| Age | Commit message (Collapse) | Author |
|---|
|
DMA engine will always write no more than dma_buf_sz bytes of a received
frame into a page buffer, the remaining spaces are unused or used by CPU
exclusively.
Setting page_pool_params.max_len to almost the full size of page(s) helps
nothing more, but wastes more CPU cycles on cache maintenance.
For a standard MTU of 1500, then dma_buf_sz is assigned to 1536, and this
patch brings ~16.9% driver performance improvement in a TCP RX
throughput test with iPerf tool on a single isolated Cortex-A65 CPU
core, from 2.43 Gbits/sec increased to 2.84 Gbits/sec.
Signed-off-by: Furong Xu <0x1207@gmail.com>
Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Reviewed-by: Yanteng Si <si.yanteng@linux.dev>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
Avoid memcpy in non-XDP RX path by marking all allocated SKBs to
be recycled in the upper network stack.
This patch brings ~11.5% driver performance improvement in a TCP RX
throughput test with iPerf tool on a single isolated Cortex-A65 CPU
core, from 2.18 Gbits/sec increased to 2.43 Gbits/sec.
Signed-off-by: Furong Xu <0x1207@gmail.com>
Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Reviewed-by: Larysa Zaremba <larysa.zaremba@intel.com>
Reviewed-by: Yanteng Si <si.yanteng@linux.dev>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
Tariq Toukan says:
====================
net/mlx5e: CT: Add support for hardware steering
This series start with one more HWS patch by Yevgeny, followed by
patches that add support for connection tracking in hardware steering
mode. It consists of:
- patch #2 hooks up the CT ops for the new mode in the right places.
- patch #3 moves a function into a common file, so it can be reused.
- patch #4 uses the HWS API to implement connection tracking.
The main advantage of hardware steering compared to software steering is
vastly improved performance when adding/removing/updating rules. Using
the T-Rex traffic generator to initiate multi-million UDP flows per
second, a kernel running with these patches was able to offload ~600K
unique UDP flows per second, a number around ~7x larger than software
steering was able to achieve on the same hardware (256-thread AMD EPYC,
512 GB RAM, ConnectX 7 b2b).
====================
Link: https://patch.msgid.link/20250114130646.1937192-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
This is modeled similar to how software steering works:
- a reference-counted matcher is maintained for each
combination of nat/no_nat x ipv4/ipv6 x tcp/udp/gre.
- adding a rule involves finding+referencing or creating a corresponding
matcher, then actually adding a rule.
- updating rules is implemented using the bwc_rule update API, which can
change a rule's actions without touching the match value.
By using a T-Rex traffic generator to initiate multi-million UDP flows
per second, a kernel running with these patches on the RX side was able
to offload ~600K flows per second, which is about ~7x larger than what
software steering could do on the same hardware (256-thread AMD EPYC,
512 GB RAM, ConnectX-7 b2b).
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Jianbo Liu <jianbol@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20250114130646.1937192-5-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
This function checks whether a flow_rule has the right flow dissector
keys and masks used for a connection tracking flow offload. It is
currently used locally by the tc_ct smfs module, but is about to be used
from another place, so this commit moves it to a better place, renames
it to mlx5e_tc_ct_is_valid_flow_rule and drops the unused fs argument.
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Jianbo Liu <jianbol@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20250114130646.1937192-4-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Connection tracking can offload tuple matches to the NIC either via
firmware commands (when the steering mode is dmfs or offload support is
disabled due to eswitch being set to legacy) or via software-managed
flow steering (smfs).
This commit adds stub operations for a third mode, hardware-managed flow
steering. This is enabled when both CONFIG_MLX5_TC_CT and
CONFIG_MLX5_HW_STEERING are enabled.
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Jianbo Liu <jianbol@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20250114130646.1937192-3-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When checking if the matcher size can be increased, check both
match and action RTCs. Also, consider the increasing step - check
that it won't cause the new matcher size to become unsupported.
Additionally, since we're using '+ 1' for action RTC size yet
again, define it as macro and use in all the required places.
Signed-off-by: Yevgeny Kliteynik <kliteyn@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20250114130646.1937192-2-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Eric Dumazet says:
====================
net: reduce RTNL pressure in unregister_netdevice()
One major source of RTNL contention resides in unregister_netdevice()
Due to RCU protection of various network structures, and
unregister_netdevice() being a synchronous function,
it is calling potentially slow functions while holding RTNL.
I think we can release RTNL in two points, so that three
slow functions are called while RTNL can be used
by other threads.
v1: https://lore.kernel.org/netdev/20250107130906.098fc8d6@kernel.org/T/#m398c95f5778e1ff70938e079d3c4c43c050ad2a6
====================
Link: https://patch.msgid.link/20250114205531.967841-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
One synchronize_net() call is currently done while holding RTNL.
This is source of RTNL contention in workloads adding and deleting
many network namespaces per second, because synchronize_rcu()
and synchronize_rcu_expedited() can use 60+ ms in some cases.
