- Add new DQO datapath structures:
- `gve_rx_buf_queue_dqo`
- `gve_rx_compl_queue_dqo`
- `gve_rx_buf_state_dqo`
- `gve_tx_desc_dqo`
- `gve_tx_pending_packet_dqo`
- Incorporate these into the existing ring data structures:
- `gve_rx_ring`
- `gve_tx_ring`
Noteworthy mentions:
- `gve_rx_buf_state` represents an RX buffer which was posted to HW.
Each RX queue has an array of these objects and the index into the
array is used as the buffer_id when posted to HW.
- `gve_tx_pending_packet_dqo` is treated similarly for TX queues. The
completion_tag is the index into the array.
- These two structures have links for linked lists which are represented
by 16b indexes into a contiguous array of these structures.
This reduces memory footprint compared to 64b pointers.
- We use unions for the writeable datapath structures to reduce cache
footprint. GQI specific members will renamed like DQO members in a
future patch.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
General description of rings and descriptors:
TX ring is used for sending TX packet buffers to the NIC. It has the
following descriptors:
- `gve_tx_pkt_desc_dqo` - Data buffer descriptor
- `gve_tx_tso_context_desc_dqo` - TSO context descriptor
- `gve_tx_general_context_desc_dqo` - Generic metadata descriptor
Metadata is a collection of 12 bytes. We define `gve_tx_metadata_dqo`
which represents the logical interpetation of the metadata bytes. It's
helpful to define this structure because the metadata bytes exist in
multiple descriptor types (including `gve_tx_tso_context_desc_dqo`),
and the device requires same field has the same value in all
descriptors.
The TX completion ring is used to receive completions from the NIC.
Having a separate ring allows for completions to be out of order. The
completion descriptor `gve_tx_compl_desc` has several different types,
most important are packet and descriptor completions. Descriptor
completions are used to notify the driver when descriptors sent on the
TX ring are done being consumed. The descriptor completion is only used
to signal that space is cleared in the TX ring. A packet completion will
be received when a packet transmitted on the TX queue is done being
transmitted.
In addition there are "miss" and "reinjection" completions. The device
implements a "flow-miss model". Most packets will simply receive a
packet completion. The flow-miss system may choose to process a packet
based on its contents. A TX packet which experiences a flow miss would
receive a miss completion followed by a later reinjection completion.
The miss-completion is received when the packet starts to be processed
by the flow-miss system and the reinjection completion is received when
the flow-miss system completes processing the packet and sends it on the
wire.
The RX buffer ring is used to send buffers to HW via the
`gve_rx_desc_dqo` descriptor.
Received packets are put into the RX queue by the device, which
populates the `gve_rx_compl_desc_dqo` descriptor. The RX descriptors
refer to buffers posted by the buffer queue. Received buffers may be
returned out of order, such as when HW LRO is enabled.
Important concepts:
- "TX" and "RX buffer" queues, which send descriptors to the device, use
MMIO doorbells to notify the device of new descriptors.
- "RX" and "TX completion" queues, which receive descriptors from the
device, use a "generation bit" to know when a descriptor was populated
by the device. The driver initializes all bits with the "current
generation". The device will populate received descriptors with the
"next generation" which is inverted from the current generation. When
the ring wraps, the current/next generation are swapped.
- It's the driver's responsibility to ensure that the RX and TX
completion queues are not overrun. This can be accomplished by
limiting the number of descriptors posted to HW.
- TX packets have a 16 bit completion_tag and RX buffers have a 16 bit
buffer_id. These will be returned on the TX completion and RX queues
respectively to let the driver know which packet/buffer was completed.
Bitfields are used to describe descriptor fields. This notation is more
concise and readable than shift-and-mask. It is possible because the
driver is restricted to little endian platforms.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Unlike GQI, DQO RX descriptors do not contain the L3 and L4 type of the
packet. L3 and L4 types are necessary in order to set the hash and csum
on RX SKBs correctly.
DQO RX descriptors instead contain a 10 bit PTYPE index. The PTYPE map
enables the device to tell the driver how to map from PTYPE index to
L3/L4 type.
The device doesn't provide any guarantees about the range of possible
PTYPEs, so we just use a 1024 entry array to implement a fast mapping
structure.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
- In addition to TX and RX queues, DQO has TX completion and RX buffer
queues.
- TX completions are received when the device has completed sending a
packet on the wire.
- RX buffers are posted on a separate queue form the RX completions.
