Connecting two bridges on the same system [1] can result in a lockdep
splat [2].
The report is a false positive. Multicast queries are built and
transmitted under the bridge multicast lock. When the outgoing port of
one bridge is configured on top of another bridge, the transmit path
re-enters bridge code and acquires the other bridge's multicast lock in
order to snoop the query. Both lock instances share a single lockdep
class, so lockdep flags the nested acquisition as an AA deadlock.
Giving each bridge its own lock class will not solve the problem: the
reverse topology would produce an ABBA splat with the same pair of
classes. It also consumes a lockdep key per bridge.
Instead, fix the problem by deferring the transmission of the queries to
a workqueue. Build the skb and update querier state under the lock as
before, then enqueue the skb on a per multicast context queue and
schedule the work.
Purge the queue when the multicast context is de-initialized. At this
stage the work cannot be requeued. There is no need to take a reference
on skb->dev since the work cannot outlive the bridge or the bridge port.
Use the high priority workqueue to reduce the delay between the enqueue
time and the transmission time. With default settings (i.e., querier
interval - 255 seconds, query interval - 125 seconds) the extra delay
should not be a problem.
Avoid the unlikely case of the queue growing endlessly by limiting it to
1,000 skbs. Use this number for the simple reason that this is the
default Tx queue length.
Use local_bh_{disable,enable}() to disable/enable softIRQs and migration
in order to avoid corrupting the multicast statistics (per-CPU
u64_stats).
[1]
ip link add name br1 up type bridge mcast_snooping 1 mcast_querier 1
ip link add name br0 up type bridge mcast_snooping 1 mcast_querier 1
ip link add link br0 name br0.10 up master br1 type vlan id 10
[2]
WARNING: possible recursive locking detected
7.0.0-virtme-gb50c64a58a90 #1 Not tainted
[...]
ip/339 is trying to acquire lock:
ffff888104f0b480 (&br->multicast_lock){+.-.}-{3:3}, at: br_ip6_multicast_query (net/bridge/br_multicast.c:3584)
but task is already holding lock:
ffff888104f03480 (&br->multicast_lock){+.-.}-{3:3}, at: br_multicast_port_query_expired (net/bridge/br_multicast.c:1904)
[...]
Call Trace:
[...]
br_ip6_multicast_query (net/bridge/br_multicast.c:3584)
br_multicast_ipv6_rcv (net/bridge/br_multicast.c:3988)
br_dev_xmit (net/bridge/br_device.c:98 (discriminator 1))
dev_hard_start_xmit (net/core/dev.c:3904)
__dev_queue_xmit (net/core/dev.c:4871)
vlan_dev_hard_start_xmit (net/8021q/vlan_dev.c:131 (discriminator 1))
dev_hard_start_xmit (net/core/dev.c:3904)
__dev_queue_xmit (net/core/dev.c:4871)
br_dev_queue_push_xmit (net/bridge/br_forward.c:60)
__br_multicast_send_query (net/bridge/br_multicast.c:1811 (discriminator 1))
br_multicast_send_query (net/bridge/br_multicast.c:1889)
br_multicast_port_query_expired (net/bridge/br_multicast.c:1914)
call_timer_fn (kernel/time/timer.c:1749)
[...]
Reported-by: syzbot+d7b7f1412c02134efa6d@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/netdev/000000000000c4c9d405f2643e01@google.com/
Reviewed-by: Petr Machata <petrm@nvidia.com>
Acked-by: Nikolay Aleksandrov <nikolay@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20260708122820.1298718-2-idosch@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Pass source address, DSCP and output interface from the tunnel
configuration to ip_route_output() in ipip_fill_forward_path(), aligning
the route lookup with the slow path in ipip_tunnel_xmit().
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/20260708-ipip-route-lookup-fill_forward_path-v1-1-b77df74822ed@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The eDMA/HDMA transfers the driver issues carry a requester function
number in their TLPs, but nothing ever programs it: eDMA v0 leaves the
FUNC_NUM field of the channel control word zero and HDMA leaves the
per-channel func_num register at its reset value, so every transfer is
attributed to function 0. That is invisible in single-function setups,
but once the DMA block serves a non-zero endpoint function, its
requests must carry that function's number for the host to attribute
and translate them correctly.
