Cross-merge networking fixes after downstream PR (net-7.1-rc8).
Conflicts:
drivers/net/ethernet/wangxun/txgbe/txgbe_aml.c
f67aead16e ("net: txgbe: rework service event handling")
57d39faed4 ("net: txgbe: improve functions of AML 40G devices")
net/rds/info.c
512db8267b ("rds: mark snapshot pages dirty in rds_info_getsockopt()")
6e94eeb2a2 ("rds: convert to getsockopt_iter")
Adjacent changes:
include/net/sock.h
1ee90b77b7 ("net: guard timestamp cmsgs to real error queue skbs")
f0de88303d ("net: make is_skb_wmem() available to modules")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Change the return type of ndo_set_rx_mode_async from void to int to
allow drivers to report failures back to the core stack. This is a
prerequisite for adding retry logic in the core when drivers fail to
program RX filters (e.g. bnxt VF when PF is unavailable).
All existing implementations return 0 for now, maintaining current
behavior.
Signed-off-by: Stanislav Fomichev <sdf@fomichev.me>
Link: https://patch.msgid.link/20260608154014.227538-2-sdf@fomichev.me
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use the GIC functions to allocate interrupt contexts for RDMA EQs. These
interrupt contexts may be shared with Ethernet EQs when MSI-X vectors
are limited.
The driver now supports allocating dedicated MSI-X for each EQ. Indicate
this capability through driver capability bits. The RDMA EQs pass
use_msi_bitmap=false to share MSI-X vectors with Ethernet, while the
capability flag advertises that the driver supports per-vPort EQ
separation when hardware has sufficient vectors.
Populate eq.irq on all RDMA EQs for consistency with the Ethernet path.
Also relocate the GDMA_DRV_CAP_FLAG_1_HW_VPORT_LINK_AWARE define to its
numeric BIT(6) position among the other capability flags.
Signed-off-by: Long Li <longli@microsoft.com>
Acked-by: Leon Romanovsky <leon@kernel.org>
Link: https://patch.msgid.link/20260605005717.2059954-7-longli@microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
To prepare for assigning vPorts to dedicated MSI-X vectors, remove EQ
sharing among the vPorts and create dedicated EQs for each vPort.
Move the EQ definition from struct mana_context to struct mana_port_context
and update related support functions. Export mana_create_eq() and
mana_destroy_eq() for use by the MANA RDMA driver.
RSS QPs now take a vport reference via pd->vport_use_count to ensure
EQs outlive all QP consumers. The vport must already be configured by
a raw QP before an RSS QP can be created. EQs are only destroyed when
the last QP (raw or RSS) on the PD releases its reference.
Restrict each vport to a single RSS QP. The hardware only supports one
steering configuration (indirection table / hash key) per vport, and
mana_disable_vport_rx() on QP destroy disables RX globally for the
vport. Previously, creating a second RSS QP would silently overwrite
the first QP's steering config and destroy would blackhole all traffic.
This is now explicitly rejected with -EBUSY. Existing applications
(DPDK being the primary RDMA consumer) always create one RSS QP per
vport, so no real-world flows are affected.
Reject cross-port PD sharing for both raw and RSS QPs. Since EQs and
vport configuration are per-port, a PD is bound to the port used by
its first raw QP. Subsequent QPs on the same PD must use the same
port or the creation fails with -EINVAL. Previously this was silently
broken: with shared EQs it appeared to work, but with per-vPort EQs
a cross-port PD would cause wrong-port EQ teardown and corruption.
DPDK creates one PD per port so no existing flows are affected.
Serialize mana_set_channels() and the async per-port queue reset
handler against RDMA vport configuration to prevent RDMA from claiming
the vport during the detach/attach window. A channel_changing flag is
set under apc->vport_mutex before detach and checked by
mana_cfg_vport() when called from the RDMA path, blocking RDMA from
grabbing the vport during the entire window. When the port is down
and RDMA already holds the vport, the channel change is rejected with
-EBUSY.
Signed-off-by: Long Li <longli@microsoft.com>
Link: https://patch.msgid.link/20260605005717.2059954-2-longli@microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
srp_process_rsp() copies sense data from rsp->data + resp_data_len,
where resp_data_len is the full 32-bit value supplied by the SRP target
and is never checked against the number of bytes actually received
(wc->byte_len). The copy length is bounded to SCSI_SENSE_BUFFERSIZE, so
at most 96 bytes are copied, but the source offset is not bounded.