- DQO descriptor rings are allowed to be smaller than PAGE_SIZE.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The currently supported queue formats are:
- GQI_RDA - GQI with raw DMA addressing
- GQI_QPL - GQI with queue page list
- DQO_RDA - DQO with raw DMA addressing
The old `gve_priv.raw_addressing` value is only used for GQI_RDA, so we
remove it in favor of just checking against GQI_RDA
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current model uses an integer ID and a fixed size struct for the
parameters of each device option.
The new model allows the device option structs to grow in size over
time. A driver may assume that changes to device option structs will
always be appended.
New device options will also generally have a
`supported_features_mask` so that the driver knows which fields within a
particular device option are enabled.
`gve_device_option.feat_mask` is changed to `required_features_mask`,
and it is a bitmask which must match the value expected by the driver.
This gives the device the ability to break backwards compatibility with
old drivers for certain features by blocking the old drivers from trying
to use the feature.
We maintain ABI compatibility with the old model for
GVE_DEV_OPT_ID_RAW_ADDRESSING in case a driver is using a device which
does not support the new model.
This patch introduces some new terminology:
RDA - Raw DMA Addressing - Buffers associated with SKBs are directly DMA
mapped and read/updated by the device.
QPL - Queue Page Lists - Driver uses bounce buffers which are DMA mapped
with the device for read/write and data is copied from/to SKBs.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using `page_offset` like a boolean means a page may only be split into
two sections. With page sizes larger than 4k, this can be very wasteful.
Future commits in this patchset use `struct gve_rx_slot_page_info` in a
way which supports a fixed buffer size and a variable page size.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Future use cases will have a different padding value.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
These functions will be shared by the GQI and DQO variants of the GVNIC
driver as of follow-up patches in this series.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
DQO is a new descriptor format for our next generation virtual NIC.
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Modify the netdev_dbg content from int to char * in usbnet_defer_kevent(),
this looks more readable.
Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
Signed-off-by: David S. Miller <davem@davemloft.net>
Steen Hegelund says:
====================
Adding the Sparx5i Switch Driver
This series provides the Microchip Sparx5i Switch Driver
The SparX-5 Enterprise Ethernet switch family provides a rich set of
Enterprise switching features such as advanced TCAM-based VLAN and QoS
processing enabling delivery of differentiated services, and security
through TCAMbased frame processing using versatile content aware processor
(VCAP). IPv4/IPv6 Layer 3 (L3) unicast and multicast routing is supported
with up to 18K IPv4/9K IPv6 unicast LPM entries and up to 9K IPv4/3K IPv6
(S,G) multicast groups. L3 security features include source guard and
reverse path forwarding (uRPF) tasks. Additional L3 features include
VRF-Lite and IP tunnels (IP over GRE/IP).
The SparX-5 switch family features a highly flexible set of Ethernet ports
with support for 10G and 25G aggregation links, QSGMII, USGMII, and
USXGMII. The device integrates a powerful 1 GHz dual-core ARM® Cortex®-A53
CPU enabling full management of the switch and advanced Enterprise
applications.
The SparX-5 switch family targets managed Layer 2 and Layer 3 equipment in
SMB, SME, and Enterprise where high port count 1G/2.5G/5G/10G switching
with 10G/25G aggregation links is required.
The SparX-5 switch family consists of following SKUs:
VSC7546 SparX-5-64 supports up to 64 Gbps of bandwidth with the following
primary port configurations.
- 6 ×10G
- 16 × 2.5G + 2 × 10G
- 24 × 1G + 4 × 10G
VSC7549 SparX-5-90 supports up to 90 Gbps of bandwidth with the following
primary port configurations.
- 9 × 10G
- 16 × 2.5G + 4 × 10G
- 48 × 1G + 4 × 10G
VSC7552 SparX-5-128 supports up to 128 Gbps of bandwidth with the
following primary port configurations.
- 12 × 10G
- 6 x 10G + 2 x 25G
- 16 × 2.5G + 8 × 10G
- 48 × 1G + 8 × 10G
VSC7556 SparX-5-160 supports up to 160 Gbps of bandwidth with the
following primary port configurations.
- 16 × 10G
- 10 × 10G + 2 × 25G
- 16 × 2.5G + 10 × 10G
- 48 × 1G + 10 × 10G
VSC7558 SparX-5-200 supports up to 200 Gbps of bandwidth with the
following primary port configurations.
- 20 × 10G
- 8 × 25G
In addition, the device supports one 10/100/1000/2500/5000 Mbps
SGMII/SerDes node processor interface (NPI) Ethernet port.
