With recent work to the doorbell workaround code a small hole was
introduced that could cause a tx_timeout. This happens if the rx
dbell_deadline goes beyond the netdev watchdog timeout set by the driver
(i.e. 2 seconds). Fix this by changing the netdev watchdog timeout to 5
seconds and reduce the max rx dbell_deadline to 4 seconds.
The test that can reproduce the issue being fixed is a multi-queue send
test via pktgen with the "burst" setting to 1. This causes the queue's
doorbell to be rung on every packet sent to the driver, which may result
in the device missing doorbells due to the high doorbell rate.
Cc: stable@vger.kernel.org
Fixes: 4ded136c78 ("ionic: add work item for missed-doorbell check")
Signed-off-by: Brett Creeley <brett.creeley@amd.com>
Reviewed-by: Shannon Nelson <shannon.nelson@amd.com>
Link: https://patch.msgid.link/20240822192557.9089-1-brett.creeley@amd.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The driver must ensure TX descriptor updates are visible
before updating TX pointer and TX clear pointer.
This resolves TX hangs observed on AST2600 when running
iperf3.
Signed-off-by: Jacky Chou <jacky_chou@aspeedtech.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The mana_hwc_rx_event_handler() / mana_hwc_handle_resp() calls
complete(&ctx->comp_event) before posting the wqe back. It's
possible that other callers, like mana_create_txq(), start the
next round of mana_hwc_send_request() before the posting of wqe.
And if the HW is fast enough to respond, it can hit no_wqe error
on the HW channel, then the response message is lost. The mana
driver may fail to create queues and open, because of waiting for
the HW response and timed out.
Sample dmesg:
[ 528.610840] mana 39d4:00:02.0: HWC: Request timed out!
[ 528.614452] mana 39d4:00:02.0: Failed to send mana message: -110, 0x0
[ 528.618326] mana 39d4:00:02.0 enP14804s2: Failed to create WQ object: -110
To fix it, move posting of rx wqe before complete(&ctx->comp_event).
Cc: stable@vger.kernel.org
Fixes: ca9c54d2d6 ("net: mana: Add a driver for Microsoft Azure Network Adapter (MANA)")
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Long Li <longli@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some CPT AF registers are per LF and others are global. Translation
of PF/VF local LF slot number to actual LF slot number is required
only for accessing perf LF registers. CPT AF global registers access
do not require any LF slot number. Also, there is no reason CPT
PF/VF to know actual lf's register offset.
Without this fix microcode loading will fail, VFs cannot be created
and hardware is not usable.
Fixes: bc35e28af7 ("octeontx2-af: replace cpt slot with lf id on reg write")
Signed-off-by: Bharat Bhushan <bbhushan2@marvell.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20240821070558.1020101-1-bbhushan2@marvell.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Tony Nguyen says:
====================
Intel Wired LAN Driver Updates 2024-08-20 (ice)
This series contains updates to ice driver only.
Maciej fixes issues with Rx data path on architectures with
PAGE_SIZE >= 8192; correcting page reuse usage and calculations for
last offset and truesize.
Michal corrects assignment of devlink port number to use PF id.
* '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue:
ice: use internal pf id instead of function number
ice: fix truesize operations for PAGE_SIZE >= 8192
ice: fix ICE_LAST_OFFSET formula
ice: fix page reuse when PAGE_SIZE is over 8k
====================
Link: https://patch.msgid.link/20240820215620.1245310-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Sabrina reports that the igb driver does not cope well with large
MAX_SKB_FRAG values: setting MAX_SKB_FRAG to 45 causes payload
corruption on TX.
An easy reproducer is to run ssh to connect to the machine. With
MAX_SKB_FRAGS=17 it works, with MAX_SKB_FRAGS=45 it fails. This has
been reported originally in
https://bugzilla.redhat.com/show_bug.cgi?id=2265320
The root cause of the issue is that the driver does not take into
account properly the (possibly large) shared info size when selecting
the ring layout, and will try to fit two packets inside the same 4K
page even when the 1st fraglist will trump over the 2nd head.
Address the issue by checking if 2K buffers are insufficient.
