- 06 Jul, 2022 27 commits
-
-
Qiao Ma authored
When using hinic device as a bond slave device, and reading device stats of master bond device, the kernel may hung. The kernel panic calltrace as follows: Kernel panic - not syncing: softlockup: hung tasks Call trace: native_queued_spin_lock_slowpath+0x1ec/0x31c dev_get_stats+0x60/0xcc dev_seq_printf_stats+0x40/0x120 dev_seq_show+0x1c/0x40 seq_read_iter+0x3c8/0x4dc seq_read+0xe0/0x130 proc_reg_read+0xa8/0xe0 vfs_read+0xb0/0x1d4 ksys_read+0x70/0xfc __arm64_sys_read+0x20/0x30 el0_svc_common+0x88/0x234 do_el0_svc+0x2c/0x90 el0_svc+0x1c/0x30 el0_sync_handler+0xa8/0xb0 el0_sync+0x148/0x180 And the calltrace of task that actually caused kernel hungs as follows: __switch_to+124 __schedule+548 schedule+72 schedule_timeout+348 __down_common+188 __down+24 down+104 hinic_get_stats64+44 [hinic] dev_get_stats+92 bond_get_stats+172 [bonding] dev_get_stats+92 dev_seq_printf_stats+60 dev_seq_show+24 seq_read_iter+964 seq_read+220 proc_reg_read+164 vfs_read+172 ksys_read+108 __arm64_sys_read+28 el0_svc_common+132 do_el0_svc+40 el0_svc+24 el0_sync_handler+164 el0_sync+324 When getting device stats from bond, kernel will call bond_get_stats(). It first holds the spinlock bond->stats_lock, and then call hinic_get_stats64() to collect hinic device's stats. However, hinic_get_stats64() calls `down(&nic_dev->mgmt_lock)` to protect its critical section, which may schedule current task out. And if system is under high pressure, the task cannot be woken up immediately, which eventually triggers kernel hung panic. Since previous patch has replaced hinic_dev.tx_stats/rx_stats with local variable in hinic_get_stats64(), there is nothing need to be protected by lock, so just removing down()/up() is ok. Fixes: edd384f6 ("net-next/hinic: Add ethtool and stats") Signed-off-by:
Qiao Ma <mqaio@linux.alibaba.com> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Qiao Ma authored
Function hinic_get_stats64() will do two operations: 1. reads stats from every hinic_rxq/txq and accumulates them 2. calls hinic_rxq/txq_clean_stats() to clean every rxq/txq's stats For hinic_get_stats64(), it could get right data, because it sums all data to nic_dev->rx_stats/tx_stats. But it is wrong for get_drv_queue_stats(), this function will read hinic_rxq's stats, which have been cleared to zero by hinic_get_stats64(). I have observed hinic's cleanup operation by using such command: > watch -n 1 "cat ethtool -S eth4 | tail -40" Result before: ... rxq7_pkts: 1 rxq7_bytes: 90 rxq7_errors: 0 rxq7_csum_errors: 0 rxq7_other_errors: 0 ... rxq9_pkts: 11 rxq9_bytes: 726 rxq9_errors: 0 rxq9_csum_errors: 0 rxq9_other_errors: 0 ... rxq11_pkts: 0 rxq11_bytes: 0 rxq11_errors: 0 rxq11_csum_errors: 0 rxq11_other_errors: 0 Result after a few seconds: ... rxq7_pkts: 0 rxq7_bytes: 0 rxq7_errors: 0 rxq7_csum_errors: 0 rxq7_other_errors: 0 ... rxq9_pkts: 2 rxq9_bytes: 132 rxq9_errors: 0 rxq9_csum_errors: 0 rxq9_other_errors: 0 ... rxq11_pkts: 1 rxq11_bytes: 170 rxq11_errors: 0 rxq11_csum_errors: 0 rxq11_other_errors: 0 To solve this problem, we just keep every queue's total stats in their own queue (aka hinic_{rxq|txq}), and simply sum all per-queue stats every time calling hinic_get_stats64(). With that solution, there is no need to clean per-queue stats now, and there is no need to maintain global hinic_dev.{tx|rx}_stats, too. Fixes: edd384f6 ("net-next/hinic: Add ethtool and stats") Signed-off-by:kernel test robot <lkp@intel.com> Signed-off-by:
-
David S. Miller authored
