When mana_per_port_queue_reset_work_handler() runs after a previous
detach succeeded but attach failed, the port is left in a detached
state with apc->tx_qp and apc->rxqs already freed. Calling
mana_detach() again unconditionally leads to NULL pointer dereferences
during queue teardown.
Add an early exit in mana_detach() when the port is already in
detached state (!netif_device_present) for non-close callers, making
it safe to call idempotently. This allows the queue reset handler and
other recovery paths to simply retry mana_attach() without redundant
teardown.
Fixes: 3b194343c2 ("net: mana: Implement ndo_tx_timeout and serialize queue resets per port.")
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Dipayaan Roy <dipayanroy@linux.microsoft.com>
Link: https://patch.msgid.link/20260525081129.1230035-3-dipayanroy@linux.microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When queue allocation fails partway through, the error cleanup frees
and NULLs apc->tx_qp and apc->rxqs. Multiple teardown paths such as
mana_remove(), mana_change_mtu() recovery, and internal error handling
in mana_alloc_queues() can subsequently call into functions that
dereference these pointers without NULL checks:
- mana_chn_setxdp() dereferences apc->rxqs[0], causing a NULL pointer
dereference panic (CR2: 0000000000000000 at mana_chn_setxdp+0x26).
- mana_destroy_vport() iterates apc->rxqs without a NULL check.
- mana_fence_rqs() iterates apc->rxqs without a NULL check.
- mana_dealloc_queues() iterates apc->tx_qp without a NULL check.
Add NULL guards for apc->rxqs in mana_fence_rqs(),
mana_destroy_vport(), and before the mana_chn_setxdp() call. Add a
NULL guard for apc->tx_qp in mana_dealloc_queues() to skip TX queue
draining when TX queues were never allocated or already freed.
Fixes: ca9c54d2d6 ("net: mana: Add a driver for Microsoft Azure Network Adapter (MANA)")
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Dipayaan Roy <dipayanroy@linux.microsoft.com>
Link: https://patch.msgid.link/20260525081129.1230035-2-dipayanroy@linux.microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Pull networking fixes from Paolo Abeni:
"This is again significantly bigger than the same point into the
previous cycle, but at least smaller than last week.
I'm not aware of any pending regression for the current cycle.
Including fixes from netfilter.
Current release - regressions:
- netfilter: walk fib6_siblings under RCU
Previous releases - regressions:
- netlink: fix sending unassigned nsid after assigned one
- bridge: fix sleep in atomic context in netlink path
- sched: fix ethx:ingress -> ethy:egress -> ethx:ingress mirred loop
- ipv4: fix net->ipv4.sysctl_local_reserved_ports UaF
- eth: tun: free page on short-frame rejection in tun_xdp_one()
Previous releases - always broken:
- skbuff: fix missing zerocopy reference in pskb_carve helpers
- handshake: drain pending requests at net namespace exit
- ethtool:
- rss: avoid modifying the RSS context response
- module: avoid leaking a netdev ref on module flash errors
- coalesce: cap profile updates at NET_DIM_PARAMS_NUM_PROFILES
- netfilter: fix dst corruption in same register operation
- nfc: hci: fix out-of-bounds read in HCP header parsing
- ipv6: exthdrs: refresh nh pointer after ipv6_hop_jumbo()
- eth:
- vti: use ip6_tnl.net in vti6_changelink().
- vxlan: do not reuse cached ip_hdr() value after
skb_tunnel_check_pmtu()"
* tag 'net-7.1-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (94 commits)
dpll: zl3073x: make frequency monitor a per-device attribute
dpll: zl3073x: use __dpll_device_change_ntf() and remove change_work
dpll: export __dpll_device_change_ntf() for use under dpll_lock
net/handshake: Drain pending requests at net namespace exit
net/handshake: Verify file-reference balance in submit paths
net/handshake: Close the submit-side sock_hold race
net/handshake: hand off the pinned file reference to accept_doit
net/handshake: Take a long-lived file reference at submit
net/handshake: Pass negative errno through handshake_complete()
nvme-tcp: store negative errno in queue->tls_err
net/handshake: Use spin_lock_bh for hn_lock
net: skbuff: fix missing zerocopy reference in pskb_carve helpers
net: hibmcge: move dma_rmb() after dma_sync_single_for_cpu() in RX path
net: hibmcge: disable Relaxed Ordering to fix RX packet corruption
selftests/tc-testing: Add netem test case exercising loops
selftests/tc-testing: Add mirred test cases exercising loops
net/sched: act_mirred: Fix return code in early mirred redirect error paths
net/sched: act_mirred: Fix blockcast recursion bypass leading to stack overflow
net/sched: Fix ethx:ingress -> ethy:egress -> ethx:ingress mirred loop
net/sched: fix packet loop on netem when duplicate is on
...
Pull gpio fixes from Bartosz Golaszewski:
- fix interrupt handling in gpio-mxc
- fix scoped_guard() usage in gpio-adnp
- don't accept partial writes in gpio-virtuser debugfs interface as
they can't really work correctly
- fix resource leaks in gpio-rockchip
- fix locking issues in remove path in shared GPIO management
- undo the vote of a GPIO shared proxy virtual device on GPIO release
* tag 'gpio-fixes-for-v7.1-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/brgl/linux:
gpio: rockchip: teardown bugs and resource leaks
gpio: rockchip: convert bank->clk to devm_clk_get_enabled()
gpio: virtuser: Fix uninitialized data bug in gpio_virtuser_direction_do_write()
gpio: shared: fix lockdep false positive by removing unneeded lock
gpio: shared: fix deadlock on shared proxy's parent removal
gpio: adnp: fix flow control regression caused by scoped_guard()
gpio: shared: undo the vote of the proxy on GPIO free
gpio: mxc: fix irq_high handling
Nicholas Carlini reports that the keyring code calls assoc_array_find()
in find_key_to_update() without holding the RCU read lock, while the
assoc_array_gc() code really is designed around removing the node from
the tree and then freeing it after an RCU grace-period.