Fixes: 3948b05950 ("net: introduce a config option to tweak MAX_SKB_FRAGS")
Reported-by: Jan Tluka <jtluka@redhat.com>
Reported-by: Jirka Hladky <jhladky@redhat.com>
Reported-by: Sabrina Dubroca <sd@queasysnail.net>
Tested-by: Sabrina Dubroca <sd@queasysnail.net>
Tested-by: Corinna Vinschen <vinschen@redhat.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Corinna Vinschen <vinschen@redhat.com>
Link: https://patch.msgid.link/20240816152034.1453285-1-vinschen@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The dpaa2_switch_add_bufs() function returns the number of bufs that it
was able to add. It returns BUFS_PER_CMD (7) for complete success or a
smaller number if there are not enough pages available. However, the
error checking is looking at the total number of bufs instead of the
number which were added on this iteration. Thus the error checking
only works correctly for the first iteration through the loop and
subsequent iterations are always counted as a success.
Fix this by checking only the bufs added in the current iteration.
Fixes: 0b1b713704 ("staging: dpaa2-switch: handle Rx path on control interface")
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Tested-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://patch.msgid.link/eec27f30-b43f-42b6-b8ee-04a6f83423b6@stanley.mountain
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When working on multi-buffer packet on arch that has PAGE_SIZE >= 8192,
truesize is calculated and stored in xdp_buff::frame_sz per each
processed Rx buffer. This means that frame_sz will contain the truesize
based on last received buffer, but commit 1dc1a7e7f4 ("ice:
Centrallize Rx buffer recycling") assumed this value will be constant
for each buffer, which breaks the page recycling scheme and mess up the
way we update the page::page_offset.
To fix this, let us work on constant truesize when PAGE_SIZE >= 8192
instead of basing this on size of a packet read from Rx descriptor. This
way we can simplify the code and avoid calculating truesize per each
received frame and on top of that when using
xdp_update_skb_shared_info(), current formula for truesize update will
be valid.
This means ice_rx_frame_truesize() can be removed altogether.
Furthermore, first call to it within ice_clean_rx_irq() for 4k PAGE_SIZE
was redundant as xdp_buff::frame_sz is initialized via xdp_init_buff()
in ice_vsi_cfg_rxq(). This should have been removed at the point where
xdp_buff struct started to be a member of ice_rx_ring and it was no
longer a stack based variable.
There are two fixes tags as my understanding is that the first one
exposed us to broken truesize and page_offset handling and then second
introduced broken skb_shared_info update in ice_{construct,build}_skb().
Reported-and-tested-by: Luiz Capitulino <luizcap@redhat.com>
Closes: https://lore.kernel.org/netdev/8f9e2a5c-fd30-4206-9311-946a06d031bb@redhat.com/
Fixes: 1dc1a7e7f4 ("ice: Centrallize Rx buffer recycling")
Fixes: 2fba7dc515 ("ice: Add support for XDP multi-buffer on Rx side")
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Chandan Kumar Rout <chandanx.rout@intel.com> (A Contingent Worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
For bigger PAGE_SIZE archs, ice driver works on 3k Rx buffers.
Therefore, ICE_LAST_OFFSET should take into account ICE_RXBUF_3072, not
ICE_RXBUF_2048.
Fixes: 7237f5b0db ("ice: introduce legacy Rx flag")
Suggested-by: Luiz Capitulino <luizcap@redhat.com>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Chandan Kumar Rout <chandanx.rout@intel.com> (A Contingent Worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Architectures that have PAGE_SIZE >= 8192 such as arm64 should act the
same as x86 currently, meaning reuse of a page should only take place
when no one else is busy with it.
Do two things independently of underlying PAGE_SIZE:
- store the page count under ice_rx_buf::pgcnt
- then act upon its value vs ice_rx_buf::pagecnt_bias when making the
decision regarding page reuse
Fixes: 2b245cb294 ("ice: Implement transmit and NAPI support")
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Chandan Kumar Rout <chandanx.rout@intel.com> (A Contingent Worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
The change in the fixes tag cleaned up too much: it removed the part
that was releasing header pages that were posted via UMR but haven't
been acknowledged yet on the ICOSQ.
This patch corrects this omission by setting the bits between pi and ci
to on when shutting down a queue with SHAMPO. To be consistent with the
Striding RQ code, this action is done in mlx5e_free_rx_missing_descs().
Fixes: e839ac9a89 ("net/mlx5e: SHAMPO, Simplify header page release in teardown")
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20240815071611.2211873-3-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When SHAMPO is used, a receive queue currently almost always leaks one
page on shutdown.