A malicious or compromised SRP target on the InfiniBand/RoCE fabric that
the initiator has logged into can return an SRP_RSP with
SRP_RSP_FLAG_SNSVALID set and a large resp_data_len. The receive buffer
is allocated at the target-chosen max_ti_iu_len, so the source of the
sense copy lands past the bytes actually received; with resp_data_len
near 0xFFFFFFFF it is gigabytes past the buffer and the read faults.
Copy the sense data only if it has not been truncated, that is, only if
the response header, the response data, and the sense region fit within
the bytes actually received; otherwise drop the sense and log. The
in-tree iSER and NVMe-RDMA receive paths already bound their parse by
wc->byte_len; this brings ib_srp into line with them.
Fixes: aef9ec39c4 ("IB: Add SCSI RDMA Protocol (SRP) initiator")
Link: https://patch.msgid.link/r/20260602220457.2542840-1-michael.bommarito@gmail.com
Cc: stable@vger.kernel.org
Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In drivers/infiniband/ulp/isert/ib_isert.c, isert_login_recv_done()
computes the login request payload length as wc->byte_len minus
ISER_HEADERS_LEN with no lower bound, and login_req_len is a signed int.
A remote iSER initiator can post a login Send work request carrying
fewer than ISER_HEADERS_LEN (76) bytes, so the subtraction underflows
and login_req_len becomes negative.
isert_rx_login_req() then reads that negative length back into a signed
int, takes size = min(rx_buflen, MAX_KEY_VALUE_PAIRS), and because the
min() is signed it keeps the negative value; the value is then passed as
the memcpy() length and sign-extended to a multi-gigabyte size_t. The
copy into the 8192-byte login->req_buf runs far out of bounds and
faults, crashing the target node. The login phase precedes iSCSI
authentication, so no credentials are required to reach this path.
Reject any login PDU shorter than ISER_HEADERS_LEN before the
subtraction, mirroring the existing early return on a failed work
completion, so login_req_len can never go negative. The upper bound was
already safe: a posted login buffer cannot deliver more than
ISER_RX_PAYLOAD_SIZE, so the difference stays at or below
MAX_KEY_VALUE_PAIRS and the existing min() clamps it; only the missing
lower bound needs to be added.
Fixes: b8d26b3be8 ("iser-target: Add iSCSI Extensions for RDMA (iSER) target driver")
Link: https://patch.msgid.link/r/20260602194642.2273217-1-michael.bommarito@gmail.com
Cc: stable@vger.kernel.org
Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
If IB_MR_REREG_ACCESS changes from RO to RW then the umem has to be
re-evaluated to ensure it is properly pinned as RW. Since the umem is
hidden inside each driver's mr struct add a ib_umem_check_rereg() function
that each driver has to call before processing IB_MR_REREG_ACCESS.
mlx4 has to retain its duplicate ib_access_writable check because it
implements IB_MR_REREG_ACCESS | IB_MR_REREG_TRANS by changing both items
in place sequentially while the MR is live, so it will continue to not
support this combination.
Cc: stable@vger.kernel.org
Fixes: b40656aa7d ("RDMA/umem: remove FOLL_FORCE usage")
Link: https://patch.msgid.link/r/0-v1-06fb1a2d6cf5+107-rereg_access_jgg@nvidia.com
Reported-by: Philip Tsukerman <philiptsukerman@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The cpu_id attribute supplied by user space through
UVERBS_ATTR_ALLOC_DMAH_CPU_ID is passed directly to cpumask_test_cpu()
without first verifying that the value is within the valid CPU range.
Passing such untrusted data to cpumask_test_cpu() may lead to an
out-of-bounds read of the underlying cpumask bitmap: the helper expands
to a test_bit() that indexes the bitmap by cpu_id / BITS_PER_LONG with
no bound check.
In addition, on kernels built with CONFIG_DEBUG_PER_CPU_MAPS it trips
the WARN_ON_ONCE() in cpumask_check(); combined with panic_on_warn this
turns a bad user input into a machine reboot.
Reject any cpu_id that is not smaller than nr_cpu_ids with -EINVAL
before it is used.
Reported by Smatch.
Fixes: d83edab562 ("RDMA/core: Introduce a DMAH object and its alloc/free APIs")
Link: https://patch.msgid.link/r/20260525142136.28165-1-yishaih@nvidia.com
Cc: stable@vger.kernel.org
Reported-by: Dan Carpenter <error27@gmail.com>
Closes: https://lore.kernel.org/r/ag68qoAW3P04J7pT@stanley.mountain/
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When the iommu is used the linearization of the mapping can give a single
block that is very large split across multiple SG entries.