Record the function number in the chip data (PCI_FUNC() of the probing
device for dw-edma-pcie) and program it per channel.
Endpoint-local chip instances keep func_no at 0, so transfers issued by
the endpoint-side driver remain PF0-attributed. Delegated channels are
programmed by the host-side dw-edma-pcie instance when it takes over the
channel, using that instance's PCI_FUNC().
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-15-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
Allow PCI ID match data to pass dw_edma_chip flags into dw_edma_probe().
This keeps per-device policy in the match data instead of open-coding it
in probe().
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-14-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
Skip data block BAR mapping and debug output when a channel has no data
block size. This lets future providers describe channels that only need
descriptor memory exposed.
No functional change intended for existing Synopsys EDDA and
AMD (Xilinx) MDB/CPM6 devices. Their static channel descriptions still
provide data block sizes where data block windows are used. A zero-sized
data block now means "not present" for future metadata providers.
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-13-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
Add an optional physical address override to struct dw_edma_block and
use a helper to compute descriptor block addresses.
No functional change intended. Existing Synopsys EDDA and AMD (Xilinx)
MDB/CPM6 block descriptors leave the override unset, so the helper still
returns the same values as before.
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-12-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
Add a match-data flag for devices whose DMA register block starts at an
offset inside the mapped BAR. Existing Synopsys EDDA and AMD (Xilinx)
MDB/CPM6 matches keep using the BAR mapping base directly.
The offsets handled here come from static, trusted match data. Later
metadata-derived offsets are validated when the endpoint DMA metadata
path parses them at that trust boundary.
No functional change intended.
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-11-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
Move the platform ops pointer into match data. Existing EDDA/MDB/CPM6
matches keep using dw_edma_pcie_plat_ops.
No functional changes intended.
Suggested-by: Frank Li <Frank.Li@nxp.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-10-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
dw_edma_pcie_probe() now obtains DMA layout data through device-specific
capability callbacks, not only from PCIe Vendor-Specific Extended
Capabilities. Rename the local data copy from vsec_data to dma_data
before adding endpoint DMA BAR metadata discovery, which does not rely
on VSEC.
No functional change intended.
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-9-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
Move device-specific capability parsing behind per-device match data.
The existing probe path mixes two decisions: which static template a PCI
ID uses, and which device-specific capability parser adjusts that
template. Split those decisions so device-specific discovery can be
added through match data instead of adding more vendor checks to
dw_edma_pcie_probe().
No functional change is intended for the existing Synopsys EDDA and
AMD (Xilinx) MDB/CPM6 matches. They still copy the same static template
data and run the same capability parsing logic before BAR mapping. The
AMD (Xilinx) MDB/CPM6 entries also keep using endpoint memory physical
addresses for descriptor windows through a new match-data flag.
Suggested-by: Frank Li <Frank.Li@nxp.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-8-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
The dw-edma-pcie driver copies static template data into a mutable
dw_edma_pcie_data instance before applying capability-derived updates.
Keep the derived non-LL mode in that copy as well, instead of only
tracking it in a local variable in dw_edma_pcie_probe().
This prepares for keeping capability parsing behind match data without a
separate non-LL output parameter.
No functional change intended.
Suggested-by: Frank Li <Frank.Li@nxp.com>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-7-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
A DesignWare eDMA instance may represent only a subset of channels that
is also initialized by another OS instance, such as an endpoint-side OS.
Add a partial ownership flag for instances that must preserve
controller-wide state owned by that peer.
In partial ownership mode, dw-edma skips the initial core reset and uses
the limited quiesce path in probe() and remove() instead of the full
core-off path. The flag also makes the driver validate the ownership
granularity required by each register layout before registering
channels.
Probe fails if the limited quiesce cannot stop its resources; remove
reports the error after completing the remaining driver teardown.
Partial instances also skip interrupt-emulation doorbell allocation: the
emulated doorbell is a controller-level resource, and a partial owner
must not claim it on behalf of the whole block.