A page has MLX5E_SHAMPO_WQ_HEADER_PER_PAGE (8) headers. These headers
are tracked in the SHAMPO bitmap. Each page is released when the last
header index in the group is processed. During header allocation, there
can be leftovers from a page that will be used in a subsequent
allocation. This is normally fine, except for the following scenario
(simplified a bit):
1) Allocate N new page fragments, showing only the relevant last 4
fragments:
0: new page
1: new page
2: new page
3: new page
4: page from previous allocation
5: page from previous allocation
6: page from previous allocation
7: page from previous allocation
2) NAPI processes header indices 4-7 because they are the oldest
allocated. Bit 7 will be set to 0.
3) Receive queue shutdown occurs. All the remaining bits are being
iterated on to release the pages. But the page assigned to header
indices 0-3 will not be freed due to what happened in step 2.
This patch fixes the issue by making sure that on allocation, header
fragments are always allocated in groups of
MLX5E_SHAMPO_WQ_HEADER_PER_PAGE so that there is never a partial page
left over between allocations.
A more appropriate fix would be a refactoring of
mlx5e_alloc_rx_hd_mpwqe() and mlx5e_build_shampo_hd_umr(). But this
refactoring is too big for net. It will be targeted for net-next.
Fixes: e839ac9a89 ("net/mlx5e: SHAMPO, Simplify header page release in teardown")
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20240815071611.2211873-2-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
I was revisiting the topic of 802.1ad treatment in the Ocelot switch [0]
and realized that not only is its basic VLAN classification pipeline
improper for offloading vlan_protocol 802.1ad bridges, but also improper
for offloading regular 802.1Q bridges already.
Namely, 802.1ad-tagged traffic should be treated as VLAN-untagged by
bridged ports, but this switch treats it as if it was 802.1Q-tagged with
the same VID as in the 802.1ad header. This is markedly different to
what the Linux bridge expects; see the "other_tpid()" function in
tools/testing/selftests/net/forwarding/bridge_vlan_aware.sh.
An idea came to me that the VCAP IS1 TCAM is more powerful than I'm
giving it credit for, and that it actually overwrites the classified VID
before the VLAN Table lookup takes place. In other words, it can be
used even to save a packet from being dropped on ingress due to VLAN
membership.
Add a sophisticated TCAM rule hardcoded into the driver to force the
switch to behave like a Linux bridge with vlan_filtering 1 vlan_protocol
802.1Q.
Regarding the lifetime of the filter: eventually the bridge will
disappear, and vlan_filtering on the port will be restored to 0 for
standalone mode. Then the filter will be deleted.
[0]: https://lore.kernel.org/netdev/20201009122947.nvhye4hvcha3tljh@skbuf/
Fixes: 7142529f16 ("net: mscc: ocelot: add VLAN filtering")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As explained by Horatiu Vultur in commit 603ead9658 ("net: sparx5: Add
spinlock for frame transmission from CPU") which is for a similar
hardware design, multiple CPUs can simultaneously perform injection
or extraction. There are only 2 register groups for injection and 2
for extraction, and the driver only uses one of each. So we'd better
serialize access using spin locks, otherwise frame corruption is
possible.
Note that unlike in sparx5, FDMA in ocelot does not have this issue
because struct ocelot_fdma_tx_ring already contains an xmit_lock.
I guess this is mostly a problem for NXP LS1028A, as that is dual core.
I don't think VSC7514 is. So I'm blaming the commit where LS1028A (aka
the felix DSA driver) started using register-based packet injection and
extraction.
Fixes: 0a6f17c6ae ("net: dsa: tag_ocelot_8021q: add support for PTP timestamping")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are 2 distinct code paths (listed below) in the source code which
set up an injection header for Ocelot(-like) switches. Code path (2)
lacks the QoS class and source port being set correctly. Especially the
improper QoS classification is a problem for the "ocelot-8021q"
alternative DSA tagging protocol, because we support tc-taprio and each
packet needs to be scheduled precisely through its time slot. This
includes PTP, which is normally assigned to a traffic class other than
0, but would be sent through TC 0 nonetheless.
The code paths are:
(1) ocelot_xmit_common() from net/dsa/tag_ocelot.c - called only by the
standard "ocelot" DSA tagging protocol which uses NPI-based
injection - sets up bit fields in the tag manually to account for
a small difference (destination port offset) between Ocelot and
Seville. Namely, ocelot_ifh_set_dest() is omitted out of
ocelot_xmit_common(), because there's also seville_ifh_set_dest().
