kernel.git/include/linux/netdevice.h, branch linux-6.11.y

net: fix races in netdev_tx_sent_queue()/dev_watchdog()

2024-11-01T01:02:36Z

[ Upstream commit 95ecba62e2fd201bcdcca636f5d774f1cd4f1458 ] Some workloads hit the infamous dev_watchdog() message: "NETDEV WATCHDOG: eth0 (xxxx): transmit queue XX timed out" It seems possible to hit this even for perfectly normal BQL enabled drivers: 1) Assume a TX queue was idle for more than dev->watchdog_timeo (5 seconds unless changed by the driver) 2) Assume a big packet is sent, exceeding current BQL limit. 3) Driver ndo_start_xmit() puts the packet in TX ring, and netdev_tx_sent_queue() is called. 4) QUEUE_STATE_STACK_XOFF could be set from netdev_tx_sent_queue() before txq->trans_start has been written. 5) txq->trans_start is written later, from netdev_start_xmit() if (rc == NETDEV_TX_OK) txq_trans_update(txq) dev_watchdog() running on another cpu could read the old txq->trans_start, and then see QUEUE_STATE_STACK_XOFF, because 5) did not happen yet. To solve the issue, write txq->trans_start right before one XOFF bit is set : - _QUEUE_STATE_DRV_XOFF from netif_tx_stop_queue() - __QUEUE_STATE_STACK_XOFF from netdev_tx_sent_queue() From dev_watchdog(), we have to read txq->state before txq->trans_start. Add memory barriers to enforce correct ordering. In the future, we could avoid writing over txq->trans_start for normal operations, and rename this field to txq->xoff_start_time. Fixes: bec251bc8b6a ("net: no longer stop all TX queues in dev_watchdog()") Signed-off-by: Eric Dumazet Reviewed-by: Willem de Bruijn Reviewed-by: Toke Høiland-Jørgensen Link: https://patch.msgid.link/20241015194118.3951657-1-edumazet@google.com Signed-off-by: Paolo Abeni Signed-off-by: Sasha Levin

net: napi: Prevent overflow of napi_defer_hard_irqs

2024-10-10T10:03:17Z

[ Upstream commit 08062af0a52107a243f7608fd972edb54ca5b7f8 ] In commit 6f8b12d661d0 ("net: napi: add hard irqs deferral feature") napi_defer_irqs was added to net_device and napi_defer_irqs_count was added to napi_struct, both as type int. This value never goes below zero, so there is not reason for it to be a signed int. Change the type for both from int to u32, and add an overflow check to sysfs to limit the value to S32_MAX. The limit of S32_MAX was chosen because the practical limit before this patch was S32_MAX (anything larger was an overflow) and thus there are no behavioral changes introduced. If the extra bit is needed in the future, the limit can be raised. Before this patch: $ sudo bash -c 'echo 2147483649 > /sys/class/net/eth4/napi_defer_hard_irqs' $ cat /sys/class/net/eth4/napi_defer_hard_irqs -2147483647 After this patch: $ sudo bash -c 'echo 2147483649 > /sys/class/net/eth4/napi_defer_hard_irqs' bash: line 0: echo: write error: Numerical result out of range Similarly, /sys/class/net/XXXXX/tx_queue_len is defined as unsigned: include/linux/netdevice.h: unsigned int tx_queue_len; And has an overflow check: dev_change_tx_queue_len(..., unsigned long new_len): if (new_len != (unsigned int)new_len) return -ERANGE; Suggested-by: Jakub Kicinski Signed-off-by: Joe Damato Reviewed-by: Eric Dumazet Link: https://patch.msgid.link/20240904153431.307932-1-jdamato@fastly.com Signed-off-by: Jakub Kicinski Signed-off-by: Sasha Levin