The regular key handling doesn't see this because holding the keyring
semaphore hides any lifetime issues, but the persistent key handling
uses a different model.
Instead of extending the keyring locking, just do the simple RCU locking
that the assoc_array was designed for.
Reported-by: Nicholas Carlini <npc@anthropic.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Jarkko Sakkinen <jarkko@kernel.org>
Cc: Paul Moore <paul@paul-moore.com>
Cc: James Morris James Morris <jmorris@namei.org>
Cc: Serge E. Hallyn <serge@hallyn.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Address several teardown issues and resource leaks in the driver's remove
path and error handling:
1. Debounce clock reference leak: The debounce clock (bank->db_clk) is
obtained using of_clk_get() which increments the clock's reference
count, but clk_put() is never called. Register a devm action to
cleanly release it on unbind. Note that of_clk_get(..., 1) remains
necessary over devm_clk_get() because the DT binding does not define
clock-names, precluding name-based lookup.
2. Unregistered chained IRQ handler: The chained IRQ handler is not
disconnected in remove(). If a stray interrupt fires after the driver
is removed, the kernel attempts to execute a stale handler, leading
to a panic. Fix this by clearing the handler in remove().
3. IRQ domain leak: The linear IRQ domain and its generic chips are
allocated manually during probe but never removed. Remove the IRQ
domain during driver teardown to free the associated generic chips
and mappings.
Fixes: 936ee2675e ("gpio/rockchip: add driver for rockchip gpio")
Assisted-by: Antigravity:gemini-3.5-flash
Signed-off-by: Marco Scardovi <scardracs@disroot.org>
Link: https://patch.msgid.link/20260526171050.12785-3-scardracs@disroot.org
[Bartosz: don't emit an error message on devres allocation failure]
Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>
The bank->clk was previously obtained via of_clk_get() and manually
prepared/enabled. However, it was missing a corresponding clk_put() in
both the error paths and the remove function, leading to a reference leak.
Convert the allocation to devm_clk_get_enabled(), which also properly
propagates failures from clk_prepare_enable() that were previously ignored.
The GPIO bank device uses the same OF node as the previous of_clk_get()
call, so devm_clk_get_enabled(dev, NULL) correctly resolves the same
clock provider entry.
Fix the reference leak and simplify the code by removing the manual
clk_disable_unprepare() calls in the probe error paths and in the
remove function.
Fixes: 936ee2675e ("gpio/rockchip: add driver for rockchip gpio")
Assisted-by: Antigravity:gemini-3.5-flash
Signed-off-by: Marco Scardovi <scardracs@disroot.org>
Link: https://patch.msgid.link/20260526171050.12785-2-scardracs@disroot.org
Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>
If *ppos is non-zero (user-space write split over multiple calls to
write()) then simple_write_to_buffer() won't initialize the start of the
buffer. Really, non-zero values for *ppos aren't going to work at all.
Check for that and return -EINVAL at the start of the function.
Fixes: 91581c4b3f ("gpio: virtuser: new virtual testing driver for the GPIO API")
Signed-off-by: Dan Carpenter <error27@gmail.com>
Link: https://patch.msgid.link/ahP3BJWWy-m_qI0X@stanley.mountain
Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>
By the time gpio_device_teardown_shared() is called, the parent device
is gone from the global list of GPIO devices and all outstanding SRCU
read-side critical sections have completed. That means that no
concurrent gpio_find_and_request() can call
gpio_shared_add_proxy_lookup() for this device at this time. There's
also no risk of the parent device being re-bound to the driver before
the unbinding completes (including the child devices).
Lockdep produces a false-positive report about a possible circular
dependency as it doesn't know the ordering guarantee. Not taking the
ref->lock in gpio_device_teardown_shared() silences it and is safe to do.
Cc: stable@vger.kernel.org
Fixes: ea513dd3c0 ("gpio: shared: make locking more fine-grained")
Reviewed-by: Linus Walleij <linusw@kernel.org>
Link: https://patch.msgid.link/20260522-gpio-shared-deadlock-v1-2-76bca088f8c0@oss.qualcomm.com
Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>
Commit 710abda580 ("gpio: shared: call gpio_chip::of_xlate() if set")
used the mutex embedded in struct gpio_shared_entry to protect the
offset field which now can be modified after assignment. The critical
section however is too wide and introduced a potential deadlock on the
removal of the shared GPIO proxy's parent.
Make the critical section shorter - only protect the offset when it's
being read.
While at it: mention the fact that the entry lock is now also used to
protect against concurrent access to the offset field in the structure's
documentation.
Cc: stable@vger.kernel.org
Fixes: 710abda580 ("gpio: shared: call gpio_chip::of_xlate() if set")
Reviewed-by: Linus Walleij <linusw@kernel.org>
Link: https://patch.msgid.link/20260522-gpio-shared-deadlock-v1-1-76bca088f8c0@oss.qualcomm.com
Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>
Ivan Vecera says:
====================
dpll: zl3073x: various fixes
Three fixes for the zl3073x DPLL driver.
