlx5 TC IP tunnel offloads to deal with changes to tunnel destination
neighbours used in offloaded flows which involved encapsulation.
In order to keep track on the validity state of such neighbours, we register
a netevent notifier callback and act on NEIGH_UPDATE events: if a neighbour
becomes valid, offload the related flows to HW (the other way around when
neigh becomes invalid) and similarly when a neigh mac addresses changes.
Since this traffic is offloaded from the host OS, the neighbour for the IP
tunnel destination can mistakenly become STALE and deleted by the kernel
since its 'used' value wasn't changed. To address that, we proactively
update the neighbour 'used' value every DELAY_PROBE_TIME seconds, using
time stamps generated by the existing driver code for HW flow counters.
We use the DELAY_PROBE_TIME_UPDATE event to adjust the frequency of the updates.
Prior to the core of the series, there's a patch from Saeed that introduces an
extendable vport representor implementation scheme. It provides a separation
between the eswitch to the netdev related aspects of the representors.
We would like to thank Ido Schimmel and Ilya Lesokhin for their coaching && advice
through the long design and review cycles while we struggled to understand and
(hopefully correctly) implement the locking around the different driver flows(..) .
- Or.
=================
Misc Updates:
From Tariq:
Some small performance and trivial code optimization for mlx5 netdev driver
- Optimize poll ICOSQ completion queue
- Use prefetchw when a write is to follow
- Use u8 as ownership type in mlx5e_get_cqe()
From Eran:
- Disable LRO by default on specific setups
From Eli:
- Small cleanup for E-Switch to avoid redundant allocation
Thanks,
Saeed.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
iQEcBAABAgAGBQJZBeGlAAoJEEg/ir3gV/o+AuMH/2OfS+diXrTMU90tQGfucurg
cPuv2qSGzSzcyFOjMUKL/hgj+y8HcUNZVSqts/LPhUj5CbR8On2jo/3XZb405SMX
1eAffpD5JxDGm4xQCADwkDMRVOuWNbXlTceRMbD+8vDtboTDEZwbaXX20En6MIEW
qcQPX0wv7/u8VpxGlhch2RRvl7+zBhhGCLNpwzkE5K+RggIPguE+F2hLpm+SH9PD
AOHSuwGWnE40In8dZIPqsEnUsmsC9LmUQ17ip8S8dEtGDHMvwDelsKGZcRGMuwoz
wSxpVTWn5JaU0EwN0t+rYtERFH18SqboG3OFwqb27qtRns6ALXIx3Dj1bb9sVBo=
=3qOg
-----END PGP SIGNATURE-----
Merge tag 'mlx5-updates-2017-04-30' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux
mlx5-updates-2017-04-30
Or says:
================
mlx5 neigh update
This series (whose code name is 'neigh update') from Hadar, enhances the
mlx5 TC IP tunnel offloads to deal with changes to tunnel destination
neighbours used in offloaded flows which involved encapsulation.
In order to keep track on the validity state of such neighbours, we register
a netevent notifier callback and act on NEIGH_UPDATE events: if a neighbour
becomes valid, offload the related flows to HW (the other way around when
neigh becomes invalid) and similarly when a neigh mac addresses changes.
Since this traffic is offloaded from the host OS, the neighbour for the IP
tunnel destination can mistakenly become STALE and deleted by the kernel
since its 'used' value wasn't changed. To address that, we proactively
update the neighbour 'used' value every DELAY_PROBE_TIME seconds, using
time stamps generated by the existing driver code for HW flow counters.
We use the DELAY_PROBE_TIME_UPDATE event to adjust the frequency of the updates.
Prior to the core of the series, there's a patch from Saeed that introduces an
extendable vport representor implementation scheme. It provides a separation
between the eswitch to the netdev related aspects of the representors.
We would like to thank Ido Schimmel and Ilya Lesokhin for their coaching && advice
through the long design and review cycles while we struggled to understand and
(hopefully correctly) implement the locking around the different driver flows(..) .
