Every devmem dmabuf binding today hands the page_pool PAGE_SIZE niovs.
This caps a single RX descriptor at PAGE_SIZE, burning CPU on buffer
churn for large flows.
Add a bind-time netlink attribute, NETDEV_A_DMABUF_RX_PAGE_SIZE, that
lets userspace request a larger niov size. The value must be a power of
two >= PAGE_SIZE.
The TX path is changed to always pass PAGE_SIZE.
Measurements:
Setup: kperf in devmem RX/TX cuda mode, 4 flows, 64 MB messages, 60s,
dctcp, num-rx-queues=4, dmabuf-rx/tx-size-mb=2048, 10 runs per niov
size, mlx5.
CPU Util:
niov net sirq % net idle % app sys % app idle %
----- ---------------- ---------------- ---------------- ----------------
4K 62.38 +/- 8.27 33.40 +/- 7.51 54.15 +/- 10.23 43.67 +/- 10.53
16K 58.91 +/- 5.35 35.23 +/- 5.88 41.05 +/- 8.87 56.42 +/- 9.24
32K 64.12 +/- 0.68 31.09 +/- 1.48 44.54 +/- 3.51 52.63 +/- 3.65
64K 54.69 +/- 5.54 39.67 +/- 5.81 35.47 +/- 3.11 61.97 +/- 3.27
RX app sys % drops ~19% from 4K to 64K.
Throughput:
niov RX dev Gbps RX flow avg Gbps
----- ---------------- -----------------
4K 300.63 +/- 53.21 75.16 +/- 13.30
16K 321.35 +/- 28.20 80.34 +/- 7.05
32K 347.63 +/- 2.20 86.91 +/- 0.55
64K 332.11 +/- 14.26 83.03 +/- 3.56
Throughput seems to increase, but the stdev is pretty wide so could just
be noise.
kperf support (not yet merged):
https://github.com/facebookexperimental/kperf/commit/8837577f920876bce6986ec18869ac04439ebcd2
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
Signed-off-by: Bobby Eshleman <bobbyeshleman@meta.com>
Link: https://patch.msgid.link/20260805-tcpdm-large-niovs-v8-1-3e0225e2808c@meta.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Cross-merge networking fixes after downstream PR (net-7.1-rc5).
No conflicts, adjacent changes:
drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c
cc199cd1b912 ("net/mlx5e: Reduce branches in napi poll")
c326f9c68921 ("net/mlx5e: xsk: Fix unlocked writing to ICOSQ")
drivers/net/ethernet/mellanox/mlx5/core/eswitch.c
c6df9a65cbb0 ("net/mlx5: Skip disabled vports when setting max TX speed")
1fba57c91416 ("net/mlx5: Add VHCA_ID page management mode support")
net/mac80211/mlme.c
a6e6ccd5bd07 ("wifi: mac80211: consume only present negotiated TTLM maps")
49e62ec6eb06 ("wifi: mac80211: move frame RX handling to type files")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net_devmem_bind_dmabuf() trusts dmabuf->size and sg_dma_len() to be
PAGE_SIZE multiples without checking:
- tx_vec is sized dmabuf->size / PAGE_SIZE, and
net_devmem_get_niov_at() only bounds-checks virt_addr < dmabuf->size
before indexing tx_vec[virt_addr / PAGE_SIZE]. With size =
N*PAGE_SIZE + r (1 <= r < PAGE_SIZE), sendmsg() at iov_base =
N*PAGE_SIZE passes the bound check and reads tx_vec[N] -- one past.
- owner->area.num_niovs = len / PAGE_SIZE while gen_pool_add_owner()
covers the full byte len, so a non-page-multiple non-final sg
desyncs num_niovs from the gen_pool region for every later sg, on
both RX and TX.
dma-buf does not require page-aligned sizes, so the bind path has to
enforce what its own indexing assumes. Reject both with -EINVAL.
The size check is TX-only (only tx_vec is sized off dmabuf->size); the
SG-length check covers both directions.
