Age | Commit message (Collapse) | Author | Files | Lines |
|
Add a function to set the dynamic queue limit minimum value.
Some specific drivers might have legitimate reasons to configure
dql.min_limit to a given value. Typically, this is the case when the
PDU of the protocol is smaller than the packet size to used to
carry those frames to the device.
Concrete example: a CAN (Control Area Network) device with an USB 2.0
interface. The PDU of classical CAN protocol are roughly 16 bytes but
the USB packet size (which is used to carry the CAN frames to the
device) might be up to 512 bytes. Wen small traffic burst occurs, BQL
algorithm is not able to immediately adjust and this would result in
having to send many small USB packets (i.e packet of 16 bytes for each
CAN frame). Filling up the USB packet with CAN frames is relatively
fast (small latency issue) but the gain of not having to send several
small USB packets is huge (big throughput increase). In this case,
forcing dql.min_limit to a given value that would allow to stuff the
USB packet is always a win.
This function is to be used by network drivers which are able to prove
through a rationale and through empirical tests on several environment
(with other applications, heavy context switching, virtualization...),
that they constantly reach better performances with a specific
predefined dql.min_limit value with no noticeable latency impact.
Signed-off-by: Vincent Mailhol <mailhol.vincent@wanadoo.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
I was working on a syzbot issue, claiming one device could not be
dismantled because its refcount was -1
unregister_netdevice: waiting for sit0 to become free. Usage count = -1
It would be nice if syzbot could trigger a warning at the time
this reference count became negative.
This patch adds CONFIG_PCPU_DEV_REFCNT options which defaults
to per cpu variables (as before this patch) on SMP builds.
v2: free_dev label in alloc_netdev_mqs() is moved to avoid
a compiler warning (-Wunused-label), as reported
by kernel test robot <lkp@intel.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
call_gro_receive() is used to limit GRO recursion, but it works only
with callback pointers.
There's a combined version of call_gro_receive() + INDIRECT_CALL_2()
in <net/inet_common.h>, but it doesn't check for IPv6 modularity.
Add a similar new helper to cover both of these. It can and will be
used to avoid retpoline overhead when IP header lies behind another
offloaded proto.
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
If some source file includes <net/gro.h>, but doesn't include
<linux/indirect_call_wrapper.h>:
In file included from net/8021q/vlan_core.c:7:
./include/net/gro.h:6:1: warning: data definition has no type or storage class
6 | INDIRECT_CALLABLE_DECLARE(struct sk_buff *ipv6_gro_receive(struct list_head *,
| ^~~~~~~~~~~~~~~~~~~~~~~~~
./include/net/gro.h:6:1: error: type defaults to ‘int’ in declaration of ‘INDIRECT_CALLABLE_DECLARE’ [-Werror=implicit-int]
[...]
Include <linux/indirect_call_wrapper.h> directly. It's small and
won't pull lots of dependencies.
Also add some incomplete struct declarations to be fully stacked.
Fixes: 04f00ab2275f ("net/core: move gro function declarations to separate header ")
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The ocelot switches are a bit odd in that they do not have an STP state
to put the ports into. Instead, the forwarding configuration is delayed
from the typical port_bridge_join into stp_state_set, when the port enters
the BR_STATE_FORWARDING state.
I can only guess that the implementation of this quirk is the reason that
led to the simplification of the driver such that only one bridge could
be offloaded at a time.
We can simplify the data structures somewhat, and introduce a per-port
bridge device pointer and STP state, similar to how the LAG offload
works now (there we have a per-port bonding device pointer and TX
enabled state). This allows offloading multiple bridges with relative
ease, while still keeping in place the quirk to delay the programming of
the PGIDs.
We actually need this change now because we need to remove the bogus
restriction from ocelot_bridge_stp_state_set that ocelot->bridge_mask
needs to contain BIT(port), otherwise that function is a no-op.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Intel mGbE variant implemented in EHL and TGL can be set to select
different clock frequency based on GPO bits in MAC_GPIO_STATUS register.
We introduce a new "void (*ptp_clk_freq_config)(void *priv)" in platform
data so that if a platform is required to configure the frequency of clock
source, in this case Intel mGBE does, the platform-specific configuration
of the PTP clock setting is done when stmmac_ptp_register() is called.
