rds_info_getsockopt() reads a callback from rds_info_funcs and invokes it
without protecting the callback's lifetime. Transport modules register
functions stored in this array. For example, rds_tcp.ko registers
rds_tcp_tc_info() for RDS_INFO_TCP_SOCKETS.
This permits the following interleaving:
CPU0 CPU1
rds_info_getsockopt()
func = rds_tcp_tc_info
rmmod rds_tcp
rds_tcp_exit()
rds_info_deregister_func()
rds_info_funcs[offset] = NULL
free rds_tcp module text
func()
The reader can therefore branch to an address in unloaded module text.
Protect callback invocation with SRCU. Enter the SRCU read-side critical
section before loading the callback and leave it only after the callback
returns. Clear the callback with WRITE_ONCE() and call
synchronize_srcu() before deregistration returns, preventing module unload
from freeing its text while an old reader is still executing it. SRCU is
required because callbacks such as RDS_INFO_COUNTERS can sleep.
Keep the callback array unannotated and use READ_ONCE() and WRITE_ONCE()
for concurrent slot access so sparse does not have to apply __rcu through
the function-pointer typedef. Replace the two callback-slot BUG_ON()
checks with WARN_ON_ONCE() and return without changing the slot on
mismatch.
Link: https://lore.kernel.org/netdev/20260720184955.3008978-1-nicoyip.dev@gmail.com/
Suggested-by: Allison Henderson <achender@kernel.org>
Suggested-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Allison Henderson <achender@kernel.org>
Signed-off-by: Chengfeng Ye <nicoyip.dev@gmail.com>
Link: https://patch.msgid.link/20260801054234.3535077-1-nicoyip.dev@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tja1102_p0_probe() schedules work to register the second port. That work
uses the Port 0 private data and phydev. The private data is
devm-allocated, but the driver does not wait for the pending work on
remove.
Store the Port 0 private data in phydev->priv and add a remove callback.
The callback cancels the registration work before devres teardown frees
the state.
This issue was found by a static analysis tool.
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Hongyan Xu <getshell@seu.edu.cn>
Link: https://patch.msgid.link/20260801140643.1871-1-getshell@seu.edu.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
pfcp_net_exit_rtnl() iterates pfcp devices whose sockets
are in the dying netns and queues them for destruction.
So the devices may reside in different netns.
Let's use unregister_netdevice_queue_net() to support per-netns
device unregistration.
list_del() is changed to list_del_init() to avoid queueing the
same device twice.
Even after pfcp_net_exit_rtnl() queues a cross-netns pfcp device,
pfcp_dellink() could be called concurrently for it (once RTNL is
removed). In such a case, __rtnl_net_unlock() will perform the
unregistration.
We can see pfcp0 below is unregistered by the per-netns work
instead of cleanup_net().
# bpftrace -e '#include <linux/netdevice.h>
kprobe:pfcp_dev_uninit {
$dev = (struct net_device *)arg0;
printf("PID: %d | DEV: %s%s\n", pid, $dev->name, kstack());
}
kprobe:pfcp_net_exit_rtnl {
printf("PID: %d%s\n", pid, kstack());
}' &
# ip netns add ns1
# ip netns add ns2
# ip -n ns1 link add pfcp0 link-netns ns2 type pfcp
# ip netns del ns2
PID: 12
pfcp_net_exit_rtnl+5
ops_undo_list+702
cleanup_net+1122
process_scheduled_works+2538
...
PID: 462 | DEV: pfcp0
pfcp_dev_uninit+5
unregister_netdevice_many_notify+7129
unregister_netdevice_many_net+1050
rtnl_net_work_func+136
process_scheduled_works+2538
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260731224406.2444121-3-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
struct pfcp_dev.net is the netns where the backend pfcp socket
resides.
struct pfcp_dev is linked to the pfcp_net.pfcp_dev_list of
the socket's netns.
During netns dismantle or module unload, pfcp_net_exit_rtnl()
iterates the list and queues devices for destruction regardless
of the devices' netns.
Thus, once RTNL is removed, the list can be modified concurrently
from different netns due to device removal.
Let's protect it with per-netns mutex.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260731224406.2444121-2-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When a function-like macro expands to an expression, that expression
doesn't need a semicolon after it. All uses have been verified to
have their own semicolons.
This was found using the following Coccinelle semantic patch:
@r@
identifier i : script:ocaml() { String.lowercase_ascii i = i };
expression e;
@@
*#define i(...) e;
Signed-off-by: Julia Lawall <Julia.Lawall@inria.fr>
Link: https://patch.msgid.link/20260801191002.1383835-10-Julia.Lawall@inria.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When a function-like macro expands to an expression, that expression
doesn't need a semicolon after it. All uses have been verified to
have their own semicolons.
This was found using the following Coccinelle semantic patch:
@r@
identifier i : script:ocaml() { String.lowercase_ascii i = i };
expression e;
@@
*#define i(...) e;
Signed-off-by: Julia Lawall <Julia.Lawall@inria.fr>
Link: https://patch.msgid.link/20260801191002.1383835-5-Julia.Lawall@inria.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
HDS mode has confusing semantics. On GET kernel reports
effective mode. On SET kernel expects explicit config.
Effective mode on GET means that we know the current
state, but we don't know if it's a driver default
or user setting. This matter because driver default
can change automatically when e.g. XDP is attached.
Explicit user setting must not be lost.
