IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
To record the stats of header split packets, three stats are added in
the driver's ethtool stats.
- rx_hsplit_pkt is the split packets count with header split
- rx_hsplit_bytes is the received header bytes count with header split
- rx_hsplit_unsplit_pkt is the unsplit packet count due to header buffer
overflow or zero header length when header split is enabled
Currently, it's entering the stats_update critical section more than
once per packet. We have plans to avoid that in the future change to let
all the stats_update happen in one place at the end of
`gve_rx_poll_dqo`.
Co-developed-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Jeroen de Borst <jeroendb@google.com>
Reviewed-by: Praveen Kaligineedi <pkaligineedi@google.com>
Reviewed-by: Harshitha Ramamurthy <hramamurthy@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add header buffers and ethtool support to enable header split via the
tcp-data-split flag in ethtool's ringparam config. A coherent dma memory
is allocated for the header buffers. There is one header buffer per ring
entry by calculating the offset to the header-buffers starting address.
The header buffer is always copied directly into the skb and payload is
always added as frags. When there is a header buffer overflow or the
header length is 0, the driver places the whole unsplit packet in frags.
When toggling header split, the driver will call gve_adjust_config to
set its queues appropriately. If header split is enabled by the user and
the max packet buffer size is no less than 4KB, driver will set the
packet buffer size as 4KB to support TCP_ZEROCOPY_RECEIVE. Otherwise the
driver will use the default 2KB as the packet buffer size.
`ethtool -G <dev> tcp-data-split on/off` is the command to toggle header
split.
`ethtool -g <dev>` will show the status of header split with the field
of `tcp-data-split`.
Co-developed-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Jeroen de Borst <jeroendb@google.com>
Reviewed-by: Praveen Kaligineedi <pkaligineedi@google.com>
Reviewed-by: Harshitha Ramamurthy <hramamurthy@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
To enable header split via ethtool, we first need to query the device to
get the max rx buffer size and header buffer size. Add a device option
to get these values and store them in the driver. If the header buffer
size received from the device is non-zero, it means header split is
supported in the device.
Currently the max rx buffer size will only be used when header split is
enabled which will set the data_buffer_size_dqo to be the max rx buffer
size. Also change the data_buffer_size_dqo from int to u16 since we are
modifying it and making it to be consistent with max_rx_buffer_size.
Co-developed-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Jeroen de Borst <jeroendb@google.com>
Reviewed-by: Praveen Kaligineedi <pkaligineedi@google.com>
Reviewed-by: Harshitha Ramamurthy <hramamurthy@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously, each caller of gve_rx_alloc_buffer had to increase counter
and as a result one caller was not tracking those failure. Increasing
counters at a common location now so callers don't have to duplicate
code or miss counter management.
Signed-off-by: Ankit Garg <nktgrg@google.com>
Link: https://lore.kernel.org/r/20240124205435.1021490-1-nktgrg@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
gve_open is rewritten to be composed of two funcs: gve_queues_mem_alloc
and gve_queues_start. The former only allocates queue resources without
doing anything to install the queues, which is taken up by the latter.
Similarly gve_close is split into gve_queues_stop and
gve_queues_mem_free.
Separating the acts of queue resource allocation and making the queue
become live help with subsequent changes that aim to not take down the
datapath when applying new configurations.
Signed-off-by: Shailend Chand <shailend@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Jeroen de Borst <jeroendb@google.com>
Link: https://lore.kernel.org/r/20240122182632.1102721-5-shailend@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The new config-aware functions will help achieve the goal of being able
to allocate resources for new queues while there already are active
queues serving traffic.
These new functions work off of arbitrary queue allocation configs
rather than just the currently active config in priv, and they return
the newly allocated resources instead of writing them into priv.
Signed-off-by: Shailend Chand <shailend@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Jeroen de Borst <jeroendb@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://lore.kernel.org/r/20240122182632.1102721-4-shailend@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Queue allocation functions currently can only allocate into priv and
free memory in priv. These new structs would be passed into the queue
functions in a subsequent change to make them capable of returning newly
allocated resources and not just writing them into priv. They also make
it possible to allocate resources for queues with a different config
than that of the currently active queues.
Signed-off-by: Shailend Chand <shailend@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Jeroen de Borst <jeroendb@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://lore.kernel.org/r/20240122182632.1102721-2-shailend@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This register is required on platforms with page sizes greater than 4k.