For EDMA_MF_EDMA_UNROLL and EDMA_MF_HDMA_COMPAT, the driver programs
per-direction registers, such as DMA_{WRITE,READ}_INT_MASK_OFF and
DMA_{WRITE,READ}_INT_CLEAR_OFF. These register layouts have at most
EDMA_MAX_{WR,RD}_CH channels per direction, so the capped hardware
channel count still represents the whole direction. A partial instance
can therefore expose write or read channels only if it owns every
channel in that direction; otherwise two OS instances could update the
same direction-wide registers without a shared locking protocol.
In contrast, HDMA native uses per-channel registers, so it can be owned
at channel granularity.
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-6-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
dw_edma_irq_request() passes struct dw_edma_irq to request_irq() before
dw_edma_channel_setup() fills the back pointer. A shared interrupt can
therefore enter the handler with dw_irq->dw still NULL, leading to a
NULL pointer dereference.
Set the back pointer before installing each handler.
Fixes: e63d79d1ff ("dmaengine: Add Synopsys eDMA IP core driver")
Cc: stable@vger.kernel.org
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-5-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
Add core operations that quiesce only the resources represented by a
dw-edma instance, separate from the existing full controller off path.
For unrolled eDMA, quiesce masks the direction, disables ENGINE_EN,
waits for it to read back clear, then clears pending interrupt status.
If ENGINE_EN does not clear, return the timeout to the caller.
HDMA compatibility mode does not implement ENGINE_EN, so quiesce clears
CHi_PWR_EN for each represented channel instead. Both register maps
share interrupt control per direction, so quiescing one channel
quiesces the whole direction. Callers must own that direction and stop
the peer from programming it first.
HDMA native has per-channel registers and can quiesce the represented
channel directly.
No caller is added yet, so this is a no-functional-change preparation
for delegated channel reclaim and partial-owned remove paths.
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-4-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
DesignWare eDMA can signal completion locally through edma_int[] and
remotely through IMWr/MSI. When channels are delegated to a remote
frontend, the local endpoint side and the remote host side must not both
service the same DONE/ABORT status.
Add channel interrupt routing state and initialize it from the controller
instance configuration. Update the eDMA and HDMA native paths so
linked-list interrupt generation, HDMA non-linked-list interrupt enables,
and DONE/ABORT masking follow the selected mode. For HDMA native
non-linked-list channels, keep the local stop/abort enables set so status
is latched. In remote mode, also enable remote signaling and mask the local
interrupt pins.
Keep the existing dw-edma-pcie host-side instances in remote interrupt
routing mode so their IMWr/MSI completion model remains unchanged after
local routing becomes the zero value.
Note:
- The routing mode describes where a channel should report completion.
It does not by itself say whether this dw-edma instance owns the
interrupt status. A local instance must ignore remote-only channels,
and a remote instance must ignore local-only channels, even if such
interrupts are unexpectedly delivered. Otherwise the non-owner side
could steal the interrupt from the owner by clearing shared DONE/ABORT
status.
Cc: Devendra K Verma <devendra.verma@amd.com>
Suggested-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-3-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
The HDMA linked-list and non-linked-list start paths both program the
stop/abort interrupt setup register using the same local/remote enable
policy. Only the interrupt-mask handling differs by transfer mode.
Factor the common setup into dw_hdma_v0_core_int_setup() before adding
per-channel interrupt routing support. No functional change intended.
Suggested-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Koichiro Den <den@valinux.co.jp>
Link: https://patch.msgid.link/20260721062815.4117887-2-den@valinux.co.jp
Signed-off-by: Vinod Koul <vkoul@kernel.org>
- Avoid redundant page-table allocations in ioremap pcache topup
- Apply SBI FWFT LOCK flag only on successful set
- Bound SBI PMU counter mask scan to BITS_PER_LONG
- Skip TLB flush when G-stage PTE becomes valid with Svvptc
- Zicbo[m|z|p] block sizes should be always present in ONE_REG
- Inject instruction access fault on unmapped guest fetch
- Serialize virtual interrupt pending state updates using raw spinlock
- Fix Spectre-v1 in vector register access via ONE_REG
-----BEGIN PGP SIGNATURE-----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=jcT0
-----END PGP SIGNATURE-----
Merge tag 'kvm-riscv-fixes-7.2-1' of https://github.com/kvm-riscv/linux into HEAD
KVM/riscv fixes for 7.2, take #1
- Avoid redundant page-table allocations in ioremap pcache topup
- Apply SBI FWFT LOCK flag only on successful set
- Bound SBI PMU counter mask scan to BITS_PER_LONG
- Skip TLB flush when G-stage PTE becomes valid with Svvptc
- Zicbo[m|z|p] block sizes should be always present in ONE_REG
- Inject instruction access fault on unmapped guest fetch
- Serialize virtual interrupt pending state updates using raw spinlock
- Fix Spectre-v1 in vector register access via ONE_REG
On Intel platforms with a VMX preemption timer and APICv, if a VMM
calls KVM_GET_LAPIC before KVM_GET_MSRS to save the vCPU state, it is
possible to lose a pending timer interrupt.