- Or.
=================
Misc Updates:
From Tariq:
Some small performance and trivial code optimization for mlx5 netdev driver
- Optimize poll ICOSQ completion queue
- Use prefetchw when a write is to follow
- Use u8 as ownership type in mlx5e_get_cqe()
From Eran:
- Disable LRO by default on specific setups
From Eli:
- Small cleanup for E-Switch to avoid redundant allocation
Signed-off-by: David S. Miller <davem@davemloft.net>
After removing the PTP related initialization from slowpath start,
the remaining PTT entry is required only in case CONFIG_RFS_ACCEL is set.
Otherwise, it leads to a warning due to it being unused.
Fixes: d179bd1699 ("qed: Acquire/release ptt_ptp lock when enabling/disabling PTP")
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Output to the RDMA driver whether DPM mode is enabled or disabled in
the HW and if so what is the number of WIDs it supports
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When calculating doorbell BAR partitioning round up the number of
CPUs to the nearest power of 2 so the size of the DPI (per user
section) configured in the hardware will be stored properly and
not truncated.
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add mechanism to verify RoCE resources are released prior to freeing the
bitmaps. If this is not the case, print what resources were not released.
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If the posting of the ramrod for the purpose of TID deregistration
fails, abort the deregistration operation without using the FW's
return code.
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The internal RoCE SQE QP state isn't being used. Instead we mark the
QP as in regular error state.
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use time_before_eq for time comparison more safe and dealing
with timer wrapping to be future-proof.
Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Try to carry error messages to the user via the netlink extended
ack message attribute.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Try to carry error messages to the user via the netlink extended
ack message attribute.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch allows users to enable/disable internal TX and/or RX
clock delay for BCM5481x series PHYs so as to satisfy RGMII timing
specifications.
On a particular platform, whether TX and/or RX clock delay is required
depends on how PHY connected to the MAC IP. This requirement can be
specified through "phy-mode" property in the platform device tree.
Signed-off-by: Abhishek Shah <abhishek.shah@broadcom.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
trivial fix to spelling mistakes in printk message.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Presumably we never hit this return, but static checkers complain that
we need to unlock so we may as well fix that.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Acked-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
My static checker complains that we're holding a mutex on this error
path. Let's goto exit instead of returning directly.
Fixes: b0bccb69eb ("qed: Change locking scheme for VF channel")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Acked-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the hip06 and hip07 SoCs, phy connect to mdio bus.The mdio
module is probed with module_init, and, as such,
is not guaranteed to probe before the HNS driver. So we need
to support deferred probe.
We check for probe deferral in the mac init, so we not init DSAF
when there is no mdio, and free all resource, to later learn that
we need to defer the probe.
Signed-off-by: lipeng <lipeng321@huawei.com>
Reviewed-by: Yisen Zhuang <yisen.zhuang@huawei.com>
Reviewed-by: Matthias Brugger <mbrugger@suse.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the hip06 and hip07 SoCs, the interrupt lines from the
DSAF controllers are connected to mbigen hw module.
The mbigen module is probed with module_init, and, as such,
is not guaranteed to probe before the HNS driver. So we need
to support deferred probe.
Signed-off-by: lipeng <lipeng321@huawei.com>
Reviewed-by: Yisen Zhuang <yisen.zhuang@huawei.com>
Reviewed-by: Matthias Brugger <mbrugger@suse.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For legacy reasons NFP FW may be compiled to DMA packets to a constant
offset into the buffer and use the space before it for metadata. This
ensures that packets data always start at a certain offset regardless of
the amount of preceding metadata.
