Add spin lock protection for irq {un}mask registers' control.
After napi_complete_done() and this protection were applied,
a lot of redundant interrupts no longer occur.
For example: when "iperf3 -c <ipaddr> -R" on R-Car S4-8 Spider
Before the patches are applied: about 800,000 times happened
After the patches were applied: about 100,000 times happened
Fixes: 3590918b5d07 ("net: ethernet: renesas: Add support for "Ethernet Switch"")
Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Fix unmasking irq condition by using napi_complete_done(). Otherwise,
redundant interrupts happen.
Fixes: 3590918b5d07 ("net: ethernet: renesas: Add support for "Ethernet Switch"")
Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
After commit 50f303496d92 ("igb: Enable SR-IOV after reinit"), removing
the igb module could hang or crash (depending on the machine) when the
module has been loaded with the max_vfs parameter set to some value != 0.
In case of one test machine with a dual port 82580, this hang occurred:
[ 232.480687] igb 0000:41:00.1: removed PHC on enp65s0f1
[ 233.093257] igb 0000:41:00.1: IOV Disabled
[ 233.329969] pcieport 0000:40:01.0: AER: Multiple Uncorrected (Non-Fatal) err0
[ 233.340302] igb 0000:41:00.0: PCIe Bus Error: severity=Uncorrected (Non-Fata)
[ 233.352248] igb 0000:41:00.0: device [8086:1516] error status/mask=00100000
[ 233.361088] igb 0000:41:00.0: [20] UnsupReq (First)
[ 233.368183] igb 0000:41:00.0: AER: TLP Header: 40000001 0000040f cdbfc00c c
[ 233.376846] igb 0000:41:00.1: PCIe Bus Error: severity=Uncorrected (Non-Fata)
[ 233.388779] igb 0000:41:00.1: device [8086:1516] error status/mask=00100000
[ 233.397629] igb 0000:41:00.1: [20] UnsupReq (First)
[ 233.404736] igb 0000:41:00.1: AER: TLP Header: 40000001 0000040f cdbfc00c c
[ 233.538214] pci 0000:41:00.1: AER: can't recover (no error_detected callback)
[ 233.538401] igb 0000:41:00.0: removed PHC on enp65s0f0
[ 233.546197] pcieport 0000:40:01.0: AER: device recovery failed
[ 234.157244] igb 0000:41:00.0: IOV Disabled
[ 371.619705] INFO: task irq/35-aerdrv:257 blocked for more than 122 seconds.
[ 371.627489] Not tainted 6.4.0-dirty #2
[ 371.632257] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this.
[ 371.641000] task:irq/35-aerdrv state:D stack:0 pid:257 ppid:2 f0
[ 371.650330] Call Trace:
[ 371.653061] <TASK>
[ 371.655407] __schedule+0x20e/0x660
[ 371.659313] schedule+0x5a/0xd0
[ 371.662824] schedule_preempt_disabled+0x11/0x20
[ 371.667983] __mutex_lock.constprop.0+0x372/0x6c0
[ 371.673237] ? __pfx_aer_root_reset+0x10/0x10
[ 371.678105] report_error_detected+0x25/0x1c0
[ 371.682974] ? __pfx_report_normal_detected+0x10/0x10
[ 371.688618] pci_walk_bus+0x72/0x90
[ 371.692519] pcie_do_recovery+0xb2/0x330
[ 371.696899] aer_process_err_devices+0x117/0x170
[ 371.702055] aer_isr+0x1c0/0x1e0
[ 371.705661] ? __set_cpus_allowed_ptr+0x54/0xa0
[ 371.710723] ? __pfx_irq_thread_fn+0x10/0x10
[ 371.715496] irq_thread_fn+0x20/0x60
[ 371.719491] irq_thread+0xe6/0x1b0
[ 371.723291] ? __pfx_irq_thread_dtor+0x10/0x10
[ 371.728255] ? __pfx_irq_thread+0x10/0x10
[ 371.732731] kthread+0xe2/0x110
[ 371.736243] ? __pfx_kthread+0x10/0x10
[ 371.740430] ret_from_fork+0x2c/0x50
[ 371.744428] </TASK>
The reproducer was a simple script:
#!/bin/sh
for i in `seq 1 5`; do
modprobe -rv igb
modprobe -v igb max_vfs=1
sleep 1
modprobe -rv igb
done
It turned out that this could only be reproduce on 82580 (quad and
dual-port), but not on 82576, i350 and i210. Further debugging showed
that igb_enable_sriov()'s call to pci_enable_sriov() is failing, because
dev->is_physfn is 0 on 82580.