Patch 1 exports __dpll_device_change_ntf() for use by drivers that
need to send device change notifications from within callbacks
already running under dpll_lock.
Patch 2 replaces the change_work workqueue mechanism with direct
calls to __dpll_device_change_ntf(), eliminating a race condition
where the work handler could dereference a freed dpll_dev pointer
during device teardown.
Patch 3 moves the freq_monitor flag from per-DPLL to per-device
scope to match the hardware behavior where frequency measurement
registers are shared across all DPLL channels.
====================
Link: https://patch.msgid.link/20260526074525.1451008-1-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The frequency monitoring feature uses shared hardware registers
that measure input reference frequencies independently of
individual DPLL channels. However, the freq_monitor flag was
incorrectly placed in the per-DPLL structure, causing each
channel to track its own enable/disable state independently.
Since the DPLL core calls measured_freq_get() only for the first
pin registration, the measured_freq_check() in the periodic worker
was gated by the per-DPLL freq_monitor flag of whichever channel
happens to be checked. If the first DPLL channel had frequency
monitoring disabled while another had it enabled, measurements
were never reported.
Move freq_monitor from struct zl3073x_dpll to struct zl3073x_dev
so all DPLL channels share a single flag, matching the hardware
behavior. Update freq_monitor_set() to notify other DPLL devices
about the change (like phase_offset_avg_factor_set() already does)
and remove the mode-dependent guard in zl3073x_dpll_changes_check()
since all input pin monitoring (pin state, phase offset, FFO, and
measured frequency) works correctly in all DPLL modes.
Fixes: bfc923b642 ("dpll: zl3073x: implement frequency monitoring")
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260526074525.1451008-4-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The change_work was introduced to send device change notifications
from DPLL device callbacks without deadlocking on dpll_lock, since
the callbacks are already invoked under that lock. Now that
__dpll_device_change_ntf() is exported for callers that already
hold dpll_lock, use it directly and remove the change_work
infrastructure entirely.
This eliminates a race condition where change_work could be
re-scheduled after cancel_work_sync() during device teardown,
potentially causing the handler to dereference a freed or NULL
dpll_dev pointer.
Fixes: 9363b48376 ("dpll: zl3073x: Allow to configure phase offset averaging factor")
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260526074525.1451008-3-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Export __dpll_device_change_ntf() so that drivers can send device
change notifications from within device callbacks, which are already
called under dpll_lock. Using dpll_device_change_ntf() in that
context would deadlock.
Add lockdep_assert_held() to catch misuse without the lock held.
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Link: https://patch.msgid.link/20260526074525.1451008-2-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Chuck Lever says:
====================
net/handshake: anchor request lifetime to a pinned file reference
handshake_nl_accept_doit() has accumulated four follow-on fixes
since 3b3009ea8a ("net/handshake: Create a NETLINK service for
handling handshake requests"): 7ea9c1ec66, 7798b59409,
fe67b063f6, and dabac51b81. Each was a local refcount or
NULL-check correction; none moved where the file reference is
owned, and the same code keeps producing the same class of bug.
Reworking the ownership is what breaks the pattern.
For the duration of a request, sock->file has no single owner.
Submit publishes the request without taking a file reference;
accept_doit acquires one inside the handler, after the request
has already left the pending list. The consumer can drop its
own reference at any time, including the moment between
handshake_req_next() popping the request and accept_doit
reaching get_file(). The submit-side sock_hold() pins only
struct sock; struct socket and sock->file remain under the
consumer's control via the file descriptor.
This series places the file reference under unambiguous
ownership. handshake_req_submit() pins it on the request and
completion or cancel drops it (patches 4-5); the submit-side
sock_hold() then becomes redundant, and dropping it also closes
a publish-before-pin race the late sock_hold itself opened
(patch 6). The handshake_complete() API and its consumers move
to a uniform negative-errno sign convention (patch 3), with the
matching sign correction in nvme-tcp (patch 2). Patch 1
hardens hn_lock for BH context, the netns-exit drain fix
builds on the new file-pin infrastructure (patch 8), and new
KUnit file-count assertions verify the refcount contract
(patch 7).
Three things in this restructuring want a careful look. In
handshake_complete(), the fput() of the request's file
reference has to come after hp_done() -- fput() can transitively
run handshake_sk_destruct() and free the request, so the patch
stashes hr_file in a local first. handshake_sk_destruct()
itself is kept on purpose: it owns rhashtable removal and
kfree, and remains the backstop if a consumer path bypasses
handshake_complete() entirely. Third, handshake_req_next() now
returns its request with an extra get_file() held under
hn_lock; accept_doit must consume that reference (FD_PREPARE on
success, explicit fput on the fdf.err path), and any future
caller has to honor the same contract.
v2: https://patch.msgid.link/20260521-handshake-file-pin-v2-0-b9dadc472840@oracle.com
v1: https://patch.msgid.link/20260518-handshake-file-pin-v1-0-4bbcb7e62fda@oracle.com
====================
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-0-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The arguments to list_splice_init() in handshake_net_exit() are
reversed. The call moves the local empty "requests" list onto
hn->hn_requests, leaving the local list empty, so the subsequent
drain loop runs zero iterations. Pending handshake requests that
had not yet been accepted are not torn down when the net namespace
is destroyed; each one keeps a reference on a socket file and on
the handshake_req allocation.