With that in mind, we can't restore the HDS setting
like we restore other NIC config. We should always
reset to default ("unknown").
This fixes an issue with tests running after the devmem
test not being able to attach XDP, e.g.
Exception| File "./xdp_metadata.py", line 105, in test_xdp_rss_hash
[...]
Exception| net.lib.py.utils.CmdExitFailure: Command failed
Exception| CMD: ip link set dev ens9np0 xdpdrv pinned /sys/fs/bpf/xdp_metadata_test/xdp_rss_hash
Exception| EXIT: 2
Exception| STDERR: Error: unable to install XDP to device using tcp-data-split.
not ok 1 xdp_metadata.test_xdp_rss_hash.tcp
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Bobby Eshleman <bobbyeshleman@meta.com>
Link: https://patch.msgid.link/20260804151040.2755153-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The external flag is part of the PCI SF port attributes, but unlike the
PCI PF and PCI VF flavours it was never filled into the port dump, so
userspace could not query it directly.
Reporting of the external flag was missed for SF ports. Hence, put
DEVLINK_ATTR_PORT_EXTERNAL for the PCI SF flavour as well, matching what
PCI PF and PCI VF ports already report.
$ devlink port show pci/0033:01:00.0/163840
pci/0033:01:00.0/163840: type eth netdev eth1 flavour pcisf controller 1 pfnum 0 sfnum 77 external true splittable false
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Shay Drory <shayd@nvidia.com>
Link: https://patch.msgid.link/20260803090012.257242-1-shayd@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Pablo Neira Ayuso says:
====================
Netfilter updates for net-next
The following patchset contains Netfilter updates for net-next:
1) Update conncount to use the original tuple after ct lookup to ensure
consistent counting, from Fernando F. Mancera.
2) Remove redundant net_device field in info structure that helps
parse the flowtable path discovery.
3) Move net_device to flowtable check to the flowtable discovery
path parser. This is preparation work to pass the tunnel dst_entry
via .fill_forward_path.
4) Update DSA .fill_forward_path to break at the user DSA, since
the conduit DSA is not used in the datapath. This slighly simplifies
the flowtable path discovery parser.
5) Do not advance index in the path stack prematurely, otherwise
it points to uninitialized slots on error. Not an issue currently
but it could be once tunnel dst_entry is passed via .fill_forward_path.
6) Pass the tunnel dst_entry via dev_fill_forward_path().
7) Update ipip and ip6ip6 tunnels to pass the dst_entry through
dev_fill_forward_path().
8) Call skb_valid_dst() before accessing skb_dst() to ensure dst_entry
is not a template.
9) Use UNACK timeout when RST packet does not match the expected
window while in ESTABLISHED state, the existing approach the CLOSE
state timeout which is only 10 seconds. Adopt a more conservative
timeout by default for this case.
* tag 'nf-next-26-07-31' of git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf-next:
netfilter: conntrack: tcp: use UNACK timeout for non-closing RST packets
netfilter: nf_tables: call skb_valid_dst() before skb_dst()
netfilter: flowtable: release tunnel route on error when building forward path
net: pass dst via net_device_path in dev_fill_forward_path()
net: do not advance stack index from dev_fwd_path()
net: dsa: stop at the user device in .fill_forward_path
netfilter: flowtable: consolidate flowtable device check
netfilter: flowtable: consolidate net_device field in nft_forward_info struct
netfilter: conncount: normalize tuple and zone on successful ct lookup
====================
Link: https://patch.msgid.link/20260731153402.851224-1-pablo@netfilter.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Vikas Gupta says:
====================
bnge: add more functionality
This patch series adds a few functionality for bnge driver:
Patch 1-2: Implements ndo_set_rx_mode_async().
Patch 3: Adds a dedicated HWRM (Hardware Resource Management) command
sequence to handle explicit interface down and up transitions cleanly.
====================
Link: https://patch.msgid.link/20260731163712.3463362-1-vikas.gupta@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Register bnge_set_rx_mode() as ndo_set_rx_mode_async to handle
unicast, multicast, broadcast, and promiscuous filter updates via
CFA_L2_SET_RX_MASK. The async variant receives pre-snapshotted address
lists from the kernel, allowing the driver to issue sleepable HWRM
firmware commands without holding the addr lock.
Move uc_update detection to the caller so the async path can compute
it directly from the snapshotted UC list before calling
bnge_cfg_rx_mode().
Handle -EAGAIN from bnge_hwrm_set_vnic_filter() and
bnge_hwrm_cfa_l2_set_rx_mask() on the open path by scheduling a retry
via netif_rx_mode_schedule_retry() rather than failing the open.
Signed-off-by: Vikas Gupta <vikas.gupta@broadcom.com>
Reviewed-by: Dharmender Garg <dharmender.garg@broadcom.com>
Reviewed-by: Rahul Gupta <rahul-rg.gupta@broadcom.com>
Link: https://patch.msgid.link/20260731163712.3463362-3-vikas.gupta@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Rename bnge_cfg_def_vnic() to bnge_cfg_rx_mode() and update
bnge_mc_list_updated() and bnge_uc_list_updated() to accept
explicit netdev_hw_addr_list pointers rather than deriving
them from the netdev.
Add a snapshot parameter to bnge_cfg_rx_mode() to skip
netif_addr_lock_bh() when the caller provides a pre-snapshotted
list. On the open path (snapshot=false), the live netdev UC list
is passed and the addr lock is taken as before.