net: Fix gso_features_check to check for both dev->gso_{ipv4_,}max_size

2024-10-10T10:02:59Z

[ Upstream commit e609c959a939660c7519895f853dfa5624c6827a ] Commit 24ab059d2ebd ("net: check dev->gso_max_size in gso_features_check()") added a dev->gso_max_size test to gso_features_check() in order to fall back to GSO when needed. This was added as it was noticed that some drivers could misbehave if TSO packets get too big. However, the check doesn't respect dev->gso_ipv4_max_size limit. For instance, a device could be configured with BIG TCP for IPv4, but not IPv6. Therefore, add a netif_get_gso_max_size() equivalent to netif_get_gro_max_size() and use the helper to respect both limits before falling back to GSO engine. Fixes: 24ab059d2ebd ("net: check dev->gso_max_size in gso_features_check()") Signed-off-by: Daniel Borkmann Cc: Eric Dumazet Cc: Paolo Abeni Reviewed-by: Eric Dumazet Link: https://patch.msgid.link/20240923212242.15669-2-daniel@iogearbox.net Signed-off-by: Paolo Abeni Signed-off-by: Sasha Levin

net: Add netif_get_gro_max_size helper for GRO

2024-10-10T10:02:59Z

[ Upstream commit e8d4d34df715133c319fabcf63fdec684be75ff8 ] Add a small netif_get_gro_max_size() helper which returns the maximum IPv4 or IPv6 GRO size of the netdevice. We later add a netif_get_gso_max_size() equivalent as well for GSO, so that these helpers can be used consistently instead of open-coded checks. Signed-off-by: Daniel Borkmann Cc: Eric Dumazet Cc: Paolo Abeni Reviewed-by: Eric Dumazet Link: https://patch.msgid.link/20240923212242.15669-1-daniel@iogearbox.net Signed-off-by: Paolo Abeni Stable-dep-of: e609c959a939 ("net: Fix gso_features_check to check for both dev->gso_{ipv4_,}max_size") Signed-off-by: Sasha Levin

netdevice: define and allocate &net_device _properly_

2024-07-12T01:11:31Z

In fact, this structure contains a flexible array at the end, but historically its size, alignment etc., is calculated manually. There are several instances of the structure embedded into other structures, but also there's ongoing effort to remove them and we could in the meantime declare &net_device properly. Declare the array explicitly, use struct_size() and store the array size inside the structure, so that __counted_by() can be applied. Don't use PTR_ALIGN(), as SLUB itself tries its best to ensure the allocated buffer is aligned to what the user expects. Also, change its alignment from %NETDEV_ALIGN to the cacheline size as per several suggestions on the netdev ML. bloat-o-meter for vmlinux: free_netdev 445 440 -5 netdev_freemem 24 - -24 alloc_netdev_mqs 1481 1450 -31 On x86_64 with several NICs of different vendors, I was never able to get a &net_device pointer not aligned to the cacheline size after the change. Signed-off-by: Alexander Lobakin Signed-off-by: Breno Leitao Reviewed-by: Przemek Kitszel Reviewed-by: Eric Dumazet Reviewed-by: Kees Cook Link: https://patch.msgid.link/20240710113036.2125584-1-leitao@debian.org Signed-off-by: Jakub Kicinski

ethtool: move firmware flashing flag to struct ethtool_netdev_state

2024-07-04T22:45:15Z

Commit 31e0aa99dc02 ("ethtool: Veto some operations during firmware flashing process") added a flag module_fw_flash_in_progress to struct net_device. As this is ethtool related state, move it to the recently created struct ethtool_netdev_state, accessed via the 'ethtool' member of struct net_device. Suggested-by: Jakub Kicinski Signed-off-by: Edward Cree Reviewed-by: Michal Kubiak Reviewed-by: Ido Schimmel Link: https://patch.msgid.link/20240703121849.652893-1-edward.cree@amd.com Signed-off-by: Jakub Kicinski

net: move ethtool-related netdev state into its own struct

2024-06-29T01:53:17Z

net_dev->ethtool is a pointer to new struct ethtool_netdev_state, which currently contains only the wol_enabled field. Suggested-by: Jakub Kicinski Signed-off-by: Edward Cree Reviewed-by: Przemek Kitszel Link: https://patch.msgid.link/293a562278371de7534ed1eb17531838ca090633.1719502239.git.ecree.xilinx@gmail.com Signed-off-by: Jakub Kicinski