Time sensitive networking (TSN) is supported through a comprehensive set of
features including frame preemption, cut-through, frame replication and
elimination for reliability, enhanced scheduling: credit-based shaping,
time-aware shaping, cyclic queuing, and forwarding, and per-stream policing
and filtering.
Together with IEEE 1588 and IEEE 802.1AS support, this guarantees
low-latency deterministic networking for Industrial Ethernet.
The Sparx5i support is developed on the PCB134 and PCB135 evaluation boards.
- PCB134 main networking features:
- 12x SFP+ front 10G module slots (connected to Sparx5i through SFI).
- 8x SFP28 front 25G module slots (connected to Sparx5i through SFI high
speed).
- Optional, one additional 10/100/1000BASE-T (RJ45) Ethernet port
(on-board VSC8211 PHY connected to Sparx5i through SGMII).
- PCB135 main networking features:
- 48x1G (10/100/1000M) RJ45 front ports using 12xVSC8514 QuadPHY’s each
connected to VSC7558 through QSGMII.
- 4x10G (1G/2.5G/5G/10G) RJ45 front ports using the AQR407 10G QuadPHY
each port connects to VSC7558 through SFI.
- 4x SFP28 25G module slots on back connected to VSC7558 through SFI high
speed.
- Optional, one additional 1G (10/100/1000M) RJ45 port using an on-board
VSC8211 PHY, which can be connected to VSC7558 NPI port through SGMII
using a loopback add-on PCB)
This series provides support for:
- SFPs and DAC cables via PHYLINK with a number of 5G, 10G and 25G
devices and media types.
- Port module configuration for 10M to 25G speeds with SGMII, QSGMII,
1000BASEX, 2500BASEX and 10GBASER as appropriate for these modes.
- SerDes configuration via the Sparx5i SerDes driver (see below).
- Host mode providing register based injection and extraction.
- Switch mode providing MAC/VLAN table learning and Layer2 switching
offloaded to the Sparx5i switch.
- STP state, VLAN support, host/bridge port mode, Forwarding DB, and
configuration and statistics via ethtool.
More support will be added at a later stage.
The Sparx5i Chip Register Model can be browsed at this location:
https://github.com/microchip-ung/sparx-5_reginfo
and the datasheet is available here:
https://ww1.microchip.com/downloads/en/DeviceDoc/SparX-5_Family_L2L3_Enterprise_10G_Ethernet_Switches_Datasheet_00003822B.pdf
The series depends on the following series currently on their way
into the kernel:
- 25G Base-R phy mode
Link: https://lore.kernel.org/r/20210611125453.313308-1-steen.hegelund@microchip.com/
- Sparx5 Reset Driver
Link: https://lore.kernel.org/r/20210416084054.2922327-1-steen.hegelund@microchip.com/
ChangeLog:
v5:
- cover letter
- updated the description to match the latest data sheets
- basic driver
- added error message in case of reset controller error
- port struct: replacing has_sfp with inband, adding pause_adv
- host mode
- port cleanup: unregisters netdevs and then removes phylink etc
- checking for pause_adv when comparing port config changes
- getting duplex and pause state in the link_up callback.
- getting inband, autoneg and pause_adv config in the pcs_config
callback.
- port
- use only the pause_adv bits when getting aneg status
- use the inband state when updating the PCS and port config
v4:
- basic driver:
Using devm_reset_control_get_optional_shared to get the reset
control, and let the reset framework check if it is valid.
- host mode (phylink):
Use the PCS operations to get state and update configuration.
Removed the setting of interface modes. Let phylink control this.
Using the new 5gbase-r and 25gbase-r modes.
Using a helper function to check if one of the 3 base-r modes has
been selected.
Currently it will not be possible to change the interface mode by
changing the speed (e.g via ethtool). This will be added later.
v3:
- basic driver:
- removed unneeded braces
- release reference to ports node after use
- use dev_err_probe to handle DEFER
- update error value when bailing out (a few cases)
- updated formatting of port struct and grouping of bool values
- simplified the spx5_rmw and spx5_inst_rmw inline functions
- host mode (netdev):
- removed lockless flag
- added port timer init
- host mode (packet - manual injection):
- updated error counters in error situations
- implemented timer handling of watermark threshold: stop and
restart netif queues.
- fixed error message handling (rate limited)
- fixed comment style error
- used DIV_ROUND_UP macro
- removed a debug message for open ports
v2:
- Updated bindings:
- drop minItems for the reg property
- Statistics implementation:
- Reorganized statistics into ethtool groups:
eth-phy, eth-mac, eth-ctrl, rmon
as defined by the IEEE 802.3 categories and RFC 2819.
- The remaining statistics are provided by the classic ethtool
statistics command.
