Currently xp_assign_dev_shared() is missing XDP_USE_SG being propagated
to flags so set it in order to preserve mtu check that is supposed to be
done only when no multi-buffer setup is in picture.
Also, this flag has the same value as XDP_UMEM_TX_SW_CSUM so we could
get unexpected SG setups for software Tx checksums. Since csum flag is
UAPI, modify value of XDP_UMEM_SG_FLAG.
Fixes: d609f3d228 ("xsk: add multi-buffer support for sockets sharing umem")
Reviewed-by: Björn Töpel <bjorn@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://patch.msgid.link/20260402154958.562179-4-maciej.fijalkowski@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch provides a setsockopt method to let applications leverage to
adjust how many descs to be handled at most in one send syscall. It
mitigates the situation where the default value (32) that is too small
leads to higher frequency of triggering send syscall.
Considering the prosperity/complexity the applications have, there is no
absolutely ideal suggestion fitting all cases. So keep 32 as its default
value like before.
The patch does the following things:
- Add XDP_MAX_TX_SKB_BUDGET socket option.
- Set max_tx_budget to 32 by default in the initialization phase as a
per-socket granular control.
- Set the range of max_tx_budget as [32, xs->tx->nentries].
The idea behind this comes out of real workloads in production. We use a
user-level stack with xsk support to accelerate sending packets and
minimize triggering syscalls. When the packets are aggregated, it's not
hard to hit the upper bound (namely, 32). The moment user-space stack
fetches the -EAGAIN error number passed from sendto(), it will loop to try
again until all the expected descs from tx ring are sent out to the driver.
Enlarging the XDP_MAX_TX_SKB_BUDGET value contributes to less frequency of
sendto() and higher throughput/PPS.
Here is what I did in production, along with some numbers as follows:
For one application I saw lately, I suggested using 128 as max_tx_budget
because I saw two limitations without changing any default configuration:
1) XDP_MAX_TX_SKB_BUDGET, 2) socket sndbuf which is 212992 decided by
net.core.wmem_default. As to XDP_MAX_TX_SKB_BUDGET, the scenario behind
this was I counted how many descs are transmitted to the driver at one
time of sendto() based on [1] patch and then I calculated the
possibility of hitting the upper bound. Finally I chose 128 as a
suitable value because 1) it covers most of the cases, 2) a higher
number would not bring evident results. After twisting the parameters,
a stable improvement of around 4% for both PPS and throughput and less
resources consumption were found to be observed by strace -c -p xxx:
1) %time was decreased by 7.8%
2) error counter was decreased from 18367 to 572
[1]: https://lore.kernel.org/all/20250619093641.70700-1-kerneljasonxing@gmail.com/
Signed-off-by: Jason Xing <kernelxing@tencent.com>
Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://patch.msgid.link/20250704160138.48677-1-kerneljasonxing@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Move rx_lock from xsk_socket to xsk_buff_pool.
Fix synchronization for shared umem mode in
generic RX path where multiple sockets share
single xsk_buff_pool.
RX queue is exclusive to xsk_socket, while FILL
queue can be shared between multiple sockets.
This could result in race condition where two
CPU cores access RX path of two different sockets
sharing the same umem.
Protect both queues by acquiring spinlock in shared
xsk_buff_pool.
Lock contention may be minimized in the future by some
per-thread FQ buffering.
It's safe and necessary to move spin_lock_bh(rx_lock)
after xsk_rcv_check():
* xs->pool and spinlock_init is synchronized by
xsk_bind() -> xsk_is_bound() memory barriers.
* xsk_rcv_check() may return true at the moment
of xsk_release() or xsk_unbind_dev(),
however this will not cause any data races or
race conditions. xsk_unbind_dev() removes xdp
socket from all maps and waits for completion
of all outstanding rx operations. Packets in
RX path will either complete safely or drop.
Signed-off-by: Eryk Kubanski <e.kubanski@partner.samsung.com>
Fixes: bf0bdd1343 ("xdp: fix race on generic receive path")
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://patch.msgid.link/20250416101908.10919-1-e.kubanski@partner.samsung.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Extend the XDP Tx metadata framework so that user can requests launch time
hardware offload, where the Ethernet device will schedule the packet for
transmission at a pre-determined time called launch time. The value of
launch time is communicated from user space to Ethernet driver via
launch_time field of struct xsk_tx_metadata.