Signed-off-by: Wong, Vee Khee <vee.khee.wong@intel.com>
Signed-off-by: Voon Weifeng <weifeng.voon@intel.com>
Co-developed-by: Ong Boon Leong <boon.leong.ong@intel.com>
Signed-off-by: Ong Boon Leong <boon.leong.ong@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
In order for a driver to be able to query a bridge for information
about itself, e.g. reading out port flags, it has to use a netdev that
is known to the bridge. In the simple case, that is just the netdev
representing the port, e.g. swp0 or swp1 in this example:
br0
/ \
swp0 swp1
But in the case of an offloaded lag, this will be the bond or team
interface, e.g. bond0 in this example:
br0
/
bond0
/ \
swp0 swp1
Add a helper that hides some of this complexity from the
drivers. Then, redefine dsa_port_offloads_bridge_port using the helper
to avoid double accounting of the set of possible offloaded uppers.
Signed-off-by: Tobias Waldekranz <tobias@waldekranz.com>
Reviewed-by: Vladimir Oltean <olteanv@gmail.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Move the xps maps (xps_cpus_map and xps_rxqs_map) to an array in
net_device. That will simplify a lot the code removing the need for lots
of if/else conditionals as the correct map will be available using its
offset in the array.
This should not modify the xps maps behaviour in any way.
Suggested-by: Alexander Duyck <alexander.duyck@gmail.com>
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Embed nr_ids (the number of cpu for the xps cpus map, and the number of
rxqs for the xps cpus map) in dev_maps. That will help not accessing out
of bound memory if those values change after dev_maps was allocated.
Suggested-by: Alexander Duyck <alexander.duyck@gmail.com>
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The xps cpus/rxqs map is accessed using dev->num_tc, which is used when
allocating the map. But later updates of dev->num_tc can lead to having
a mismatch between the maps and how they're accessed. In such cases the
map values do not make any sense and out of bound accesses can occur
(that can be easily seen using KASAN).
This patch aims at fixing this by embedding num_tc into the maps, using
the value at the time the map is created. This brings two improvements:
- The maps can be accessed using the embedded num_tc, so we know for
sure we won't have out of bound accesses.
- Checks can be made before accessing the maps so we know the values
retrieved will make sense.
We also update __netif_set_xps_queue to conditionally copy old maps from
dev_maps in the new one only if the number of traffic classes from both
maps match.
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Add support for legacy Broadcom tags, which are similar to DSA_TAG_PROTO_BRCM.
These tags are used on BCM5325, BCM5365 and BCM63xx switches.
Signed-off-by: Álvaro Fernández Rojas <noltari@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Add a function to handle the common pattern of printing a string into the
ethtool strings interface and incrementing the string pointer by the
ETH_GSTRING_LEN. Most of the drivers end up doing this and several have
implemented their own versions of this function so it would make sense to
consolidate on one implementation.
Signed-off-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux
Saeed Mahameed says:
====================
mlx5-updates-2021-03-16
mlx5 uplink representor netdev persistence.
Before this patchset we used to have separate netdevs for Native NIC mode
and Switchdev mode (uplink representor netdev), meaning that if user
switches modes between Native to Switchdev and vice versa, the driver
would cleanup the current netdev representor and create a new one for the
new mode, such behavior created an administrative nightmare for users,
where users need to be aware of such loss of both data path and control
path configurations, e.g. netdev attributes and arp/route tables,
where the later is more painful.
A simple solution for this is not to replace the netdev in first place
and use a single netdev to serve the uplink/physical port whether it is
in switchdev mode or native mode.
We already have different HW profiles for each netdev mode, in this series
we just replace the HW profile on the fly and we keep the same netdev
attached.
Refactoring: Some refactoring has been made to overcome some technical
difficulties
1) The netdev is created with the maximum amount of tx/rx queues to serve
the two profiles.
2) Some ndos are not supported in some modes, so we added a mode check for
such cases, e.g legacy sriov ndos must be blocked in switchdev mode.
3) Some mlx5 netdev private attributes need to be moved out of profiles
and kept in a persistent place, where the netdev is created
e.g devlink port and other global HW resources
4) The netdev devlink port is now always registered with the switch id
Implementation: the last three patches implement the mechanism now as the
netdev can be shared.