- Hostmode support:
- Removed netdev renaming
- Validate ethernet address in sparx5_set_mac_address()
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This provides the configuration for the currently available evaluation
boards PCB134 and PCB135.
The series depends on the following series currently on its way
into the kernel:
- Sparx5 Reset Driver
Link: https://lore.kernel.org/r/20210416084054.2922327-1-steen.hegelund@microchip.com/
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds statistic counters for the network interfaces provided
by the driver. It also adds CPU port counters (which are not
exposed by ethtool).
This also adds support for configuring the network interface
parameters via ethtool: speed, duplex, aneg etc.
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This configures the Sparx5 calendars according to the bandwidth
requested in the Device Tree nodes.
It also checks if the total requested bandwidth is within the
specs of the detected Sparx5 models limits.
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds SwitchDev support by hardware offloading the
software bridge.
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds Sparx5 VLAN support.
Sparx5 has more VLAN features than provided here, but these will be added
in later series. For now we only add the basic L2 features.
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds the Sparx5 MAC tables: listening for MAC table updates and
updating on request.
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This add configuration of the Sparx5 port module instances.
Sparx5 has in total 65 logical ports (denoted D0 to D64) and 33
physical SerDes connections (S0 to S32). The 65th port (D64) is fixed
allocated to SerDes0 (S0). The remaining 64 ports can in various
multiplexing scenarios be connected to the remaining 32 SerDes using
QSGMII, or USGMII or USXGMII extenders. 32 of the ports can have a 1:1
mapping to the 32 SerDes.
Some additional ports (D65 to D69) are internal to the device and do not
connect to port modules or SerDes macros. For example, internal ports are
used for frame injection and extraction to the CPU queues.
The 65 logical ports are split up into the following blocks.
- 13 x 5G ports (D0-D11, D64)
- 32 x 2G5 ports (D16-D47)
- 12 x 10G ports (D12-D15, D48-D55)
- 8 x 25G ports (D56-D63)
Each logical port supports different line speeds, and depending on the
speeds supported, different port modules (MAC+PCS) are needed. A port
supporting 5 Gbps, 10 Gbps, or 25 Gbps as maximum line speed, will have a
DEV5G, DEV10G, or DEV25G module to support the 5 Gbps, 10 Gbps (incl 5
Gbps), or 25 Gbps (including 10 Gbps and 5 Gbps) speeds. As well as, it
will have a shadow DEV2G5 port module to support the lower speeds
(10/100/1000/2500Mbps). When a port needs to operate at lower speed and the
shadow DEV2G5 needs to be connected to its corresponding SerDes
Not all interface modes are supported in this series, but will be added at
a later stage.
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds netdevs and phylink support for the ports in the switch.
It also adds register based injection and extraction for these ports.
Frame DMA support for injection and extraction will be added in a later
series.
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds the Sparx5 basic SwitchDev driver framework with IO range
mapping, switch device detection and core clock configuration.
Support for ports, phylink, netdev, mactable etc. are in the following
patches.
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Reviewed-by: Philipp Zabel <p.zabel@pengutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Document the Sparx5 switch device driver bindings
Signed-off-by: Steen Hegelund <steen.hegelund@microchip.com>
Signed-off-by: Lars Povlsen <lars.povlsen@microchip.com>
Signed-off-by: Bjarni Jonasson <bjarni.jonasson@microchip.com>
Reviewed-by: Rob Herring <robh@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The fallback case of fwnode_mdbiobus_register()
(relevant for !CONFIG_FWNODE_MDIO) was defined with wrong
argument name, causing a compilation error. Fix that.
Signed-off-by: Marcin Wojtas <mw@semihalf.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Dave observed number of machines hitting OOM on the UDP send
path. The workload seems to be sending large UDP packets over
loopback. Since loopback has MTU of 64k kernel will try to
allocate an skb with up to 64k of head space. This has a good
chance of failing under memory pressure. What's worse if
the message length is <32k the allocation may trigger an
OOM killer.
This is entirely avoidable, we can use an skb with page frags.
af_unix solves a similar problem by limiting the head
length to SKB_MAX_ALLOC. This seems like a good and simple
approach. It means that UDP messages > 16kB will now
use fragments if underlying device supports SG, if extra
allocator pressure causes regressions in real workloads
we can switch to trying the large allocation first and
falling back.
v4: pre-calculate all the additions to alloclen so
we can be sure it won't go over order-2
Reported-by: Dave Jones <dsj@fb.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make sure that SO_NETNS_COOKIE returns a non-zero value, and
that sockets from different namespaces have a distinct cookie
value.