When __rdma_block_iter_next() reassembles the split SG entries it is
overflowing the 32 bit stack values and computed the wrong DMA addresses
for blocks after the truncation.
Use the right types to hold DMA addresses.
Link: https://patch.msgid.link/r/1-v1-88303e9e509f+f7-ib_umem_types_jgg@nvidia.com
Cc: stable@vger.kernel.org
Fixes: a808273a49 ("RDMA/verbs: Add a DMA iterator to return aligned contiguous memory blocks")
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Validate the SQ ring size against the device's max LLQ size. This
ensures that when using 128-byte WQEs, userspace cannot exceed the queue
limits.
On create QP, userspace provides the SQ ring size (depth x WQE size)
which is validated against the max LLQ size.
Fixes: 40909f664d ("RDMA/efa: Add EFA verbs implementation")
Link: https://patch.msgid.link/r/20260526081536.1203553-1-ynachum@amazon.com
Reviewed-by: Michael Margolin <mrgolin@amazon.com>
Signed-off-by: Yonatan Nachum <ynachum@amazon.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Sashiko pointed out it is not safe to rely only on the devt because
char/block alias so if the user finds a block device with the same dev_t
it can masquerade as a ucap cdev fd.
Test the f_ops to only accept authentic cdevs.
Link: https://patch.msgid.link/r/0-v1-fd9482545e37+1e25-ib_ucaps_fd_ops_jgg@nvidia.com
Cc: stable@vger.kernel.org
Fixes: 61e5168281 ("RDMA/uverbs: Introduce UCAP (User CAPabilities) API")
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
- syzbot triggred crash in rxe due to concurrent plug/unplug
- Possible non-zero'd memory exposed to userspace in bnxt_re
- Malicous 'magic packet' with SIW causes a buffer overflow
- Tighten the new uAPI validation code to not crash in debugging prints
and have the right module dependencies in drivers
- mana was missing the max_msg_sz report to userspace
- UAF in rtrs on an error path
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQRRRCHOFoQz/8F5bUaFwuHvBreFYQUCahDmpgAKCRCFwuHvBreF
YRRTAP0YXV95Y5gcli55IhetKjJUzQbaREz2NueqIpf1IorMbAD+Lns4DgZCU0KW
bC81x7cGHBSyCju9zogIdBFJhsbxeQ4=
=MxJz
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
Pull rdma fixes from Jason Gunthorpe:
- syzbot triggred crash in rxe due to concurrent plug/unplug
- Possible non-zero'd memory exposed to userspace in bnxt_re
- Malicous 'magic packet' with SIW causes a buffer overflow
- Tighten the new uAPI validation code to not crash in debugging prints
and have the right module dependencies in drivers
- mana was missing the max_msg_sz report to userspace
- UAF in rtrs on an error path
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
RDMA/rtrs: Fix use-after-free in path file creation cleanup
RDMA/mana_ib: Report max_msg_sz in mana_ib_query_port
RDMA/core: Do not read wild stack memory in uverbs_get_handler_fn()
RDMA/core: Move the _ib_copy_validate_udata* functions to ib_core_uverbs
RDMA/siw: Reject MPA FPDU length underflow before signed receive math
RDMA/bnxt_re: zero shared page before exposing to userspace
selftests/rdma: explicitly skip tests when required modules are missing
RDMA/nldev: Add mutual exclusion in nldev_dellink()
In the error path of rtrs_srv_create_path_files(), the sysfs root folders
may already have been created and srv_path->kobj may already have been
initialized. If a later step fails, the cleanup currently calls
kobject_put(&srv_path->kobj) before
rtrs_srv_destroy_once_sysfs_root_folders(srv_path).
kobject_put() may drop the last reference to srv_path->kobj and invoke the
release callback, rtrs_srv_release(), which frees srv_path. The following
call to rtrs_srv_destroy_once_sysfs_root_folders(srv_path) then
dereferences srv_path internally to access srv_path->srv, resulting in a
use-after-free.
This failure path is reached before rtrs_srv_create_path_files() returns
success, so the successful-path lifetime handling is not involved.
Fix this by destroying the sysfs root folders before calling
kobject_put(&srv_path->kobj), so srv_path is still valid while the helper
accesses it.
This issue was found by a static analysis tool I am developing.