If rx offset is set to 0 there may still be up to 64 bytes of metadata
but metadata will start at the beginning of the buffer, instead of:
data_start_offset = rx_offset - meta_len
Even though we make the buffers larger to accommodate up to 64 bytes of
metadata, if there is only N bytes of metadata, we will end up with
N bytes of headroom and 64 - N bytes of tailroom. Therefore we can't
rely on that space for XDP headroom. Make sure we always allocate
full 256 bytes. This, unfortunately, means we can't fit the headroom
on an u8 any more.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Right now the required Service Process ABI version is still tied
to max ID of known commands. For new NSP commands we are adding
we are checking if NSP version is recent enough on command-by-command
basis. The driver doesn't have to force the device to have the
very latest flash, anything newer than 0.8 should do.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reading TX queue indexes from the device memory on each interrupt
is expensive. It's doubly expensive with XDP running since we have
two TX rings to check there. If the software indexes indicate that
the TX queue is completely empty, however, we don't need to look at
the device completion index at all.
The queuing CPU is doing a wmb() before kicking the device TX so
we should be safe to assume on the CPU handling the completions will
never see old value of the software copy of the index.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
On the RX path we follow the "drop if allocation of replacement
buffer fails" rule. With XDP we extended that to the TX action,
so if XDP prog returned TX but allocation of replacement RX buffer
failed, we will drop the packet.
To improve our XDP TX performance extend the idea of rings being
always full to XDP TX rings. Pre-fill the XDP TX rings with RX
buffers, and when XDP prog returns TX action swap the RX buffer
with the next buffer from the TX ring.
XDP TX complete will no longer free the buffers but let them
sit on the TX ring and wait for swap with RX buffer, instead.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We will soon allocate RX buffers for caching on XDP TX rings.
The rx_ring parameter passed to nfp_net_rx_alloc_one() is not
actually used, remove it.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As Or points out in commit 423b3aecf2 ("net/mlx4: Change ENOTSUPP
to EOPNOTSUPP"), ENOTSUPP is NFS specific error. Replace it with
EOPNOTSUPP.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Avoid hashing the tx napi struct into napi_hash[], which is used for
busy polling receive queues.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Creating a geneve link with 'udpcsum' set results in a creation of link
for which UDP checksum will NOT be computed on outbound packets, as can
be seen below.
11: gen0: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN
link/ether c2:85:27:b6:b4:15 brd ff:ff:ff:ff:ff:ff promiscuity 0
geneve id 200 remote 192.168.13.1 dstport 6081 noudpcsum
Similarly, creating a link with 'noudpcsum' set results in a creation
of link for which UDP checksum will be computed on outbound packets.
Fixes: 9b4437a5b8 ("geneve: Unify LWT and netdev handling.")
Signed-off-by: Girish Moodalbail <girish.moodalbail@oracle.com>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
Acked-by: Lance Richardson <lrichard@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The message "Cannot bind port X, err=Y" creates only confusion. In metadata
based mode, failure of IPv6 socket creation is okay if IPv6 is disabled and
no error message should be printed. But when IPv6 tunnel was requested, such
failure is fatal. The vxlan_socket_create does not know when the error is
harmless and when it's not.
Instead of passing such information down to vxlan_socket_create, remove the
message completely. It's not useful. We propagate the error code up to the
user space and the port number comes from the user space. There's nothing in
the message that the process creating vxlan interface does not know.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When IPv6 is compiled but disabled at runtime, __vxlan_sock_add returns
-EAFNOSUPPORT. For metadata based tunnels, this causes failure of the whole
operation of bringing up the tunnel.
Ignore failure of IPv6 socket creation for metadata based tunnels caused by
IPv6 not being available.
Fixes: b1be00a6c3 ("vxlan: support both IPv4 and IPv6 sockets in a single vxlan device")
Signed-off-by: Jiri Benc <jbenc@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace pattern
int status;
...
status = func(...);
return status;
by
return func(...);
No functional change intented.
Signed-off-by: Andy Shevchenko <andy.shevchenko@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reuse bnx2x_null_format_ver() in functions where it's appropriated
instead of open coded variant.
Signed-off-by: Andy Shevchenko <andy.shevchenko@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use scnprintf() when printing version instead of custom open coded variants.