(2) ocelot_ifh_set_basic(), called by:
- ocelot_fdma_prepare_skb() for FDMA transmission of the ocelot
switchdev driver
- ocelot_port_xmit() -> ocelot_port_inject_frame() for
register-based transmission of the ocelot switchdev driver
- felix_port_deferred_xmit() -> ocelot_port_inject_frame() for the
DSA tagger ocelot-8021q when it must transmit PTP frames (also
through register-based injection).
sets the bit fields according to its own logic.
The problem is that (2) doesn't call ocelot_ifh_set_qos_class().
Copying that logic from ocelot_xmit_common() fixes that.
Unfortunately, although desirable, it is not easily possible to
de-duplicate code paths (1) and (2), and make net/dsa/tag_ocelot.c
directly call ocelot_ifh_set_basic()), because of the ocelot/seville
difference. This is the "minimal" fix with some logic duplicated (but
at least more consolidated).
Fixes: 0a6f17c6ae ("net: dsa: tag_ocelot_8021q: add support for PTP timestamping")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Problem description
-------------------
On an NXP LS1028A (felix DSA driver) with the following configuration:
- ocelot-8021q tagging protocol
- VLAN-aware bridge (with STP) spanning at least swp0 and swp1
- 8021q VLAN upper interfaces on swp0 and swp1: swp0.700, swp1.700
- ptp4l on swp0.700 and swp1.700
we see that the ptp4l instances do not see each other's traffic,
and they all go to the grand master state due to the
ANNOUNCE_RECEIPT_TIMEOUT_EXPIRES condition.
Jumping to the conclusion for the impatient
-------------------------------------------
There is a zero-day bug in the ocelot switchdev driver in the way it
handles VLAN-tagged packet injection. The correct logic already exists in
the source code, in function ocelot_xmit_get_vlan_info() added by commit
5ca721c54d ("net: dsa: tag_ocelot: set the classified VLAN during xmit").
But it is used only for normal NPI-based injection with the DSA "ocelot"
tagging protocol. The other injection code paths (register-based and
FDMA-based) roll their own wrong logic. This affects and was noticed on
the DSA "ocelot-8021q" protocol because it uses register-based injection.
By moving ocelot_xmit_get_vlan_info() to a place that's common for both
the DSA tagger and the ocelot switch library, it can also be called from
ocelot_port_inject_frame() in ocelot.c.
We need to touch the lines with ocelot_ifh_port_set()'s prototype
anyway, so let's rename it to something clearer regarding what it does,
and add a kernel-doc. ocelot_ifh_set_basic() should do.
Investigation notes
-------------------
Debugging reveals that PTP event (aka those carrying timestamps, like
Sync) frames injected into swp0.700 (but also swp1.700) hit the wire
with two VLAN tags:
00000000: 01 1b 19 00 00 00 00 01 02 03 04 05 81 00 02 bc
~~~~~~~~~~~
00000010: 81 00 02 bc 88 f7 00 12 00 2c 00 00 02 00 00 00
~~~~~~~~~~~
00000020: 00 00 00 00 00 00 00 00 00 00 00 01 02 ff fe 03
00000030: 04 05 00 01 00 04 00 00 00 00 00 00 00 00 00 00
00000040: 00 00
The second (unexpected) VLAN tag makes felix_check_xtr_pkt() ->
ptp_classify_raw() fail to see these as PTP packets at the link
partner's receiving end, and return PTP_CLASS_NONE (because the BPF
classifier is not written to expect 2 VLAN tags).
The reason why packets have 2 VLAN tags is because the transmission
code treats VLAN incorrectly.
Neither ocelot switchdev, nor felix DSA, declare the NETIF_F_HW_VLAN_CTAG_TX
feature. Therefore, at xmit time, all VLANs should be in the skb head,
and none should be in the hwaccel area. This is done by:
static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
netdev_features_t features)
{
if (skb_vlan_tag_present(skb) &&
!vlan_hw_offload_capable(features, skb->vlan_proto))
skb = __vlan_hwaccel_push_inside(skb);
return skb;
}
But ocelot_port_inject_frame() handles things incorrectly:
ocelot_ifh_port_set(ifh, port, rew_op, skb_vlan_tag_get(skb));
void ocelot_ifh_port_set(struct sk_buff *skb, void *ifh, int port, u32 rew_op)
{
(...)
if (vlan_tag)
ocelot_ifh_set_vlan_tci(ifh, vlan_tag);
(...)