Signed-off-by: Lorenz Bauer <lmb@cloudflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It's getting more common to run nested container environments for
testing cloud software. One of such examples is Kind [1] which runs a
Kubernetes cluster in Docker containers on a single host. Each container
acts as a Kubernetes node, and thus can run any Pod (aka container)
inside the former. This approach simplifies testing a lot, as it
eliminates complicated VM setups.
Unfortunately, such a setup breaks some functionality when cgroupv2 BPF
programs are used for load-balancing. The load-balancer BPF program
needs to detect whether a request originates from the host netns or a
container netns in order to allow some access, e.g. to a service via a
loopback IP address. Typically, the programs detect this by comparing
netns cookies with the one of the init ns via a call to
bpf_get_netns_cookie(NULL). However, in nested environments the latter
cannot be used given the Kubernetes node's netns is outside the init ns.
To fix this, we need to pass the Kubernetes node netns cookie to the
program in a different way: by extending getsockopt() with a
SO_NETNS_COOKIE option, the orchestrator which runs in the Kubernetes
node netns can retrieve the cookie and pass it to the program instead.
Thus, this is following up on Eric's commit 3d368ab87c ("net:
initialize net->net_cookie at netns setup") to allow retrieval via
SO_NETNS_COOKIE. This is also in line in how we retrieve socket cookie
via SO_COOKIE.
[1] https://kind.sigs.k8s.io/
Signed-off-by: Lorenz Bauer <lmb@cloudflare.com>
Signed-off-by: Martynas Pumputis <m@lambda.lt>
Cc: Eric Dumazet <edumazet@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The cpsw driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Grygorii Strashko <grygorii.strashko@ti.com>
Reviewed-by: Grygorii Strashko <grygorii.strashko@ti.com>
Cc: linux-omap@vger.kernel.org
Link: https://lore.kernel.org/bpf/20210624160609.292325-20-toke@redhat.com
The stmmac driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@foss.st.com>
Cc: Jose Abreu <joabreu@synopsys.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-19-toke@redhat.com
The netsec driver has a rcu_read_lock()/rcu_read_unlock() pair around the
full RX loop, covering everything up to and including xdp_do_flush(). This
is actually the correct behaviour, but because it all happens in a single
NAPI poll cycle (and thus under local_bh_disable()), it is also technically
redundant.
With the addition of RCU annotations to the XDP_REDIRECT map types that
take bh execution into account, lockdep even understands this to be safe,
so there's really no reason to keep the rcu_read_lock() around anymore, so
let's just remove it.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Cc: Jassi Brar <jaswinder.singh@linaro.org>
Link: https://lore.kernel.org/bpf/20210624160609.292325-18-toke@redhat.com
The sfc driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Edward Cree <ecree.xilinx@gmail.com>
Cc: Martin Habets <habetsm.xilinx@gmail.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-17-toke@redhat.com
The qede driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Ariel Elior <aelior@marvell.com>
Cc: gr-everest-linux-l2@marvell.com
Link: https://lore.kernel.org/bpf/20210624160609.292325-16-toke@redhat.com
The nfp driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small.
While this is not actually an issue for the nfp driver because it doesn't
support XDP_REDIRECT (and thus doesn't call xdp_do_flush()), the
rcu_read_lock() is still unneeded. And With the addition of RCU annotations
to the XDP_REDIRECT map types that take bh execution into account, lockdep
even understands this to be safe, so there's really no reason to keep it
around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Cc: oss-drivers@netronome.com
Link: https://lore.kernel.org/bpf/20210624160609.292325-15-toke@redhat.com
The mlx4 driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around. Also switch the RCU
dereferences in the driver loop itself to the _bh variants.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-14-toke@redhat.com
The mvneta and mvpp2 drivers have rcu_read_lock()/rcu_read_unlock() pairs
around XDP program invocations. However, the actual lifetime of the objects
referred by the XDP program invocation is longer, all the way through to
the call to xdp_do_flush(), making the scope of the rcu_read_lock() too
small. This turns out to be harmless because it all happens in a single
NAPI poll cycle (and thus under local_bh_disable()), but it makes the
rcu_read_lock() misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Marcin Wojtas <mw@semihalf.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-13-toke@redhat.com
The Intel drivers all have rcu_read_lock()/rcu_read_unlock() pairs around
XDP program invocations. However, the actual lifetime of the objects
referred by the XDP program invocation is longer, all the way through to
the call to xdp_do_flush(), making the scope of the rcu_read_lock() too
small. This turns out to be harmless because it all happens in a single
NAPI poll cycle (and thus under local_bh_disable()), but it makes the