If the thread running these ioctls is migrated to another core after
calling KVM_GET_LAPIC but before KVM_GET_MSRS and the guest is using
their LAPIC timer in TSC-deadline mode, not only does the save LAPIC
state not carry the pending interrupt, the TSCDEADLINE MSR will be
zeroed.
After migration across CPUs, KVM_GET_MSRS calls vcpu_load, posting the
interrupt and clearing the MSR:
vcpu_load() ->
kvm_arch_vcpu_load() ->
kvm_lapic_restart_hv_timer() ->
start_hv_timer() ->
apic_timer_expired() ->
kvm_apic_inject_pending_timer_irqs()
. post interrupt into the LAPIC state
. clear IA32_TSCDEADLINE
The saved LAPIC state will be missing the pending interrupt and the saved
MSR will be zero. Oops.
Fix by only posting an interrupt when we're attempting to enter the guest
(vcpu->wants_to_run == true), not for vcpu_load from other paths.
Assisted-by: gemini:gemini-3.1-pro-preview
Debugged-by: David Matlack <dmatlack@google.com>
Debugged-by: Sean Christopherson <seanjc@google.com>
Debugged-by: Jim Mattson <jmattson@google.com>
Debugged-by: James Houghton <jthoughton@google.com>
Signed-off-by: Venkatesh Srinivas <venkateshs@chromium.org>
Message-ID: <20260715234234.15382-2-venkateshs@chromium.org>
Reviewed-by: James Houghton <jthoughton@google.com>
Reviewed-by: Chao Gao <chao.gao@intel.com>
Cc: stable@vger.kernel.org
Fixes: ae95f566b3 ("KVM: X86: TSCDEADLINE MSR emulation fastpath", 2020-05-15)
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The test asserted that the X86_FEATURE_SEV CPUID bit exactly matches
whether KVM offers KVM_X86_SEV_VM. That is not an invariant: when all
SEV ASIDs are assigned to SEV-SNP, KVM does not offer the SEV VM type
even though CPUID reports SEV, so the test aborts on an SNP-only host.
Derive SEV availability from KVM_CAP_VM_TYPES (as already done for SEV-ES
and SNP), assert only the one-way implication that a type offered by KVM
is also reported in CPUID, and TEST_REQUIRE() the SEV VM type so the test
skips cleanly when it is unavailable.
Reviewed-by: Tycho Andersen (AMD) <tycho@kernel.org>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Message-ID: <5d3c345113748f39b7982e365d241abaf3e11086.1784545391.git.dwmw@amazon.co.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
sev_smoke_test ran the plain SEV subtest unconditionally, gated only on
the X86_FEATURE_SEV CPUID bit, while gating SEV-ES and SNP on the
KVM_CAP_VM_TYPES bits. CPUID reporting SEV does not mean KVM offers the
SEV VM type: when all SEV ASIDs are assigned to SEV-SNP, KVM_X86_SEV_VM
is unavailable even though X86_FEATURE_SEV is set. On such a host the
test aborts in KVM_CREATE_VM instead of exercising the available modes.
Gate the SEV subtest on KVM_CAP_VM_TYPES like the others, so the test
runs the VM types the host actually offers.