Jakub Kicinski says: ==================== tls: rx: nopad and backlog flushing This small series contains the two changes I've been working towards in the previous ~50 patches a couple of months ago. The first major change is the optional "nopad" optimization. Currently TLS 1.3 Rx performs quite poorly because it does not support the "zero-copy" or rather direct decrypt to a user space buffer. Because of TLS 1.3 record padding we don't know if a record contains data or a control message until we decrypt it. Most records will contain data, tho, so the optimization is to try the decryption hoping its data and retry if it wasn't. The performance gain from doing that is significant (~40%) but if I'm completely honest the major reason is that we call skb_cow_data() on the non-"zc" path. The next series will remove the CoW, dropping the gain to only ~10%. The second change is to flush the backlog every 128kB. ==================== Signed-off-by: -
Jakub Kicinski authored
We continuously hold the socket lock during large reads and writes. This may inflate RTT and negatively impact TCP performance. Flush the backlog periodically. I tried to pick a flush period (128kB) which gives significant benefit but the max Bps rate is not yet visibly impacted. Signed-off-by:
Jakub Kicinski <kuba@kernel.org> Signed-off-by:
-
Jakub Kicinski authored
Add a self-test variant with TLS 1.3 nopad set. Signed-off-by:
-
Jakub Kicinski authored
Since optimisitic decrypt may add extra load in case of retries require socket owner to explicitly opt-in. Signed-off-by:
-
Jakub Kicinski authored
We currently don't support decrypt to user buffer with TLS 1.3 because we don't know the record type and how much padding record contains before decryption. In practice data records are by far most common and padding gets used rarely so we can assume data record, no padding, and if we find out that wasn't the case - retry the crypto in place (decrypt to skb). To safeguard from user overwriting content type and padding before we can check it attach a 1B sg entry where last byte of the record will land. Signed-off-by:
-
Jakub Kicinski authored
To make future patches easier to review make data_len contain the length of the data, without the tail. Signed-off-by:
-
David S. Miller authored
Ratheesh Kannoth says: ==================== octeontx2: *** Exact Match Table and Field hash *** *** Exact match table and Field hash support for CN10KB silicon *** Ratheesh Kannoth (11): These patch series enables exact match table in CN10KB silicon. Legacy silicon used NPC mcam to do packet fields/channel matching for NPC rules. NPC mcam resources exahausted as customer use case increased. Supporting many DMAC filter becomes a challenge, as RPM based filter count is less. Exact match table has 4way 2K entry table and a 32 entry fully associative cam table. Second table is to handle hash table collision overflows in 4way 2K entry table. Enabling exact match table results in KEX key to be appended with Hit/Miss status. This can be used to match in NPC mcam for a more generic rule and drop those packets than having DMAC drop rules for each DMAC entry in NPC mcam. octeontx2-af: Exact match support octeontx2-af: Exact match scan from kex profile octeontx2-af: devlink configuration support octeontx2-af: FLR handler for exact match table. octeontx2-af: Drop rules for NPC MCAM octeontx2-af: Debugsfs support for exact match. octeontx2: Modify mbox request and response structures octeontx2-af: Wrapper functions for mac addr add/del/update/reset octeontx2-af: Invoke exact match functions if supported octeontx2-pf: Add support for exact match table. octeontx2-af: Enable Exact match flag in kex profile Suman Ghosh (1): CN10KB variant of CN10K series of silicons supports a new feature where in a large protocol field (eg 128bit IPv6 DIP) can be condensed into a small hashed 32bit data. This saves a lot of space in MCAM key and allows user to add more protocol fields into the filter. A max of two such protocol data can be hashed. This patch adds support for hashing IPv6 SIP and/or DIP. ==================== Signed-off-by: -