For cleanup_net() use, temporarily release RTNL
while calling the last synchronize_net().
This should be safe, because devices are no longer visible
to other threads after unlist_netdevice() call
and setting dev->reg_state to NETREG_UNREGISTERING.
In any case, the new netdev_lock() / netdev_unlock()
infrastructure that we are adding should allow
to fix potential issues, with a combination
of a per-device mutex and dev->reg_state awareness.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jesse Brandeburg <jbrandeburg@cloudflare.com>
Link: https://patch.msgid.link/20250114205531.967841-6-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Two synchronize_net() calls are currently done while holding RTNL.
This is source of RTNL contention in workloads adding and deleting
many network namespaces per second, because synchronize_rcu()
and synchronize_rcu_expedited() can use 60+ ms in some cases.
For cleanup_net() use, temporarily release RTNL
while calling the last synchronize_net().
This should be safe, because devices are no longer visible
to other threads at this point.
In any case, the new netdev_lock() / netdev_unlock()
infrastructure that we are adding should allow
to fix potential issues, with a combination
of a per-device mutex and dev->reg_state awareness.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jesse Brandeburg <jbrandeburg@cloudflare.com>
Link: https://patch.msgid.link/20250114205531.967841-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
flush_all_backlogs() is called from unregister_netdevice_many_notify()
as part of netdevice dismantles.
This is currently called under RTNL, and can last up to 50 ms
on busy hosts.
There is no reason to hold RTNL at this stage, if our caller
is cleanup_net() : netns are no more visible, devices
are in NETREG_UNREGISTERING state and no other thread
could mess our state while RTNL is temporarily released.
In order to provide isolation, this patch provides a separate
'net_todo_list' for cleanup_net().
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jesse Brandeburg <jbrandeburg@cloudflare.com>
Link: https://patch.msgid.link/20250114205531.967841-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
flush_all_backlogs() uses per-cpu and static data to hold its
temporary data, on the assumption it is called under RTNL
protection.
Following patch in the series will break this assumption.
Use instead a dynamically allocated piece of memory.
In the unlikely case the allocation fails,
use a boot-time allocated memory.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jesse Brandeburg <jbrandeburg@cloudflare.com>
Link: https://patch.msgid.link/20250114205531.967841-3-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
cleanup_net() is the single thread responsible
for netns dismantles, and a serious bottleneck.
Before we can get per-netns RTNL, make sure
all synchronize_net() called from this thread
are using rcu_synchronize_expedited().
v3: deal with CONFIG_NET_NS=n
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jesse Brandeburg <jbrandeburg@cloudflare.com>
Link: https://patch.msgid.link/20250114205531.967841-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Jakub Kicinski says:
====================
net: use netdev->lock to protect NAPI
We recently added a lock member to struct net_device, with a vague
plan to start using it to protect netdev-local state, removing
the need to take rtnl_lock for new configuration APIs.
Lay some groundwork and use this lock for protecting NAPI APIs.
v1: https://lore.kernel.org/20250114035118.110297-1-kuba@kernel.org
====================
Link: https://patch.msgid.link/20250115035319.559603-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
NAPI lifetime, visibility and config are all fully under
netdev_lock protection now.
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-12-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Protect the following members of netdev and napi by netdev_lock:
- defer_hard_irqs,
- gro_flush_timeout,
- irq_suspend_timeout.
The first two are written via sysfs (which this patch switches
to new lock), and netdev genl which holds both netdev and rtnl locks.
irq_suspend_timeout is only written by netdev genl.
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-11-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Take netdev_lock() in netif_napi_set_irq(). All NAPI "control fields"
are now protected by that lock (most of the other ones are set during
napi add/del). The napi_hash_node is fully protected by the hash
spin lock, but close enough for the kdoc...
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-10-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Now that NAPI instances can't come and go without holding
netdev->lock we can trivially switch from rtnl_lock() to
netdev_lock() for setting netdev->threaded via sysfs.
Note that since we do not lock netdev_lock around sysfs
calls in the core we don't have to "trylock" like we do
with rtnl_lock.
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-9-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
In prep for dropping rtnl_lock, start locking netdev->lock in netlink
genl ops. We need to be using netdev->up instead of flags & IFF_UP.
We can remove the RCU lock protection for the NAPI since NAPI list
is protected by netdev->lock already.
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-8-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Wrap napi_enable() / napi_disable() with netdev_lock().
Provide the "already locked" flavor of the API.
iavf needs the usual adjustment. A number of drivers call
napi_enable() under a spin lock, so they have to be modified
to take netdev_lock() first, then spin lock then call
napi_enable_locked().
Protecting napi_enable() implies that napi->napi_id is protected
by netdev_lock().
Acked-by: Francois Romieu <romieu@fr.zoreil.com> # via-velocity
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-7-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Hold netdev->lock when NAPIs are getting added or removed.