Fixes: ae4c81644e ("RDMA/rtrs-srv: Rename rtrs_srv_sess to rtrs_srv_path")
Signed-off-by: Guangshuo Li <lgs201920130244@gmail.com>
Link: https://patch.msgid.link/20260514113834.865530-1-lgs201920130244@gmail.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Report max_msg_sz for mana_ib, which is 16MB.
Fixes: 4bda1d5332 ("RDMA/mana_ib: Implement port parameters")
Signed-off-by: Shiraz Saleem <shirazsaleem@microsoft.com>
Signed-off-by: Konstantin Taranov <kotaranov@microsoft.com>
Link: https://patch.msgid.link/20260512094209.264955-1-kotaranov@linux.microsoft.com
Reviewed-by: Long Li <longli@microsoft.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Sashiko points out the legacy write path in ib_uverbs_write() does
allocate a struct uverbs_attr_bundle, but it doesn't wrap it in a
bundle_priv so downcasting here isn't safe.
Instead lift the method_elm out of the bundle_priv and use it for the
debug function. The legacy write path will leave it set as NULL since the
write method_elm uses a different type.
Cc: stable@vger.kernel.org
Fixes: 1de9287ece ("RDMA: Add ib_copy_validate_udata_in()")
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
It was incorrect to place them in uverbs_ioctl because that makes every
driver depends on ib_uverbs.ko, which is undesired. ib_core_uverbs.c is
for functions used by alot of drivers that are linked into ib_core
instead.
Fixes: 1de9287ece ("RDMA: Add ib_copy_validate_udata_in()")
Link: https://patch.msgid.link/r/1-v1-045258567bd6+9fe-ib_uverbs_support_ko_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The commit in the fixes tag added a warning for devices
that are netdev ops locked that they should be converted
to .ndo_set_rx_mode_async. IPoIB for mlx5 is such a
driver which was missed during the conversion because the
flow is more complex:
- mlx5 part of IPoIB device was converted to ops-lock in commit [1].
- ipoib_intf_init() then overrides netdev_ops with
ipoib_netdev_ops_{pf,vf}, which still wired ndo_set_rx_mode to the
legacy sync path -- tripping the new warning on every probe.
So now we have the following splat:
netdevice: ib0 (uninitialized): ops-locked drivers should use ndo_set_rx_mode_async
WARNING: net/core/dev.c:11366 at register_netdevice+0x83c/0x21d0
...
register_netdev+0x1f/0x40
ipoib_add_one+0x35c/0x880 [ib_ipoib]
This patch implements .ndo_set_rx_mode_async but it simply schedules the
multicast restart task like before. This is done to maintain the
assumption that this task and others [2] must run on the same order
workqueue to avoid racing with themselves. The race between
ipoib_mcast_join_task() and ipoib_mcast_restart_task() would be the most
obvious example.
[1] 8f7b00307b, "net/mlx5e: Convert mlx5 netdevs to instance locking")
[2] ipoib_mcast_join_task, ipoib_mcast_restart_task,
ipoib_mcast_carrier_on_task, ipoib_reap_ah, ipoib_reap_neigh
Fixes: 3cbd229388 ("net: warn ops-locked drivers still using ndo_set_rx_mode")
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
Acked-by: Jason Gunthorpe <jgg@nvidia.com>
Link: https://patch.msgid.link/20260513124519.3357165-1-dtatulea@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
A malicious connected siw peer can send an iWARP FPDU whose MPA length
field (c_hdr->mpa_len, 16 bit big-endian, peer-controlled) is smaller
than the fixed DDP/RDMAP header for the announced opcode. Soft-iWARP
parses the full header in siw_get_hdr() based on iwarp_pktinfo[opcode]
.hdr_len, but never compares mpa_len against that header length.
siw_tcp_rx_data() then derives
srx->fpdu_part_rem = be16_to_cpu(mpa_len) - fpdu_part_rcvd
+ MPA_HDR_SIZE;
where fpdu_part_rcvd equals iwarp_pktinfo[opcode].hdr_len at this
point. For a tagged WRITE (hdr_len 16, MPA_HDR_SIZE 2) the smallest
on-wire mpa_len of 0 yields fpdu_part_rem = -14, and any mpa_len below
hdr_len - MPA_HDR_SIZE underflows to a negative int.
The signed value then flows into siw_proc_write()/siw_proc_rresp() as
bytes = min(srx->fpdu_part_rem, srx->skb_new);
is handed to siw_check_mem() as an int len (whose interval check
addr + len > mem->va + mem->len is satisfied for a valid base when
len is negative), and reaches siw_rx_data() -> siw_rx_kva() /
siw_rx_umem() -> skb_copy_bits() as a signed copy length. The header
copy branch in skb_copy_bits() promotes that to size_t, producing a
multi-gigabyte read.