Signed-off-by: Vikas Gupta <vikas.gupta@broadcom.com>
Reviewed-by: Dharmender Garg <dharmender.garg@broadcom.com>
Reviewed-by: Rahul Gupta <rahul-rg.gupta@broadcom.com>
Link: https://patch.msgid.link/20260731163712.3463362-2-vikas.gupta@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
bonding_init() ignores register_netdevice_notifier() errors and still
returns success, which can leave the bonding module loaded without its
netdev notifier registered.
Check the error and unwind prior initialization on failure.
This is a future looking check, register_netdevice_notifier()
only fails on double registration or if the registered notifier
itself returns an error.
Signed-off-by: Minhong He <heminhong@kylinos.cn>
Acked-by: Jay Vosburgh <jv@jvosburgh.net>
Link: https://patch.msgid.link/20260803090012.142638-1-heminhong@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
vrf_init_module() ignores register_netdevice_notifier() errors and
continues module initialization, which can leave VRF loaded without its
netdev notifier registered.
Check the error and fail module initialization early.
This is a future looking check, register_netdevice_notifier()
only fails on double registration or if the registered notifier
itself returns an error.
Signed-off-by: Minhong He <heminhong@kylinos.cn>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/20260803090002.142453-1-heminhong@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
macvlan_init_module() ignores register_netdevice_notifier() errors and
continues module initialization, which can leave macvlan loaded without
its netdev notifier registered.
Check the error and fail module initialization early.
This is a future looking check, register_netdevice_notifier()
only fails on double registration or if the registered notifier
itself returns an error.
Signed-off-by: Minhong He <heminhong@kylinos.cn>
Link: https://patch.msgid.link/20260803085950.142325-1-heminhong@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
team_module_init() ignores register_netdevice_notifier() errors and
continues module initialization, which can leave the team module loaded
without its netdev notifier registered.
Check the error and fail module initialization early.
This is a future looking check, register_netdevice_notifier()
only fails on double registration or if the registered notifier
itself returns an error.
Signed-off-by: Minhong He <heminhong@kylinos.cn>
Link: https://patch.msgid.link/20260803085943.142261-1-heminhong@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
lapbeth_init_driver() ignores register_netdevice_notifier() errors and
always returns success, which can leave the module loaded without its
netdev notifier registered.
Check the error and remove the packet type on failure.
This is a future looking check, register_netdevice_notifier()
only fails on double registration or if the registered notifier
itself returns an error.
Signed-off-by: Minhong He <heminhong@kylinos.cn>
Link: https://patch.msgid.link/20260803085936.142160-1-heminhong@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
failover_init() ignores register_netdevice_notifier() errors and always
returns success, which can leave the failover module loaded without its
netdev notifier registered.
Return the notifier registration result directly so module initialization
fails when registration fails.
This is a future looking check, register_netdevice_notifier()
only fails on double registration or if the registered notifier
itself returns an error.
Signed-off-by: Minhong He <heminhong@kylinos.cn>
Link: https://patch.msgid.link/20260731030338.82508-1-heminhong@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
nexthop_init() ignores errors from register_pernet_subsys() and
register_netdevice_notifier(), so a partial initialization can appear
successful.
Check those steps and unwind prior registrations on failure.
Do not check rtnl_register_many(): for built-in code it panics on
failure, so the call cannot return an error to nexthop_init().
Cc: stable+noautosel@kernel.org # untested fix to unlikely error path
Signed-off-by: Minhong He <heminhong@kylinos.cn>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20260731025249.80026-1-heminhong@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The memory allocated for struct vcap_rule_internal, keyfields and
actionfields inside vcap_dup_rule() are not freed in some of the error
paths in vcap_decode_rule(). Fix that by calling vcap_free_rule().
Compile tested only. Issue found using a prototype static analysis tool
built on top of the LLVM compiler infrastructure.
Cc: stable+noautosel@kernel.org # untested fix to unlikely driver error path
Reviewed-by: Joe Damato <joe@dama.to>
Signed-off-by: Abdun Nihaal <nihaal@cse.iitm.ac.in>
Link: https://patch.msgid.link/20260801055507.47534-1-nihaal@cse.iitm.ac.in
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Allison Henderson says:
====================
net/rds: Bug fix ports
This is a small set of net/rds bug fixes and ports from uek to upstream
rds. I've been working on extending the rds selftest case, but need to
stabilize a few more bugs and the first few fall into net with Fixes
tags. I decided to leverage fable for this set and I thought the ports we
clean and well explained.
This series fixes a sleeping-in-softirq bug in the RDS message free
path, a use-after-free of the RDS socket through long-lived MR
references, a message leak in the rds_send_xmit() drop path, and - new
in v4 - a pinned-page leak in the IB transport's MR teardown.
The first three patches are ports of fixes carried in the Oracle UEK
kernel, reworked where the UEK approach no longer applies upstream.
====================
Link: https://patch.msgid.link/20260730041629.3512480-1-achender@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The pages backing an RDS memory region are pinned in __rds_rdma_map()
with rds_pin_pages(), which uses pin_user_pages_fast(): each page's
refcount is biased by GUP_PIN_COUNTING_BIAS to account the pin. The
scatterlist is then handed to the IB transport, and the transport
releases the pages in __rds_ib_teardown_mr() with
set_page_dirty(page);
put_page(page);
put_page() drops a single reference instead of removing the pin bias,
so every MR teardown permanently strands the remaining references and
the pages are never freed - a userspace-triggerable memory leak of up
to RDS_MAX_MSG_SIZE per RDS_GET_MR/RDS_GET_MR_FOR_DEST call.