Ratheesh Kannoth authored
Enabled EXACT match flag in Kex default profile. Since there is no space in key, NPC_PARSE_NIBBLE_ERRCODE is removed Signed-off-by:
Ratheesh Kannoth <rkannoth@marvell.com> Signed-off-by:
-
Ratheesh Kannoth authored
NPC exact match table can support more entries than RPM dmac filters. This requires field size of DMAC filter count and index to be increased. Signed-off-by:
-
Ratheesh Kannoth authored
If exact match table is suppoted, call functions to add/del/update entries in exact match table instead of RPM dmac filters Signed-off-by:
-
Ratheesh Kannoth authored
These functions are wrappers for mac add/addr/del/update in exact match table. These will be invoked from mbox handler routines if exact matct table is supported and enabled. Signed-off-by:
-
Ratheesh Kannoth authored
Exact match table modification requires wider fields as it has more number of slots to fill in. Modifying an entry in exact match table may cause hash collision and may be required to delete entry from 4-way 2K table and add to fully associative 32 entry CAM table. Signed-off-by:
-
Ratheesh Kannoth authored
There debugfs files created. 1. General information on exact match table 2. Exact match table entries. 3. NPC mcam drop on hit count stats. Signed-off-by:
-
Ratheesh Kannoth authored
NPC exact match table installs drop on hit rules in NPC mcam for each channel. This rule has broadcast and multicast bits cleared. Exact match bit cleared and channel bits set. If exact match table hit bit is 0, corresponding NPC mcam drop rule will be hit for the packet and will be dropped. Signed-off-by:
-
Ratheesh Kannoth authored
FLR handler should remove/free all exact match table resources corresponding to each interface. Signed-off-by:
-
Ratheesh Kannoth authored
CN10KB silicon supports Exact match feature. This feature can be disabled through devlink configuration. Devlink command fails if DMAC filter rules are already present. Once disabled, legacy RPM based DMAC filters will be configured. Signed-off-by:
-
Ratheesh Kannoth authored
CN10KB silicon supports exact match table. Scanning KEX profile should check for exact match feature is enabled and then set profile masks properly. These kex profile masks are required to configure NPC MCAM drop rules. If there is a miss in exact match table, these drop rules will drop those packets. Signed-off-by:
-
Ratheesh Kannoth authored
CN10KB silicon has support for exact match table. This table can be used to match maimum 64 bit value of KPU parsed output. Hit/non hit in exact match table can be used as a KEX key to NPC mcam. This patch makes use of Exact match table to increase number of DMAC filters supported. NPC mcam is no more need for each of these DMAC entries as will be populated in Exact match table. This patch implements following 1. Initialization of exact match table only for CN10KB. 2. Add/del/update interface function for exact match table. Signed-off-by:
-
Ratheesh Kannoth authored
CN10KB variant of CN10K series of silicons supports a new feature where in a large protocol field (eg 128bit IPv6 DIP) can be condensed into a small hashed 32bit data. This saves a lot of space in MCAM key and allows user to add more protocol fields into the filter. A max of two such protocol data can be hashed. This patch adds support for hashing IPv6 SIP and/or DIP. Signed-off-by:
Suman Ghosh <sumang@marvell.com> Signed-off-by:
-
David S. Miller authored
