We (Paolo and I) noticed that in the sending path touching an extra
cacheline due to cq_cached_prod_lock will impact the performance. After
moving the lock from struct xsk_buff_pool to struct xsk_queue, the
performance is increased by ~5% which can be observed by xdpsock.
An alternative approach [1] can be using atomic_try_cmpxchg() to have the
same effect. But unfortunately I don't have evident performance numbers to
prove the atomic approach is better than the current patch. The advantage
is to save the contention time among multiple xsks sharing the same pool
while the disadvantage is losing good maintenance. The full discussion can
be found at the following link.
[1]: https://lore.kernel.org/all/20251128134601.54678-1-kerneljasonxing@gmail.com/
Suggested-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Jason Xing <kernelxing@tencent.com>
Link: https://patch.msgid.link/20260104012125.44003-3-kerneljasonxing@gmail.com
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
- Split cq_lock into two smaller locks: cq_prod_lock and
cq_cached_prod_lock
- Avoid disabling/enabling interrupts in the hot xmit path
In either xsk_cq_cancel_locked() or xsk_cq_reserve_locked() function,
the race condition is only between multiple xsks sharing the same
pool. They are all in the process context rather than interrupt context,
so now the small lock named cq_cached_prod_lock can be used without
handling interrupts.
While cq_cached_prod_lock ensures the exclusive modification of
@cached_prod, cq_prod_lock in xsk_cq_submit_addr_locked() only cares
about @producer and corresponding @desc. Both of them don't necessarily
be consistent with @cached_prod protected by cq_cached_prod_lock.
That's the reason why the previous big lock can be split into two
smaller ones. Please note that SPSC rule is all about the global state
of producer and consumer that can affect both layers instead of local
or cached ones.
Frequently disabling and enabling interrupt are very time consuming
in some cases, especially in a per-descriptor granularity, which now
can be avoided after this optimization, even when the pool is shared by
multiple xsks.
With this patch, the performance number[1] could go from 1,872,565 pps
to 1,961,009 pps. It's a minor rise of around 5%.
[1]: taskset -c 1 ./xdpsock -i enp2s0f1 -q 0 -t -S -s 64
Signed-off-by: Jason Xing <kernelxing@tencent.com>
Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://patch.msgid.link/20251030000646.18859-3-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>
Currently, when your driver supports XSk Tx metadata and you want to
send an XSk frame, you need to do the following:
* call external xsk_buff_raw_get_dma();
* call inline xsk_buff_get_metadata(), which calls external
xsk_buff_raw_get_data() and then do some inline checks.
This effectively means that the following piece:
addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
is done twice per frame, plus you have 2 external calls per frame, plus
this:
meta = pool->addrs + addr - pool->tx_metadata_len;
if (unlikely(!xsk_buff_valid_tx_metadata(meta)))
is always inlined, even if there's no meta or it's invalid.
Add xsk_buff_raw_get_ctx() (xp_raw_get_ctx() to be precise) to do that
in one go. It returns a small structure with 2 fields: DMA address,
filled unconditionally, and metadata pointer, non-NULL only if it's
present and valid. The address correction is performed only once and
you also have only 1 external call per XSk frame, which does all the
calculations and checks outside of your hotpath. You only need to
check `if (ctx.meta)` for the metadata presence.
To not copy any existing code, derive address correction and getting
virtual and DMA address into small helpers. bloat-o-meter reports no
object code changes for the existing functionality.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Link: https://patch.msgid.link/20250206182630.3914318-5-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
After the series "XSk buff on a diet" by Maciej, the greatest pow-2
which &xdp_buff_xsk can be divided got reduced from 16 to 8 on x86_64.
Also, sizeof(xdp_buff_xsk) now is 120 bytes, which, taking the previous
sentence into account, leads to that it leaves 8 bytes at the end of
cacheline, which means an array of buffs will have its elements
messed between the cachelines chaotically.
