Commit c80794323e ("net: Fix packet reordering caused by GRO and
listified RX cooperation") had the unfortunate effect of adding
latencies in common workloads.
Before the patch, GRO packets were immediately passed to
upper stacks.
After the patch, we can accumulate quite a lot of GRO
packets (depdending on NAPI budget).
My fix is counting in napi->rx_count number of segments
instead of number of logical packets.
Fixes: c80794323e ("net: Fix packet reordering caused by GRO and listified RX cooperation")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Bisected-by: John Sperbeck <jsperbeck@google.com>
Tested-by: Jian Yang <jianyang@google.com>
Cc: Maxim Mikityanskiy <maximmi@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Edward Cree <ecree.xilinx@gmail.com>
Reviewed-by: Alexander Lobakin <alobakin@pm.me>
Link: https://lore.kernel.org/r/20210204213146.4192368-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Following race condition was detected:
<CPU A, t0> - neigh_flush_dev() is under execution and calls
neigh_mark_dead(n) marking the neighbour entry 'n' as dead.
<CPU B, t1> - Executing: __netif_receive_skb() ->
__netif_receive_skb_core() -> arp_rcv() -> arp_process().arp_process()
calls __neigh_lookup() which takes a reference on neighbour entry 'n'.
<CPU A, t2> - Moves further along neigh_flush_dev() and calls
neigh_cleanup_and_release(n), but since reference count increased in t2,
'n' couldn't be destroyed.
<CPU B, t3> - Moves further along, arp_process() and calls
neigh_update()-> __neigh_update() -> neigh_update_gc_list(), which adds
the neighbour entry back in gc_list(neigh_mark_dead(), removed it
earlier in t0 from gc_list)
<CPU B, t4> - arp_process() finally calls neigh_release(n), destroying
the neighbour entry.
This leads to 'n' still being part of gc_list, but the actual
neighbour structure has been freed.
The situation can be prevented from happening if we disallow a dead
entry to have any possibility of updating gc_list. This is what the
patch intends to achieve.
Fixes: 9c29a2f55e ("neighbor: Fix locking order for gc_list changes")
Signed-off-by: Chinmay Agarwal <chinagar@codeaurora.org>
Reviewed-by: Cong Wang <xiyou.wangcong@gmail.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/r/20210127165453.GA20514@chinagar-linux.qualcomm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit 3226b158e6 ("net: avoid 32 x truesize under-estimation for
tiny skbs") ensured that skbs with data size lower than 1025 bytes
will be kmalloc'ed to avoid excessive page cache fragmentation and
memory consumption.
However, the fix adressed only __napi_alloc_skb() (primarily for
virtio_net and napi_get_frags()), but the issue can still be achieved
through __netdev_alloc_skb(), which is still used by several drivers.
Drivers often allocate a tiny skb for headers and place the rest of
the frame to frags (so-called copybreak).
Mirror the condition to __netdev_alloc_skb() to handle this case too.
Since v1 [0]:
- fix "Fixes:" tag;
- refine commit message (mention copybreak usecase).
[0] https://lore.kernel.org/netdev/20210114235423.232737-1-alobakin@pm.me
Fixes: a1c7fff7e1 ("net: netdev_alloc_skb() use build_skb()")
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Link: https://lore.kernel.org/r/20210115150354.85967-1-alobakin@pm.me
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Cited patch below blocked the TLS TX device offload unless HW_CSUM
is set. This broke devices that use IP_CSUM && IP6_CSUM.
Here we fix it.
Note that the single HW_TLS_TX feature flag indicates support for
both IPv4/6, hence it should still be disabled in case only one of
(IP_CSUM | IPV6_CSUM) is set.
Fixes: ae0b04b238 ("net: Disable NETIF_F_HW_TLS_TX when HW_CSUM is disabled")
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Reported-by: Rohit Maheshwari <rohitm@chelsio.com>
Reviewed-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Link: https://lore.kernel.org/r/20210114151215.7061-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Both virtio net and napi_get_frags() allocate skbs
with a very small skb->head
While using page fragments instead of a kmalloc backed skb->head might give
a small performance improvement in some cases, there is a huge risk of
under estimating memory usage.