This will allow safe access to NAPI instances of a net_device
without rtnl_lock.
Create a family of helpers which assume the lock is already taken.
Switch iavf to them, as it makes extensive use of netdev->lock,
already.
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-6-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Some uAPI (netdev netlink) hide net_device's sub-objects while
the interface is down to ensure uniform behavior across drivers.
To remove the rtnl_lock dependency from those uAPIs we need a way
to safely tell if the device is down or up.
Add an indication of whether device is open or closed, protected
by netdev->lock. The semantics are the same as IFF_UP, but taking
netdev_lock around every write to ->flags would be a lot of code
churn.
We don't want to blanket the entire open / close path by netdev_lock,
because it will prevent us from applying it to specific structures -
core helpers won't be able to take that lock from any function
called by the drivers on open/close paths.
So the state of the flag is "pessimistic", as in it may report false
negatives, but never false positives.
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-5-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Add helpers for accessing netdevs under netdev_lock().
There's some careful handling needed to find the device and lock it
safely, without it getting unregistered, and without taking rtnl_lock
(the latter being the whole point of the new locking, after all).
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-4-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Protect writes to netdev->reg_state with netdev_lock().
From now on holding netdev_lock() is sufficient to prevent
the net_device from getting unregistered, so code which
wants to hold just a single netdev around no longer needs
to hold rtnl_lock.
We do not protect the NETREG_UNREGISTERED -> NETREG_RELEASED
transition. We'd need to move mutex_destroy(netdev->lock)
to .release, but the real reason is that trying to stop
the unregistration process mid-way would be unsafe / crazy.
Taking references on such devices is not safe, either.
So the intended semantics are to lock REGISTERED devices.
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250115035319.559603-3-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Add helpers for locking the netdev instance, use it in drivers
and the shaper code. This will make grepping for the lock usage
much easier, as we extend the lock to cover more fields.
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Link: https://patch.msgid.link/20250115035319.559603-2-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Currently, the driver is seriously broken with respect to the
hibernation (S4): after image restore the device is back into
IPC_MEM_EXEC_STAGE_BOOT (which AFAIK means bootloader stage) and needs
full re-launch of the rest of its firmware, but the driver restore
handler treats the device as merely sleeping and just sends it a
wake-up command.
This wake-up command times out but device nodes (/dev/wwan*) remain
accessible.
However attempting to use them causes the bootloader to crash and
enter IPC_MEM_EXEC_STAGE_CD_READY stage (which apparently means "a crash
dump is ready").
It seems that the device cannot be re-initialized from this crashed
stage without toggling some reset pin (on my test platform that's
apparently what the device _RST ACPI method does).
While it would theoretically be possible to rewrite the driver to tear
down the whole MUX / IPC layers on hibernation (so the bootloader does
not crash from improper access) and then re-launch the device on
restore this would require significant refactoring of the driver
(believe me, I've tried), since there are quite a few assumptions
hard-coded in the driver about the device never being partially
de-initialized (like channels other than devlink cannot be closed,
for example).
Probably this would also need some programming guide for this hardware.
Considering that the driver seems orphaned [1] and other people are
hitting this issue too [2] fix it by simply unbinding the PCI driver
before hibernation and re-binding it after restore, much like
USB_QUIRK_RESET_RESUME does for USB devices that exhibit a similar
problem.
Tested on XMM7360 in HP EliteBook 855 G7 both with s2idle (which uses
the existing suspend / resume handlers) and S4 (which uses the new code).
[1]: https://lore.kernel.org/all/c248f0b4-2114-4c61-905f-466a786bdebb@leemhuis.info/
[2]:
https://github.com/xmm7360/xmm7360-pci/issues/211#issuecomment-1804139413
Reviewed-by: Sergey Ryazanov <ryazanov.s.a@gmail.com>
Signed-off-by: Maciej S. Szmigiero <mail@maciej.szmigiero.name>
Link: https://patch.msgid.link/e60287ebdb0ab54c4075071b72568a40a75d0205.1736372610.git.mail@maciej.szmigiero.name
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queue
Tony Nguyen says:
====================
Intel Wired LAN Driver Updates 2025-01-08 (ice)
This series contains updates to ice driver only.
Przemek reworks implementation so that ice_init_hw() is called before
ice_adapter initialization. The motivation is to have ability to act
on the number of PFs in ice_adapter initialization. This is not done
here but the code is also a bit cleaner.
Michal adds priority to be considered when matching recipes for proper
differentiation.
Konrad adds devlink health reporting for firmware generated events.
R Sundar utilizes string helpers over open coded versions.
Jake adds implementation to utilize a lower latency interface to program
PHY timer when supported.
Additional information can be found on the original cover letter:
https://lore.kernel.org/intel-wired-lan/20241216145453.333745-1-anton.nadezhdin@intel.com/
Karol adds and allows for different PTP delay values to be used per pin.