Signed-off-by: Andy Shevchenko <andy.shevchenko@gmail.com>
Acked-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since 83a77e9ec4, the phydev irq is explicitly set to PHY_POLL when
there is no pdata. It doesn't work on DT enabled platforms because the
phydev irq is already set by libphy before.
Fixes: 83a77e9ec4 ("net: macb: Added PCI wrapper for Platform Driver.")
Signed-off-by: Alexandre Belloni <alexandre.belloni@free-electrons.com>
Acked-by: Nicolas Ferre <nicolas.ferre@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jeff Kirsher says:
====================
40GbE Intel Wired LAN Driver Updates 2017-04-30
This series contains updates to i40e and i40evf only.
Jake provides majority of the changes in this series, starting with the
renaming of a flag to avoid confusion. Then renamed a variable to a
more meaningful name to clarify what is actually being done and to
reduce confusion. Amortizes the wait time when initializing or disabling
lots of VFs by using i40e_reset_all_vfs() and
i40e_vsi_stop_rings_no_wait(). Cleaned up a unnecessary delay since
pci_disable_sriov() already has its own delay, so need to add a additional
delay when removing VFs. Avoid using the same name flags for both
vsi->state and pf->state, to make code review easier and assist future
work to use the correct state field when checking bits. Use
DECLARE_BITMAP() to ensure that we always allocate enough space for flags.
Replace hw_disabled_flags with the new _AUTO_DISABLED flags, which are
more readable because we are not setting an *_ENABLED flag to
disable the feature.
Alex corrects a oversight where we were not reprogramming the ports
after a reset, which was causing us to lose all of the receive tunnel
offloads.
Arnd Bergmann moves the declaration of a local variable to avoid a
warning seen on architectures with larger pages about an unused variable.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
struct esw_mc_addr is a small struct that can be part of struct
mlx5_eswitch. Define it as a field and not as a pointer and save the
kzalloc call and then error flow handling.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
We will activate the HW LRO only on servers with PCI BW > MAX LINK BW,
or when PCI BW > 16Gbps. On other cases we do not want LRO by default as
LRO sessions might get timeout and add redundant software overhead.
Tested:
ethtool -k <ifs-name> | grep large-receive-offload
On systems with and without the limitations.
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Cc: kernel-team@fb.com
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
CQE ownership indication is as small as a single bit.
Use u8 to speedup the comparison.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Cc: kernel-team@fb.com
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
"prefetchw()" prefetches the cacheline for write. Use it for
skb->data, as soon we'll be copying the packet header there.
Performance:
Single-stream packet-rate tested with pktgen.
Packets are dropped in tc level to zoom into driver data-path.
Larger gain is expected for smaller packets, as less time
is spent on handling SKB fragments, making the path shorter
and the improvement more significant.
---------------------------------------------
packet size | before | after | gain |
64B | 4,113,306 | 4,778,720 | 16% |
1024B | 3,633,819 | 3,950,593 | 8.7% |
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Cc: kernel-team@fb.com
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
UMR operations are more frequent and important.
Check them first, and add a compiler branch predictor hint.
According to current design, ICOSQ CQ can contain at most one
pending CQE per napi. Poll function is optimized accordingly.
Performance:
Single-stream packet-rate tested with pktgen.
Packets are dropped in tc level to zoom into driver data-path.
Larger gain is expected for larger packet sizes, as BW is higher
and UMR posts are more frequent.
---------------------------------------------
packet size | before | after | gain |
64B | 4,092,370 | 4,113,306 | 0.5% |
1024B | 3,421,435 | 3,633,819 | 6.2% |
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Cc: kernel-team@fb.com
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
The user can change delay_first_probe_time parameter through sysctl.
Listen to NETEVENT_DELAY_PROBE_TIME_UPDATE notifications and update the
intervals for updating the neighbours 'used' value periodic task and
for flow HW counters query periodic task.
Both of the intervals will be update only in case the new delay prob
time value is lower the current interval.
Since the driver saves only one min interval value and not per device,
the users will be able to set lower interval value for updating
neighbour 'used' value periodic task but they won't be able to schedule
a higher interval for this periodic task.