Prior to commit 50f303496d92 ("igb: Enable SR-IOV after reinit"),
igb_enable_sriov() jumped into the "err_out" cleanup branch. After this
commit it only returned the error code.
So the cleanup didn't take place, and the incorrect VF setup in the
igb_adapter structure fooled the igb driver into assuming that VFs have
been set up where no VF actually existed.
Fix this problem by cleaning up again if pci_enable_sriov() fails.
Fixes: 50f303496d92 ("igb: Enable SR-IOV after reinit")
Signed-off-by: Corinna Vinschen <vinschen@redhat.com>
Reviewed-by: Akihiko Odaki <akihiko.odaki@daynix.com>
Tested-by: Rafal Romanowski <rafal.romanowski@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit in fixes introduced flags to control the status of hardware
configuration while processing packets. At the same time another structure
is used to provide configuration of timestamper to user-space applications.
The way it was coded makes this structures go out of sync easily. The
repro is easy for 82599 chips:
[root@hostname ~]# hwstamp_ctl -i eth0 -r 12 -t 1
current settings:
tx_type 0
rx_filter 0
new settings:
tx_type 1
rx_filter 12
The eth0 device is properly configured to timestamp any PTPv2 events.
[root@hostname ~]# hwstamp_ctl -i eth0 -r 1 -t 1
current settings:
tx_type 1
rx_filter 12
SIOCSHWTSTAMP failed: Numerical result out of range
The requested time stamping mode is not supported by the hardware.
The error is properly returned because HW doesn't support all packets
timestamping. But the adapter->flags is cleared of timestamp flags
even though no HW configuration was done. From that point no RX timestamps
are received by user-space application. But configuration shows good
values:
[root@hostname ~]# hwstamp_ctl -i eth0
current settings:
tx_type 1
rx_filter 12
Fix the issue by applying new flags only when the HW was actually
configured.
Fixes: a9763f3cb54c ("ixgbe: Update PTP to support X550EM_x devices")
Signed-off-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Reviewed-by: Simon Horman <horms@kernel.org>
Tested-by: Pucha Himasekhar Reddy <himasekharx.reddy.pucha@intel.com> (A Contingent worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There's an early return in veth_set_features() if the device is in a down
state, which leads to the XDP feature flags not being updated when enabling
GRO while the device is down. Which in turn leads to XDP_REDIRECT not
working, because the redirect code now checks the flags.
Fix this by updating the feature flags after bringing the device up.
Before this patch:
NETDEV_XDP_ACT_BASIC: yes
NETDEV_XDP_ACT_REDIRECT: yes
NETDEV_XDP_ACT_NDO_XMIT: no
NETDEV_XDP_ACT_XSK_ZEROCOPY: no
NETDEV_XDP_ACT_HW_OFFLOAD: no
NETDEV_XDP_ACT_RX_SG: yes
NETDEV_XDP_ACT_NDO_XMIT_SG: no
After this patch:
NETDEV_XDP_ACT_BASIC: yes
NETDEV_XDP_ACT_REDIRECT: yes
NETDEV_XDP_ACT_NDO_XMIT: yes
NETDEV_XDP_ACT_XSK_ZEROCOPY: no
NETDEV_XDP_ACT_HW_OFFLOAD: no
NETDEV_XDP_ACT_RX_SG: yes
NETDEV_XDP_ACT_NDO_XMIT_SG: yes
Fixes: fccca038f300 ("veth: take into account device reconfiguration for xdp_features flag")
Fixes: 66c0e13ad236 ("drivers: net: turn on XDP features")
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/r/20230911135826.722295-1-toke@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
macb_set_tx_clk() is called under a spinlock but itself calls clk_set_rate()
which can sleep. This results in:
| BUG: sleeping function called from invalid context at kernel/locking/mutex.c:580
| pps pps1: new PPS source ptp1
| in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 40, name: kworker/u4:3
| preempt_count: 1, expected: 0
| RCU nest depth: 0, expected: 0
| 4 locks held by kworker/u4:3/40:
| #0: ffff000003409148
| macb ff0c0000.ethernet: gem-ptp-timer ptp clock registered.