5) Don't recreate the netdev on switchdev mode changes
6) Prevent changing switchdev mode when some netdev operations
are active, mostly when TC rules are being processed.
This is required since the netdev is kept registered while switchdev mode
can be changed.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When switching modes between legacy and switchdev and back, do not
reload ethernet interfaces. just change the profile from nic profile
to uplink rep profile in switchdev mode.
Signed-off-by: Roi Dayan <roid@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
|
|
We re-use the native NIC port net device instance for the Uplink
representor, and the devlink port.
When changing profiles we reset the mlx5e priv but we should still
use the devlink port so move it to mlx5e resources.
Signed-off-by: Roi Dayan <roid@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
|
|
This is to separate between resources attributes and other
attributes we will want to use.
Signed-off-by: Roi Dayan <roid@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
|
|
Not all ndos check the present bit before calling the ndo and the driver
may want to check it. Sometimes the dev parameter passed as const so we
pass it to netif_device_present() as const.
Since netif_device_present() doesn't modify dev parameter anyway, declare
it as const.
Signed-off-by: Roi Dayan <roid@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
|
|
This reverts commit 0bb3262c0248d44aea3be31076f44beb82a7b120.
Breaks things on mips64/qemu
Reported-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Now when extracting frames from CPU the cpuq is not used anymore so
remove it.
Signed-off-by: Horatiu Vultur <horatiu.vultur@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch extends MRP support for Ocelot. It allows to have multiple
rings and when the node has the MRC role it forwards MRP Test frames in
HW. For MRM there is no change.
Signed-off-by: Horatiu Vultur <horatiu.vultur@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Add a new PGID that is used not to forward frames anywhere. It is used
by MRP to make sure that MRP Test frames will not reach CPU port.
Signed-off-by: Horatiu Vultur <horatiu.vultur@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Replace the use of C bit-fields in the rmnet_map_ul_csum_header
structure with a single two-byte (big endian) structure member,
and use masks to encode or get values within it. The content of
these fields can be accessed using simple bitwise AND and OR
operations on the (host byte order) value of the new structure
member.
Previously rmnet_map_ipv4_ul_csum_header() would update C bit-field
values in host byte order, then forcibly fix their byte order using
a combination of byte swap operations and types.
Instead, just compute the value that needs to go into the new
structure member and save it with a simple byte-order conversion.
Make similar simplifications in rmnet_map_ipv6_ul_csum_header().
Finally, in rmnet_map_checksum_uplink_packet() a set of assignments
zeroes every field in the upload checksum header. Replace that with
a single memset() operation.
Signed-off-by: Alex Elder <elder@linaro.org>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Replace the use of C bit-fields in the rmnet_map_dl_csum_trailer
structure with a single one-byte field, using constant field masks
to encode or get at embedded values.
Signed-off-by: Alex Elder <elder@linaro.org>
Reviewed-by: Bjorn Andersson <bjorn.andersson@linaro.org>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The actual layout of bits defined in C bit-fields (e.g. int foo : 3)
is implementation-defined. Structures defined in <linux/if_rmnet.h>
address this by specifying all bit-fields twice, to cover two
possible layouts.
I think this pattern is repetitive and noisy, and I find the whole
notion of compiler "bitfield endianness" to be non-intuitive.
Stop using C bit-fields for the command/data flag and the pad length
fields in the rmnet_map structure, and define a single-byte flags
field instead. Define a mask for the single-bit "command" flag,
and another mask for the encoded pad length. The content of both
fields can be accessed using a simple bitwise AND operation.
Signed-off-by: Alex Elder <elder@linaro.org>
Reviewed-by: Bjorn Andersson <bjorn.andersson@linaro.org>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The fields in the checksum trailer structure used for QMAP protocol
RX packets are all big-endian format, so define them that way.
It turns out these fields are never actually used by the RMNet code.
The start offset is always assumed to be zero, and the length is
taken from the other packet headers. So making these fields
explicitly big endian has no effect on the behavior of the code.
Signed-off-by: Alex Elder <elder@linaro.org>
Reviewed-by: Bjorn Andersson <bjorn.andersson@linaro.org>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Add basic support for the Marvell 88X2222 multi-speed ethernet
transceiver.