KASAN under a KUnit harness that drives the real kernel TCP receive
path -- a loopback AF_INET socketpair, the malformed FPDU written via
kernel_sendmsg, sk_data_ready firing in softirq, tcp_read_sock
dispatching to siw_tcp_rx_data -- reports:
BUG: KASAN: use-after-free in skb_copy_bits+0x284/0x480
Read of size 4294967295 at addr ffff888...
Call Trace:
skb_copy_bits
siw_rx_kva
siw_rx_data
siw_check_mem
siw_proc_write
siw_tcp_rx_data
__tcp_read_sock
siw_qp_llp_data_ready
tcp_data_ready
tcp_data_queue
Add the missing invariant at the earliest point where the peer header
is fully assembled. iwarp_pktinfo[*].hdr_len - MPA_HDR_SIZE is exactly
the value the siw transmitter uses as the minimum mpa_len for each
opcode (drivers/infiniband/sw/siw/siw_qp.c:33), so this matches the
protocol contract. Out-of-range FPDUs terminate the connection with
TERM_ERROR_LAYER_LLP / LLP_ETYPE_MPA / LLP_ECODE_FPDU_START -- which
is RFC 5044 Section 8 error code 3 ("Marker and ULPDU Length fields
do not agree on the start of an FPDU"), the correct framing-error
class for this inconsistency.
Fixes: 8b6a361b8c ("rdma/siw: receive path")
Link: https://patch.msgid.link/r/20260513175325.2042630-2-michael.bommarito@gmail.com
Cc: stable@vger.kernel.org
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
Assisted-by: Claude:claude-opus-4-7
Acked-by: Bernard Metzler <bernard.metzler@linux.dev>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
bnxt_re_alloc_ucontext() allocates uctx->shpg via
__get_free_page(GFP_KERNEL). The buddy allocator does not zero pages
without __GFP_ZERO, so the page contains stale kernel data from
whatever object most recently freed it.
The page is then mapped into userspace via vm_insert_page() under
BNXT_RE_MMAP_SH_PAGE in bnxt_re_mmap(). The driver only ever writes
4 bytes (a u32 AVID) at offset BNXT_RE_AVID_OFFT (0x10) inside
bnxt_re_create_ah(); the remaining 4092 bytes of the page are exposed
to userspace unsanitised, leaking kernel memory contents.
Any user with access to /dev/infiniband/uverbsX on a host with a
bnxt_re device (typically rdma group membership) can read this data
via a single mmap() at pgoff 0 after IB_USER_VERBS_CMD_GET_CONTEXT.
Other shared pages in the same file already use get_zeroed_page()
correctly:
drivers/infiniband/hw/bnxt_re/ib_verbs.c
srq->uctx_srq_page = (void *)get_zeroed_page(GFP_KERNEL);
cq->uctx_cq_page = (void *)get_zeroed_page(GFP_KERNEL);
uctx->shpg is the only outlier. Bring it in line with the existing
convention by switching to get_zeroed_page().
Fixes: 1ac5a40479 ("RDMA/bnxt_re: Add bnxt_re RoCE driver")
Signed-off-by: Lord Ulf Henrik Holmberg <henrik.holmberg@defensify.se>
Link: https://patch.msgid.link/20260509084011.11971-1-pomzm67@gmail.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Representor callbacks can be registered and unregistered while the
E-Switch is already in switchdev mode, and the same E-Switch may also be
reconfigured by devlink, VF changes and SF changes. Serialize these paths
with the per-E-Switch representor mutex instead of relying on ad-hoc bit
state and wait queues.
Take the representor lock around the mode transition, VF/SF representor
changes and representor ops registration. Keep mode_lock and the
representor lock unnested by using the operation flag while the mode lock
is dropped. During mode changes, drop the representor lock around the
auxiliary bus rescan because driver bind/unbind may register or unregister
representor ops.
Split representor ops registration into locked public wrappers and blocked
internal helpers, clear the ops pointer on unregister, and add nested
wrappers for the shared-FDB master IB path that registers peer
representor ops while another E-Switch representor lock is already held.
On unregister, always call __unload_reps_all_vport() before marking reps
unregistered and clearing rep_ops. The per-representor state check makes
this a no-op for types that were not loaded, so unregister no longer has
to infer load state from esw->mode.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260503202726.266415-6-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
None of the udp_tunnel users need struct socket in their
fast paths; it is only used for tunnel setup / teardown.