The used interval for scheduling neighbour 'used' value periodic task is
the minimal delay prob time parameter ever seen by the driver.
Signed-off-by: Hadar Hen Zion <hadarh@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
When IP tunnel encapsulation rules are offloaded, the kernel can't see
the traffic of the offloaded flow. The neighbour for the IP tunnel
destination of the offloaded flow can mistakenly become STALE and
deleted by the kernel since its 'used' value wasn't changed.
To make sure that a neighbour which is used by the HW won't become
STALE, we proactively update the neighbour 'used' value every
DELAY_PROBE_TIME period, when packets were matched and counted by the HW
for one of the tunnel encap flows related to this neighbour.
The periodic task that updates the used neighbours is scheduled when a
tunnel encap rule is successfully offloaded into HW and keeps re-scheduling
itself as long as the representor's neighbours list isn't empty.
Add, remove, lookup and status change operations done over the
representor's neighbours list or the neighbour hash entry encaps list
are all serialized by RTNL lock.
Signed-off-by: Hadar Hen Zion <hadarh@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
In order to offload TC encap rules, the driver does a lookup for the IP
tunnel neighbour according to the output device and the destination IP
given by the user.
To keep tracking after the validity state of such neighbours, we keep
the neighbours information (pair of device pointer and destination IP)
in a hash table maintained at the relevant egress representor and
register to get NETEVENT_NEIGH_UPDATE events. When getting neighbour update
netevent, we search for a match among the cached neighbours entries used for
encapsulation.
In case the neighbour isn't valid, we can't offload the flow into the
HW. We cache the flow (requested matching and actions) in the driver and
offload the rule later, when the neighbour is resolved and becomes
valid.
When a flow is only cached in the driver and not offloaded into HW
yet, we use EAGAIN return value to mark it internally, the TC ndo still
returns success.
Listen to kernel neighbour update netevents to trace relevant neighbours
validity state:
1. If a neighbour becomes valid, offload the related rules to HW.
2. If the neighbour becomes invalid, remove the related rules from HW.
3. If the neighbour mac address was changed, update the encap header.
Remove all the offloaded rules using the old encap header from the HW
and insert new rules to HW with updated encap header.
Access to the neighbors hash table is protected by RTNL lock of its
caller or by the table's spinlock.
Details of the locking/synchronization among the different actions
applied on the neighbour table:
Add/remove operations - protected by RTNL lock of its caller (all TC
commands are protected by RTNL lock). Add and remove operations are
initiated only when the user inserts/removes a TC rule into/from the driver.
Lookup/remove operations - since the lookup operation is done from
netevent notifier block, RTNL lock can't be used (atomic context).
Use the table's spin lock to protect lookups from TC user removal operation.
bh is used since netevent can be called from a softirq context.
Lookup/add operations - The hash table access functions are taking
care of the protection between lookup and add operations.
When adding/removing encap headers and rules to/from the HW, RTNL lock
is used. It can happen when:
1. The user inserts/removes a TC rule into/from the driver (TC commands
are protected by RTNL lock of it's caller).
2. The driver gets neighbour notification event, which reports about
neighbour validity status change. Before adding/removing encap headers
and rules to/from the HW, RTNL lock is taken.
A neighbour hash table entry should be freed when its encap list is empty.
Since The neighbour update netevent notification schedules a neighbour
update work that uses the neighbour hash entry, it can't be freed
unconditionally when the encap list becomes empty during TC delete rule flow.
Use reference count to protect from freeing neighbour hash table entry
while it's still in use.
When the user asks to unregister a netdvice used by one of the neigbours,
neighbour removal notification is received. Then we take a reference on the
neighbour and don't free it until the relevant encap entries (and flows) are
marked as invalid (not offloaded) and removed from HW.
As long as the encap entry is still valid (checked under RTNL lock) we
can safely access the neighbour device saved on mlx5e_neigh struct.
Signed-off-by: Hadar Hen Zion <hadarh@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Vivien Didelot