* '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queue:
ice: Add in/out PTP pin delays
ice: implement low latency PHY timer updates
ice: check low latency PHY timer update firmware capability
ice: add lock to protect low latency interface
ice: rename TS_LL_READ* macros to REG_LL_PROXY_H_*
ice: use read_poll_timeout_atomic in ice_read_phy_tstamp_ll_e810
ice: use string choice helpers
ice: add fw and port health reporters
ice: add recipe priority check in search
ice: ice_probe: init ice_adapter after HW init
ice: minor: rename goto labels from err to unroll
ice: split ice_init_hw() out from ice_init_dev()
ice: c827: move wait for FW to ice_init_hw()
====================
Link: https://patch.msgid.link/20250115000844.714530-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Following fields of 'struct mr_mfc' can be updated
concurrently (no lock protection) from ip_mr_forward()
and ip6_mr_forward()
- bytes
- pkt
- wrong_if
- lastuse
They also can be read from other functions.
Convert bytes, pkt and wrong_if to atomic_long_t,
and use READ_ONCE()/WRITE_ONCE() for lastuse.
Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/20250114221049.1190631-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Taehee Yoo says:
====================
bnxt_en: implement tcp-data-split and thresh option
This series implements hds-thresh ethtool command.
This series also implements backend of tcp-data-split and
hds-thresh ethtool command for bnxt_en driver.
These ethtool commands are mandatory options for device memory TCP.
NICs that use the bnxt_en driver support tcp-data-split feature named
HDS(header-data-split).
But there is no implementation for the HDS to enable by ethtool.
Only getting the current HDS status is implemented and the HDS is just
automatically enabled only when either LRO, HW-GRO, or JUMBO is enabled.
The hds_threshold follows the rx-copybreak value but it wasn't
changeable.
Currently, bnxt_en driver enables tcp-data-split by default but not
always work.
There is hds_threshold value, which indicates that a packet size is
larger than this value, a packet will be split into header and data.
hds_threshold value has been 256, which is a default value of
rx-copybreak value too.
The rx-copybreak value hasn't been allowed to change so the
hds_threshold too.
This patchset decouples hds_threshold and rx-copybreak first.
and make tcp-data-split, rx-copybreak, and
hds-thresh configurable independently.
But the default configuration is the same.
The default value of rx-copybreak is 256 and default
hds-thresh is also 256.
The behavior of rx-copybreak will probably be changed in almost all
drivers. If HDS is not enabled, rx-copybreak copies both header and
payload from a page.
But if HDS is enabled, rx-copybreak copies only header from the first
page.
Due to this change, it may need to disable(set to 0) rx-copybreak when
the HDS is required.
There are several related options.
TPA(HW-GRO, LRO), JUMBO, jumbo_thresh(firmware command), and Aggregation
Ring.
The aggregation ring is fundamental to these all features.
When gro/lro/jumbo packets are received, NIC receives the first packet
from the normal ring.
follow packets come from the aggregation ring.
These features are working regardless of HDS.
If HDS is enabled, the first packet contains the header only, and the
following packets contain only payload.
So, HW-GRO/LRO is working regardless of HDS.
There is another threshold value, which is jumbo_thresh.
This is very similar to hds_thresh, but jumbo thresh doesn't split
header and data.
It just split the first and following data based on length.
When NIC receives 1500 sized packet, and jumbo_thresh is 256(default, but
follows rx-copybreak),
the first data is 256 and the following packet size is 1500-256.
Before this patch, at least if one of GRO, LRO, and JUMBO flags is
enabled, the Aggregation ring will be enabled.
If the Aggregation ring is enabled, both hds_threshold and
jumbo_thresh are set to the default value of rx-copybreak.
So, GRO, LRO, JUMBO frames, they larger than 256 bytes, they will
be split into header and data if the protocol is TCP or UDP.
for the other protocol, jumbo_thresh works instead of hds_thresh.
This means that tcp-data-split relies on the GRO, LRO, and JUMBO flags.
But by this patch, tcp-data-split no longer relies on these flags.
If the tcp-data-split is enabled, the Aggregation ring will be
enabled.
Also, hds_threshold no longer follows rx-copybreak value, it will
be set to the hds-thresh value by user-space, but the
default value is still 256.
If the protocol is TCP or UDP and the HDS is disabled and Aggregation
ring is enabled, a packet will be split into several pieces due to
jumbo_thresh.
When single buffer XDP is attached, tcp-data-split is automatically
disabled.
LRO, GRO, and JUMBO are tested with BCM57414, BCM57504 and the firmware
version is 230.0.157.0.
I couldn't find any specification about minimum and maximum value
of hds_threshold, but from my test result, it was about 0 ~ 1023.
It means, over 1023 sized packets will be split into header and data if
tcp-data-split is enabled regardless of hds_treshold value.
When hds_threshold is 1500 and received packet size is 1400, HDS should
not be activated, but it is activated.