Suggested-by: Stanislav Fomichev <sdf@fomichev.me>
Signed-off-by: Song Yoong Siang <yoong.siang.song@intel.com>
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Link: https://patch.msgid.link/20250216093430.957880-2-yoong.siang.song@intel.com
Every NIC driver utilizing XDP should invoke xdp_do_flush() after
processing all packages. With the introduction of the bpf_net_context
logic the flush lists (for dev, CPU-map and xsk) are lazy initialized
only if used. However xdp_do_flush() tries to flush all three of them so
all three lists are always initialized and the likely empty lists are
"iterated".
Without the usage of XDP but with CONFIG_DEBUG_NET the lists are also
initialized due to xdp_do_check_flushed().
Jakub suggest to utilize the hints in bpf_net_context and avoid invoking
the flush function. This will also avoiding initializing the lists which
are otherwise unused.
Introduce bpf_net_ctx_get_all_used_flush_lists() to return the
individual list if not-empty. Use the logic in xdp_do_flush() and
xdp_do_check_flushed(). Remove the not needed .*_check_flush().
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This change actually defines the (initial) metadata layout
that should be used by AF_XDP userspace (xsk_tx_metadata).
The first field is flags which requests appropriate offloads,
followed by the offload-specific fields. The supported per-device
offloads are exported via netlink (new xsk-flags).
The offloads themselves are still implemented in a bit of a
framework-y fashion that's left from my initial kfunc attempt.
I'm introducing new xsk_tx_metadata_ops which drivers are
supposed to implement. The drivers are also supposed
to call xsk_tx_metadata_request/xsk_tx_metadata_complete in
the right places. Since xsk_tx_metadata_{request,_complete}
are static inline, we don't incur any extra overhead doing
indirect calls.
The benefit of this scheme is as follows:
- keeps all metadata layout parsing away from driver code
- makes it easy to grep and see which drivers implement what
- don't need any extra flags to maintain to keep track of what
offloads are implemented; if the callback is implemented - the offload
is supported (used by netlink reporting code)
Two offloads are defined right now:
1. XDP_TXMD_FLAGS_CHECKSUM: skb-style csum_start+csum_offset
2. XDP_TXMD_FLAGS_TIMESTAMP: writes TX timestamp back into metadata
area upon completion (tx_timestamp field)
XDP_TXMD_FLAGS_TIMESTAMP is also implemented for XDP_COPY mode: it writes
SW timestamp from the skb destructor (note I'm reusing hwtstamps to pass
metadata pointer).
The struct is forward-compatible and can be extended in the future
by appending more fields.
Reviewed-by: Song Yoong Siang <yoong.siang.song@intel.com>
Signed-off-by: Stanislav Fomichev <sdf@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20231127190319.1190813-3-sdf@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
For zerocopy mode, tx_desc->addr can point to an arbitrary offset
and carry some TX metadata in the headroom. For copy mode, there
is no way currently to populate skb metadata.
Introduce new tx_metadata_len umem config option that indicates how many
bytes to treat as metadata. Metadata bytes come prior to tx_desc address
(same as in RX case).
The size of the metadata has mostly the same constraints as XDP:
- less than 256 bytes
- 8-byte aligned (compared to 4-byte alignment on xdp, due to 8-byte
timestamp in the completion)
- non-zero
This data is not interpreted in any way right now.
Reviewed-by: Song Yoong Siang <yoong.siang.song@intel.com>
Signed-off-by: Stanislav Fomichev <sdf@google.com>
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20231127190319.1190813-2-sdf@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
In the previous implementation, when multiple xsk sockets were
associated with a single xsk_buff_pool, a situation could arise
where the xsk_tx_list maintained data at the front for one xsk
socket while starving the xsk sockets at the back of the list.