The conversion to the pin API updated the unpin sites in rdma.c but
missed this one on the transport side. Release the pages with
unpin_user_pages_dirty_lock(), which removes the pin bias and also
dirties the page under the folio lock, closing the truncation race
that a bare set_page_dirty() leaves open.
Dirtying under the folio lock can sleep, which is safe in every path
that reaches __rds_ib_teardown_mr(): the registration-reuse path
(rds_ib_map_frmr()) runs in syscall context, and the pool flush
(rds_ib_unreg_frmr()) runs under pool->flush_lock, a mutex, and
already sleeps in rds_ib_post_inv(). The WARN_ON that guarded the
old irq-context set_page_dirty() case is dropped along with it.
Signed-off-by: Allison Henderson <achender@kernel.org>
Link: https://patch.msgid.link/20260730041629.3512480-5-achender@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When rds_send_xmit() picks the next message off cp_send_queue it takes
its own reference with rds_message_addref(). If the message then hits
the never-retransmit check (RDS_MSG_FLUSH, or an RDMA op that was
already retransmitted), it is moved to the local to_be_dropped list and
that reference is dropped after the batch.
However, if RDS_MSG_ON_CONN has already been cleared, the message is
not added to to_be_dropped and the reference taken above is never
dropped: cp_xmit_rm has not been set at this point, so the loop simply
abandons rm and the rds_message (and everything it pins: pages, MRs,
notifiers) leaks after an RDMA error.
The only other places that clear RDS_MSG_ON_CONN are
rds_send_path_drop_acked() and rds_send_drop_to(), and both can run
while rds_send_xmit() has dropped cp_lock between moving the message
to cp_retrans and re-taking the lock in the never-retransmit check:
rds_send_path_drop_acked() can ack away a message that already sat on
cp_retrans - the RDS_MSG_RETRANSMITTED case above - and
rds_send_drop_to() runs on socket close. Both unlink the message
under cp_lock and put their own reference, leaving the xmit-path
reference stranded.
Drop the reference directly in that case.
This mirrors Oracle UEK commit "net/rds: fix rds_message memleak in
rds_send_xmit".
Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com>
Signed-off-by: Sharath Srinivasan <sharath.srinivasan@oracle.com>
[achender: port to net-next; update commit message, checkpatch nits]
Signed-off-by: Allison Henderson <achender@kernel.org>
Link: https://patch.msgid.link/20260730041629.3512480-4-achender@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Each rds_mr stores a bare back pointer to the socket that created it
(mr->r_sock) but takes no reference on it. When the mr is destroyed it
references the rs. Hence, provisions must be made to avoid the rs
being destroyed before all mrs referencing it have been destroyed.
The MR itself is refcounted, and in-flight messages legitimately hold
MR krefs that can outlive the socket: rds_release() drops the rb-tree
references via rds_rdma_drop_keys(), but a send completion arriving
afterwards drops the final message reference from the CQ handler and
ends up in
rds_message_purge()
__rds_put_mr_final()
rds_destroy_mr() -> takes rs->rs_rdma_lock
dereferencing a socket that may already have been freed.
Oracle UEK fixed the same use-after-free ("rds: Add proper refcnt when
an RDS MR references an RDS Socket") after seeing crashes of the form:
PF: supervisor write access in kernel mode
_raw_spin_lock_irqsave+0x4a/0x6a
__rds_put_mr_final+0x2c/0xe0 [rds]
rds_message_purge+0x13c/0x150 [rds]
rds_message_put+0x39/0x54 [rds]
rds_ib_send_cqe_handler+0x147/0x3dd [rds_rdma]
To fix this, take a socket reference when an MR is created and drop it
when the final MR kref goes away. The reference cycle is broken by
rds_release(), which always runs rds_rdma_drop_keys() on close. So the
socket reference held by an MR never prevents release, it only delays
sk_free() until the last MR user is done.
The hold sits next to kref_init() at both allocation sites -
__rds_rdma_map() and the on-demand-paging path in
rds_cmsg_rdma_args() - so every MR owns exactly one socket reference
from the moment it becomes kref-managed. For that to work on the ODP
path, its get_mr() error handling is converted from a bare kfree() to
kref_put(..., __rds_put_mr_final), with r_trans_private cleared first
since it holds an ERR_PTR there; both sites then tear down through
the same path and a future error-path change cannot silently leak or
double-drop the reference.
Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
[achender: port to net-next (sock_hold/sock_put in place of the UEK
rds_sock_addref/rds_sock_put helpers); also balance the reference on
the rds_cmsg_rdma_args() ODP path and unify its error path with
__rds_put_mr_final(); update commit message]
Signed-off-by: Allison Henderson <achender@kernel.org>
Link: https://patch.msgid.link/20260730041629.3512480-3-achender@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
rds_rdma_free_op() and rds_atomic_free_op() are reached from the IB
send completion path via
rds_ib_tasklet_fn_send()
rds_ib_send_cqe_handler()
rds_message_put()
rds_message_purge()
rds_rdma_free_op() / rds_atomic_free_op()
which runs in tasklet (softirq) context. Both functions unpin the
user pages of the op with unpin_user_pages_dirty_lock(), which uses
set_page_dirty_lock() and thus may take the folio lock and sleep.