Fixes: bd61848900bf ("net: devmem: Implement TX path")
Cc: stable@vger.kernel.org
Signed-off-by: David Carlier <devnexen@gmail.com>
Reviewed-by: Bobby Eshleman <bobbyeshleman@meta.com>
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Reviewed-by: Mina Almasry <almasrymina@google.com>
Link: https://patch.msgid.link/20260519203530.66310-1-devnexen@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When a netkit virtual device leases queues from a physical NIC, devmem TX bindings created on the netkit device must still result in the dmabuf being mapped for dma by the physical device. This patch accomplishes this by teaching the bind handler to search for the underlying DMA-capable device by looking it up via leased rx queues. The function netdev_find_netmem_tx_dev(), used for finding the underlying DMA-capable device, can be extended to support other non-netkit NETMEM_TX_NO_DMA devices in the future if needed. Additionally, this patch extends validate_xmit_unreadable_skb() to support the netkit case, where the skb is validated twice: once on the netkit guest device and again on the physical NIC after BPF redirect or ip forwarding. Acked-by: Stanislav Fomichev <sdf@fomichev.me> Signed-off-by: Bobby Eshleman <bobbyeshleman@meta.com> Link: https://patch.msgid.link/20260514-tcp-dm-netkit-v5-3-408c59b91e66@meta.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit db359fccf212 ("mm: introduce a new page type for page pool in
page type") added a page_type field to struct net_iov at the same
offset as struct page::page_type, so that page_pool_set_pp_info() can
call __SetPageNetpp() uniformly on both pages and net_iovs.
The page-type API requires the field to hold the UINT_MAX "no type"
sentinel before a type can be set; for real struct page that invariant
is established by the page allocator on free. struct net_iov is not
allocated through the page allocator, so the field is left as zero
(io_uring zcrx, which uses __GFP_ZERO) or as slab garbage (devmem,
which uses kvmalloc_objs() without zeroing). When the page pool then
calls page_pool_set_pp_info() on a freshly-bound niov,
__SetPageNetpp()'s VM_BUG_ON_PAGE(page->page_type != UINT_MAX) fires
and the kernel BUGs. Triggered in selftests by io_uring zcrx setup
through the fbnic queue restart path:
kernel BUG at ./include/linux/page-flags.h:1062!
RIP: 0010:page_pool_set_pp_info (./include/linux/page-flags.h:1062
net/core/page_pool.c:716)
Call Trace:
<TASK>
net_mp_niov_set_page_pool (net/core/page_pool.c:1360)
io_pp_zc_alloc_netmems (io_uring/zcrx.c:1089 io_uring/zcrx.c:1110)
fbnic_fill_bdq (./include/net/page_pool/helpers.h:160
drivers/net/ethernet/meta/fbnic/fbnic_txrx.c:906)
__fbnic_nv_restart (drivers/net/ethernet/meta/fbnic/fbnic_txrx.c:2470
drivers/net/ethernet/meta/fbnic/fbnic_txrx.c:2874)
fbnic_queue_start (drivers/net/ethernet/meta/fbnic/fbnic_txrx.c:2903)
netdev_rx_queue_reconfig (net/core/netdev_rx_queue.c:137)
__netif_mp_open_rxq (net/core/netdev_rx_queue.c:234)
io_register_zcrx (io_uring/zcrx.c:818 io_uring/zcrx.c:903)
__io_uring_register (io_uring/register.c:931)
__do_sys_io_uring_register (io_uring/register.c:1029)
do_syscall_64 (arch/x86/entry/syscall_64.c:63
arch/x86/entry/syscall_64.c:94)
</TASK>
The same path is reachable through devmem dmabuf binding via
netdev_nl_bind_rx_doit() -> net_devmem_bind_dmabuf_to_queue().
Add a net_iov_init() helper that stamps ->owner, ->type and the
->page_type sentinel, and use it from both the devmem and io_uring
zcrx niov init loops.
Fixes: db359fccf212 ("mm: introduce a new page type for page pool in page type")
Acked-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
Acked-by: Byungchul Park <byungchul@sk.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Acked-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://patch.msgid.link/20260428025320.853452-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net_mp_open_rxq is currently not used in the tree as all callers are
using __net_mp_open_rxq directly, and net_mp_close_rxq is only used
once while all other locations use __net_mp_close_rxq.