The maximum value of hds-thresh value is 256 because it
has been working. It was decided very conservatively.
I checked out the tcp-data-split(HDS) works independently of GRO, LRO,
JUMBO.
Also, I checked out tcp-data-split should be disabled automatically
when XDP is attached and disallowed to enable it again while XDP is
attached. I tested ranged values from min to max for
hds-thresh and rx-copybreak, and it works.
hds-thresh from 0 to 256, and rx-copybreak 0 to 256.
When testing this patchset, I checked skb->data, skb->data_len, and
nr_frags values.
By this patchset, bnxt_en driver supports a force enable tcp-data-split,
but it doesn't support for disable tcp-data-split.
When tcp-data-split is explicitly enabled, HDS works always.
When tcp-data-split is unknown, it depends on the current
configuration of LRO/GRO/JUMBO.
1/10 patch adds a new hds_config member in the ethtool_netdev_state.
It indicates that what tcp-data-split value is really updated from
userspace.
So the driver can distinguish a passed tcp-data-split value is
came from user or driver itself.
2/10 patch adds hds-thresh command in the ethtool.
This threshold value indicates if a received packet size is larger
than this threshold, the packet's header and payload will be split.
Example:
# ethtool -G <interface name> hds-thresh <value>
This option can not be used when tcp-data-split is disabled or not
supported.
# ethtool -G enp14s0f0np0 tcp-data-split on hds-thresh 256
# ethtool -g enp14s0f0np0
Ring parameters for enp14s0f0np0:
Pre-set maximums:
...
Current hardware settings:
...
TCP data split: on
HDS thresh: 256
3/10, 4/10 add condition checks for devmem and ethtool.
If tcp-data-split is disabled or threshold value is not zero, setup of
devmem will be failed.
Also, tcp-data-split and hds-thresh will not be changed
while devmem is running.
5/10 add condition checks for netdev core.
It disallows setup single buffer XDP program when tcp-data-split is
enabled.
6/10 patch implements .{set, get}_tunable() in the bnxt_en.
The bnxt_en driver has been supporting the rx-copybreak feature but is
not configurable, Only the default rx-copybreak value has been working.
So, it changes the bnxt_en driver to be able to configure
the rx-copybreak value.
7/10 patch adds an implementation of tcp-data-split ethtool
command.
The HDS relies on the Aggregation ring, which is automatically enabled
when either LRO, GRO, or large mtu is configured.
So, if the Aggregation ring is enabled, HDS is automatically enabled by
it.
8/10 patch adds the implementation of hds-thresh logic
in the bnxt_en driver.
The default value is 256, which used to be the default rx-copybreak
value.
9/10 add HDS feature implementation for netdevsim.
HDS feature is not common so far. Only a few NICs support this feature.
There is no way to test HDS core-API unless we have proper hw NIC.
In order to test HDS core-API without hw NIC, netdevsim can be used.
It implements HDS control and data plane for netdevsim.
10/10 add selftest for HDS(tcp-data-split and HDS-thresh).
The tcp-data-split tests are the same with
`ethtool -G tcp-data-split <on | auto>`
HDS-thresh tests are same with `ethtool -G eth0 hds-thresh <0 - MAX>`
This series is tested with BCM57504 and netdevsim.
====================
Link: https://patch.msgid.link/20250114142852.3364986-1-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
HDS/HDS-thresh features were updated/implemented. so add some tests for
these features.
HDS tests are the same with `ethtool -G eth0 tcp-data-split <on | off |
auto >` but `auto` depends on driver specification.
So, it doesn't include `auto` case.
HDS-thresh tests are same with `ethtool -G eth0 hds-thresh <0 - MAX>`
It includes both 0 and MAX cases. It also includes exceed case, MAX + 1.
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Link: https://patch.msgid.link/20250114142852.3364986-11-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
HDS options(tcp-data-split, hds-thresh) have dependencies between other
features like XDP. Basic dependencies are checked in the core API.
netdevsim is very useful to check basic dependencies.
The default tcp-data-split mode is UNKNOWN but netdevsim driver
returns ENABLED when ethtool dumps tcp-data-split mode.
The default value of HDS threshold is 0 and the maximum value is 1024.
ethtool shows like this.
ethtool -g eni1np1
Ring parameters for eni1np1:
Pre-set maximums:
...
HDS thresh: 1024
Current hardware settings:
...
TCP data split: on
HDS thresh: 0
ethtool -G eni1np1 tcp-data-split on hds-thresh 1024
ethtool -g eni1np1
Ring parameters for eni1np1:
Pre-set maximums:
...
HDS thresh: 1024
Current hardware settings:
...
TCP data split: on
HDS thresh: 1024
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Link: https://patch.msgid.link/20250114142852.3364986-10-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The bnxt_en driver has configured the hds_threshold value automatically
when TPA is enabled based on the rx-copybreak default value.
Now the hds-thresh ethtool command is added, so it adds an
implementation of hds-thresh option.