This PHY provides data transmission over fiber-optic as well as Twinax
copper links. The 88X2222 supports 2 ports of 10GBase-R and 1000Base-X
on the line-side interface. The host-side interface supports 4 ports of
10GBase-R, RXAUI, 1000Base-X and 2 ports of XAUI.
This driver, however, supports only XAUI on the host-side and
1000Base-X/10GBase-R on the line-side, for now. The SGMII is also
supported over 1000Base-X. Interrupts are not supported.
Internal registers access compliant with the Clause 45 specification.
Signed-off-by: Ivan Bornyakov <i.bornyakov@metrotek.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch intends to add platform level clocks management. Some
platforms may have their own special clocks, they also need to be
managed dynamically. If you want to manage such clocks, please implement
clks_config callback.
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Joakim Zhang <qiangqing.zhang@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Not all platform uses DT, so phylink_parse_mode() will skip in-band setup
of pl->supported and pl->link_config.advertising entirely. So, we add the
setting of ovr_an_inband flag to make it works for non-DT platform.
Signed-off-by: Ong Boon Leong <boon.leong.ong@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Certain platform does not support DT, so we make phylink_parse_mode() to
allow non-DT platform to use it to setup in-band AN advertising.
Signed-off-by: Ong Boon Leong <boon.leong.ong@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
XPCS IP supports C37 SGMII AN process and it is used in intel multi-GbE
controller as MAC-side SGMII.
Signed-off-by: Ong Boon Leong <boon.leong.ong@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The current implementation for XPCS is validated for C73, so we rename them
to have _c73 suffix and introduce a set of functions to use an_mode flag
to switch between C73 and C37 AN later.
Signed-off-by: Ong Boon Leong <boon.leong.ong@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This file has been effectively empty since 2.3.99-pre3 !
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Extend psample to report the following attributes when available:
* Output traffic class as a 16-bit value
* Output traffic class occupancy in bytes as a 64-bit value
* End-to-end latency of the packet in nanoseconds resolution
* Software timestamp in nanoseconds resolution (always available)
* Packet's protocol. Needed for packet dissection in user space (always
available)
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Currently, callers of psample_sample_packet() pass three metadata
attributes: Ingress port, egress port and truncated size. Subsequent
patches are going to add more attributes (e.g., egress queue occupancy),
which also need an indication whether they are valid or not.
Encapsulate packet metadata in a struct in order to keep the number of
arguments reasonable.
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
{,__}skb_header_pointer() helpers exist mainly for preventing
accesses-beyond-end of the linear data.
In the vast majorify of cases, they bail out on the first condition.
All code going after is mostly a fallback.
Mark the most common branch as 'likely' one to move it in-line.
Also, skb_copy_bits() can return negative values only when the input
arguments are invalid, e.g. offset is greater than skb->len. It can
be safely marked as 'unlikely' branch, assuming that hotpath code
provides sane input to not fail here.
These two bump the throughput with a single Flow Dissector pass on
every packet (e.g. with RPS or driver that uses eth_get_headlen())
on 20 Mbps per flow/core.
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
It's used only for flow dissection, which now takes constant data
pointers.
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Caught by the text editor. Fix it separately from the actual changes.
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Flow Dissector code never modifies the input buffer, neither skb nor
raw data.
Make 'data' argument const for all of the Flow dissector's functions.
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The function never modifies the input buffer, so 'data' argument
can be marked as const.
This implies one harmless cast-away.
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
BPF Flow dissection programs are read-only and don't touch input
buffers.
Mark 'data' and 'data_end' in struct bpf_flow_dissector as const
in preparation for global input constifying.
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Allow a policer action to enforce a rate-limit based on packets-per-second,
configurable using a packet-per-second rate and burst parameters.
e.g.
tc filter add dev tap1 parent ffff: u32 match \
u32 0 0 police pkts_rate 3000 pkts_burst 1000
Testing was unable to uncover a performance impact of this change on
existing features.
Signed-off-by: Baowen Zheng <baowen.zheng@corigine.com>
Signed-off-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: Louis Peens <louis.peens@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Allow flow_offload API to configure packet-per-second policing using rate
and burst parameters.