Sleeping in softirq context is not allowed and can deadlock or crash.
Dirtying the pages with the non-sleeping set_page_dirty() instead
would just trade one bug for another, as pointed out during review:
the pinned range can be file-backed. rds_pin_pages() pins with
FOLL_LONGTERM, which refuses fs-dax but takes the page-cache pages
of a MAP_SHARED file mapping just fine, and RDS does not restrict
what memory the caller registers as an RDMA destination.
For a file-backed page, set_page_dirty() from a tasklet can take
non-irq-safe filesystem locks (e.g. mapping->i_private_lock and
inode->i_lock in block_dirty_folio()) and deadlock against the task
it interrupted. Without the folio lock, it races with truncation
clearing folio->mapping, which is the race set_page_dirty_lock()
exists to close. The pre-pin_user_pages() version of this code
dirtied pages that way from the tasklet, so that bug is older than
the sleeping unpin.
The page dirtying therefore has to move to process context, not
merely avoid the folio lock. When the final rds_message_put() runs
in atomic context, rds_rdma_free_op() and rds_atomic_free_op() now
leave the op's pages pinned and flag the op. Later, rds_message_put()
hands the message to a work item that unpins the flagged ops' pages
and frees the message from process context. Here,
unpin_user_pages_dirty_lock() is safe outside the atomic context.
Everything else keeps running in the caller's context exactly as
before: the rest of the purge - the zerocopy completion, the socket
put and the MR reference drops - as well as RDMA writes, whose pages
the remote side only reads and which unpin without dirtying,
everything on rds_tcp, and final puts that already happen in process
context (socket close, connection teardown).
Deferring only the unpin means the work item touches nothing but the
pinned pages and the rds module's own memory: it cannot call back
into a transport module, so it changes nothing about the transports'
shutdown and unload ordering. rds_exit() drains any pending unpin
work via destroy_workqueue(rds_wq) before the module goes away.
The Oracle UEK kernel avoids the sleeping unpin by calling
set_page_dirty() directly from the tasklet, which is subject to the
file-backed page problem above, so this deliberately does not follow
UEK here.
Signed-off-by: Allison Henderson <achender@kernel.org>
Link: https://patch.msgid.link/20260730041629.3512480-2-achender@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
airoha_alloc_gdm_device() initializes the txq_lock[] array iterating
over ARRAY_SIZE(dev->txq_lock). ARRAY_SIZE() expands to
sizeof(dev->txq_lock) / sizeof((dev->txq_lock)[0]), but on UP builds
(CONFIG_SMP unset, CONFIG_DEBUG_SPINLOCK unset) arch_spinlock_t is an
empty struct, so sizeof(spinlock_t) is zero and the expression is a
compile-time division by zero (undefined behavior), reported by clang
as "division by zero is undefined [-Wdivision-by-zero]".
Since the array is statically sized with AIROHA_NUM_NETDEV_TX_RINGS,
use the named constant as loop bound instead of ARRAY_SIZE().
Fixes: 78a35725e5 ("net: airoha: defer GDM3/GDM4 WAN mode and GDM2 loopback to QoS offload")
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202607311850.6p0ZUVq4-lkp@intel.com/
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
Link: https://patch.msgid.link/20260731-airoha-spinlock-array-fix-v1-1-863a7e239a5f@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
IFLA_MACVLAN_BC_QUEUE_LEN accepts any u32 and becomes
port->bc_queue_len_used, the only bound on port->bc_queue. rtnetlink checks
CAP_NET_ADMIN against the target netns only, so a user who unshares a
user+net namespace, creates a veth and puts a macvlan on it can set the
backlog to 0xffffffff and flood broadcast frames until the host dies:
Out of memory: Killed process 141 (su) UID:0
Kernel panic - not syncing: System is deadlocked on memory
Call Trace:
vpanic (kernel/panic.c:650)
panic (kernel/panic.c:787)
out_of_memory (mm/oom_kill.c:1166)
__alloc_frozen_pages_noprof (mm/page_alloc.c:4914)
alloc_pages_mpol (mm/mempolicy.c:2490)
folio_alloc_noprof (mm/mempolicy.c:2591)
filemap_fault (mm/filemap.c:3565)
A fixed upper bound does not work. Deployments carrying 600-800 real-time
audio streams run bc_queue_len=100000, and no constant serves both cases:
the queue counts skbs, not bytes, and the frame size is attacker-chosen too
(up to ETH_MAX_MTU on a veth the caller creates).
Gate the elevated range on CAP_NET_ADMIN in the initial user namespace
instead. A backlog of that size is a host-wide tuning decision, and an
unprivileged owner of a namespace it created itself should not be able to
make it; privileged configurations keep working unchanged..
Cc: stable+noautosel@kernel.org # local DoS by userns are a dime a dozen
Reported-by: AutonomousCodeSecurity@microsoft.com
Link: https://lore.kernel.org/r/20260706212556.3199234-1-xmei5@asu.edu
Signed-off-by: Xiang Mei (Microsoft) <xmei5@asu.edu>
Link: https://patch.msgid.link/20260729200621.2521588-1-xmei5@asu.edu
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Kuniyuki Iwashima says:
====================
geneve: Support per-netns device unregistration.