Use __aligned_largest for this struct. This alignment is usually 16
bytes, which makes it fill two full cachelines and align an array
nicely. ___cacheline_aligned may be excessive here, especially on
arches with 128-256 byte CLs, as well as 32-bit arches (76 -> 96
bytes on MIPS32R2), while not doing better than _largest.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://patch.msgid.link/20241203173733.3181246-2-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Continue the process of dieting xdp_buff_xsk by removing orig_addr
member. It can be calculated from xdp->data_hard_start where it was
previously used, so it is not anything that has to be carried around in
struct used widely in hot path.
This has been used for initializing xdp_buff_xsk::frame_dma during pool
setup and as a shortcut in xp_get_handle() to retrieve address provided
to xsk Rx queue.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20241007122458.282590-4-maciej.fijalkowski@intel.com
XSk infra's been using its own DMA sync shortcut to try avoiding
redundant function calls. Now that there is a generic one, remove
the custom implementation and rely on the generic helpers.
xsk_buff_dma_sync_for_cpu() doesn't need the second argument anymore,
remove it.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
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>
Given that skb_shared_info relies on skb_frag_t, in order to support
xskb chaining, introduce xdp_buff_xsk::xskb_list_node and
xsk_buff_pool::xskb_list.
This is needed so ZC drivers can add frags as xskb nodes which will make
it possible to handle it both when producing AF_XDP Rx descriptors as
well as freeing/recycling all the frags that a single frame carries.
Speaking of latter, update xsk_buff_free() to take care of list nodes.
For the former (adding as frags), introduce xsk_buff_add_frag() for ZC
drivers usage that is going to be used to add a frag to xskb list from
pool.
xsk_buff_get_frag() will be utilized by XDP_TX and, on contrary, will
return xdp_buff.
One of the previous patches added a wrapper for ZC Rx so implement xskb
list walk and production of Rx descriptors there.
On bind() path, bail out if socket wants to use ZC multi-buffer but
underlying netdev does not support it.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-12-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Drivers are used to check for EOP bit whereas AF_XDP operates on
inverted logic - user space indicates that current frag is not the last
one and packet continues. For AF_XDP core needs, add xp_mb_desc() that
will simply test XDP_PKT_CONTD from xdp_desc::options, but in order to
preserve drivers default behavior, introduce an interface for ZC drivers
that will negate xp_mb_desc() result and therefore make it easier to
test EOP bit from during production of HW Tx descriptors.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-8-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Make sure unaligned descriptors that straddle the end of the UMEM are
considered invalid. Currently, descriptor validation is broken for
zero-copy mode which only checks descriptors at page granularity.
For example, descriptors in zero-copy mode that overrun the end of the
UMEM but not a page boundary are (incorrectly) considered valid. The
UMEM boundary check needs to happen before the page boundary and
contiguity checks in xp_desc_crosses_non_contig_pg(). Do this check in
xp_unaligned_validate_desc() instead like xp_check_unaligned() already
does.
Fixes: 2b43470add ("xsk: Introduce AF_XDP buffer allocation API")
Signed-off-by: Kal Conley <kal.conley@dectris.com>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/r/20230405235920.7305-2-kal.conley@dectris.com
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Move desc_array from the driver to the pool. The reason behind this is
that we can then reuse this array as a temporary storage for descriptors
in all zero-copy drivers that use the batched interface. This will make
it easier to add batching to more drivers.
i40e is the only driver that has a batched Tx zero-copy
implementation, so no need to touch any other driver.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Alexander Lobakin <alexandr.lobakin@intel.com>
Link: https://lore.kernel.org/bpf/20220125160446.78976-6-maciej.fijalkowski@intel.com
Maciej Fijalkowski
Jason Xing
e.kubanski
Alexander Lobakin
Stanislav Fomichev
Kal Conley
Toke Høiland-Jørgensen
Jalal Mostafa
Magnus Karlsson