ethtool: Veto some operations during firmware flashing process

2024-06-28T09:48:22Z

Some operations cannot be performed during the firmware flashing process. For example: - Port must be down during the whole flashing process to avoid packet loss while committing reset for example. - Writing to EEPROM interrupts the flashing process, so operations like ethtool dump, module reset, get and set power mode should be vetoed. - Split port firmware flashing should be vetoed. In order to veto those scenarios, add a flag in 'struct net_device' that indicates when a firmware flash is taking place on the module and use it to prevent interruptions during the process. Signed-off-by: Danielle Ratson Reviewed-by: Petr Machata Reviewed-by: Andrew Lunn Signed-off-by: David S. Miller

ethtool: provide customized dim profile management

2024-06-26T00:15:06Z

The NetDIM library, currently leveraged by an array of NICs, delivers excellent acceleration benefits. Nevertheless, NICs vary significantly in their dim profile list prerequisites. Specifically, virtio-net backends may present diverse sw or hw device implementation, making a one-size-fits-all parameter list impractical. On Alibaba Cloud, the virtio DPU's performance under the default DIM profile falls short of expectations, partly due to a mismatch in parameter configuration. I also noticed that ice/idpf/ena and other NICs have customized profilelist or placed some restrictions on dim capabilities. Motivated by this, I tried adding new params for "ethtool -C" that provides a per-device control to modify and access a device's interrupt parameters. Usage ======== The target NIC is named ethx. Assume that ethx only declares support for rx profile setting (with DIM_PROFILE_RX flag set in profile_flags) and supports modification of usec and pkt fields. 1. Query the currently customized list of the device $ ethtool -c ethx ... rx-profile: {.usec = 1, .pkts = 256, .comps = n/a,}, {.usec = 8, .pkts = 256, .comps = n/a,}, {.usec = 64, .pkts = 256, .comps = n/a,}, {.usec = 128, .pkts = 256, .comps = n/a,}, {.usec = 256, .pkts = 256, .comps = n/a,} tx-profile: n/a 2. Tune $ ethtool -C ethx rx-profile 1,1,n_2,n,n_3,3,n_4,4,n_n,5,n "n" means do not modify this field. $ ethtool -c ethx ... rx-profile: {.usec = 1, .pkts = 1, .comps = n/a,}, {.usec = 2, .pkts = 256, .comps = n/a,}, {.usec = 3, .pkts = 3, .comps = n/a,}, {.usec = 4, .pkts = 4, .comps = n/a,}, {.usec = 256, .pkts = 5, .comps = n/a,} tx-profile: n/a 3. Hint If the device does not support some type of customized dim profiles, the corresponding "n/a" will display. If the "n/a" field is being modified, -EOPNOTSUPP will be reported. Signed-off-by: Heng Qi Reviewed-by: Simon Horman Link: https://patch.msgid.link/20240621101353.107425-4-hengqi@linux.alibaba.com Signed-off-by: Jakub Kicinski

dev: Use nested-BH locking for softnet_data.process_queue.

2024-06-24T23:41:23Z

softnet_data::process_queue is a per-CPU variable and relies on disabled BH for its locking. Without per-CPU locking in local_bh_disable() on PREEMPT_RT this data structure requires explicit locking. softnet_data::input_queue_head can be updated lockless. This is fine because this value is only update CPU local by the local backlog_napi thread. Add a local_lock_t to softnet_data and use local_lock_nested_bh() for locking of process_queue. This change adds only lockdep coverage and does not alter the functional behaviour for !PREEMPT_RT. Signed-off-by: Sebastian Andrzej Siewior Link: https://patch.msgid.link/20240620132727.660738-11-bigeasy@linutronix.de Signed-off-by: Jakub Kicinski