Dummy implementations of tcf_police_rate_pkt_ps() and
tcf_police_burst_pkt() are supplied which return 0, the unconfigured state.
This is to facilitate splitting the offload, driver, and TC code portion of
this feature into separate patches with the aim of providing a logical flow
for review. And the implementation of these helpers will be filled out by a
follow-up patch.
Signed-off-by: Xingfeng Hu <xingfeng.hu@corigine.com>
Signed-off-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: Louis Peens <louis.peens@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch defined a new struct mptcp_rm_list, the ids field was an
array of the removing address ids, the nr field was the valid number of
removing address ids in the array. The array size was definced as a new
macro MPTCP_RM_IDS_MAX. Changed the member rm_id of struct
mptcp_out_options to rm_list.
In mptcp_established_options_rm_addr, invoked mptcp_pm_rm_addr_signal to
get the rm_list. According the number of addresses in it, calculated
the padded RM_ADDR suboption length. And saved the ids array in struct
mptcp_out_options's rm_list member.
In mptcp_write_options, iterated each address id from struct
mptcp_out_options's rm_list member, set the invalid ones as TCPOPT_NOP,
then filled them into the RM_ADDR suboption.
Changed TCPOLEN_MPTCP_RM_ADDR_BASE from 4 to 3.
Signed-off-by: Geliang Tang <geliangtang@gmail.com>
Signed-off-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Fixes the following W=1 kernel build warning(s):
drivers/ptp/ptp_pch.c:193:6: warning: no previous prototype for ‘pch_ch_control_write’ [-Wmissing-prototypes]
drivers/ptp/ptp_pch.c:201:5: warning: no previous prototype for ‘pch_ch_event_read’ [-Wmissing-prototypes]
drivers/ptp/ptp_pch.c:212:6: warning: no previous prototype for ‘pch_ch_event_write’ [-Wmissing-prototypes]
drivers/ptp/ptp_pch.c:220:5: warning: no previous prototype for ‘pch_src_uuid_lo_read’ [-Wmissing-prototypes]
drivers/ptp/ptp_pch.c:231:5: warning: no previous prototype for ‘pch_src_uuid_hi_read’ [-Wmissing-prototypes]
drivers/ptp/ptp_pch.c:242:5: warning: no previous prototype for ‘pch_rx_snap_read’ [-Wmissing-prototypes]
drivers/ptp/ptp_pch.c:259:5: warning: no previous prototype for ‘pch_tx_snap_read’ [-Wmissing-prototypes]
drivers/ptp/ptp_pch.c:300:5: warning: no previous prototype for ‘pch_set_station_address’ [-Wmissing-prototypes]
Cc: Richard Cochran <richardcochran@gmail.com> (maintainer:PTP HARDWARE CLOCK SUPPORT)
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Jakub Kicinski <kuba@kernel.org>
Cc: Flavio Suligoi <f.suligoi@asem.it>
Cc: netdev@vger.kernel.org
Signed-off-by: Lee Jones <lee.jones@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Support matching on ICMPv4/6 type and code parameters using misc3
section of match parameters.
Signed-off-by: Maor Dickman <maord@nvidia.com>
Reviewed-by: Roi Dayan <roid@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
|
|
Jakub and Neil reported an increase of RTO timers whenever
TX completions are delayed a bit more (by increasing
NIC TX coalescing parameters)
Main issue is that TCP stack has a logic preventing a packet
being retransmit if the prior clone has not yet been
orphaned or freed.
This logic came with commit 1f3279ae0c13 ("tcp: avoid
retransmits of TCP packets hanging in host queues")
Thankfully, in the case skb_still_in_host_queue() detects
the initial clone is still in flight, it can use TSQ logic
that will eventually retry later, at the moment the clone
is freed or orphaned.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Neil Spring <ntspring@fb.com>
Reported-by: Jakub Kicinski <kuba@kernel.org>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The kernel periodically checks the idle time of nexthop buckets to
determine if they are idle and can be re-populated with a new nexthop.
When the resilient nexthop group is offloaded to hardware, the kernel
will not see activity on nexthop buckets unless it is reported from
hardware.
Add a function that can be periodically called by device drivers to
report activity on nexthop buckets after querying it from the underlying
device.
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Add a function that can be called by device drivers to set "offload" or
"trap" indication on nexthop buckets following nexthop notifications and
other changes such as a neighbour becoming invalid.