rcu_read_lock() misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> # i40e
Cc: Jesse Brandeburg <jesse.brandeburg@intel.com>
Cc: Tony Nguyen <anthony.l.nguyen@intel.com>
Cc: intel-wired-lan@lists.osuosl.org
Link: https://lore.kernel.org/bpf/20210624160609.292325-12-toke@redhat.com
The dpaa and dpaa2 drivers have rcu_read_lock()/rcu_read_unlock() pairs
around XDP program invocations. However, the actual lifetime of the objects
referred by the XDP program invocation is longer, all the way through to
the call to xdp_do_flush(), making the scope of the rcu_read_lock() too
small. This turns out to be harmless because it all happens in a single
NAPI poll cycle (and thus under local_bh_disable()), but it makes the
rcu_read_lock() misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Camelia Groza <camelia.groza@nxp.com>
Cc: Ioana Radulescu <ruxandra.radulescu@nxp.com>
Cc: Madalin Bucur <madalin.bucur@nxp.com>
Cc: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-11-toke@redhat.com
The thunderx driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Sunil Goutham <sgoutham@marvell.com>
Cc: linux-arm-kernel@lists.infradead.org
Link: https://lore.kernel.org/bpf/20210624160609.292325-10-toke@redhat.com
The bnxt driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Michael Chan <michael.chan@broadcom.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-9-toke@redhat.com
The ena driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Saeed Bishara <saeedb@amazon.com>
Cc: Guy Tzalik <gtzalik@amazon.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-8-toke@redhat.com
The rcu_read_lock() call in cls_bpf and act_bpf are redundant: on the TX
side, there's already a call to rcu_read_lock_bh() in __dev_queue_xmit(),
and on RX there's a covering rcu_read_lock() in
netif_receive_skb{,_list}_internal().
With the previous patches we also amended the lockdep checks in the map
code to not require any particular RCU flavour, so we can just get rid of
the rcu_read_lock()s.
Suggested-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210624160609.292325-7-toke@redhat.com
XDP_REDIRECT works by a three-step process: the bpf_redirect() and
bpf_redirect_map() helpers will lookup the target of the redirect and store
it (along with some other metadata) in a per-CPU struct bpf_redirect_info.
Next, when the program returns the XDP_REDIRECT return code, the driver
will call xdp_do_redirect() which will use the information thus stored to
actually enqueue the frame into a bulk queue structure (that differs
slightly by map type, but shares the same principle). Finally, before
exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will
flush all the different bulk queues, thus completing the redirect.
Pointers to the map entries will be kept around for this whole sequence of
steps, protected by RCU. However, there is no top-level rcu_read_lock() in
the core code; instead drivers add their own rcu_read_lock() around the XDP
portions of the code, but somewhat inconsistently as Martin discovered[0].
However, things still work because everything happens inside a single NAPI
poll sequence, which means it's between a pair of calls to
local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could
document this intention by using rcu_dereference_check() with
rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and
lockdep to verify that everything is done correctly.
This patch does just that: we add an __rcu annotation to the map entry
pointers and remove the various comments explaining the NAPI poll assurance
strewn through devmap.c in favour of a longer explanation in filter.c. The
goal is to have one coherent documentation of the entire flow, and rely on
the RCU annotations as a "standard" way of communicating the flow in the
map code (which can additionally be understood by sparse and lockdep).
The RCU annotation replacements result in a fairly straight-forward
replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE()
becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the
proper constructs to cast the pointer back and forth between __rcu and
__kernel address space (for the benefit of sparse). The one complication is
that xskmap has a few constructions where double-pointers are passed back
and forth; these simply all gain __rcu annotations, and only the final
reference/dereference to the inner-most pointer gets changed.
With this, everything can be run through sparse without eliciting
complaints, and lockdep can verify correctness even without the use of
rcu_read_lock() in the drivers. Subsequent patches will clean these up from
the drivers.
[0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/
[1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
XDP programs are called from a NAPI poll context, which means the RCU
reference liveness is ensured by local_bh_disable(). Add
rcu_read_lock_bh_held() as a condition to the RCU checks for map lookups so
lockdep understands that the dereferences are safe from inside *either* an
rcu_read_lock() section *or* a local_bh_disable() section. While both
bh_disabled and rcu_read_lock() provide RCU protection, they are
semantically distinct, so we need both conditions to prevent lockdep
complaints.