Pass the source and destination in the documented order
(list_splice_init(list, head) moves list onto head) so the pending
list is transferred to the local scratch list and drained through
handshake_complete().
Fixing the splice direction exposes a list-corruption race. After
the splice each req->hr_list still has non-empty link pointers,
threading the stack-local scratch list rather than hn_requests.
A concurrent handshake_req_cancel() -- for example, from sunrpc's
TLS timeout on a kernel socket whose netns reference was not
taken -- finds the request through the rhashtable, calls
remove_pending(), and sees !list_empty(&req->hr_list).
__remove_pending_locked() then list_del_init()s an entry off the
scratch list while the drain iterates, corrupting it. The same
call arriving after the drain loop has run list_del() on an
entry hits LIST_POISON instead.
Have remove_pending() check HANDSHAKE_F_NET_DRAINING under
hn_lock and report not-found when drain is in progress. The
drain has already taken ownership; handshake_complete()'s existing
test_and_set on HANDSHAKE_F_REQ_COMPLETED still arbitrates
between drain and cancel for who calls the consumer's hp_done. Use
list_del_init() rather than list_del() in the drain so req->hr_list
does not carry LIST_POISON after drain releases the entry.
The DRAINING guard in remove_pending() makes cancel return false,
but cancel still falls through to test_and_set_bit on
HANDSHAKE_F_REQ_COMPLETED and drops the request's hr_file reference.
Without another pin, if that is the last reference, sk_destruct frees
the request while it is still linked on the drain loop's local list.
Pin each request's hr_file under hn_lock before releasing the list,
and drop that drain pin after the loop finishes with the request.
Fixes: 3b3009ea8a ("net/handshake: Create a NETLINK service for handling handshake requests")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-8-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The new file-reference contract on struct handshake_req is silently
breakable: a missing get_file() at submit or a missing fput() on an
error path leaves the file leaked but does not crash the test, so
the existing absence-of-crash checks pass either way.
Snapshot file_count(filp) before each handshake_req_submit() in
the submit-success, EAGAIN, EBUSY, and cancel tests, and assert
the expected balance after submit and again after cancel. The
already-completed cancel test also asserts the post-complete
balance, which pins down that handshake_complete() drops the
reference and that the subsequent cancel does not double-fput.
The destroy test gets the same treatment before __fput_sync(),
which double-checks that cancel's fput() ran and the only
remaining reference is the one sock_alloc_file() established.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-7-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
handshake_req_submit() publishes the request via
handshake_req_hash_add() and __add_pending_locked(), drops
hn_lock, and calls handshake_genl_notify() (which can sleep)
before taking sock_hold() on req->hr_sk. A fast tlshd ACCEPT
followed by DONE can drive handshake_complete()'s sock_put()
into the window between the spin_unlock and the late
sock_hold(); on a system where the consumer's fd held the
only sk reference, the late sock_hold() then operates on an
sk whose refcount has reached zero.
The preceding two patches install an explicit file reference
on struct handshake_req. That file pins sock->file, which
pins the embedded struct socket, which defers inet_release()'s
sock_put(). As long as hr_file is held, sk cannot reach refcount
zero from the consumer side, and the submit-side sock_hold()
with its matching sock_put() calls in handshake_complete() and
handshake_req_cancel() is now redundant.
Drop all three. The file reference already keeps each request's
socket alive, and the lifetime story is contained in a single
get_file()/fput() pair.
Fixes: 3b3009ea8a ("net/handshake: Create a NETLINK service for handling handshake requests")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-6-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
handshake_req_next() removes the request from the per-net
pending list and drops hn_lock before handshake_nl_accept_doit()
reads req->hr_sk->sk_socket and dereferences sock->file (once in
FD_PREPARE() and again in get_file()). In that window a
consumer running tls_handshake_cancel() followed by sockfd_put()
(svc_sock_free) or __fput_sync() (xs_reset_transport) releases
sock->file. sock_release() then runs sock_orphan(), zeroing
sk_socket, and frees the struct socket. The accept-side code
either reads NULL through sk_socket or chases freed memory.
The submit-side sock_hold() does not prevent this. sk_refcnt
protects struct sock, but struct socket and sock->file are
independently refcounted via the file descriptor the consumer
owns. Pinning sk leaves sock and sock->file unprotected.
Retarget the accept-side dereferences at req->hr_file, which was
pinned at submit time, instead of req->hr_sk->sk_socket->file.
Pinning on its own is not sufficient: a consumer that cancels
between handshake_req_next() returning and accept_doit reaching
FD_PREPARE() takes the !remove_pending() branch in
handshake_req_cancel() and drops hr_file before the accept side
takes its own reference. Hand off an additional file reference
inside handshake_req_next(), under hn_lock, so the accept side
operates on a reference that no concurrent handshake_req_cancel()
can revoke. FD_PREPARE() consumes that handed-off reference,
either by transferring it to the new fd in fd_publish() or by
dropping it in the cleanup destructor on error; the explicit
get_file() that previously balanced FD_PREPARE() is therefore
redundant and goes away.
Update handshake_req_cancel_test2 and _test3 to simulate the
FD_PREPARE() consumption with an fput() so the kunit file-count
assertions stay balanced.