}
The way __vlan_hwaccel_push_inside() pushes the tag inside the skb head
is by calling:
static inline void __vlan_hwaccel_clear_tag(struct sk_buff *skb)
{
skb->vlan_present = 0;
}
which does _not_ zero out skb->vlan_tci as seen by skb_vlan_tag_get().
This means that ocelot, when it calls skb_vlan_tag_get(), sees
(and uses) a residual skb->vlan_tci, while the same VLAN tag is
_already_ in the skb head.
The trivial fix for double VLAN headers is to replace the content of
ocelot_ifh_port_set() with:
if (skb_vlan_tag_present(skb))
ocelot_ifh_set_vlan_tci(ifh, skb_vlan_tag_get(skb));
but this would not be correct either, because, as mentioned,
vlan_hw_offload_capable() is false for us, so we'd be inserting dead
code and we'd always transmit packets with VID=0 in the injection frame
header.
I can't actually test the ocelot switchdev driver and rely exclusively
on code inspection, but I don't think traffic from 8021q uppers has ever
been injected properly, and not double-tagged. Thus I'm blaming the
introduction of VLAN fields in the injection header - early driver code.
As hinted at in the early conclusion, what we _want_ to happen for
VLAN transmission was already described once in commit 5ca721c54d
("net: dsa: tag_ocelot: set the classified VLAN during xmit").
ocelot_xmit_get_vlan_info() intends to ensure that if the port through
which we're transmitting is under a VLAN-aware bridge, the outer VLAN
tag from the skb head is stripped from there and inserted into the
injection frame header (so that the packet is processed in hardware
through that actual VLAN). And in all other cases, the packet is sent
with VID=0 in the injection frame header, since the port is VLAN-unaware
and has logic to strip this VID on egress (making it invisible to the
wire).
Fixes: 08d02364b1 ("net: mscc: fix the injection header")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The driver currently blindly deletes its cache of RSS cotexts and
ntuple filters when the ethtool channel count is changing. It also
deletes the ntuple filters cache when the default indirection table
is changing.
The core will not allow ethtool channels to drop below any that
have been configured as ntuple destinations since this commit from 2022:
47f3ecf476 ("ethtool: Fail number of channels change when it conflicts with rxnfc")
So there is absolutely no need to delete the ntuple filters and
RSS contexts when changing ethtool channels.
It is also unnecessary to delete ntuple filters when the default
RSS indirection table is changing.
Remove bnxt_clear_usr_fltrs() and bnxt_clear_rss_ctxis() from the
ethtool ops and change them to static functions.
This bug will cause confusion to the end user and causes failure when
running the rss_ctx.py selftest.
Fixes: 1018319f94 ("bnxt_en: Invalidate user filters when needed")
Reported-by: Jakub Kicinski <kuba@kernel.org>
Closes: https://lore.kernel.org/netdev/20240725111912.7bc17cf6@kernel.org/
Reviewed-by: Andy Gospodarek <andrew.gospodarek@broadcom.com>
Signed-off-by: Pavan Chebbi <pavan.chebbi@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Link: https://patch.msgid.link/20240814225429.199280-1-michael.chan@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
pci_request_regions is called to apply for PCI I/O and memory resources
when the driver is initialized, Therefore, when the driver is uninstalled,
pci_release_regions should be used to release PCI I/O and memory resources
instead of pci_release_mem_regions is used to release memory reasouces
only.
Signed-off-by: Peiyang Wang <wangpeiyang1@huawei.com>
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When query reg inf of SSU, it loops tnl_num times. However, tnl_num comes
from hardware and the length of array is a fixed value. To void array out
of bound, make sure the loop time is not greater than the length of array
Signed-off-by: Peiyang Wang <wangpeiyang1@huawei.com>
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When config TC during the reset process, may cause a deadlock, the flow is
as below:
pf reset start
│
▼
......
setup tc │
│ ▼
▼ DOWN: napi_disable()
napi_disable()(skip) │
│ │
▼ ▼
...... ......
│ │
▼ │
napi_enable() │
▼
UINIT: netif_napi_del()
│
▼
......
│
▼
INIT: netif_napi_add()
│
▼
...... global reset start
│ │
▼ ▼
UP: napi_enable()(skip) ......