Patch 1 is a prep patch to make patch 2 clean, which
adds a per-netns mutex for geneve linked lists.
Patch 3 supports per-netns netdev unreg by using
unregister_netdevice_queue_net().
====================
Link: https://patch.msgid.link/20260731164612.2148830-1-kuniyu@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
geneve_exit_rtnl_net() iterates geneve devices whose sockets
are in the dying netns and queues them for destruction.
So the devices may reside in different netns.
Let's use unregister_netdevice_queue_net() to support per-netns
device unregistration.
list_del() is changed to list_del_init() to avoid queueing the
same device twice.
Even after geneve_exit_rtnl_net() queues a cross-netns geneve
device, geneve_dellink() can be called concurrently for it.
In such a case, __rtnl_net_unlock() will perform the unregistration.
Note that geneve uses register_pernet_subsys() instead of _device(),
so default_device_exit_batch() guarantees that the async per-netns
works are flushed before ->exit().
Tested:
1. Create geneve device across two netns.
# ip netns add ns1
# ip netns add ns2
# ip -n ns1 link add geneve0 link-netns ns2 type geneve external
2. Run bpftrace to check that geneve_uninit() is called between
->exit_rtnl() and ->exit().
# bpftrace -e '#include <linux/netdevice.h>
kprobe:geneve_uninit {
$dev = (struct net_device *)arg0;
printf("PID: %d | DEV: %s%s\n", pid, $dev->name, kstack());
}
kprobe:geneve_exit_rtnl_net,
kprobe:geneve_exit_net {
printf("PID: %d%s\n", pid, kstack());
}'
3. Remove the netns where the geneve socket resides
# ip netns del ns2
Now, we can see geneve0 is unregistered by per-netns work
instead of cleanup_net() and it finishes before ->exit() to
avoid WARN_ON_ONCE(!list_empty(&gn->sock_list)) there.
PID: 571
geneve_exit_rtnl_net+5
ops_undo_list+702
cleanup_net+1122
process_scheduled_works+2538
...
PID: 1047 | DEV: geneve0
geneve_uninit+5
unregister_netdevice_many_notify+7129
unregister_netdevice_many_net+1050
rtnl_net_work_func+136
process_scheduled_works+2538
...
PID: 571
geneve_exit_net+5
ops_undo_list+1064
cleanup_net+1122
process_scheduled_works+2538
...
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260731164612.2148830-4-kuniyu@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
struct geneve_dev.net is the netns where the backend geneve
socket resides.
struct geneve_dev is linked to the geneve_net.geneve_list of
the socket's netns.
During netns dismantle or module unload, geneve_exit_rtnl_net()
iterates the list and queues devices for destruction regardless
of devices' netns.
Moreover, a socket can be shared by multiple geneve devices in
different netns, and geneve_open() and geneve_stop() modify
geneve_sock.vni_list and geneve_net.sock_list.
Thus, once RTNL is removed, the three lists can be modified
concurrently from different netns due to device removal and
link-up/down.
Let's protect them with per-netns mutex.
geneve_newlink() is still protected by rtnl_net_lock()s, so
acquiring gn->lock twice in geneve_find_dev() and
geneve_configure() is not a problem.
Note that udp_tunnel_notify_add_rx_port() is moved outside of
the mutex, otherwise gn->lock -> utn->lock ordering would trigger
AB-BA deadlock in geneve_offload_rx_ports(), which acquires
gn->lock under utn->lock. Even without gn->lock, geneve_sock_add()
and geneve_offload_rx_ports() are still serialised with (per-netns)
RTNL, so there is no race.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260731164612.2148830-3-kuniyu@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Currently, geneve->sock[46].hlist[46] is unliked from
geneve_sock.vni_list in geneve_stop() and geneve_sock.refcnt is
decremented for each socket later in __geneve_sock_release().
The following patch will introduce a mutex in geneve_net to
protect geneve_sock.{refcnt,vni_list}.
However, udp_tunnel_notify_del_rx_port() must be outside of the
lock to avoid AB-BA deadlock.
To make the change cleaner, let's move hlist_del_init_rcu()
from geneve_stop() to __geneve_sock_release().
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260731164612.2148830-2-kuniyu@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Dipayaan Roy says:
====================
net: mana: add ethtool private flag for full-page RX buffers
On some ARM64 platforms with 4K PAGE_SIZE, utilizing page_pool
fragments for allocation in the RX refill path (~2kB buffer per
fragment) causes 15-20% throughput regression under high connection
counts (>16 TCP streams at 180+ Gbps). Using full-page buffers on
these platforms shows no regression and restores line-rate
performance.
This behavior is observed on a single platform; other platforms
perform better with page_pool fragments, indicating this is not a
page_pool issue but platform-specific.
This series adds an ethtool private flag "full-page-rx" to let the
user opt in to one RX buffer per page:
ethtool --set-priv-flags eth0 full-page-rx on
There is no behavioral change by default. The flag can be persisted
via udev rule for affected platforms.
====================
Link: https://patch.msgid.link/20260729063347.3388035-1-dipayanroy@linux.microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
On some ARM64 platforms with 4K PAGE_SIZE, page_pool fragment
allocation in the RX refill path can cause 15-20% throughput
regression under high connection counts (>16 TCP streams).
Add an ethtool private flag "full-page-rx" that allows the user to
force one RX buffer per page, bypassing the page_pool fragment path.