For both GOOD_COPY_LEN and GRO_MAX_HEAD, we can fit at least 32 allocations
per page (order-3 page in x86), or even 64 on PowerPC
We have been tracking OOM issues on GKE hosts hitting tcp_mem limits
but consuming far more memory for TCP buffers than instructed in tcp_mem[2]
Even if we force napi_alloc_skb() to only use order-0 pages, the issue
would still be there on arches with PAGE_SIZE >= 32768
This patch makes sure that small skb head are kmalloc backed, so that
other objects in the slab page can be reused instead of being held as long
as skbs are sitting in socket queues.
Note that we might in the future use the sk_buff napi cache,
instead of going through a more expensive __alloc_skb()
Another idea would be to use separate page sizes depending
on the allocated length (to never have more than 4 frags per page)
I would like to thank Greg Thelen for his precious help on this matter,
analysing crash dumps is always a time consuming task.
Fixes: fd11a83dd3 ("net: Pull out core bits of __netdev_alloc_skb and add __napi_alloc_skb")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Paolo Abeni <pabeni@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Link: https://lore.kernel.org/r/20210113161819.1155526-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
skb_seq_read iterates over an skb, returning pointer and length of
the next data range with each call.
It relies on kmap_atomic to access highmem pages when needed.
An skb frag may be backed by a compound page, but kmap_atomic maps
only a single page. There are not enough kmap slots to always map all
pages concurrently.
Instead, if kmap_atomic is needed, iterate over each page.
As this increases the number of calls, avoid this unless needed.
The necessary condition is captured in skb_frag_must_loop.
I tried to make the change as obvious as possible. It should be easy
to verify that nothing changes if skb_frag_must_loop returns false.
Tested:
On an x86 platform with
CONFIG_HIGHMEM=y
CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP=y
CONFIG_NETFILTER_XT_MATCH_STRING=y
Run
ip link set dev lo mtu 1500
iptables -A OUTPUT -m string --string 'badstring' -algo bm -j ACCEPT
dd if=/dev/urandom of=in bs=1M count=20
nc -l -p 8000 > /dev/null &
nc -w 1 -q 0 localhost 8000 < in
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
If register_netdevice() fails at the very last stage - the
notifier call - some subsystems may have already seen it and
grabbed a reference. struct net_device can't be freed right
away without calling netdev_wait_all_refs().
Now that we have a clean interface in form of dev->needs_free_netdev
and lenient free_netdev() we can undo what commit 93ee31f14f ("[NET]:
Fix free_netdev on register_netdev failure.") has done and complete
the unregistration path by bringing the net_set_todo() call back.
After registration fails user is still expected to explicitly
free the net_device, so make sure ->needs_free_netdev is cleared,
otherwise rolling back the registration will cause the old double
free for callers who release rtnl_lock before the free.
This also solves the problem of priv_destructor not being called
on notifier error.
net_set_todo() will be moved back into unregister_netdevice_queue()
in a follow up.
Reported-by: Hulk Robot <hulkci@huawei.com>
Reported-by: Yang Yingliang <yangyingliang@huawei.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
There are two flavors of handling netdev registration:
- ones called without holding rtnl_lock: register_netdev() and
unregister_netdev(); and
- those called with rtnl_lock held: register_netdevice() and
unregister_netdevice().
While the semantics of the former are pretty clear, the same can't
be said about the latter. The netdev_todo mechanism is utilized to
perform some of the device unregistering tasks and it hooks into
rtnl_unlock() so the locked variants can't actually finish the work.
In general free_netdev() does not mix well with locked calls. Most
drivers operating under rtnl_lock set dev->needs_free_netdev to true
and expect core to make the free_netdev() call some time later.
The part where this becomes most problematic is error paths. There is
no way to unwind the state cleanly after a call to register_netdevice(),
since unreg can't be performed fully without dropping locks.
Make free_netdev() more lenient, and defer the freeing if device
is being unregistered. This allows error paths to simply call
free_netdev() both after register_netdevice() failed, and after
a call to unregister_netdevice() but before dropping rtnl_lock.