Even setup_udp_tunnel_sock() does not need struct socket.
Let's change setup_udp_tunnel_sock() to take struct sock
instead of struct socket.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260502031401.3557229-3-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
None of the udp_tunnel users need struct socket in their
fast paths; it is only used for tunnel setup / teardown.
While the UDP tunnel interface accepts struct socket, this
encourages users to store the pointer unnecessarily. This
leads to extra dereferences when accessing struct sock fields
(e.g., sk->sk_user_data instead of sock->sk->sk_user_data).
Furthermore, these dereferences necessitate synchronize_rcu()
in udp_tunnel_sock_release() to protect the fast paths from
sock_orphan() setting sk->sk_socket to NULL.
This overhead can be avoided if users store the struct sock
pointer directly in their private structures.
As a prep, let's change udp_tunnel_sock_release() to take
struct sock instead of struct socket.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260502031401.3557229-2-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Similar to the SRQ case the hr_qp is stored in the xarray before it is
fully initialized. Unlike the SRQ case the error unwinds do not wait for
the completion so keep the refcount 0 until the function succeeds.
Fixes: 9a4435375c ("IB/hns: Add driver files for hns RoCE driver")
Link: https://patch.msgid.link/r/14-v1-41f3135e5565+9d2-rdma_ai_fixes1_jgg@nvidia.com
Suggested-by: Junxian Huang <huangjunxian6@hisilicon.com>
Reviewed-by: Junxian Huang <huangjunxian6@hisilicon.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Sashiko points out that once the srq memory is stored into the xarray by
alloc_srqc() it can immediately be looked up by:
xa_lock(&srq_table->xa);
srq = xa_load(&srq_table->xa, srqn & (hr_dev->caps.num_srqs - 1));
if (srq)
refcount_inc(&srq->refcount);
xa_unlock(&srq_table->xa);
Which will fail refcount debug because the refcount is 0 and then crash:
srq->event(srq, event_type);
Because event is NULL.
Use refcount_inc_not_zero() instead to ensure a partially prepared srq is
never retrieved from the event handler and fix the ordering of the
initialization so refcount becomes 1 only after it is fully ready.
All the initialization must be done before calling free_srqc() since it
depends on the completion and refcount.
Fixes: 9a4435375c ("IB/hns: Add driver files for hns RoCE driver")
Link: https://sashiko.dev/#/patchset/0-v1-e911b76a94d1%2B65d95-rdma_udata_rep_jgg%40nvidia.com?part=3
Link: https://patch.msgid.link/r/13-v1-41f3135e5565+9d2-rdma_ai_fixes1_jgg@nvidia.com
Reviewed-by: Junxian Huang <huangjunxian6@hisilicon.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
mlx5 has a common pattern implementing a device-global singleton resource
where it checks the resource pointer for !NULL and then skips obtaining
the lock.
This is not ordered properly as observing !NULL doesn't mean that all the
data under that pointer is also visible on this CPU when the lock is not
taken.
Use a release/acquire pairing to explicitly manage this.
Pointed out by sashiko, Codex found more cases.
Fixes: 5895e70f2e ("IB/mlx5: Allocate resources just before first QP/SRQ is created")
Fixes: 638420115c ("IB/mlx5: Create UMR QP just before first reg_mr occurs")
Link: https://sashiko.dev/#/patchset/SYBPR01MB7881E1E0970268BD69C0BA75AF2B2%40SYBPR01MB7881.ausprd01.prod.outlook.com
Link: https://patch.msgid.link/r/3-v1-41f3135e5565+9d2-rdma_ai_fixes1_jgg@nvidia.com
Assisted-by: Codex:GPT-5.5
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Rename the vport number enums MLX5_VPORT_PF to MLX5_VPORT_HOST_PF and
MLX5_VPORT_FIRST_VF to MLX5_VPORT_FIRST_HOST_VF to indicate that these
vport indices represent the host PF and its VFs. This prepares the code
for upcoming support of an additional PF type.
Signed-off-by: Moshe Shemesh <moshe@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260428053851.220089-2-tariqt@nvidia.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
When a driver creates a new MR during rereg_user_mr, a race window
exists between rdma_alloc_commit_uobject() for the new MR and the point
where the code reads that MR to populate the response keys.
A concurrent rereg_mr or destroy_mr could destroy the MR in this window
and cause UAF in the first thread.