This change is done in preparation for removing the redundant
rcu_read_lock()s from drivers.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-5-toke@redhat.com
This commit gives an example of non-obvious RCU reader/updater pairing
in the guise of the XDP feature in networking, which calls BPF programs
from network-driver NAPI (softirq) context.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210624160609.292325-4-toke@redhat.com
This commit clarifies which primitives readers can use given that the
corresponding updaters have made a specific choice. This commit also adds
this information for the various RCU Tasks flavors. While in the area, it
removes a paragraph that no longer applies in any straightforward manner.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210624160609.292325-3-toke@redhat.com
The xchg() and cmpxchg() functions are sometimes used to carry out RCU
updates. Unfortunately, this can result in sparse warnings for both
the old-value and new-value arguments, as well as for the return value.
The arguments can be dealt with using RCU_INITIALIZER():
old_p = xchg(&p, RCU_INITIALIZER(new_p));
But a sparse warning still remains due to assigning the __rcu pointer
returned from xchg to the (most likely) non-__rcu pointer old_p.
This commit therefore provides an unrcu_pointer() macro that strips
the __rcu. This macro can be used as follows:
old_p = unrcu_pointer(xchg(&p, RCU_INITIALIZER(new_p)));
Reported-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210624160609.292325-2-toke@redhat.com
Stephen reported a warning:
drivers/net/wireless/intel/iwlwifi/fw/acpi.c:720:12: warning: 'iwl_acpi_eval_dsm_func' defined but not used [-Wunused-function]
The warning is correct and the function is not used anywhere, so let's
just remove it.
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Fixes: 7119f02b5d ("iwlwifi: mvm: support BIOS enable/disable for 11ax in Russia")
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Acked-by: Luca Coelho <luciano.coelho@intel.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Link: https://lore.kernel.org/r/20210624052918.4946-1-kvalo@codeaurora.org
Fix erroneous code that leads to unreferenced objects. During H2C
operations, some functions returned without freeing the memory that only
the function have access to. Release these objects when they're no longer
needed to avoid potentially memory leaks.
Signed-off-by: Po-Hao Huang <phhuang@realtek.com>
Signed-off-by: Ping-Ke Shih <pkshih@realtek.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Link: https://lore.kernel.org/r/20210624023459.10294-1-pkshih@realtek.com
With any regulatory domain requests coming from either user space or
802.11 IE (Information Element), the country is coded in ISO3166
standard. It needs to be translated to firmware country code and
revision with the mapping info in settings->country_codes table.
Support populate country_codes table by parsing the mapping from DT.
The BRCMF_BUSTYPE_SDIO bus_type check gets separated from general DT
validation, so that country code can be handled as general part rather
than SDIO bus specific one.
Signed-off-by: Shawn Guo <shawn.guo@linaro.org>
Reviewed-by: Arend van Spriel <arend.vanspriel@broadcom.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Link: https://lore.kernel.org/r/20210417075428.2671-1-shawn.guo@linaro.org
Since we no longer modify gso_size, it is now theoretically
safe to not set SKB_GSO_DODGY and reset gso_segs to zero.
This also means the skb_is_gso_tcp() check should no longer
be necessary.
Unfortunately we cannot remove the skb_{decrease,increase}_gso_size()
helpers, as they are still used elsewhere:
bpf_skb_net_grow() without BPF_F_ADJ_ROOM_FIXED_GSO
bpf_skb_net_shrink() without BPF_F_ADJ_ROOM_FIXED_GSO
net/core/lwt_bpf.c's handle_gso_type()
Signed-off-by: Maciej Żenczykowski <maze@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Dongseok Yi <dseok.yi@samsung.com>
Cc: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/bpf/20210617000953.2787453-3-zenczykowski@gmail.com
This is technically a backwards incompatible change in behaviour, but I'm
going to argue that it is very unlikely to break things, and likely to fix
*far* more then it breaks.
In no particular order, various reasons follow:
(a) I've long had a bug assigned to myself to debug a super rare kernel crash
on Android Pixel phones which can (per stacktrace) be traced back to BPF clat
IPv6 to IPv4 protocol conversion causing some sort of ugly failure much later
on during transmit deep in the GSO engine, AFAICT precisely because of this
change to gso_size, though I've never been able to manually reproduce it. I
believe it may be related to the particular network offload support of attached
USB ethernet dongle being used for tethering off of an IPv6-only cellular
connection. The reason might be we end up with more segments than max permitted,
or with a GSO packet with only one segment... (either way we break some
assumption and hit a BUG_ON)
(b) There is no check that the gso_size is > 20 when reducing it by 20, so we
might end up with a negative (or underflowing) gso_size or a gso_size of 0.