Simplify the error paths which are currently doing gymnastics
around free_netdev() handling.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
skbs in fraglist could be shared by a BPF filter loaded at TC. If TC
writes, it will call skb_ensure_writable -> pskb_expand_head to create
a private linear section for the head_skb. And then call
skb_clone_fraglist -> skb_get on each skb in the fraglist.
skb_segment_list overwrites part of the skb linear section of each
fragment itself. Even after skb_clone, the frag_skbs share their
linear section with their clone in PF_PACKET.
Both sk_receive_queue of PF_PACKET and PF_INET (or PF_INET6) can have
a link for the same frag_skbs chain. If a new skb (not frags) is
queued to one of the sk_receive_queue, multiple ptypes can see and
release this. It causes use-after-free.
[ 4443.426215] ------------[ cut here ]------------
[ 4443.426222] refcount_t: underflow; use-after-free.
[ 4443.426291] WARNING: CPU: 7 PID: 28161 at lib/refcount.c:190
refcount_dec_and_test_checked+0xa4/0xc8
[ 4443.426726] pstate: 60400005 (nZCv daif +PAN -UAO)
[ 4443.426732] pc : refcount_dec_and_test_checked+0xa4/0xc8
[ 4443.426737] lr : refcount_dec_and_test_checked+0xa0/0xc8
[ 4443.426808] Call trace:
[ 4443.426813] refcount_dec_and_test_checked+0xa4/0xc8
[ 4443.426823] skb_release_data+0x144/0x264
[ 4443.426828] kfree_skb+0x58/0xc4
[ 4443.426832] skb_queue_purge+0x64/0x9c
[ 4443.426844] packet_set_ring+0x5f0/0x820
[ 4443.426849] packet_setsockopt+0x5a4/0xcd0
[ 4443.426853] __sys_setsockopt+0x188/0x278
[ 4443.426858] __arm64_sys_setsockopt+0x28/0x38
[ 4443.426869] el0_svc_common+0xf0/0x1d0
[ 4443.426873] el0_svc_handler+0x74/0x98
[ 4443.426880] el0_svc+0x8/0xc
Fixes: 3a1296a38d (net: Support GRO/GSO fraglist chaining.)
Signed-off-by: Dongseok Yi <dseok.yi@samsung.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/r/1610072918-174177-1-git-send-email-dseok.yi@samsung.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
pneigh_enqueue() tries to obtain a random delay by mod
NEIGH_VAR(p, PROXY_DELAY). However, NEIGH_VAR(p, PROXY_DELAY)
migth be zero at that point because someone could write zero
to /proc/sys/net/ipv4/neigh/[device]/proxy_delay after the
callers check it.
This patch uses prandom_u32_max() to get a random delay instead
which avoids potential division by zero.
Signed-off-by: weichenchen <weichen.chen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Accesses to dev->xps_rxqs_map (when using dev->num_tc) should be
protected by the rtnl lock, like we do for netif_set_xps_queue. I didn't
see an actual bug being triggered, but let's be safe here and take the
rtnl lock while accessing the map in sysfs.
Fixes: 8af2c06ff4 ("net-sysfs: Add interface for Rx queue(s) map per Tx queue")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Two race conditions can be triggered when storing xps rxqs, resulting in
various oops and invalid memory accesses:
1. Calling netdev_set_num_tc while netif_set_xps_queue:
- netif_set_xps_queue uses dev->tc_num as one of the parameters to
compute the size of new_dev_maps when allocating it. dev->tc_num is
also used to access the map, and the compiler may generate code to
retrieve this field multiple times in the function.
- netdev_set_num_tc sets dev->tc_num.
If new_dev_maps is allocated using dev->tc_num and then dev->tc_num
is set to a higher value through netdev_set_num_tc, later accesses to
new_dev_maps in netif_set_xps_queue could lead to accessing memory
outside of new_dev_maps; triggering an oops.
2. Calling netif_set_xps_queue while netdev_set_num_tc is running:
2.1. netdev_set_num_tc starts by resetting the xps queues,
dev->tc_num isn't updated yet.
2.2. netif_set_xps_queue is called, setting up the map with the
*old* dev->num_tc.
2.3. netdev_set_num_tc updates dev->tc_num.
2.4. Later accesses to the map lead to out of bound accesses and
oops.
A similar issue can be found with netdev_reset_tc.
One way of triggering this is to set an iface up (for which the driver
uses netdev_set_num_tc in the open path, such as bnx2x) and writing to
xps_rxqs in a concurrent thread. With the right timing an oops is
triggered.