This is because the tx side of the driver vmaps the entire queue page
list of pages into a single flat address space, then uses the entire
space. Without communicating the guest page size to the backend, the
backend will only access the first 4k of each page in the queue page list.
Signed-off-by: Jordan Kimbrough <jrkim@google.com>
Signed-off-by: John Fraker <jfraker@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/r/20231128002648.320892-5-jfraker@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
adminq_pfn assumes a page size of 4k, causing this mechanism to break in
kernels compiled with different page sizes. A new PCI device revision was
needed for the device to be able to communicate with the driver how to
set up the admin queue prior to having access to the admin queue.
Signed-off-by: Jordan Kimbrough <jrkim@google.com>
Signed-off-by: John Fraker <jfraker@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/r/20231128002648.320892-3-jfraker@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
It is suboptimal to attempt skb linearization from ndo_start_xmit()
if a gso skb has pathological layout, or if host stack does not have
access to the payload (TCP direct). Linearization of large skbs
can also fail under memory pressure.
We should instead have an ndo_features_check() so that we can
fallback to GSO, which is supported even for TCP direct,
and generally much more efficient (no payload copy).
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Bailey Forrest <bcf@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Jeroen de Borst <jeroendb@google.com>
Cc: Praveen Kaligineedi <pkaligineedi@google.com>
Cc: Shailend Chand <shailend@google.com>
Cc: Ziwei Xiao <ziweixiao@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Netpoll will explicilty pass the polling call with a budget of 0 to
indicate it's clearing the Tx path only. For the gve_rx_poll and
gve_xdp_poll, they were mistakenly taking the 0 budget as the indication
to do all the work. Add check to avoid the rx path and xdp path being
called when budget is 0. And also avoid napi_complete_done being called
when budget is 0 for netpoll.
Fixes: f5cedc84a3 ("gve: Add transmit and receive support")
Signed-off-by: Ziwei Xiao <ziweixiao@google.com>
Link: https://lore.kernel.org/r/20231114004144.2022268-1-ziweixiao@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
If, for any reason, `tx_stats_num + rx_stats_num` wraps around, the
protection that struct_size() adds against potential integer overflows
is defeated. Fix this by hardening call to struct_size() with size_add().
Fixes: 691f4077d5 ("gve: Replace zero-length array with flexible-array member")
Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
gve_rx_append_frags() is able to build skbs chained with frag_list,
like GRO engine.
Problem is that shinfo->frag_list should only be used
for the head of the chain.
All other links should use skb->next pointer.
Otherwise, built skbs are not valid and can cause crashes.
Equivalent code in GRO (skb_gro_receive()) is:
if (NAPI_GRO_CB(p)->last == p)
skb_shinfo(p)->frag_list = skb;
else
NAPI_GRO_CB(p)->last->next = skb;
NAPI_GRO_CB(p)->last = skb;
Fixes: 9b8dd5e5ea ("gve: DQO: Add RX path")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Bailey Forrest <bcf@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Catherine Sullivan <csully@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The RX path allocates the QPL page pool at queue creation, and
tries to reuse these pages through page recycling. This patch
ensures that on refill no non-QPL pages are posted to the device.
When the driver is running low on free buffers, an ondemand
allocation step kicks in that allocates a non-qpl page for
SKB business to free up the QPL page in use.
gve_try_recycle_buf was moved to gve_rx_append_frags so that driver does
not attempt to mark buffer as used if a non-qpl page was allocated
ondemand.
Signed-off-by: Rushil Gupta <rushilg@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Praveen Kaligineedi <pkaligineedi@google.com>
Signed-off-by: Bailey Forrest <bcf@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Each QPL page is divided into GVE_TX_BUFS_PER_PAGE_DQO buffers.
When a packet needs to be transmitted, we break the packet into max
GVE_TX_BUF_SIZE_DQO sized chunks and transmit each chunk using a TX
descriptor.
We allocate the TX buffers from the free list in dqo_tx.
We store these TX buffer indices in an array in the pending_packet
structure.
The TX buffers are returned to the free list in dqo_compl after
receiving packet completion or when removing packets from miss
completions list.
Signed-off-by: Rushil Gupta <rushilg@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Praveen Kaligineedi <pkaligineedi@google.com>
Signed-off-by: Bailey Forrest <bcf@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
GVE supports QPL ("queue-page-list") mode where
all data is communicated through a set of pre-registered
pages. Adding this mode to DQO descriptor format.