This restores line-rate (180+ Gbps) performance on affected platforms.
Usage:
ethtool --set-priv-flags eth0 full-page-rx on
There is no behavioral change by default. The flag must be explicitly
enabled by the user or udev rule.
The existing single-buffer-per-page logic for XDP and jumbo frames is
consolidated into a new helper mana_use_single_rxbuf_per_page() which
is now the single decision point for both the automatic and
user-controlled paths.
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Dipayaan Roy <dipayanroy@linux.microsoft.com>
Link: https://patch.msgid.link/20260729063347.3388035-3-dipayanroy@linux.microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Maxime Chevallier says:
====================
net: stmmac: Cleanup rx coalescing computation when using RIWT
Currently when configuring interrupt coalescing on devices that relies
on the Receive Interrupt Watchdog Timer feature of dwmac, the
computation of the RIWT timings leads to off-by-one values when
reporting the timings back to userspace.
RIWT works by arming a watchdog timer upon receiving frames with the RI
bit not set in the descriptor. The timer duration is expressed in units
of 256 stmmac clock ticks, and therefore requires a bit of computation
to derive it :
riwt = (rx_usecs * n_clk_ticks_per_usec) / 256
and conversely
rx_usecs = (riwt * 256) / n_clk_ticks_per_usec
This computation as-is leads to a consistent off-by-one when setting
then getting back the rx-usecs value due to rounding errors (by truncation):
ethtool -C eth1 rx-usecs 42
ethtool -c eth1
-> reports rx-usecs: 41
Let's use DIV_ROUND_CLOSEST instead for the computations. It does have
one side effect, the accepted boundaries for rx-usecs also shifts by one
now, going from [16us, 246us] to [15us, 245us]. For that reason, I'm not
targeting the net tree here, and it's overall a very small issue.
====================
Link: https://patch.msgid.link/20260802114015.214212-1-maxime.chevallier@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When reading the rx-usecs coalescing parameters on a dwmac variant that
uses the RIWT for RX interrupt coalescing, we convert the riwt value to
usecs :
- One riwt cycle is 256 clock ticks, we compute how many ticks in $riwt
cycles
- divide that by how many ticks in a microsecond, and we get the
rx-usecs.
The opposite computation is done when setting the rx-usecs param.
Because of the 256 ratio, we're subjected to off-by-one errors in the
value read-back, which can be reliably measured on i.mx8MP :
$ ethtool -C eth1 rx-usecs 102
$ ethtool -c eth1
Coalesce parameters for eth1:
[...]
rx-usecs: 101
Let's be more explicit about the rounding for the riwt to usec
computations by using DIV_ROUND_CLOSEST, which solves the off-by-one.
This does change the boundaries of accepted rx-usecs parameters, as the
previously accepted values were in the 16-246 us range, and now fall
into the 15-245 range on imx8mp.
Signed-off-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/20260802114015.214212-3-maxime.chevallier@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Receive Interrupt Watchdog Timer is an RX interrupt coalescing mechanism
used by some variants of dwmac. It allows waiting a bit before
triggering the rx interrupts, allowing for batch processing.
The RIWT is configured with a granularity of 256 stmmac clk ticks. Let's
add a comment for that and wrap the raw "1000000" into USEC_PER_SEC, as
we're computing "how many clock cycles in one microsec" with that step.
Signed-off-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/20260802114015.214212-2-maxime.chevallier@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add a reconfig_tx_stall test that detects the possibility of a TX stall
after ring reconfiguration. The key observation is that drivers using
netif_tx_start_all_queues() are prone to experiencing a stall when
reconfiguration completes compared to drivers using
netif_tx_wake_all_queues(). start_all_queues only clears DRV_XOFF, while
wake_all_queues also calls __netif_schedule() to kick the qdisc. Without
the kick, qdisc backlog present at reconfig time can stay stuck until a
new trigger is issued.
The test caps the TX ring at 64 entries so it fills quickly, then
installs FQ on a target TX queue and sends UDP packets with SO_TXTIME
scheduled in the future. With napi_defer_hard_irqs slowing completions,
the small ring can fill when FQ releases the burst, leaving requeued
qdisc backlog with no FQ timer to rescue it. A subsequent ring reconfig
must wake the queues to drain the backlog. Simply starting the queues can
leave it stuck.
Some drivers lack backpressure on the TX path and may not be able to
build up the qdisc backlog the test relies on. In that case report an
expected failure (xfail) instead of a hard failure.
Testing on some of the existing drivers: Driver-A does not have the bug,
Driver-B has the bug, Driver-C had the bug but it is fixed now.