Configuration of the hds-thresh is applied only when
the tcp-data-split is enabled. The default value of
hds-thresh is 256, which is the default value of
rx-copybreak, which used to be the hds_thresh value.
The maximum hds-thresh is 1023.
# Example:
# ethtool -G enp14s0f0np0 tcp-data-split on hds-thresh 256
# ethtool -g enp14s0f0np0
Ring parameters for enp14s0f0np0:
Pre-set maximums:
...
HDS thresh: 1023
Current hardware settings:
...
TCP data split: on
HDS thresh: 256
Tested-by: Stanislav Fomichev <sdf@fomichev.me>
Tested-by: Andy Gospodarek <gospo@broadcom.com>
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Reviewed-by: Michael Chan <michael.chan@broadcom.com>
Link: https://patch.msgid.link/20250114142852.3364986-9-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
NICs that uses bnxt_en driver supports tcp-data-split feature by the
name of HDS(header-data-split).
But there is no implementation for the HDS to enable by ethtool.
Only getting the current HDS status is implemented and The HDS is just
automatically enabled only when either LRO, HW-GRO, or JUMBO is enabled.
The hds_threshold follows rx-copybreak value. and it was unchangeable.
This implements `ethtool -G <interface name> tcp-data-split <value>`
command option.
The value can be <on> and <auto>.
The value is <auto> and one of LRO/GRO/JUMBO is enabled, HDS is
automatically enabled and all LRO/GRO/JUMBO are disabled, HDS is
automatically disabled.
HDS feature relies on the aggregation ring.
So, if HDS is enabled, the bnxt_en driver initializes the aggregation ring.
This is the reason why BNXT_FLAG_AGG_RINGS contains HDS condition.
Acked-by: Jakub Kicinski <kuba@kernel.org>
Tested-by: Stanislav Fomichev <sdf@fomichev.me>
Tested-by: Andy Gospodarek <gospo@broadcom.com>
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Reviewed-by: Michael Chan <michael.chan@broadcom.com>
Link: https://patch.msgid.link/20250114142852.3364986-8-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The bnxt_en driver supports rx-copybreak, but it couldn't be set by
userspace. Only the default value(256) has worked.
This patch makes the bnxt_en driver support following command.
`ethtool --set-tunable <devname> rx-copybreak <value> ` and
`ethtool --get-tunable <devname> rx-copybreak`.
By this patch, hds_threshol is set to the rx-copybreak value.
But it will be set by `ethtool -G eth0 hds-thresh N`
in the next patch.
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Reviewed-by: Brett Creeley <brett.creeley@amd.com>
Tested-by: Stanislav Fomichev <sdf@fomichev.me>
Tested-by: Andy Gospodarek <gospo@broadcom.com>
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Reviewed-by: Michael Chan <michael.chan@broadcom.com>
Link: https://patch.msgid.link/20250114142852.3364986-7-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When a single buffer XDP is attached, NIC should guarantee only single
page packets will be received.
tcp-data-split feature splits packets into header and payload. single
buffer XDP can't handle it properly.
So attaching single buffer XDP should be disallowed when tcp-data-split
is enabled.
Acked-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Link: https://patch.msgid.link/20250114142852.3364986-6-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
While the devmem is running, the tcp-data-split and
hds-thresh configuration should not be changed.
If user tries to change tcp-data-split and threshold value while the
devmem is running, it fails and shows extack message.
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Tested-by: Stanislav Fomichev <sdf@fomichev.me>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Link: https://patch.msgid.link/20250114142852.3364986-5-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
If driver doesn't support ring parameter or tcp-data-split configuration
is not sufficient, the devmem should not be set up.
Before setup the devmem, tcp-data-split should be ON and hds-thresh
value should be 0.
Tested-by: Stanislav Fomichev <sdf@fomichev.me>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Link: https://patch.msgid.link/20250114142852.3364986-4-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The hds-thresh option configures the threshold value of
the header-data-split.
If a received packet size is larger than this threshold value, a packet
will be split into header and payload.
The header indicates TCP and UDP header, but it depends on driver spec.
The bnxt_en driver supports HDS(Header-Data-Split) configuration at
FW level, affecting TCP and UDP too.
So, If hds-thresh is set, it affects UDP and TCP packets.
Example:
# ethtool -G <interface name> hds-thresh <value>
# ethtool -G enp14s0f0np0 tcp-data-split on hds-thresh 256
# ethtool -g enp14s0f0np0
Ring parameters for enp14s0f0np0:
Pre-set maximums:
...
HDS thresh: 1023
Current hardware settings:
...
TCP data split: on
HDS thresh: 256
The default/min/max values are not defined in the ethtool so the drivers
should define themself.
The 0 value means that all TCP/UDP packets' header and payload
will be split.
Tested-by: Stanislav Fomichev <sdf@fomichev.me>
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Link: https://patch.msgid.link/20250114142852.3364986-3-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When tcp-data-split is UNKNOWN mode, drivers arbitrarily handle it.