Reported-by: Chris Mason <clm@meta.com>
Fixes: 3b3009ea8a ("net/handshake: Create a NETLINK service for handling handshake requests")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-5-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
handshake_nl_accept_doit() needs the file pointer backing
req->hr_sk->sk_socket to survive the window between
handshake_req_next() and the subsequent FD_PREPARE() and get_file().
The submit-side sock_hold() does not provide that. sk_refcnt keeps
struct sock alive, but struct socket is owned by sock->file: when
the consumer fputs the last file reference, sock_release() tears
the socket down regardless of any sock_hold.
Add an hr_file pointer to struct handshake_req and acquire an
explicit reference on sock->file during handshake_req_submit().
handshake_complete() and handshake_req_cancel() release the
reference on the completion-bit-winning path.
The submit error path must also release the file reference, but
after rhashtable insertion a concurrent handshake_req_cancel() can
discover the request and race the error path. Gate the error-path
cleanup -- sk_destruct restoration, fput, and request destruction
-- with test_and_set_bit(HANDSHAKE_F_REQ_COMPLETED), the same
serialization handshake_complete() and handshake_req_cancel()
already use. When cancel has already claimed ownership, the submit
error path returns without touching the request; socket teardown
handles final destruction.
The accept-side dereferences are not yet retargeted; that change
comes in the next patch.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-4-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
handshake_complete() declares status as unsigned int and
tls_handshake_done() negates that value (-status) before handing
it to the TLS consumer. Consumers match on negative errno
constants -- xs_tls_handshake_done() has
switch (status) {
case 0:
case -EACCES:
case -ETIMEDOUT:
lower_transport->xprt_err = status;
break;
default:
lower_transport->xprt_err = -EACCES;
}
so the API as designed expects callers to pass positive errno
values that the tlshd shim then negates.
Three internal callers in handshake_nl_accept_doit(), the
net-exit drain, and a kunit test follow kernel convention and
pass negative errnos -- -EIO, -ETIMEDOUT, -ETIMEDOUT. The
implicit conversion to unsigned int turns -ETIMEDOUT into
0xFFFFFF92; the subsequent -status in tls_handshake_done()
wraps back to 110, the consumer's switch falls through, and
the xprt reports -EACCES on what should be -ETIMEDOUT or -EIO.
Fix the API rather than the call sites. The natural kernel
convention is negative errno in, negative errno out. Change
handshake_complete() and hp_done to take int status, drop the
negation in tls_handshake_done(), and negate once in
handshake_nl_done_doit() where status arrives from the wire
as an unsigned netlink attribute. The three internal callers
were already correct under that convention and need no change.
At the same wire boundary, declare MAX_ERRNO as the netlink
policy upper bound for HANDSHAKE_A_DONE_STATUS. Attribute
validation rejects out-of-range values before
handshake_nl_done_doit() runs, and negating a bounded u32 there
stays within int range -- closing the UBSAN-visible signed-
integer overflow that an unconstrained u32 would invoke.
Fixes: 3b3009ea8a ("net/handshake: Create a NETLINK service for handling handshake requests")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-3-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
nvme_tcp_tls_done() assigns queue->tls_err in three branches. The
ENOKEY lookup failure and the EOPNOTSUPP initializer both store
negative errnos. The third branch, reached when the handshake
layer reports a non-zero status, stores -status.
The handshake layer delivers status to the consumer callback as a
negative errno; the other in-tree consumers --
xs_tls_handshake_done() and the nvmet target callback -- treat
their status argument that way. The extra negation in
nvme_tcp_tls_done() flips the sign, leaving tls_err as a positive
value (for instance, +EIO), which nvme_tcp_start_tls() then
returns to its caller.
Drop the extra negation so queue->tls_err uniformly carries a
negative errno on failure.
Fixes: be8e82caa6 ("nvme-tcp: enable TLS handshake upcall")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Reviewed-by: Alistair Francis <alistair.francis@wdc.com>
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-2-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
nvmet_tcp_state_change(), a socket callback that runs in BH context,
can reach handshake_req_cancel() via nvmet_tcp_schedule_release_queue()
and tls_handshake_cancel(). handshake_req_cancel() acquires
hn->hn_lock with plain spin_lock(). If a process-context thread on
the same CPU holds hn->hn_lock when a softirq invokes the cancel path,
the lock attempt deadlocks. This is the only caller that invokes
tls_handshake_cancel() from BH context; every other consumer calls it
from process context.
Deferring the cancel to process context in the NVMe target is not
straightforward: nvmet_tcp_schedule_release_queue() must call
tls_handshake_cancel() atomically with its state transition to
DISCONNECTING. If the cancel were deferred, the handshake completion
callback could fire in the window before the cancel runs, observe the
unexpected state, and return without dropping its kref on the queue.
Reworking that interlock is considerably more invasive than hardening
the handshake lock. Convert all hn->hn_lock acquisitions from
spin_lock/spin_unlock to spin_lock_bh/spin_unlock_bh so the lock is
never taken with softirqs enabled.
Fixes: 675b453e02 ("nvmet-tcp: enable TLS handshake upcall")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Hannes Reinecke <hare@kernel.org>
Link: https://patch.msgid.link/20260525-handshake-file-pin-v3-1-66c616906ead@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
pskb_carve_inside_header() and pskb_carve_inside_nonlinear() both copy
the old skb_shared_info header into a new buffer via memcpy(), which
includes the destructor_arg pointer (uarg) for MSG_ZEROCOPY skbs.