This could result in issues such as the inability to transmit packets,
increased latency, and jitter. To address this problem, we introduce
a new variable called tx_budget_spent, which limits each xsk to transmit
a maximum of MAX_PER_SOCKET_BUDGET tx descriptors. This allocation ensures
equitable opportunities for subsequent xsk sockets to send tx descriptors.
The value of MAX_PER_SOCKET_BUDGET is set to 32.
Signed-off-by: Albert Huang <huangjie.albert@bytedance.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20231023125732.82261-1-huangjie.albert@bytedance.com
A few drivers were missing a xdp_do_flush() invocation after
XDP_REDIRECT.
Add three helper functions each for one of the per-CPU lists. Return
true if the per-CPU list is non-empty and flush the list.
Add xdp_do_check_flushed() which invokes each helper functions and
creates a warning if one of the functions had a non-empty list.
Hide everything behind CONFIG_DEBUG_NET.
Suggested-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20231016125738.Yt79p1uF@linutronix.de
Userspace applications indicate their multi-buffer capability to xsk
using XSK_USE_SG socket bind flag. For sockets using shared umem the
bind flag may contain XSK_USE_SG only for the first socket. For any
subsequent socket the only option supported is XDP_SHARED_UMEM.
Add option XDP_UMEM_SG_FLAG in umem config flags to store the
multi-buffer handling capability when indicated by XSK_USE_SG option in
bing flag by the first socket. Use this to derive multi-buffer capability
for subsequent sockets in xsk core.
Signed-off-by: Tirthendu Sarkar <tirthendu.sarkar@intel.com>
Fixes: 81470b5c3c ("xsk: introduce XSK_USE_SG bind flag for xsk socket")
Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/r/20230907035032.2627879-1-tirthendu.sarkar@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
In Tx path, xsk core reserves space for each desc to be transmitted in
the completion queue and it's address contained in it is stored in the
skb destructor arg. After successful transmission the skb destructor
submits the addr marking completion.
To handle multiple descriptors per packet, now along with reserving
space for each descriptor, the corresponding address is also stored in
completion queue. The number of pending descriptors are stored in skb
destructor arg and is used by the skb destructor to update completions.
Introduce 'skb' in xdp_sock to store a partially built packet when
__xsk_generic_xmit() must return before it sees the EOP descriptor for
the current packet so that packet building can resume in next call of
__xsk_generic_xmit().
Helper functions are introduced to set and get the pending descriptors
in the skb destructor arg. Also, wrappers are introduced for storing
descriptor addresses, submitting and cancelling (for unsuccessful
transmissions) the number of completions.
Signed-off-by: Tirthendu Sarkar <tirthendu.sarkar@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-7-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
As of now xsk core drops any xdp_buff with data size greater than the
xsk frame_size as set by the af_xdp application. With multi-buffer
support introduced in the next patch xsk core can now split those
buffers into multiple descriptors provided the af_xdp application can
handle them. Such capability of the application needs to be independent
of the xdp_prog's frag support capability since there are cases where
even a single xdp_buffer may need to be split into multiple descriptors
owing to a smaller xsk frame size.
For e.g., with NIC rx_buffer size set to 4kB, a 3kB packet will
constitute of a single buffer and so will be sent as such to AF_XDP layer
irrespective of 'xdp.frags' capability of the XDP program. Now if the xsk
frame size is set to 2kB by the AF_XDP application, then the packet will
need to be split into 2 descriptors if AF_XDP application can handle
multi-buffer, else it needs to be dropped.
Applications can now advertise their frag handling capability to xsk core
so that xsk core can decide if it should drop or split xdp_buffs that
exceed xsk frame size. This is done using a new 'XSK_USE_SG' bind flag
for the xdp socket.
Signed-off-by: Tirthendu Sarkar <tirthendu.sarkar@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-3-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
A new helper is introduced to calculate xskmap memory usage.
The xfsmap memory usage can be dynamically changed when we add or remove
a xsk_map_node. Hence we need to track the count of xsk_map_node to get
its memory usage.