This can't possibly be good. Indeed this is probably somehow exploitable (or
at least can result in a kernel crash) by delivering crafted packets and perhaps
triggering an infinite loop or a divide by zero... As a reminder: gso_size (MSS)
is related to MTU, but not directly derived from it: gso_size/MSS may be
significantly smaller then one would get by deriving from local MTU. And on
some NICs (which do loose MTU checking on receive, it may even potentially be
larger, for example my work pc with 1500 MTU can receive 1520 byte frames [and
sometimes does due to bugs in a vendor plat46 implementation]). Indeed even just
going from 21 to 1 is potentially problematic because it increases the number
of segments by a factor of 21 (think DoS, or some other crash due to too many
segments).
(c) It's always safe to not increase the gso_size, because it doesn't result in
the max packet size increasing. So the skb_increase_gso_size() call was always
unnecessary for correctness (and outright undesirable, see later). As such the
only part which is potentially dangerous (ie. could cause backwards compatibility
issues) is the removal of the skb_decrease_gso_size() call.
(d) If the packets are ultimately destined to the local device, then there is
absolutely no benefit to playing around with gso_size. It only matters if the
packets will egress the device. ie. we're either forwarding, or transmitting
from the device.
(e) This logic only triggers for packets which are GSO. It does not trigger for
skbs which are not GSO. It will not convert a non-GSO MTU sized packet into a
GSO packet (and you don't even know what the MTU is, so you can't even fix it).
As such your transmit path must *already* be able to handle an MTU 20 bytes
larger then your receive path (for IPv4 to IPv6 translation) - and indeed 28
bytes larger due to IPv4 fragments. Thus removing the skb_decrease_gso_size()
call doesn't actually increase the size of the packets your transmit side must
be able to handle. ie. to handle non-GSO max-MTU packets, the IPv4/IPv6 device/
route MTUs must already be set correctly. Since for example with an IPv4 egress
MTU of 1500, IPv4 to IPv6 translation will already build 1520 byte IPv6 frames,
so you need a 1520 byte device MTU. This means if your IPv6 device's egress
MTU is 1280, your IPv4 route must be 1260 (and actually 1252, because of the
need to handle fragments). This is to handle normal non-GSO packets. Thus the
reduction is simply not needed for GSO packets, because when they're correctly
built, they will already be the right size.
(f) TSO/GSO should be able to exactly undo GRO: the number of packets (TCP
segments) should not be modified, so that TCP's MSS counting works correctly
(this matters for congestion control). If protocol conversion changes the
gso_size, then the number of TCP segments may increase or decrease. Packet loss
after protocol conversion can result in partial loss of MSS segments that the
sender sent. How's the sending TCP stack going to react to receiving ACKs/SACKs
in the middle of the segments it sent?
(g) skb_{decrease,increase}_gso_size() are already no-ops for GSO_BY_FRAGS
case (besides triggering WARN_ON_ONCE). This means you already cannot guarantee
that gso_size (and thus resulting packet MTU) is changed. ie. you must assume
it won't be changed.
(h) changing gso_size is outright buggy for UDP GSO packets, where framing
matters (I believe that's also the case for SCTP, but it's already excluded
by [g]). So the only remaining case is TCP, which also doesn't want it
(see [f]).
(i) see also the reasoning on the previous attempt at fixing this
(commit fa7b83bf3b) which shows that the current
behaviour causes TCP packet loss:
In the forwarding path GRO -> BPF 6 to 4 -> GSO for TCP traffic, the
coalesced packet payload can be > MSS, but < MSS + 20.
bpf_skb_proto_6_to_4() will upgrade the MSS and it can be > the payload
length. After then tcp_gso_segment checks for the payload length if it
is <= MSS. The condition is causing the packet to be dropped.
tcp_gso_segment():
[...]
mss = skb_shinfo(skb)->gso_size;
if (unlikely(skb->len <= mss)) goto out;
[...]
Thus changing the gso_size is simply a very bad idea. Increasing is unnecessary
and buggy, and decreasing can go negative.
Fixes: 6578171a7f ("bpf: add bpf_skb_change_proto helper")
Signed-off-by: Maciej Żenczykowski <maze@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Dongseok Yi <dseok.yi@samsung.com>
Cc: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/bpf/CANP3RGfjLikQ6dg=YpBU0OeHvyv7JOki7CyOUS9modaXAi-9vQ@mail.gmail.com
Link: https://lore.kernel.org/bpf/20210617000953.2787453-2-zenczykowski@gmail.com