│ │
▼ ▼
...... napi_disable()
In reset process, the driver will DOWN the port and then UINIT, in this
case, the setup tc process will UP the port before UINIT, so cause the
problem. Adds a DOWN process in UINIT to fix it.
Fixes: bb6b94a896 ("net: hns3: Add reset interface implementation in client")
Signed-off-by: Jie Wang <wangjie125@huawei.com>
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Consider the followed case that the user change speed and reset the net
interface. Before the hw change speed successfully, the driver get old
old speed from hw by timer task. After reset, the previous speed is config
to hw. As a result, the new speed is configed successfully but lost after
PF reset. The followed pictured shows more dirrectly.
+------+ +----+ +----+
| USER | | PF | | HW |
+---+--+ +-+--+ +-+--+
| ethtool -s 100G | |
+------------------>| set speed 100G |
| +--------------------->|
| | set successfully |
| |<---------------------+---+
| |query cfg (timer task)| |
| +--------------------->| | handle speed
| | return 200G | | changing event
| ethtool --reset |<---------------------+ | (100G)
+------------------>| cfg previous speed |<--+
| | after reset (200G) |
| +--------------------->|
| | +---+
| |query cfg (timer task)| |
| +--------------------->| | handle speed
| | return 100G | | changing event
| |<---------------------+ | (200G)
| | |<--+
| |query cfg (timer task)|
| +--------------------->|
| | return 200G |
| |<---------------------+
| | |
v v v
This patch save new speed if hw change speed successfully, which will be
used after reset successfully.
Fixes: 2d03eacc0b ("net: hns3: Only update mac configuation when necessary")
Signed-off-by: Peiyang Wang <wangpeiyang1@huawei.com>
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Currently, if hns3 PF or VF FLR reset failed after five times retry,
the reset done process will directly release the semaphore
which has already released in hclge_reset_prepare_general.
This will cause down operation fail.
So this patch fixes it by adding reset state judgement. The up operation is
only called after successful PF FLR reset.
Fixes: 8627bdedc4 ("net: hns3: refactor the precedure of PF FLR")
Fixes: f28368bb45 ("net: hns3: refactor the procedure of VF FLR")
Signed-off-by: Jie Wang <wangjie125@huawei.com>
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Commit 94833addfa ("net: thunderx: Unembed netdev structure") had
a go at dynamically allocating the netdev structures for the thunderx_bgx
driver. This change results in my ThunderX box catching fire (to be fair,
it is what it does best).
The issues with this change are that:
- bgx_lmac_enable() is called *after* bgx_acpi_register_phy() and
bgx_init_of_phy(), both expecting netdev to be a valid pointer.
- bgx_init_of_phy() populates the MAC addresses for *all* LMACs
attached to a given BGX instance, and thus needs netdev for each of
them to have been allocated.
There is a few things to be said about how the driver mixes LMAC and
BGX states which leads to this sorry state, but that's beside the point.
To address this, go back to a situation where all netdev structures
are allocated before the driver starts relying on them, and move the
freeing of these structures to driver removal. Someone brave enough
can always go and restructure the driver if they want.
Fixes: 94833addfa ("net: thunderx: Unembed netdev structure")
Signed-off-by: Marc Zyngier <maz@kernel.org>
Cc: Breno Leitao <leitao@debian.org>
Cc: Sunil Goutham <sgoutham@marvell.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Jakub Kicinski <kuba@kernel.org>
Cc: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20240812141322.1742918-1-maz@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
A recent change to the driver exposed a bug where the MAC RX
filters (unicast MAC, broadcast MAC, and multicast MAC) are
configured and enabled before the RX path is fully initialized.
The result of this bug is that after the PHY is started packets
that match these MAC RX filters start to flow into the RX FIFO.
And then, after rx_init() is completed, these packets will go
into the driver RX ring as well. If enough packets are received
to fill the RX ring (default size is 128 packets) before the call
to request_irq() completes, the driver RX function becomes stuck.
This bug is intermittent but is most likely to be seen where the
oob_net0 interface is connected to a busy network with lots of
broadcast and multicast traffic.
All the MAC RX filters must be disabled until the RX path is ready,
i.e. all initialization is done and all the IRQs are installed.