For example, bnxt_en driver automatically enables if at least one of
LRO/GRO/JUMBO is enabled.
If tcp-data-split is UNKNOWN and LRO is enabled, a driver returns
ENABLES of tcp-data-split, not UNKNOWN.
So, `ethtool -g eth0` shows tcp-data-split is enabled.
The problem is in the setting situation.
In the ethnl_set_rings(), it first calls get_ringparam() to get the
current driver's config.
At that moment, if driver's tcp-data-split config is UNKNOWN, it returns
ENABLE if LRO/GRO/JUMBO is enabled.
Then, it sets values from the user and driver's current config to
kernel_ethtool_ringparam.
Last it calls .set_ringparam().
The driver, especially bnxt_en driver receives
ETHTOOL_TCP_DATA_SPLIT_ENABLED.
But it can't distinguish whether it is set by the user or just the
current config.
When user updates ring parameter, the new hds_config value is updated
and current hds_config value is stored to old_hdsconfig.
Driver's .set_ringparam() callback can distinguish a passed
tcp-data-split value is came from user explicitly.
If .set_ringparam() is failed, hds_config is rollbacked immediately.
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Link: https://patch.msgid.link/20250114142852.3364986-2-ap420073@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
blackhole_netdev is the global device in init_net.
Let's hold rtnl_net_lock(&init_net) in blackhole_netdev_init().
While at it, the unnecessary dev_net_set() call is removed, which
is done in alloc_netdev_mqs().
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250114081352.47404-1-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Running "make kselftest TARGETS=net/forwarding" results in
multiple ccurrences of the same error:
- ./lib.sh: line 787: teamd: command not found
This patch adds the variable $REQUIRE_TEAMD in every test that uses the
command teamd and checks the $REQUIRE_TEAMD variable in the file "lib.sh"
to skip the test if the command is not installed.
Signed-off-by: Alessandro Zanni <alessandro.zanni87@gmail.com>
Link: https://patch.msgid.link/20250114003323.97207-1-alessandro.zanni87@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Sanman Pradhan says:
====================
eth: fbnic: Add hardware monitoring support
This patch series adds hardware monitoring support to the fbnic driver.
It implements support for reading temperature and voltage sensors via
firmware requests, and exposes this data through the HWMON interface.
The series is structured as follows:
Patch 1: Adds completion infrastructure for firmware requests
Patch 2: Implements TSENE sensor message handling
Patch 3: Adds HWMON interface support
Output:
$ ls -l /sys/class/hwmon/hwmon1/
total 0
lrwxrwxrwx 1 root root 0 Sep 10 00:00 device -> ../../../0000:01:00.0
-r--r--r-- 1 root root 4096 Sep 10 00:00 in0_input
-r--r--r-- 1 root root 4096 Sep 10 00:00 name
lrwxrwxrwx 1 root root 0 Sep 10 00:00 subsystem -> ../../../../../../class/hwmon
-r--r--r-- 1 root root 4096 Sep 10 00:00 temp1_input
-rw-r--r-- 1 root root 4096 Sep 10 00:00 uevent
$ cat /sys/class/hwmon/hwmon1/temp1_input
40480
$ cat /sys/class/hwmon/hwmon1/in0_input
750
====================
Link: https://patch.msgid.link/20250114000705.2081288-1-sanman.p211993@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
This patch adds support for hardware monitoring to the fbnic driver,
allowing for temperature and voltage sensor data to be exposed to
userspace via the HWMON interface. The driver registers a HWMON device
and provides callbacks for reading sensor data, enabling system
admins to monitor the health and operating conditions of fbnic.
Signed-off-by: Sanman Pradhan <sanman.p211993@gmail.com>
Reviewed-by: Michal Swiatkowski <michal.swiatkowski@linux.intel.com>
Link: https://patch.msgid.link/20250114000705.2081288-4-sanman.p211993@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Add support for reading temperature and voltage sensor data from firmware
by implementing a new TSENE message type and response parsing. This adds
message handler infrastructure to transmit sensor read requests and parse
responses. The sensor data will be exposed through the driver's hwmon interface.
Signed-off-by: Sanman Pradhan <sanman.p211993@gmail.com>
Reviewed-by: Michal Swiatkowski <michal.swiatkowski@linux.intel.com>
Link: https://patch.msgid.link/20250114000705.2081288-3-sanman.p211993@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Add infrastructure to support firmware request/response handling with
completions. Add a completion structure to track message state including
message type for matching, completion for waiting for response, and
result for error propagation. Use existing spinlock to protect the writes.
The data from the various response types will be added to the "union u"
by subsequent commits.
Signed-off-by: Sanman Pradhan <sanman.p211993@gmail.com>
Reviewed-by: Michal Swiatkowski <michal.swiatkowski@linux.intel.com>
Link: https://patch.msgid.link/20250114000705.2081288-2-sanman.p211993@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
Daniel Machon says:
====================
net: lan969x: add FDMA support
== Description:
This series is the last of a multi-part series, that prepares and adds
support for the new lan969x switch driver.