Merge branch 'nfp-tso' Simon Horman says: ==================== nfp: enable TSO by default this short series enables TSO by default on all NICs supported by the NFP driver. ==================== Signed-off-by: -
Simon Horman authored
We can benefit from TSO when the host CPU is not powerful enough, so enable it by default now. Signed-off-by:
Yinjun Zhang <yinjun.zhang@corigine.com> Reviewed-by:
Louis Peens <louis.peens@corigine.com> Signed-off-by:
Simon Horman <simon.horman@corigine.com> Signed-off-by:
-
Yinjun Zhang authored
Packets with metadata prepended can be correctly handled in firmware when TSO is enabled, now remove the error path and related comments. Since there's no existing firmware that uses prepended metadata, no need to add compatibility check here. Signed-off-by:
-
Zhengchao Shao authored
The functions of asix_set_sw/hw_mii are not called in other files, so change them to static. Signed-off-by:
Zhengchao Shao <shaozhengchao@huawei.com> Link: https://lore.kernel.org/r/20220704123448.128980-1-shaozhengchao@huawei.comSigned-off-by:
-
Vladimir Oltean authored
felix_vsc9959.c calls taprio_offload_get() and taprio_offload_free(), symbols exported by net/sched/sch_taprio.c. As such, we must disallow building the Felix driver as built-in when the symbol exported by tc-taprio isn't present in the kernel image. Fixes: 1c9017e4 ("net: dsa: felix: keep reference on entire tc-taprio config") Signed-off-by:
Vladimir Oltean <vladimir.oltean@nxp.com> Link: https://lore.kernel.org/r/20220704190241.1288847-2-vladimir.oltean@nxp.comSigned-off-by:
-
Vladimir Oltean authored
All callers of taprio_offload_get() and taprio_offload_free() prior to the blamed commit are conditionally compiled based on CONFIG_NET_SCH_TAPRIO. felix_vsc9959.c is different; it provides vsc9959_qos_port_tas_set() even when taprio is compiled out. Provide shim definitions for the functions exported by taprio so that felix_vsc9959.c is able to compile. vsc9959_qos_port_tas_set() in that case is dead code anyway, and ocelot_port->taprio remains NULL, which is fine for the rest of the logic. Fixes: 1c9017e4 ("net: dsa: felix: keep reference on entire tc-taprio config") Reported-by:
Colin Foster <colin.foster@in-advantage.com> Signed-off-by:
Vinicius Costa Gomes <vinicius.gomes@intel.com> Link: https://lore.kernel.org/r/20220704190241.1288847-1-vladimir.oltean@nxp.comSigned-off-by:
-
- 05 Jul, 2022 7 commits
-
-
Jakub Kicinski authored
The last meaningful change to this driver was made by Jon in 2011. As much as we'd like to believe that this is because the code is perfect the chances are nobody is using this hardware. Because of the size of this driver there is a nontrivial maintenance cost to keeping this code around, in the last 2 years we're averaging more than 1 change a month. Some of which require nontrivial review effort, see commit 877fe9d4 ("Revert "drivers/net/ethernet/neterion/vxge: Fix a use-after-free bug in vxge-main.c"") for example. Let's try to remove this driver. In general, IMHO, we need to establish a clear path for shedding dead code. It will be hard to unless we have some experience trying to delete stuff. Link: https://lore.kernel.org/r/20220701044234.706229-1-kuba@kernel.orgSigned-off-by:
-
Rob Herring authored
The 'reg' property is missing from the mediatek,mt7530 schema which results in the following warning once 'unevaluatedProperties' is fixed: Documentation/devicetree/bindings/net/dsa/mediatek,mt7530.example.dtb: switch@0: Unevaluated properties are not allowed ('reg' was unexpected) Fixes: e0dda311 ("dt-bindings: net: dsa: convert binding for mediatek switches") Signed-off-by:Rob Herring <robh@kernel.org> Reviewed-by:
Florian Fainelli <f.fainelli@gmail.com> Link: https://lore.kernel.org/r/20220701222240.1706272-1-robh@kernel.orgSigned-off-by:
-
Christophe JAILLET authored
Use bitmap_zalloc()/bitmap_free() instead of hand-writing them. It is less verbose and it improves the semantic. While at it, remove a useless bitmap_zero(). The bitmap is already zeroed when allocated. Signed-off-by:
Christophe JAILLET <christophe.jaillet@wanadoo.fr> Link: https://lore.kernel.org/r/8a2168ef9871bd9c4f1cf19b8d5f7530662a5d15.1656866770.git.christophe.jaillet@wanadoo.frSigned-off-by:
Paolo Abeni <pabeni@redhat.com>
-
Paolo Abeni authored
We can't use the division operator on 64 bits integers, that breaks 32 bits build. Instead use the relevant helper. Fixes: 6ddac26c ("net/mlx5e: Add support to modify hardware flow meter parameters") Acked-by:
Saeed Mahameed <saeedm@nvidia.com> Link: https://lore.kernel.org/r/ecb00ddd1197b4f8a4882090206bd2eee1eb8b5b.1657005206.git.pabeni@redhat.comSigned-off-by:
-
Paolo Abeni authored
Kuniyuki Iwashima says: ==================== af_unix: Fix regression by the per-netns hash table series. The series 6dd4142f ("Merge branch 'af_unix-per-netns-socket-hash'") replaced a global hash table with per-netns tables, which caused regression reported in the links below. [0][1] When a pathname socket is visible, any socket with the same type has to be able to connect to it even in different netns. The series puts all sockets into each namespace's hash table, making it impossible to look up a visible socket in different netns. On the other hand, while dumping sockets, they are filtered by netns. To keep such code simple, let's add a new global hash table only for pathname sockets and link them with sk_bind_node. Then we can keep all sockets in each per-netns table and look up pathname sockets via the global table. [0]: https://lore.kernel.org/netdev/B2AA3091-796D-475E-9A11-0021996E1C00@linux.ibm.com/ [1]: https://lore.kernel.org/netdev/5fb8d86f-b633-7552-8ba9-41e42f07c02a@gmail.com/ Changes: v3: * 1st: Update changelog s/named/pathname/ * 2nd: Fix checkpatch.pl CHECK by --strict option v2: https://lore.kernel.org/netdev/20220702014447.93746-1-kuniyu@amazon.com/ * Add selftest v1: https://lore.kernel.org/netdev/20220701072519.96097-1-kuniyu@amazon.com/ ==================== Link: https://lore.kernel.org/r/20220702154818.66761-1-kuniyu@amazon.comSigned-off-by:
-
Kuniyuki Iwashima authored
This patch add a test that checks connect()ivity between two sockets: unnamed socket -> bound socket * SOCK_STREAM or SOCK_DGRAM * pathname or abstract * same or different netns Signed-off-by:Kuniyuki Iwashima <kuniyu@amazon.com> Signed-off-by:
-
Kuniyuki Iwashima authored
Commit cf2f225e ("af_unix: Put a socket into a per-netns hash table.") accidentally broke user API for pathname sockets. A socket was able to connect() to a pathname socket whose file was visible even if they were in different network namespaces. The commit puts all sockets into a per-netns hash table. As a result, connect() to a pathname socket in a different netns fails to find it in the caller's per-netns hash table and returns -ECONNREFUSED even when the task can view the peer socket file. We can reproduce this issue by: Console A: # python3 >>> from socket import * >>> s = socket(AF_UNIX, SOCK_STREAM, 0) >>> s.bind('test') >>> s.listen(32) Console B: # ip netns add test # ip netns exec test sh # python3 >>> from socket import * >>> s = socket(AF_UNIX, SOCK_STREAM, 0) >>> s.connect('test') Note when dumping sockets by sock_diag, procfs, and bpf_iter, they are filtered only by netns. In other words, even if they are visible and connect()able, all sockets in different netns are skipped while iterating sockets. Thus, we need a fix only for finding a peer pathname socket. This patch adds a global hash table for pathname sockets, links them with sk_bind_node, and uses it in unix_find_socket_byinode(). By doing so, we can keep sockets in per-netns hash tables and dump them easily. Thanks to Sachin Sant and Leonard Crestez for reports, logs and a reproducer. Fixes: cf2f225e ("af_unix: Put a socket into a per-netns hash table.") Reported-by:
Sachin Sant <sachinp@linux.ibm.com> Reported-by:
Leonard Crestez <cdleonard@gmail.com> Tested-by:
Nathan Chancellor <nathan@kernel.org> Signed-off-by:
-
- 04 Jul, 2022 6 commits
-
-
Zhang Jiaming authored
Fix spelling of 'waitting' in comments. remove unnecessary space of 'MDIO_COMMAND_REG 's'. Signed-off-by:
Zhang Jiaming <jiaming@nfschina.com> Signed-off-by:
-
David S. Miller authored
Simon Horman says: ==================== nfp: support VLAN strip and insert this series adds support to the NFP driver for HW offload of both: * RX VLAN ctag/stag strip * TX VLAN ctag insert ==================== Signed-off-by: -
Diana Wang authored
Add support for TX VLAN ctag insert which may be configured via ethtool. e.g. # ethtool -K $DEV tx-vlan-offload on The NIC supplies VLAN insert information as packet metadata. The fields of this VLAN metadata are gotten from sk_buff, including vlan_proto and vlan tag. Configuration control bit NFP_NET_CFG_CTRL_TXVLAN_V2 is to signal availability of ctag-insert features of the firmware. NFDK is used to communicate via PCIE to NFP-3800 based NICs while NFD3 is used for other NICs supported by the NFP driver. The metadata format on tx side of NFD3 is different from NFDK. This feature is not currently implemented for NFDK. Signed-off-by:Diana Wang <na.wang@corigine.com> Reviewed-by:
-
Diana Wang authored
Add support for RX VLAN ctag/stag strip which may be configured via ethtool. e.g. # ethtool -K $DEV rx-vlan-offload on # ethtool -K $DEV rx-vlan-stag-hw-parse on Ctag-stripped and stag-stripped cannot be enabled at the same time because currently the kernel supports only one layer of VLAN stripping. The NIC supplies VLAN strip information as packet metadata. The fields of this VLAN metadata are: * strip flag: 1 for stripped; 0 for unstripped * tci: VLAN TCI ID * tpid: 1 for ETH_P_8021AD; 0 for ETH_P_8021Q Configuration control bits NFP_NET_CFG_CTRL_RXVLAN_V2 and NFP_NET_CFG_CTRL_RXQINQ are to signal availability of ctag-strip and stag-strip features of the firmware. Signed-off-by: -
David S. Miller authored
Lukas Wunner says: ==================== Deadlock no more in LAN95xx Second attempt at fixing a runtime resume deadlock in the LAN95xx USB driver: In short, the driver isn't using the "nopm" register accessors in portions of its runtime resume path, causing a deadlock. I'm fixing that by auto-detecting whether nopm accessors shall be used, instead of having to explicitly call them wherever it's necessary. As a byproduct, code size shrinks significantly (see diffstat below). Back in April I submitted a first attempt which was rejected by Alan Stern: https://lore.kernel.org/all/6710d8c18ff54139cdc538763ba544187c5a0cee.1651041411.git.lukas@wunner.de/ That approach only detected whether a PM callback is running concurrently, not whether the access is performed by the PM callback. I've come up with a different approach which should resolve the objection (see patch [1/3]). ==================== Signed-off-by:
-
Lukas Wunner authored
smsc95xx_read_reg() and smsc95xx_write_reg() call BUG_ON() if the struct usbnet pointer passed in is NULL. The functions have just been amended to dereference the pointer on entry. So the kernel now oopses if the pointer is NULL, eliminating the need for an explicit BUG_ON(). Signed-off-by:
Lukas Wunner <lukas@wunner.de> Signed-off-by:
-