| ((wq_completion)events_power_efficient){+.+.}-{0:0}, at: process_one_work+0x14c/0x51c
| #1: ffff8000833cbdd8 ((work_completion)(&pl->resolve)){+.+.}-{0:0}, at: process_one_work+0x14c/0x51c
| #2: ffff000004f01578 (&pl->state_mutex){+.+.}-{4:4}, at: phylink_resolve+0x44/0x4e8
| #3: ffff000004f06f50 (&bp->lock){....}-{3:3}, at: macb_mac_link_up+0x40/0x2ac
| irq event stamp: 113998
| hardirqs last enabled at (113997): [<ffff800080e8503c>] _raw_spin_unlock_irq+0x30/0x64
| hardirqs last disabled at (113998): [<ffff800080e84478>] _raw_spin_lock_irqsave+0xac/0xc8
| softirqs last enabled at (113608): [<ffff800080010630>] __do_softirq+0x430/0x4e4
| softirqs last disabled at (113597): [<ffff80008001614c>] ____do_softirq+0x10/0x1c
| CPU: 0 PID: 40 Comm: kworker/u4:3 Not tainted 6.5.0-11717-g9355ce8b2f50-dirty #368
| Hardware name: ... ZynqMP ... (DT)
| Workqueue: events_power_efficient phylink_resolve
| Call trace:
| dump_backtrace+0x98/0xf0
| show_stack+0x18/0x24
| dump_stack_lvl+0x60/0xac
| dump_stack+0x18/0x24
| __might_resched+0x144/0x24c
| __might_sleep+0x48/0x98
| __mutex_lock+0x58/0x7b0
| mutex_lock_nested+0x24/0x30
| clk_prepare_lock+0x4c/0xa8
| clk_set_rate+0x24/0x8c
| macb_mac_link_up+0x25c/0x2ac
| phylink_resolve+0x178/0x4e8
| process_one_work+0x1ec/0x51c
| worker_thread+0x1ec/0x3e4
| kthread+0x120/0x124
| ret_from_fork+0x10/0x20
The obvious fix is to move the call to macb_set_tx_clk() out of the
protected area. This seems safe as rx and tx are both disabled anyway at
this point.
It is however not entirely clear what the spinlock shall protect. It
could be the read-modify-write access to the NCFGR register, but this
is accessed in macb_set_rx_mode() and macb_set_rxcsum_feature() as well
without holding the spinlock. It could also be the register accesses
done in mog_init_rings() or macb_init_buffers(), but again these
functions are called without holding the spinlock in macb_hresp_error_task().
The locking seems fishy in this driver and it might deserve another look
before this patch is applied.
Fixes: 633e98a711ac0 ("net: macb: use resolved link config in mac_link_up()")
Signed-off-by: Sascha Hauer <s.hauer@pengutronix.de>
Link: https://lore.kernel.org/r/20230908112913.1701766-1-s.hauer@pengutronix.de
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
MT7988 SoC support 3 NICs. Fix pse_port configuration in
mtk_flow_set_output_device routine if the traffic is offloaded to eth2.
Rely on mtk_pse_port definitions.
Fixes: 88efedf517e6 ("net: ethernet: mtk_eth_soc: enable nft hw flowtable_offload for MT7988 SoC")
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Variable dma_addr in function mtk_poll_rx can be uninitialized on some of the error paths. In practise this doesn't matter, even random data present in uninitialized stack memory can safely be used in the way it happens in the error path. However, in order to make Smatch happy make sure the variable is always initialized. Signed-off-by: Daniel Golle <daniel@makrotopia.org> Signed-off-by: David S. Miller <davem@davemloft.net>
According to the document of napi, there is no rx process when the
budget is 0. Therefore, r8152_poll() has to return 0 directly when the
budget is equal to 0.
Fixes: d2187f8e4454 ("r8152: divide the tx and rx bottom functions")
Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, when we add the first sja1105 port to a bridge with
vlan_filtering 1, then we sometimes see this output:
sja1105 spi2.2: port 4 failed to read back entry for be:79:b4:9e:9e:96 vid 3088: -ENOENT
sja1105 spi2.2: Reset switch and programmed static config. Reason: VLAN filtering
sja1105 spi2.2: port 0 failed to add be:79:b4:9e:9e:96 vid 0 to fdb: -2
It is because sja1105_fdb_add() runs from the dsa_owq which is no longer
serialized with switch resets since it dropped the rtnl_lock() in the
blamed commit.
Either performing the FDB accesses before the reset, or after the reset,
is equally fine, because sja1105_static_fdb_change() backs up those
changes in the static config, but FDB access during reset isn't ok.
Make sja1105_static_config_reload() take the fdb_lock to fix that.
Fixes: 0faf890fc519 ("net: dsa: drop rtnl_lock from dsa_slave_switchdev_event_work")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>