Fixes: f7442a634a ("mlxbf_gige: call request_irq() after NAPI initialized")
Reviewed-by: Asmaa Mnebhi <asmaa@nvidia.com>
Signed-off-by: David Thompson <davthompson@nvidia.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20240809163612.12852-1-davthompson@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
After napi_complete_done() is called when NAPI is polling in the current
process context, another NAPI may be scheduled and start running in
softirq on another CPU and may ring the doorbell before the current CPU
does. When combined with unnecessary rings when there is no need to arm
the CQ, it triggers error paths in the hardware.
This patch fixes this by calling napi_complete_done() after doorbell
rings. It limits the number of unnecessary rings when there is
no need to arm. MANA hardware specifies that there must be one doorbell
ring every 8 CQ wraparounds. This driver guarantees one doorbell ring as
soon as the number of consumed CQEs exceeds 4 CQ wraparounds. In practical
workloads, the 4 CQ wraparounds proves to be big enough that it rarely
exceeds this limit before all the napi weight is consumed.
To implement this, add a per-CQ counter cq->work_done_since_doorbell,
and make sure the CQ is armed as soon as passing 4 wraparounds of the CQ.
Cc: stable@vger.kernel.org
Fixes: e1b5683ff6 ("net: mana: Move NAPI from EQ to CQ")
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Long Li <longli@microsoft.com>
Link: https://patch.msgid.link/1723219138-29887-1-git-send-email-longli@linuxonhyperv.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When there are multiple ap interfaces on one band and with WED on,
turning the interface down will cause a kernel panic on MT798X.
Previously, cb_priv was freed in mtk_wed_setup_tc_block() without
marking NULL,and mtk_wed_setup_tc_block_cb() didn't check the value, too.
Assign NULL after free cb_priv in mtk_wed_setup_tc_block() and check NULL
in mtk_wed_setup_tc_block_cb().
----------
Unable to handle kernel paging request at virtual address 0072460bca32b4f5
Call trace:
mtk_wed_setup_tc_block_cb+0x4/0x38
0xffffffc0794084bc
tcf_block_playback_offloads+0x70/0x1e8
tcf_block_unbind+0x6c/0xc8
...
---------
Fixes: 799684448e ("net: ethernet: mtk_wed: introduce wed wo support")
Signed-off-by: Zheng Zhang <everything411@qq.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The MANA driver's RX buffer alloc_size is passed into napi_build_skb() to
create SKB. skb_shinfo(skb) is located at the end of skb, and its alignment
is affected by the alloc_size passed into napi_build_skb(). The size needs
to be aligned properly for better performance and atomic operations.
Otherwise, on ARM64 CPU, for certain MTU settings like 4000, atomic
operations may panic on the skb_shinfo(skb)->dataref due to alignment fault.
To fix this bug, add proper alignment to the alloc_size calculation.
Sample panic info:
[ 253.298819] Unable to handle kernel paging request at virtual address ffff000129ba5cce
[ 253.300900] Mem abort info:
[ 253.301760] ESR = 0x0000000096000021
[ 253.302825] EC = 0x25: DABT (current EL), IL = 32 bits
[ 253.304268] SET = 0, FnV = 0
[ 253.305172] EA = 0, S1PTW = 0
[ 253.306103] FSC = 0x21: alignment fault
Call trace:
__skb_clone+0xfc/0x198
skb_clone+0x78/0xe0
raw6_local_deliver+0xfc/0x228
ip6_protocol_deliver_rcu+0x80/0x500
ip6_input_finish+0x48/0x80
ip6_input+0x48/0xc0
ip6_sublist_rcv_finish+0x50/0x78
ip6_sublist_rcv+0x1cc/0x2b8
ipv6_list_rcv+0x100/0x150
__netif_receive_skb_list_core+0x180/0x220
netif_receive_skb_list_internal+0x198/0x2a8
__napi_poll+0x138/0x250
net_rx_action+0x148/0x330
handle_softirqs+0x12c/0x3a0
Cc: stable@vger.kernel.org
Fixes: 80f6215b45 ("net: mana: Add support for jumbo frame")
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Long Li <longli@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In axiethernet header fix register defines comment description to be
inline with IP documentation. It updates MAC configuration register,
MDIO configuration register and frame filter control description.
Fixes: 8a3b7a252d ("drivers/net/ethernet/xilinx: added Xilinx AXI Ethernet driver")
Signed-off-by: Radhey Shyam Pandey <radhey.shyam.pandey@amd.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
ueue
Tony Nguyen says:
====================
Intel Wired LAN Driver Updates 2024-08-07 (igc)
This series contains updates to igc driver only.