The result as follows,
- before
10: xskmap name count_map flags 0x0
key 4B value 4B max_entries 65536 memlock 524288B
- after
10: xskmap name count_map flags 0x0 <<< no elements case
key 4B value 4B max_entries 65536 memlock 524608B
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Link: https://lore.kernel.org/r/20230305124615.12358-17-laoar.shao@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
sock.h is pretty heavily used (5k objects rebuilt on x86 after
it's touched). We can drop the include of filter.h from it and
add a forward declaration of struct sk_filter instead.
This decreases the number of rebuilt objects when bpf.h
is touched from ~5k to ~1k.
There's a lot of missing includes this was masking. Primarily
in networking tho, this time.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Marc Kleine-Budde <mkl@pengutronix.de>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Acked-by: Nikolay Aleksandrov <nikolay@nvidia.com>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Link: https://lore.kernel.org/bpf/20211229004913.513372-1-kuba@kernel.org
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
Currently the bpf_redirect_map() implementation dispatches to the
correct map-lookup function via a switch-statement. To avoid the
dispatching, this change adds bpf_redirect_map() as a map
operation. Each map provides its bpf_redirect_map() version, and
correct function is automatically selected by the BPF verifier.
A nice side-effect of the code movement is that the map lookup
functions are now local to the map implementation files, which removes
one additional function call.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/bpf/20210308112907.559576-2-bjorn.topel@gmail.com
Fix a race when multiple sockets are simultaneously calling sendto()
when the completion ring is shared in the SKB case. This is the case
when you share the same netdev and queue id through the
XDP_SHARED_UMEM bind flag. The problem is that multiple processes can
be in xsk_generic_xmit() and call the backpressure mechanism in
xskq_prod_reserve(xs->pool->cq). As this is a shared resource in this
specific scenario, a race might occur since the rings are
single-producer single-consumer.
Fix this by moving the tx_completion_lock from the socket to the pool
as the pool is shared between the sockets that share the completion
ring. (The pool is not shared when this is not the case.) And then
protect the accesses to xskq_prod_reserve() with this lock. The
tx_completion_lock is renamed cq_lock to better reflect that it
protects accesses to the potentially shared completion ring.
Fixes: 35fcde7f8d ("xsk: support for Tx")
Reported-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/20201218134525.13119-2-magnus.karlsson@gmail.com
Fix a bug that is triggered when a partially setup socket is
destroyed. For a fully setup socket, a socket that has been bound to a
device, the cleanup of the umem is performed at the end of the buffer
pool's cleanup work queue item. This has to be performed in a work
queue, and not in RCU cleanup, as it is doing a vunmap that cannot
execute in interrupt context. However, when a socket has only been
partially set up so that a umem has been created but the buffer pool
has not, the code erroneously directly calls the umem cleanup function
instead of using a work queue, and this leads to a BUG_ON() in
vunmap().
As there in this case is no buffer pool, we cannot use its work queue,
so we need to introduce a work queue for the umem and schedule this for
the cleanup. So in the case there is no pool, we are going to use the
umem's own work queue to schedule the cleanup. But if there is a
pool, the cleanup of the umem is still being performed by the pool's
work queue, as it is important that the umem is cleaned up after the
pool.
Fixes: e5e1a4bc91 ("xsk: Fix possible memory leak at socket close")
Reported-by: Marek Majtyka <marekx.majtyka@intel.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Marek Majtyka <marekx.majtyka@intel.com>
Link: https://lore.kernel.org/bpf/1605873219-21629-1-git-send-email-magnus.karlsson@gmail.com
Move queue_id, dev, and need_wakeup from the umem to the
buffer pool. This so that we in a later commit can share the umem
between multiple HW queues. There is one buffer pool per dev and
queue id, so these variables should belong to the buffer pool, not
the umem. Need_wakeup is also something that is set on a per napi
level, so there is usually one per device and queue id. So move
this to the buffer pool too.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/1598603189-32145-6-git-send-email-magnus.karlsson@intel.com
Maciej Fijalkowski
Jason Xing
e.kubanski
Song Yoong Siang
Sebastian Andrzej Siewior
Stanislav Fomichev
Albert Huang
Tirthendu Sarkar
Yafang Shao
Jakub Kicinski
Toke Høiland-Jørgensen
Björn Töpel
Magnus Karlsson