Neither function calls net_zcopy_get() for the new shinfo, creating an
unaccounted holder: every skb_shared_info with destructor_arg set will
call skb_zcopy_clear() once when freed, but the corresponding
net_zcopy_get() was never called for the new copy. Repeated calls
drive uarg->refcnt to zero prematurely, freeing ubuf_info_msgzc while
TX skbs still hold live destructor_arg pointers.
KASAN reports use-after-free on a freed ubuf_info_msgzc:
BUG: KASAN: slab-use-after-free in skb_release_data+0x77b/0x810
Read of size 8 at addr ffff88801574d3e8 by task poc/220
Call Trace:
skb_release_data+0x77b/0x810
kfree_skb_list_reason+0x13e/0x610
skb_release_data+0x4cd/0x810
sk_skb_reason_drop+0xf3/0x340
skb_queue_purge_reason+0x282/0x440
rds_tcp_inc_free+0x1e/0x30
rds_recvmsg+0x354/0x1780
__sys_recvmsg+0xdf/0x180
Allocated by task 219:
msg_zerocopy_realloc+0x157/0x7b0
tcp_sendmsg_locked+0x2892/0x3ba0
Freed by task 219:
ip_recv_error+0x74a/0xb10
tcp_recvmsg+0x475/0x530
The skb consuming the late access still referenced the same uarg via
shinfo->destructor_arg copied by pskb_carve_inside_nonlinear() without
a refcount bump. This has been verified to be reliably exploitable: a
working proof-of-concept achieves full root privilege escalation from
an unprivileged local user on a default kernel configuration.
The fix follows the pattern of pskb_expand_head() which has the same
memcpy/cloned structure. For pskb_carve_inside_header(), net_zcopy_get()
is placed after skb_orphan_frags() succeeds, so the orphan error path
needs no cleanup. For pskb_carve_inside_nonlinear(), net_zcopy_get() is
placed after all failure points and just before skb_release_data(), so
no error path needs cleanup at all -- matching pskb_expand_head() more
closely and avoiding the need for a balancing net_zcopy_put().
Fixes: 6fa01ccd88 ("skbuff: Add pskb_extract() helper function")
Cc: stable@vger.kernel.org
Assisted-by: Claude:claude-sonnet-4-6
Signed-off-by: Minh Nguyen <minhnguyen.080505@gmail.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260526041240.329462-1-minhnguyen.080505@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Jijie Shao says:
====================
hibmcge: fix RX packet corruption issue
This series fixes an RX packet corruption issue observed when SMMU is
disabled on the hibmcge driver. The fixes include disabling PCI Relaxed
Ordering and correcting the order of DMA barrier operations in the RX
data sync path.
====================
Link: https://patch.msgid.link/20260525144525.94884-1-shaojijie@huawei.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The dma_rmb() barrier was placed before dma_sync_single_for_cpu(), which
is incorrect. DMA sync must complete first to make the buffer accessible
to the CPU, then the rmb barrier ensures subsequent descriptor reads
observe the latest data written by the hardware.
Reorder the operations so dma_sync_single_for_cpu() is called before
dma_rmb() to guarantee the driver reads consistent data from the DMA
buffer.
Fixes: f72e255940 ("net: hibmcge: Implement rx_poll function to receive packets")
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Link: https://patch.msgid.link/20260525144525.94884-3-shaojijie@huawei.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When SMMU is disabled, the hibmcge driver may receive corrupted packets.
The hardware writes packet data and descriptors to the same page, but
with Relaxed Ordering enabled, PCI write transactions may not be
strictly ordered. This can cause the driver to observe a valid
descriptor before the corresponding packet data is fully written.
Fix this by clearing PCI_EXP_DEVCTL_RELAX_EN in the PCI bridge control
register to ensure strict write ordering between packet data and
descriptors.
Fixes: f72e255940 ("net: hibmcge: Implement rx_poll function to receive packets")
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Link: https://patch.msgid.link/20260525144525.94884-2-shaojijie@huawei.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Jamal Hadi Salim says:
====================
net/sched: Fix packet loops in mirred and netem
This patchset adds a 2-bit per-skb tc_depth counter that travels with
the packet. The existing per-CPU mirred nest tracking loses state
when a packet is deferred through the backlog or moves between CPUs
via XPS/RPS. A per-skb field covers both cases.
Patch 1 adds the tc_depth field in a padding hole in sk_buff.
Patches 2-3 revert the check_netem_in_tree() fix and its tests,
which broke legitimate multi-netem configurations.
Patch 4 uses tc_depth to stop netem duplicate recursion.
Patch 5 uses tc_depth to catch mirred ingress redirect loops.
Patch 6 fixes the infinite loop in the mirred egress blockcast case.
Patch 7 fixes drop stats in early return error scenarios in tcf_mirred_act
for redirect (caught by Sashiko [1]).
Patches 8-9 add mirred and netem test cases.
[1] https://sashiko.dev/#/patchset/20260413082027.2244884-1-hxzene%40gmail.com
====================
Link: https://patch.msgid.link/20260525122556.973584-1-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
tcf_mirred_act() checks sched_mirred_nest against MIRRED_NEST_LIMIT (4)
to prevent deep recursion. However, when the action uses blockcast
(tcfm_blockid != 0), the function returns at the tcf_blockcast() call
BEFORE reaching the counter increment. As a result, the recursion
counter never advances and the limit check is entirely bypassed.