The upstreaming efforts has been split into multiple series:
1) Prepare the Sparx5 driver for lan969x (merged)
2) Add support for lan969x (same basic features as Sparx5
provides excl. FDMA and VCAP, merged).
3) Add lan969x VCAP functionality (merged).
4) Add RGMII support (merged).
--> 5) Add FDMA support.
== FDMA support:
The lan969x switch device uses the same FDMA engine as the Sparx5 switch
device, with the same number of channels etc. This means we can utilize
the newly added FDMA library, that is already in use by the lan966x and
sparx5 drivers.
As previous lan969x series, the FDMA implementation will hook into the
Sparx5 implementation where possible, however both RX and TX handling
will be done differently on lan969x and therefore requires a separate
implementation of the RX and TX path.
Details are in the commit description of the individual patches
== Patch breakdown:
Patch #1: Enable FDMA support on lan969x
Patch #2: Split start()/stop() functions
Patch #3: Activate TX FDMA in start()
Patch #4: Ops out a few functions that differ on the two platforms
Patch #5: Add FDMA implementation for lan969x
v1: https://lore.kernel.org/20250109-sparx5-lan969x-switch-driver-5-v1-0-13d6d8451e63@microchip.com
====================
Link: https://patch.msgid.link/20250113-sparx5-lan969x-switch-driver-5-v2-0-c468f02fd623@microchip.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The lan969x switch device supports manual frame injection and extraction
to and from the switch core, using a number of injection and extraction
queues. This technique is currently supported, but delivers poor
performance compared to Frame DMA (FDMA).
This lan969x implementation of FDMA, hooks into the existing FDMA for
Sparx5, but requires its own RX and TX handling, as lan969x does not
support the same native cache coherency that Sparx5 does. Effectively,
this means that we are going to use the DMA mapping API for mapping and
unmapping TX buffers. The RX loop will utilize the page pool API for
efficient RX handling. Other than that, the implementation is largely
the same, and utilizes the FDMA library for DCB and DB handling.
Some numbers:
Manual injection/extraction (before this series):
// iperf3 -c 1.0.1.1
[ ID] Interval Transfer Bitrate
[ 5] 0.00-10.02 sec 345 MBytes 289 Mbits/sec sender
[ 5] 0.00-10.06 sec 345 MBytes 288 Mbits/sec receiver
FDMA (after this series):
// iperf3 -c 1.0.1.1
[ ID] Interval Transfer Bitrate
[ 5] 0.00-10.03 sec 1.10 GBytes 940 Mbits/sec sender
[ 5] 0.00-10.07 sec 1.10 GBytes 936 Mbits/sec receiver
Reviewed-by: Steen Hegelund <Steen.Hegelund@microchip.com>
Signed-off-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20250113-sparx5-lan969x-switch-driver-5-v2-5-c468f02fd623@microchip.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
We are going to implement the RX and TX paths a bit differently on
lan969x and therefore need to introduce new ops for FDMA functions:
init, deinit, xmit and poll. Assign the Sparx5 equivalents for these and
update the code throughout. Also add a 'struct net_device' argument to
the xmit() function, as we will be needing that for lan969x.
Reviewed-by: Steen Hegelund <Steen.Hegelund@microchip.com>
Signed-off-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20250113-sparx5-lan969x-switch-driver-5-v2-4-c468f02fd623@microchip.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The function sparx5_fdma_tx_activate() is responsible for configuring
the TX FDMA instance and activating the channel. TX activation has
previously been done in the xmit() function, when the first frame is
transmitted. Now that we have separate functions for starting and
stopping the FDMA, it seems reasonable to move the TX activation to the
start function. This change has no implications on the functionality.
Reviewed-by: Steen Hegelund <Steen.Hegelund@microchip.com>
Signed-off-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20250113-sparx5-lan969x-switch-driver-5-v2-3-c468f02fd623@microchip.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
The two functions: sparx5_fdma_{start(),stop()} are responsible for a
number of things, namely: allocation and initialization of FDMA buffers,
activation FDMA channels in hardware and activation of the NAPI
instance.
This patch splits the buffer allocation and initialization into init and
deinit functions, and the channel and NAPI activation into start and
stop functions. This serves two purposes: 1) the start() and stop()
functions can be reused for lan969x and 2) prepares for future MTU
change support, where we must be able to stop and start the FDMA
channels and NAPI instance, without free'ing and reallocating the FDMA
buffers.
Reviewed-by: Steen Hegelund <Steen.Hegelund@microchip.com>
Signed-off-by: Daniel Machon <daniel.machon@microchip.com>
Link: https://patch.msgid.link/20250113-sparx5-lan969x-switch-driver-5-v2-2-c468f02fd623@microchip.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