Faizal adjusts the size of the MAC internal buffer on i226 devices to
resolve an errata for leaking packet transmits. He also corrects a
condition in which qbv_config_change_errors are incorrectly counted.
Lastly, he adjusts the conditions for resetting the adapter when
changing TSN Tx mode and corrects the conditions in which gtxoffset
register is set.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
`ip_hdr(skb)->ihl << 2` is the same as `ip_hdrlen(skb)`
Therefore, we should use a well-defined function not a bit shift
to find the header length.
It also compresses two lines to a single line.
Signed-off-by: Moon Yeounsu <yyyynoom@gmail.com>
Reviewed-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously, an ethtool rx flow with no attrs would not be added to the
NIC as it has no rules to configure the hw with, but it would be
reported as successful to the caller (return code 0). This is confusing
for the user as ethtool then reports "Added rule $num", but no rule was
actually added.
This change corrects that by instead reporting these wrong rules as
-EINVAL.
Fixes: b29c61dac3 ("net/mlx5e: Ethtool steering flow validation refactoring")
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20240808144107.2095424-5-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
During latency tests (netperf TCP_RR) a 30% degradation of HW GRO vs SW
GRO was observed. This is due to SHAMPO triggering timeout filler CQEs
instead of delivering the CQE for the packet.
Having a short timeout for SHAMPO doesn't bring any benefits as it is
the driver that does the merging, not the hardware. On the contrary, it
can have a negative impact: additional filler CQEs are generated due to
the timeout. As there is no way to disable this timeout, this change
sets it to the maximum value.
Instead of using the packet_merge.timeout parameter which is also used
for LRO, set the value directly when filling in the rest of the SHAMPO
parameters in mlx5e_build_rq_param().
Fixes: 99be56171f ("net/mlx5e: SHAMPO, Re-enable HW-GRO")
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20240808144107.2095424-3-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Unconditionally calling the MPIR query on BF separate mode yields the FW
syndrome below [1]. Do not call it unless admin clearly specified the SD
group, i.e. expressing the intention of using the multi-PF netdev
feature.
This fix covers cases not covered in
commit fca3b47918 ("net/mlx5: Do not query MPIR on embedded CPU function").
[1]
mlx5_cmd_out_err:808:(pid 8267): ACCESS_REG(0x805) op_mod(0x1) failed,
status bad system state(0x4), syndrome (0x685f19), err(-5)
Fixes: 678eb44805 ("net/mlx5: SD, Implement basic query and instantiation")
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Gal Pressman <gal@nvidia.com>
Link: https://patch.msgid.link/20240808144107.2095424-2-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Tony Nguyen says:
====================
Intel Wired LAN Driver Updates 2024-08-07 (ice)
This series contains updates to ice driver only.
Grzegorz adds IRQ synchronization call before performing reset and
prevents writing to hardware when it is resetting.
Mateusz swaps incorrect assignment of FEC statistics.
* '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/net-queue:
ice: Fix incorrect assigns of FEC counts
ice: Skip PTP HW writes during PTP reset procedure
ice: Fix reset handler
====================
Link: https://patch.msgid.link/20240807224521.3819189-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Both ethtool_ops.rxfh_max_context_id and the default value used when
it's not specified are supposed to be exclusive maxima (the former
is documented as such; the latter, U32_MAX, cannot be used as an ID
since it equals ETH_RXFH_CONTEXT_ALLOC), but xa_alloc() expects an
inclusive maximum.
Subtract one from 'limit' to produce an inclusive maximum, and pass
that to xa_alloc().
Increase bnxt's max by one to prevent a (very minor) regression, as
BNXT_MAX_ETH_RSS_CTX is an inclusive max. This is safe since bnxt
is not actually hard-limited; BNXT_MAX_ETH_RSS_CTX is just a
leftover from old driver code that managed context IDs itself.
Rename rxfh_max_context_id to rxfh_max_num_contexts to make its
semantics (hopefully) more obvious.
Fixes: 847a8ab186 ("net: ethtool: let the core choose RSS context IDs")
Signed-off-by: Edward Cree <ecree.xilinx@gmail.com>
Link: https://patch.msgid.link/5a2d11a599aa5b0cc6141072c01accfb7758650c.1723045898.git.ecree.xilinx@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