Add checks, abi-changes and device options to support
QPL mode for DQO in addition to GQI. Also, use
pages-per-qpl supplied by device-option to control the
size of the "queue-page-list".
Signed-off-by: Rushil Gupta <rushilg@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Praveen Kaligineedi <pkaligineedi@google.com>
Signed-off-by: Bailey Forrest <bcf@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Spotted this trivial spell mistake while casually reading
the google GVE driver code.
Signed-off-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current codebase contained the usage of two different names for this
driver (i.e., `gvnic` and `gve`), which is quite unfriendly for users
to use, especially when trying to bind or unbind the driver manually.
The corresponding kernel module is registered with the name of `gve`.
It's more reasonable to align the name of the driver with the module.
Fixes: 893ce44df5 ("gve: Add basic driver framework for Compute Engine Virtual NIC")
Cc: csully@google.com
Signed-off-by: Junfeng Guo <junfeng.guo@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current duplex mode was unset in the driver, resulting in the default
parameter being set to 0, which corresponds to half duplex. It might
mislead users to have incorrect expectation about the driver's
transmission capabilities.
Set the default duplex configuration to full, as the driver runs in
full duplex mode at this point.
Fixes: 7e074d5a76 ("gve: Enable Link Speed Reporting in the driver.")
Signed-off-by: Junfeng Guo <junfeng.guo@intel.com>
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>
Message-ID: <20230706044128.2726747-1-junfeng.guo@intel.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add support for using IPv6 Big TCP on DQ which can handle large TSO/GRO
packets. See https://lwn.net/Articles/895398/. This can improve the
throughput and CPU usage.
Perf test result:
ip -d link show $DEV
gso_max_size 185000 gso_max_segs 65535 tso_max_size 262143 tso_max_segs 65535 gro_max_size 185000
For performance, tested with neper using 9k MTU on hardware that supports 200Gb/s line rate.
In single streams when line rate is not saturated, we expect throughput improvements.
When the networking is performing at line rate, we expect cpu usage improvements.
Tcp_stream (unidirectional stream test, T=thread, F=flow):
skb=180kb, T=1, F=1, no zerocopy: throughput average=64576.88 Mb/s, sender stime=8.3, receiver stime=10.68
skb=64kb, T=1, F=1, no zerocopy: throughput average=64862.54 Mb/s, sender stime=9.96, receiver stime=12.67
skb=180kb, T=1, F=1, yes zerocopy: throughput average=146604.97 Mb/s, sender stime=10.61, receiver stime=5.52
skb=64kb, T=1, F=1, yes zerocopy: throughput average=131357.78 Mb/s, sender stime=12.11, receiver stime=12.25
skb=180kb, T=20, F=100, no zerocopy: throughput average=182411.37 Mb/s, sender stime=41.62, receiver stime=79.4
skb=64kb, T=20, F=100, no zerocopy: throughput average=182892.02 Mb/s, sender stime=57.39, receiver stime=72.69
skb=180kb, T=20, F=100, yes zerocopy: throughput average=182337.65 Mb/s, sender stime=27.94, receiver stime=39.7
skb=64kb, T=20, F=100, yes zerocopy: throughput average=182144.20 Mb/s, sender stime=47.06, receiver stime=39.01
Signed-off-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Coco Li <lixiaoyan@google.com>
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Link: https://lore.kernel.org/r/20230522201552.3585421-1-ziweixiao@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Clearing the PBA bit from the driver is race prone and it may lead to
dropped interrupt events. This could potentially lead to the traffic
being completely halted.
Fixes: 5e8c5adf95 ("gve: DQO: Add core netdev features")
Signed-off-by: Ziwei Xiao <ziweixiao@google.com>
Signed-off-by: Bailey Forrest <bcf@google.com>
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Non-GSO TCP packets whose SKBs' linear portion did not include the
entire TCP header were not populating the first Tx descriptor with
as many bytes as the vNIC expected. This change ensures that all
TCP packets populate the first descriptor with the correct number of
bytes.
Fixes: 893ce44df5 ("gve: Add basic driver framework for Compute Engine Virtual NIC")
Signed-off-by: Shailend Chand <shailend@google.com>
Link: https://lore.kernel.org/r/20230403172809.2939306-1-shailend@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The link speed is never changed for the uptime of a VM, and the current
implementation sends an admin queue command for each call. Admin queue
command invocations have nontrivial overhead (e.g., VM exits), which can
be disruptive to users if triggered frequently. Our telemetry data shows
that there are VMs that make frequent calls to this admin queue command.