Racing flow between two rereg_mr calls:
CPU0 CPU1
---- ----
rereg_user_mr(mr_handle)
uobj_get_write(mr_handle) -> mr0
mr1 = driver→rereg()
rdma_alloc_commit_uobject(mr1)
// mr1 replaced mr0 and is unlocked
uobj_put_destroy(mr0)
rereg_user_mr(mr_handle)
uobj_get_write(mr_handle) -> mr1
mr2 = driver→rereg()
rdma_alloc_commit_uobject(mr2)
// mr2 replaced mr1 and is unlocked
uobj_put_destroy(mr1)
// Destroys mr1!
resp.lkey = mr1->lkey; // UAF - mr1 was freed!
resp.rkey = mr1->rkey; // UAF - mr1 was freed!
Fix by storing lkey/rkey in local variables before the new MR is
unlocked and using the local variables to set the user response.
Fixes: 6e0954b11c ("RDMA/uverbs: Allow drivers to create a new HW object during rereg_mr")
Link: https://patch.msgid.link/r/20260427-security-bug-fixes-v3-4-4621fa52de0e@nvidia.com
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When resolving an RDMA-CM IPv6 address, ib_nl_ip_send_msg() sends a
netlink request to the userspace daemon to perform IP-to-GID
resolution in certain cases. The function allocates the netlink message
buffer using nla_total_size(sizeof(size)), which passes 8 bytes (the
size of size_t) instead of 16 bytes (the size of an IPv6 address).
This results in an 8-byte under-allocation.
This is currently masked by nlmsg_new() over-allocation of the skb
in its internal logic. However, the code remains incorrect.
Fix the issue by supplying the proper IPv6 address length to
nla_total_size().
Fixes: ae43f82867 ("IB/core: Add IP to GID netlink offload")
Link: https://patch.msgid.link/r/20260427-security-bug-fixes-v3-3-4621fa52de0e@nvidia.com
Signed-off-by: Maher Sanalla <msanalla@nvidia.com>
Reviewed-by: Patrisious Haddad <phaddad@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
A potential race condition exists between mlx5_core_destroy_dct() and
mlx5_core_create_dct() that can lead to a use-after-free.
After _mlx5_core_destroy_dct() releases the DCT to firmware, the DCTN
can be immediately reallocated for a new DCT being created concurrently.
If the create path stores the new DCT in the xarray before the destroy path
erases it, the destroy will incorrectly delete the new DCT's entry.
Later accesses then hit freed memory.
Fix by replacing the unconditional xa_erase_irq() with xa_cmpxchg_irq()
that only erases the entry if it hasn't already been replaced (still
contains XA_ZERO_ENTRY), preserving any newly created DCT.
Fixes: afff248998 ("RDMA/mlx5: Handle DCT QP logic separately from low level QP interface")
Link: https://patch.msgid.link/r/20260427-security-bug-fixes-v3-2-4621fa52de0e@nvidia.com
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Reviewed-by: Michael Guralnik <michaelgur@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
mlx5_ib_dev_res_srq_init() allocates two SRQs, s0 and s1. When
ib_create_srq() fails for s1, the error branch destroys s0 but falls
through and unconditionally assigns the freed s0 and the ERR_PTR s1 to
devr->s0 and devr->s1.
This leads to several problems: the lock-free fast path checks
"if (devr->s1) return 0;" and treats the ERR_PTR as already initialised;
users in mlx5_ib_create_qp() dereference the freed SRQ or ERR_PTR via
to_msrq(devr->s0)->msrq.srqn; and mlx5_ib_dev_res_cleanup() dereferences
the ERR_PTR and double-frees s0 on teardown.
Fix by adding the same `goto unlock` in the s1 failure path.
Cc: stable@vger.kernel.org
Fixes: 5895e70f2e ("IB/mlx5: Allocate resources just before first QP/SRQ is created")
Link: https://patch.msgid.link/r/SYBPR01MB7881E1E0970268BD69C0BA75AF2B2@SYBPR01MB7881.ausprd01.prod.outlook.com
Reported-by: Yuhao Jiang <danisjiang@gmail.com>
Signed-off-by: Junrui Luo <moonafterrain@outlook.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
atomic_write_reply() at drivers/infiniband/sw/rxe/rxe_resp.c
unconditionally dereferences 8 bytes at payload_addr(pkt):
value = *(u64 *)payload_addr(pkt);
check_rkey() previously accepted an ATOMIC_WRITE request with pktlen ==
resid == 0 because the length validation only compared pktlen against
resid. A remote initiator that sets the RETH length to 0 therefore reaches
atomic_write_reply() with a zero-byte logical payload, and the responder
reads sizeof(u64) bytes from past the logical end of the packet into
skb->head tailroom, then writes those 8 bytes into the attacker's MR via
rxe_mr_do_atomic_write(). That is a remote disclosure of 4 bytes of kernel
tailroom per probe (the other 4 bytes are the packet's own trailing ICRC).