When two devices share a TC egress block with a mirred blockcast rule,
a packet egressing on device A is mirrored to device B via blockcast;
device B's egress TC re-enters tcf_mirred_act() via blockcast and
mirrors back to A, creating an unbounded recursion loop:
tcf_mirred_act -> tcf_blockcast -> tcf_mirred_to_dev -> dev_queue_xmit
-> sch_handle_egress -> tcf_classify -> tcf_mirred_act -> (repeat)
This recursion continues until the kernel stack overflows.
The bug is reachable from an unprivileged user via
unshare(CLONE_NEWUSER | CLONE_NEWNET): user namespaces grant
CAP_NET_ADMIN in the new network namespace, which is sufficient to
create dummy devices, attach clsact qdiscs with shared blocks, and
install mirred blockcast filters.
BUG: TASK stack guard page was hit at ffffc90000b7fff8
Oops: stack guard page: 0000 [#1] SMP KASAN NOPTI
CPU: 2 UID: 1000 PID: 169 Comm: poc Not tainted 7.0.0-rc7-next-20260410
RIP: 0010:xas_find+0x17/0x480
Call Trace:
xa_find+0x17b/0x1d0
tcf_mirred_act+0x640/0x1060
tcf_action_exec+0x400/0x530
basic_classify+0x128/0x1d0
tcf_classify+0xd83/0x1150
tc_run+0x328/0x620
__dev_queue_xmit+0x797/0x3100
tcf_mirred_to_dev+0x7b1/0xf70
tcf_mirred_act+0x68a/0x1060
[repeating ~30+ times until stack overflow]
Kernel panic - not syncing: Fatal exception in interrupt
Fix this by incrementing sched_mirred_nest before calling
tcf_blockcast() and decrementing it on return, mirroring the
non-blockcast path. This ensures subsequent recursive entries see the
updated counter and are correctly limited by MIRRED_NEST_LIMIT.
Fixes: fe946a751d ("net/sched: act_mirred: add loop detection")
Signed-off-by: Kito Xu (veritas501) <hxzene@gmail.com>
Link: https://patch.msgid.link/20260525122556.973584-7-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When mirred redirects to ingress (from either ingress or egress) the loop
state from sched_mirred_dev array dev is lost because of 1) the packet
deferral into the backlog and 2) the fact the sched_mirred_dev array is
cleared. In such cases, if there was a loop we won't discover it.
Here's a simple test to reproduce:
ip a add dev port0 10.10.10.11/24
tc qdisc add dev port0 clsact
tc filter add dev port0 egress protocol ip \
prio 10 matchall action mirred ingress redirect dev port1
tc qdisc add dev port1 clsact
tc filter add dev port1 ingress protocol ip \
prio 10 matchall action mirred egress redirect dev port0
ping -c 1 -W0.01 10.10.10.10
Fixes: fe946a751d ("net/sched: act_mirred: add loop detection")
Tested-by: Victor Nogueira <victor@mojatatu.com>
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260525122556.973584-6-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add a 2-bit per-skb tc depth field to track packet loops across the stack.
The previous per-CPU loop counters like MIRRED_NEST_LIMIT
assume a single call stack and lose state in two cases:
1) When a packet is queued and reprocessed later (e.g., egress->ingress
via backlog), the per-cpu state is gone by the time it is dequeued.
2) With XPS/RPS a packet may arrive on one CPU and be processed on
another.
A per-skb field solves both by travelling with the packet itself.
The field fits in existing padding, using 2 bits that were previously a
hole:
pahole before(-) and after (+) diff looks like:
__u8 slow_gro:1; /* 132: 3 1 */
__u8 csum_not_inet:1; /* 132: 4 1 */
__u8 unreadable:1; /* 132: 5 1 */
+ __u8 tc_depth:2; /* 132: 6 1 */
- /* XXX 2 bits hole, try to pack */
/* XXX 1 byte hole, try to pack */
__u16 tc_index; /* 134 2 */
There used to be a ttl field which was removed as part of tc_verd in commit
aec745e2c5 ("net-tc: remove unused tc_verd fields"). It was already
unused by that time, due to remove earlier in commit c19ae86a51 ("tc: remove
unused redirect ttl").
The first user of this field is netem, which increments tc_depth on
duplicated packets before re-enqueueing them at the root qdisc. On
re-entry, netem skips duplication for any skb with tc_depth already set,
bounding recursion to a single level regardless of tree topology.
The other user is mirred which increments it on each pass
and limits to depth to MIRRED_DEFER_LIMIT (3).
The new field was called ttl in earlier versions of this patch
but renamed to tc_depth to avoid confusion with IP ttl.
Note (looking at you Sashiko! Dont ignore me and continue bringing this up):
1. Since both mirred and netem utilize the same 2-bit tc_depth field it is
possible when netem and mirred are used together that netem qdisc to skip
the duplication step. This is a known trade-off, as a 2-bit field cannot
independently track both features' recursion depths and it is not considered
sane to have a setup that addresses both features on at the same time.
2. skb_scrub_packet does not clear tc_depth. This means a packet's loop history
is preserved even across namespaces. While this might be restrictive for
some topologies, it is also design intent to provide robustness against loops
across namespaces.