Both issues have the same fix: netif_set_xps_queue, netdev_set_num_tc
and netdev_reset_tc should be mutually exclusive. We do that by taking
the rtnl lock in xps_rxqs_store.
Fixes: 8af2c06ff4 ("net-sysfs: Add interface for Rx queue(s) map per Tx queue")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Accesses to dev->xps_cpus_map (when using dev->num_tc) should be
protected by the rtnl lock, like we do for netif_set_xps_queue. I didn't
see an actual bug being triggered, but let's be safe here and take the
rtnl lock while accessing the map in sysfs.
Fixes: 184c449f91 ("net: Add support for XPS with QoS via traffic classes")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Two race conditions can be triggered when storing xps cpus, resulting in
various oops and invalid memory accesses:
1. Calling netdev_set_num_tc while netif_set_xps_queue:
- netif_set_xps_queue uses dev->tc_num as one of the parameters to
compute the size of new_dev_maps when allocating it. dev->tc_num is
also used to access the map, and the compiler may generate code to
retrieve this field multiple times in the function.
- netdev_set_num_tc sets dev->tc_num.
If new_dev_maps is allocated using dev->tc_num and then dev->tc_num
is set to a higher value through netdev_set_num_tc, later accesses to
new_dev_maps in netif_set_xps_queue could lead to accessing memory
outside of new_dev_maps; triggering an oops.
2. Calling netif_set_xps_queue while netdev_set_num_tc is running:
2.1. netdev_set_num_tc starts by resetting the xps queues,
dev->tc_num isn't updated yet.
2.2. netif_set_xps_queue is called, setting up the map with the
*old* dev->num_tc.
2.3. netdev_set_num_tc updates dev->tc_num.
2.4. Later accesses to the map lead to out of bound accesses and
oops.
A similar issue can be found with netdev_reset_tc.
One way of triggering this is to set an iface up (for which the driver
uses netdev_set_num_tc in the open path, such as bnx2x) and writing to
xps_cpus in a concurrent thread. With the right timing an oops is
triggered.
Both issues have the same fix: netif_set_xps_queue, netdev_set_num_tc
and netdev_reset_tc should be mutually exclusive. We do that by taking
the rtnl lock in xps_cpus_store.
Fixes: 184c449f91 ("net: Add support for XPS with QoS via traffic classes")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
There are some use cases for netdev_notify_peers in the context
when rtnl lock is already held. Introduce lockless version
of netdev_notify_peers call to save the extra code to call
call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
After that, convert netdev_notify_peers to call the new helper.
Suggested-by: Nathan Lynch <nathanl@linux.ibm.com>
Signed-off-by: Lijun Pan <ljp@linux.ibm.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Pull networking updates from Jakub Kicinski:
"Core:
- support "prefer busy polling" NAPI operation mode, where we defer
softirq for some time expecting applications to periodically busy
poll
- AF_XDP: improve efficiency by more batching and hindering the
adjacency cache prefetcher
- af_packet: make packet_fanout.arr size configurable up to 64K
- tcp: optimize TCP zero copy receive in presence of partial or
unaligned reads making zero copy a performance win for much smaller
messages
- XDP: add bulk APIs for returning / freeing frames
- sched: support fragmenting IP packets as they come out of conntrack
- net: allow virtual netdevs to forward UDP L4 and fraglist GSO skbs
BPF:
- BPF switch from crude rlimit-based to memcg-based memory accounting
- BPF type format information for kernel modules and related tracing
enhancements
- BPF implement task local storage for BPF LSM
- allow the FENTRY/FEXIT/RAW_TP tracing programs to use
bpf_sk_storage
Protocols:
- mptcp: improve multiple xmit streams support, memory accounting and
many smaller improvements
- TLS: support CHACHA20-POLY1305 cipher
- seg6: add support for SRv6 End.DT4/DT6 behavior
- sctp: Implement RFC 6951: UDP Encapsulation of SCTP
- ppp_generic: add ability to bridge channels directly
- bridge: Connectivity Fault Management (CFM) support as is defined
in IEEE 802.1Q section 12.14.