Driver-A:
./drivers/net/ring_reconfig.py -t reconfig_tx_stall
TAP version 13
1..1
Sent 1024 SO_TXTIME packets (+100ms)
Backlog before reconfig: 1176378 bytes
ok 1 ring_reconfig.reconfig_tx_stall
Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
Driver-B:
TAP version 13
1..1
Sent 128 SO_TXTIME packets (+100ms)
Sent 128 SO_TXTIME packets (+200ms)
Backlog before reconfig: 148372 bytes
Check| At ./drivers/net/ring_reconfig.py, line 397, in reconfig_tx_stall:
Check| ksft_eq(0, backlog,
Check failed 0 != 148372 qdisc backlog stuck on queue 1 after ring ....
not ok 1 ring_reconfig.reconfig_tx_stall
Totals: pass:0 fail:1 xfail:0 xpass:0 skip:0 error:0
Driver-C:
TAP version 13
1..1
Sent 128 SO_TXTIME packets (+100ms)
Backlog before reconfig: 192278 bytes
ok 1 ring_reconfig.reconfig_tx_stall
Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
Signed-off-by: Mohsin Bashir <hmohsin@meta.com>
Link: https://patch.msgid.link/20260731021543.1058526-1-mohsin.bashr@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Maxime Chevallier says:
====================
net: stmmac: only use MAC loopback for selftests
stmmac selftests currently use PHY loopback when a PHY is attached, and
fallback to MAC loopback otherwise. PHY loopback however isn't ideal nor
necessary for the tests we are running, that only stress the internal
stmmac features.
Some PHYs bring the carrier state down when in loopback mode, which will
prevent any packet transmission even for ourselves, making selftests
fail for non-stmmac related reasons.
Let's rely only on MAC-side loopback for selftests, making it clear that
any problem found with stmmac selftests are indeed caused by the stmmac
driver, and not external factors.
This was tested on :
- Cyclone V with RGMII link to KSZ9031
- Cyclone V with 1000BaseX
- imx8mp with RGMII link to KSZ9131
- stm32mp157& with RGMII link to RTL8211F
- Allwinner H2+ with an internal PHY
====================
Link: https://patch.msgid.link/20260728155728.1193169-1-maxime.chevallier@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For flow-control testing in loopback mode, we don't need to ask what the
PHY is currently using as pause/asym settings. The PHY is no longer
involved in selftest, we rely strictly on MAC loopback. We therefore
only need to know if the MAC supports Symmetric pause for the test, as
we exercise both TX and RX pause support in the selftest.
Remove phydev requirement for flowcontrol selftest as well as the
AsymPause requirement.
With that, we can also drop the linux/phy.h include.
Signed-off-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Reviewed-by: Oleksij Rempel <o.rempel@pengutronix.de>
Link: https://patch.msgid.link/20260728155728.1193169-3-maxime.chevallier@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Stmmac selftests validate the internal behaviour of the various IPs,
using local loopback. The current logic is relies on PHY-side local
loopback if a PHY is attached, with a fallback to MAC loopback
otherwise.
However, PHY loopback is currently fragile especially for stmmac that
may require RXC to be provided from the PHY. Some PHYs shutdown RXC
while in loopback, while others will report carrier off when in local
loopback. This also fails when using SFP setup with a module that embeds
a PHY, that may also fail to enter loopback.
MAC loopback is done at the GMII level on dwmac, allowing the internal
to be just as meaningful as PHY-loopback testing.
Let's simplify stmmac selftests by only relying on MAC-side local
loopback, which makes the selftests runnable on a wider HW variety.
Signed-off-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260728155728.1193169-2-maxime.chevallier@bootlin.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Breno Leitao says:
====================
net: convert rawv6, ieee802154, phonet and tls getsockopt to sockopt_t
Now that sockopt_init_user() was already merged, builds a user-backed
sockopt_t from the __user pair. A getsockopt leaf can then take
a sockopt_t behind a thin __user wrapper: the wrapper builds it, calls
the leaf, and writes the length back to optlen. The leaf copies with
copy_to_iter() instead of copy_to_user().
Convert four more leaves the way udp and raw already were: ipv6 raw
(do_rawv6_getsockopt), ieee802154 dgram, phonet pep, and tls
(do_tls_getsockopt and its per-option helpers).
Converting phonet surfaced a pre-existing bug: pep_getsockopt() clamps the
length it reports but writes a full int with put_user(), overrunning an
optval buffer shorter than sizeof(int). It is fixed in its own patch, with
a Fixes: tag, before the phonet conversion, so it can be backported alone.
The last patch adds getsockopt_iter selftest fixtures for rawv6,
ieee802154, phonet and tls, checking the returned length and errno across
exact, oversized and short buffers, an unknown optname and a bad level.
For full motivation about these changes, please check the initial thread
at link
https://lore.kernel.org/all/20260401-getsockopt-v2-0-611df6771aff@debian.org/#t
====================
Link: https://patch.msgid.link/20260729-getsockopt_phase4-v4-0-c44576757c17@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add fixtures for the newly converted getsockopt leaves:
- rawv6: IPV6_HDRINCL / IPV6_CHECKSUM int paths + a SOL_RAW
unknown-optname case that reaches do_rawv6_getsockopt().
- tls: TLS_TX_ZEROCOPY_RO, the TLS_TX crypto_info round-trip at
the base and full cipher sizes, the NULL-optval and short
buffer EINVAL paths, and an unknown optname. It skips when
the kernel lacks TLS or AES-GCM.
Each fixture pins the returned-length / errno semantics across exact,
oversized and short buffers and an unknown optname. The semantics are
unchanged by the sockopt_t conversion, so the tests pass both before and
after the leaf conversions.
ieee802154 and phonet are not covered: their CONFIG options are absent
from the net selftest target config, so the cases would only ever skip.
Acked-by: Rémi Denis-Courmont <remi@remlab.net>
Reviewed-by: Sabrina Dubroca <sd@queasysnail.net>
Reviewed-by: Joe Damato <joe@dama.to>
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20260729-getsockopt_phase4-v4-7-c44576757c17@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