Caching the result of the original admin queue command would eliminate
the need to send multiple admin queue commands on subsequent calls to
retrieve link speed.
Fixes: 7e074d5a76 ("gve: Enable Link Speed Reporting in the driver.")
Signed-off-by: Joshua Washington <joshwash@google.com>
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Link: https://lore.kernel.org/r/20230321172332.91678-1-joshwash@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Adding AF_XDP zero-copy support.
Note: Although these changes support AF_XDP socket in zero-copy
mode, there is still a copy happening within the driver between
XSK buffer pool and QPL bounce buffers in GQI-QPL format.
In GQI-QPL queue format, the driver needs to allocate a fixed size
memory, the size specified by vNIC device, for RX/TX and register this
memory as a bounce buffer with the vNIC device when a queue is
created. The number of pages in the bounce buffer is limited and the
pages need to be made available to the vNIC by copying the RX data out
to prevent head-of-line blocking. Therefore, we cannot pass the XSK
buffer pool to the vNIC.
The number of copies on RX path from the bounce buffer to XSK buffer is 2
for AF_XDP copy mode (bounce buffer -> allocated page frag -> XSK buffer)
and 1 for AF_XDP zero-copy mode (bounce buffer -> XSK buffer).
This patch contains the following changes:
1) Enable and disable XSK buffer pool
2) Copy XDP packets from QPL bounce buffers to XSK buffer on rx
3) Copy XDP packets from XSK buffer to QPL bounce buffers and
ring the doorbell as part of XDP TX napi poll
4) ndo_xsk_wakeup callback support
Signed-off-by: Praveen Kaligineedi <pkaligineedi@google.com>
Reviewed-by: Jeroen de Borst <jeroendb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch contains the following changes:
1) Support for XDP REDIRECT action on rx
2) ndo_xdp_xmit callback support
In GQI-QPL queue format, the driver needs to allocate a fixed size
memory, the size specified by vNIC device, for RX/TX and register this
memory as a bounce buffer with the vNIC device when a queue is created.
The number of pages in the bounce buffer is limited and the pages need to
be made available to the vNIC by copying the RX data out to prevent
head-of-line blocking. The XDP_REDIRECT packets are therefore immediately
copied to a newly allocated page.
Signed-off-by: Praveen Kaligineedi <pkaligineedi@google.com>
Reviewed-by: Jeroen de Borst <jeroendb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for XDP PASS, DROP and TX actions.
This patch contains the following changes:
1) Support installing/uninstalling XDP program
2) Add dedicated XDP TX queues
3) Add support for XDP DROP action
4) Add support for XDP TX action
Signed-off-by: Praveen Kaligineedi <pkaligineedi@google.com>
Reviewed-by: Jeroen de Borst <jeroendb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Changes to enable adding and removing TX queues without calling
gve_close() and gve_open().
Made the following changes:
1) priv->tx, priv->rx and priv->qpls arrays are allocated based on
max tx queues and max rx queues
2) Changed gve_adminq_create_tx_queues(), gve_adminq_destroy_tx_queues(),
gve_tx_alloc_rings() and gve_tx_free_rings() functions to add/remove a
subset of TX queues rather than all the TX queues.
Signed-off-by: Praveen Kaligineedi <pkaligineedi@google.com>
Reviewed-by: Jeroen de Borst <jeroendb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
IRQs are currently requested before the netdevice is registered
and a proper name is assigned to the device. Changing interrupt
name to avoid using the format string in the name.
Interrupt name before change: eth%d-ntfy-block.<blk_id>
Interrupt name after change: gve-ntfy-blk<blk_id>@pci:<pci_name>
Signed-off-by: Praveen Kaligineedi <pkaligineedi@google.com>
Reviewed-by: Jeroen de Borst <jeroendb@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The virtual NIC has 2 ways of indicating a miss-path
completion. This handles the alternate.
Signed-off-by: Jeroen de Borst <jeroendb@google.com>
Reviewed-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously, even if just one of the many fragments of a 9k packet
required a copy, we'd copy the whole packet into a freshly-allocated
9k-sized linear SKB, and this led to performance issues.
By having a pool of pages to copy into, each fragment can be
independently handled, leading to a reduced incidence of
allocation and copy.
Signed-off-by: Shailend Chand <shailend@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