Drivers:
- mlx5: make use of the new auxiliary bus to organize the driver
internals
- mlx5: more accurate port TX timestamping support
- mlxsw:
- improve the efficiency of offloaded next hop updates by using
the new nexthop object API
- support blackhole nexthops
- support IEEE 802.1ad (Q-in-Q) bridging
- rtw88: major bluetooth co-existance improvements
- iwlwifi: support new 6 GHz frequency band
- ath11k: Fast Initial Link Setup (FILS)
- mt7915: dual band concurrent (DBDC) support
- net: ipa: add basic support for IPA v4.5
Refactor:
- a few pieces of in_interrupt() cleanup work from Sebastian Andrzej
Siewior
- phy: add support for shared interrupts; get rid of multiple driver
APIs and have the drivers write a full IRQ handler, slight growth
of driver code should be compensated by the simpler API which also
allows shared IRQs
- add common code for handling netdev per-cpu counters
- move TX packet re-allocation from Ethernet switch tag drivers to a
central place
- improve efficiency and rename nla_strlcpy
- number of W=1 warning cleanups as we now catch those in a patchwork
build bot
Old code removal:
- wan: delete the DLCI / SDLA drivers
- wimax: move to staging
- wifi: remove old WDS wifi bridging support"
* tag 'net-next-5.11' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next: (1922 commits)
net: hns3: fix expression that is currently always true
net: fix proc_fs init handling in af_packet and tls
nfc: pn533: convert comma to semicolon
af_vsock: Assign the vsock transport considering the vsock address flags
af_vsock: Set VMADDR_FLAG_TO_HOST flag on the receive path
vsock_addr: Check for supported flag values
vm_sockets: Add VMADDR_FLAG_TO_HOST vsock flag
vm_sockets: Add flags field in the vsock address data structure
net: Disable NETIF_F_HW_TLS_TX when HW_CSUM is disabled
tcp: Add logic to check for SYN w/ data in tcp_simple_retransmit
net: mscc: ocelot: install MAC addresses in .ndo_set_rx_mode from process context
nfc: s3fwrn5: Release the nfc firmware
net: vxget: clean up sparse warnings
mlxsw: spectrum_router: Use eXtended mezzanine to offload IPv4 router
mlxsw: spectrum: Set KVH XLT cache mode for Spectrum2/3
mlxsw: spectrum_router_xm: Introduce basic XM cache flushing
mlxsw: reg: Add Router LPM Cache Enable Register
mlxsw: reg: Add Router LPM Cache ML Delete Register
mlxsw: spectrum_router_xm: Implement L-value tracking for M-index
mlxsw: reg: Add XM Router M Table Register
...
Pull scheduler updates from Thomas Gleixner:
- migrate_disable/enable() support which originates from the RT tree
and is now a prerequisite for the new preemptible kmap_local() API
which aims to replace kmap_atomic().
- A fair amount of topology and NUMA related improvements
- Improvements for the frequency invariant calculations
- Enhanced robustness for the global CPU priority tracking and decision
making
- The usual small fixes and enhancements all over the place
* tag 'sched-core-2020-12-14' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (61 commits)
sched/fair: Trivial correction of the newidle_balance() comment
sched/fair: Clear SMT siblings after determining the core is not idle
sched: Fix kernel-doc markup
x86: Print ratio freq_max/freq_base used in frequency invariance calculations
x86, sched: Use midpoint of max_boost and max_P for frequency invariance on AMD EPYC
x86, sched: Calculate frequency invariance for AMD systems
irq_work: Optimize irq_work_single()
smp: Cleanup smp_call_function*()
irq_work: Cleanup
sched: Limit the amount of NUMA imbalance that can exist at fork time
sched/numa: Allow a floating imbalance between NUMA nodes
sched: Avoid unnecessary calculation of load imbalance at clone time
sched/numa: Rename nr_running and break out the magic number
sched: Make migrate_disable/enable() independent of RT
sched/topology: Condition EAS enablement on FIE support
arm64: Rebuild sched domains on invariance status changes
sched/topology,schedutil: Wrap sched domains rebuild
sched/uclamp: Allow to reset a task uclamp constraint value
sched/core: Fix typos in comments
Documentation: scheduler: fix information on arch SD flags, sched_domain and sched_debug
...