IBA oA19-28 defines ATOMIC_WRITE as exactly 8 bytes. Anything else is
protocol-invalid. Hoist a strict length check into check_rkey() so the
responder never reaches the unchecked dereference, and keep the existing
WRITE-family length logic for the normal RDMA WRITE path.
Reproduced on mainline with an unmodified rxe driver: a sustained
zero-length ATOMIC_WRITE probe repeatedly leaks adjacent skb head-buffer
bytes into the attacker's MR, including recognisable kernel strings and
partial kernel-direct-map pointer words. With this patch applied the
responder rejects the PDU and the MR stays all-zero.
Cc: stable@vger.kernel.org
Fixes: 034e285f8b ("RDMA/rxe: Make responder support atomic write on RC service")
Link: https://patch.msgid.link/r/20260418162141.3610201-1-michael.bommarito@gmail.com
Assisted-by: Claude:claude-opus-4-7
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Even after applying commit 7244491dab ("RDMA/rxe: Validate pad and ICRC
before payload_size() in rxe_rcv"), a single unauthenticated UDP packet
can still trigger panic. That patch handled payload_size() underflow only
for valid opcodes with short packets, not for packets carrying an unknown
opcode. The unknown-opcode OOB read described below predates that commit
and reaches back to the initial Soft RoCE driver.
The check added there reads
pkt->paylen < header_size(pkt) + bth_pad(pkt) + RXE_ICRC_SIZE
where header_size(pkt) expands to rxe_opcode[pkt->opcode].length. The
rxe_opcode[] array has 256 entries but is only populated for defined IB
opcodes; any other entry (for example opcode 0xff) is zero-initialized, so
length == 0 and the check degenerates to
pkt->paylen < 0 + bth_pad(pkt) + RXE_ICRC_SIZE
which does not constrain pkt->paylen enough. rxe_icrc_hdr() then computes
rxe_opcode[pkt->opcode].length - RXE_BTH_BYTES
which underflows when length == 0 and passes a huge value to rxe_crc32(),
causing an out-of-bounds read of the skb payload.
Reproduced on v7.0-rc7 with that fix applied, QEMU/KVM with
CONFIG_RDMA_RXE=y and CONFIG_KASAN=y, after
rdma link add rxe0 type rxe netdev eth0
A single 48-byte UDP packet to port 4791 with BTH opcode=0xff and
QPN=IB_MULTICAST_QPN triggers:
BUG: KASAN: slab-out-of-bounds in crc32_le+0x115/0x170
Read of size 1 at addr ...
The buggy address is located 0 bytes to the right of
allocated 704-byte region
Call Trace:
crc32_le+0x115/0x170
rxe_icrc_hdr.isra.0+0x226/0x300
rxe_icrc_check+0x13f/0x3a0
rxe_rcv+0x6e1/0x16e0
rxe_udp_encap_recv+0x20a/0x320
udp_queue_rcv_one_skb+0x7ed/0x12c0
Subsequent packets with the same shape fault on unmapped memory and panic
the kernel. The trigger requires only module load and "rdma link add"; no
QP, no connection, and no authentication.
Fix this by rejecting packets whose opcode has no rxe_opcode[] entry,
detected via the zero mask or zero length, before any length arithmetic
runs.
Cc: stable@vger.kernel.org
Fixes: 8700e3e7c4 ("Soft RoCE driver")
Link: https://patch.msgid.link/r/20260414111555.3386793-1-michael.bommarito@gmail.com
Assisted-by: Claude:claude-opus-4-6
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The current teardown logic for dd->pio_map and dd->sdma_map frees the
structures while they might still be accessed by RCU readers. Although the
pointer is nulled under a spinlock, the memory is reclaimed before waiting
for the grace period to end.
This patch fixes the sequence by:
1. Extracting the pointer under the lock.
2. Clearing the RCU-protected pointer.
3. Waiting for readers to finish with synchronize_rcu().
4. Finally freeing the memory.
Fixes: 7724105686 ("IB/hfi1: add driver files")
Link: https://patch.msgid.link/r/20260206050836.5890-1-lirongqing@baidu.com
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