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Add data structures that will be used for in-kernel notifications about
addition / deletion of a resilient nexthop group and about changes to a
hash bucket within a resilient group.
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
At this moment, there is only one type of next-hop group: an mpath group,
which implements the hash-threshold algorithm.
To select a next hop, hash-threshold algorithm first assigns a range of
hashes to each next hop in the group, and then selects the next hop by
comparing the SKB hash with the individual ranges. When a next hop is
removed from the group, the ranges are recomputed, which leads to
reassignment of parts of hash space from one next hop to another. While
there will usually be some overlap between the previous and the new
distribution, some traffic flows change the next hop that they resolve to.
That causes problems e.g. as established TCP connections are reset, because
the traffic is forwarded to a server that is not familiar with the
connection.
Resilient hashing is a technique to address the above problem. Resilient
next-hop group has another layer of indirection between the group itself
and its constituent next hops: a hash table. The selection algorithm uses a
straightforward modulo operation to choose a hash bucket, and then reads
the next hop that this bucket contains, and forwards traffic there.
This indirection brings an important feature. In the hash-threshold
algorithm, the range of hashes associated with a next hop must be
continuous. With a hash table, mapping between the hash table buckets and
the individual next hops is arbitrary. Therefore when a next hop is deleted
the buckets that held it are simply reassigned to other next hops. When
weights of next hops in a group are altered, it may be possible to choose a
subset of buckets that are currently not used for forwarding traffic, and
use those to satisfy the new next-hop distribution demands, keeping the
"busy" buckets intact. This way, established flows are ideally kept being
forwarded to the same endpoints through the same paths as before the
next-hop group change.
In a nutshell, the algorithm works as follows. Each next hop has a number
of buckets that it wants to have, according to its weight and the number of
buckets in the hash table. In case of an event that might cause bucket
allocation change, the numbers for individual next hops are updated,
similarly to how ranges are updated for mpath group next hops. Following
that, a new "upkeep" algorithm runs, and for idle buckets that belong to a
next hop that is currently occupying more buckets than it wants (it is
"overweight"), it migrates the buckets to one of the next hops that has
fewer buckets than it wants (it is "underweight"). If, after this, there
are still underweight next hops, another upkeep run is scheduled to a
future time.
Chances are there are not enough "idle" buckets to satisfy the new demands.
The algorithm has knobs to select both what it means for a bucket to be
idle, and for whether and when to forcefully migrate buckets if there keeps
being an insufficient number of idle buckets.
There are three users of the resilient data structures.
- The forwarding code accesses them under RCU, and does not modify them
except for updating the time a selected bucket was last used.
- Netlink code, running under RTNL, which may modify the data.
- The delayed upkeep code, which may modify the data. This runs unlocked,
and mutual exclusion between the RTNL code and the delayed upkeep is
maintained by canceling the delayed work synchronously before the RTNL
code touches anything. Later it restarts the delayed work if necessary.
The RTNL code has to implement next-hop group replacement, next hop
removal, etc. For removal, the mpath code uses a neat trick of having a
backup next hop group structure, doing the necessary changes offline, and
then RCU-swapping them in. However, the hash tables for resilient hashing
are about an order of magnitude larger than the groups themselves (the size
might be e.g. 4K entries), and it was felt that keeping two of them is an
overkill. Both the primary next-hop group and the spare therefore use the
same resilient table, and writers are careful to keep all references valid
for the forwarding code. The hash table references next-hop group entries
from the next-hop group that is currently in the primary role (i.e. not
spare). During the transition from primary to spare, the table references a
mix of both the primary group and the spare. When a next hop is deleted,
the corresponding buckets are not set to NULL, but instead marked as empty,
so that the pointer is valid and can be used by the forwarding code. The
buckets are then migrated to a new next-hop group entry during upkeep. The
only times that the hash table is invalid is the very beginning and very
end of its lifetime. Between those points, it is always kept valid.
This patch introduces the core support code itself. It does not handle
notifications towards drivers, which are kept as if the group were an mpath
one. It does not handle netlink either. The only bit currently exposed to
user space is the new next-hop group type, and that is currently bounced.
There is therefore no way to actually access this code.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|