Consolidate into a single API, netif_mp_{open,close}_rxq, using the
netif_ prefix to indicate that the caller is responsible for locking.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Co-developed-by: David Wei <dw@davidwei.uk>
Signed-off-by: David Wei <dw@davidwei.uk>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
Link: https://patch.msgid.link/20260402231031.447597-6-daniel@iogearbox.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
binding->dev is protected on the write-side in
mp_dmabuf_devmem_uninstall() against concurrent writes, but due to the
concurrent bare reads in net_devmem_get_binding() and
validate_xmit_unreadable_skb() it should be wrapped in a
READ_ONCE/WRITE_ONCE pair to make sure no compiler optimizations play
with the underlying register in unforeseen ways.
Doesn't present a critical bug because the known compiler optimizations
don't result in bad behavior. There is no tearing on u64, and load
omissions/invented loads would only break if additional binding->dev
references were inlined together (they aren't right now).
This just more strictly follows the linux memory model (i.e.,
"Lock-Protected Writes With Lockless Reads" in
tools/memory-model/Documentation/access-marking.txt).
Fixes: bd61848900bf ("net: devmem: Implement TX path")
Signed-off-by: Bobby Eshleman <bobbyeshleman@meta.com>
Link: https://patch.msgid.link/20260302-devmem-membar-fix-v2-1-5b33c9cbc28b@meta.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Conversion performed via this Coccinelle script:
// SPDX-License-Identifier: GPL-2.0-only
// Options: --include-headers-for-types --all-includes --include-headers --keep-comments
virtual patch
@gfp depends on patch && !(file in "tools") && !(file in "samples")@
identifier ALLOC = {kmalloc_obj,kmalloc_objs,kmalloc_flex,
kzalloc_obj,kzalloc_objs,kzalloc_flex,
kvmalloc_obj,kvmalloc_objs,kvmalloc_flex,
kvzalloc_obj,kvzalloc_objs,kvzalloc_flex};
@@
ALLOC(...
- , GFP_KERNEL
)
$ make coccicheck MODE=patch COCCI=gfp.cocci
Build and boot tested x86_64 with Fedora 42's GCC and Clang:
Linux version 6.19.0+ (user@host) (gcc (GCC) 15.2.1 20260123 (Red Hat 15.2.1-7), GNU ld version 2.44-12.fc42) #1 SMP PREEMPT_DYNAMIC 1970-01-01
Linux version 6.19.0+ (user@host) (clang version 20.1.8 (Fedora 20.1.8-4.fc42), LLD 20.1.8) #1 SMP PREEMPT_DYNAMIC 1970-01-01
Signed-off-by: Kees Cook <kees@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is the result of running the Coccinelle script from scripts/coccinelle/api/kmalloc_objs.cocci. The script is designed to avoid scalar types (which need careful case-by-case checking), and instead replace kmalloc-family calls that allocate struct or union object instances: Single allocations: kmalloc(sizeof(TYPE), ...) are replaced with: kmalloc_obj(TYPE, ...) Array allocations: kmalloc_array(COUNT, sizeof(TYPE), ...) are replaced with: kmalloc_objs(TYPE, COUNT, ...) Flex array allocations: kmalloc(struct_size(PTR, FAM, COUNT), ...) are replaced with: kmalloc_flex(*PTR, FAM, COUNT, ...) (where TYPE may also be *VAR) The resulting allocations no longer return "void *", instead returning "TYPE *". Signed-off-by: Kees Cook <kees@kernel.org>
1) Inline this small helper to reduce code size and decrease cpu costs. 2) Constify its argument. 3) Move it to include/net/netmem.h, as a prereq for the following patch. $ scripts/bloat-o-meter -t vmlinux.2 vmlinux.3 add/remove: 0/2 grow/shrink: 0/4 up/down: 0/-158 (-158) Function old new delta validate_xmit_skb 866 857 -9 __pfx_net_is_devmem_iov 16 - -16 net_is_devmem_iov 22 - -22 get_netmem 152 130 -22 put_netmem 140 114 -26 tcp_recvmsg_locked 3860 3797 -63 Total: Before=22566015, After=22565857, chg -0.00% Signed-off-by: Eric Dumazet <edumazet@google.com> Link: https://patch.msgid.link/20260122045720.1221017-4-edumazet@google.com Signed-off-by: Jakub Kicinski <kuba@kernel.org>