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260525122556.973584-2-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
ipv6_rpl_srh_decompress() computes:
outhdr->hdrlen = (((n + 1) * sizeof(struct in6_addr)) >> 3);
hdrlen is __u8. For n >= 127 the result exceeds 255 and silently
truncates. With n=127 (cmpri=15, cmpre=15, pad=0, hdrlen=16):
(128 * 16) >> 3 = 256, truncated to 0 as __u8
The caller in ipv6_rpl_srh_rcv() then places the compressed header
at buf + ((ohdr->hdrlen + 1) << 3). With hdrlen=0 this is buf + 8,
but the decompressed region occupies buf[0..2055] (8-byte header
plus 128 full addresses). The compressed header overlaps the
decompressed data, and ipv6_rpl_srh_compress() writes into this
overlap, corrupting the routing header of the forwarded packet.
The existing guard at exthdrs.c:546 checks (n + 1) > 255, which
prevents n+1 from overflowing unsigned char (the segments_left
field), but does not prevent the computed hdrlen from overflowing
__u8. n=127 passes because 128 <= 255, yet hdrlen=256 does not
fit.
Tighten the bound to (n + 1) > 127. This caps n at 126, giving
hdrlen = (127 * 16) >> 3 = 254, which fits in __u8. The compressed
header then lands at buf + ((254 + 1) << 3) = buf + 2040, exactly
past the decompressed region (buf[0..2039]). No overlap. 127
segments is well beyond any realistic RPL deployment.
Fixes: 8610c7c6e3 ("net: ipv6: add support for rpl sr exthdr")
Signed-off-by: Rahul Chandelkar <rc@rexion.ai>
Link: https://patch.msgid.link/20260525154031.2290876-1-rc@rexion.ai
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jakub Kicinski says:
====================
ethtool: more bug fixes
Last week I sent two patch sets - one fixing bugs in RSS handling,
and one fixing CMIS / module handling. This set contains the remaining
fixes. There's a concentration of fixes around PHY and timestamp config
handling but not enough to break those out as separate sets.
====================
Link: https://patch.msgid.link/20260526153533.2779187-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The Netlink fallback path for reading module EEPROM
(fallback_set_params()) validates that offset < eeprom_len,
but does not check that offset + length stays within eeprom_len.
The ioctl equivalent (ethtool_get_any_eeprom() in ioctl.c) has
always enforced both bounds:
if (eeprom.offset + eeprom.len > total_len)
return -EINVAL;
This could lead to surprises in both drivers and device FW.
Add the missing offset + length validation to fallback_set_params(),
mirroring the ioctl.
Similarly - ethtool core in general, and ethtool_get_any_eeprom()
in particular tries to zero-init all buffers passed to the drivers
to avoid any extra work of zeroing things out. eeprom_fallback()
uses a plain kmalloc(), change it to zalloc.
Fixes: 96d971e307 ("ethtool: Add fallback to get_module_eeprom from netlink command")
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260526153533.2779187-11-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
All ethtool driver op calls should be sandwiched between
ethnl_ops_begin() / ethnl_ops_complete(). In Netlink eeprom code,
if the paged access failed we fall back to old API, but we
first call _complete() and the fallback never does its own
ethnl_ops_begin(). Move the fallback into the _begin() / _complete()
section.
Fixes: 96d971e307 ("ethtool: Add fallback to get_module_eeprom from netlink command")
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260526153533.2779187-10-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
strset_prepare_data() passes ETHTOOL_A_HEADER_FLAGS (3) as the header
attribute to ethnl_req_get_phydev(). This is incorrect, in the main
attr space 3 is ETHTOOL_A_STRSET_COUNTS_ONLY, not the request
header attr. The correct constant is ETHTOOL_A_STRSET_HEADER (1).
ethnl_req_get_phydev() only uses this value for the extack,
so this is not a "functionally visible"(?) bug.
Fixes: e96c93aa4b ("net: ethtool: strset: Allow querying phy stats by index")
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260526153533.2779187-9-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tsinfo_prepare_data() has two code paths: a "by-PHC" path for
user-specified hardware timestamping providers, and the old path.
Commit 89e281ebff ("ethtool: init tsinfo stats if requested") added
ethtool_stats_init() to mark stat slots as ETHTOOL_STAT_NOT_SET before
the driver callback populates them, but placed the call inside the
old-path block.
When commit b9e3f7dc9e ("net: ethtool: tsinfo: Enhance tsinfo to
support several hwtstamp by net topology") added the by-PHC early
return, it landed above the stats initialization. On that path
the stats array retains the zero-fill from ethnl_init_reply_data()'s
zalloc. This leads to the reply including a stats nest with four
zero-valued attributes that should have been absent.
Reject GET requests for stats with HWTSTAMP_PROVIDER or dump.
Fixes: b9e3f7dc9e ("net: ethtool: tsinfo: Enhance tsinfo to support several hwtstamp by net topology")
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260526153533.2779187-7-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
pse_prepare_data() is missing ethnl_ops_complete() if
ethnl_req_get_phydev() returned an error. Move getting
phydev up so that we don't have to worry about this
(similar order to linkstate_prepare_data()).
Note that phydev may still be NULL (this is checked in
pse_get_pse_attributes()), the goal isn't really to avoid
the _begin() / _complete() calls, only to simplify the error
handling.
While at it propagate the original error. Why this code
overrides the error with -ENODEV but !phydev generates
-EOPNOTSUPP is unclear to me...
Fixes: 31748765be ("net: ethtool: pse-pd: Target the command to the requested PHY")
Reviewed-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
Link: https://patch.msgid.link/20260526153533.2779187-5-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