Reviewed-by: Tycho Andersen (AMD) <tycho@kernel.org>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Message-ID: <2b5e7a83d277134294199a455469bb436196b902.1784545391.git.dwmw@amazon.co.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
mmu_destroy_caches() destroys pte_list_desc_cache and
mmu_page_header_cache, but leaves both pointers unchanged. The pointers
live in kvm.ko, and therefore survive when a vendor module is unloaded
while kvm.ko remains loaded.
If creation of pte_list_desc_cache fails during a subsequent vendor
module load, its assignment sets pte_list_desc_cache to NULL and the
error path calls mmu_destroy_caches(). mmu_page_header_cache still
points to the cache destroyed during the preceding vendor module
unload. Passing that stale pointer to kmem_cache_destroy() causes a
slab use-after-free.
Reproduce the issue on a v7.1.3 kernel with CONFIG_KASAN=y,
CONFIG_KASAN_GENERIC=y, CONFIG_KVM=m, and CONFIG_KVM_INTEL=m. A
one-shot test hook forces pte_list_desc_cache to NULL on the second
invocation of kvm_mmu_vendor_module_init():
1. Load kvm.ko and kvm-intel.ko, creating both caches.
2. Unload only kvm_intel, leaving kvm.ko loaded.
3. Reload kvm_intel and force initialization through the -ENOMEM path.
KASAN reports:
BUG: KASAN: slab-use-after-free in
kvm_mmu_vendor_module_init+0x5b/0x170 [kvm]
...
kmem_cache_destroy+0x21/0x1d0
kvm_mmu_vendor_module_init+0x5b/0x170 [kvm]
...
Allocated by task 16817:
__kmem_cache_create_args+0x12c/0x3b0
__kmem_cache_create.constprop.0+0xb6/0xf0 [kvm]
kvm_mmu_vendor_module_init+0x13b/0x170 [kvm]
...
Freed by task 16820:
kmem_cache_destroy+0x117/0x1d0
kvm_mmu_vendor_module_exit+0x21/0x30 [kvm]
Clear both pointers immediately after destroying their caches so that
the stored state reflects the caches' lifetime and repeated cleanup is
safe.
With the fix applied, the same injected vendor module reload fails with
-ENOMEM as expected and produces no KASAN report.
Fixes: cb498ea2ce ("KVM: Portability: Combine kvm_init and kvm_init_x86")
Cc: stable@vger.kernel.org
Signed-off-by: Phil Rosenthal <phil@phil.gs>
Message-ID: <20260718-kvm-mmu-cache-uaf-v3-1-e103b93c74e1@phil.gs>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
kvm_mmu_zap_oldest_mmu_pages() excludes a shadow page whose root_count
is non-zero from top-level reclaim, because such a page cannot be
freed. The path in mmu_page_zap_pte() that recursively zaps a parentless
nested TDP child has no such check. As a result, a shadow page can
be zapped even if the page itself can't be freed; as the comment in
kvm_mmu_zap_oldest_mmu_pages() notes, zapping it will just force vCPUs
to rebuild the page.
As in top-level reclaim, do not recursively prepare zapping of a
nested TDP child whose root_count is non-zero.
Fixes: 2de4085ccc ("KVM: x86/MMU: Recursively zap nested TDP SPs when zapping last/only parent")
Signed-off-by: Hyunwoo Kim <imv4bel@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Check for a "stale" page fault, i.e. for an invalid and/or obsolete root,
after making MMU pages available for the shadow MMU. If reclaiming shadow
pages zaps an in-use root, i.e. marks it invalid, then KVM will attempt to
map memory into an invalid root. On its own, populating an invalid root is
"fine", but because child shadow pages inherit their parent's role, any
children created during the map/fetch will be created as invalid pages,
thus violating KVM's invariant that invalid pages are never on the list of
active MMU pages.
Note, the underlying flaw has existed since KVM first started tracking
invalid roots in 2008 (commit 2e53d63acb, "KVM: MMU: ignore zapped root
pagetables"), but the true badness only came along in 2020 (Linux 5.9)
with the invariant that invalid shadow pages can't be on the list of
active pages.
Note #2, inheriting role.invalid when creating child shadow pages is also
far from ideal; that flaw will be addressed separately.
Reported-by: Hyunwoo Kim <imv4bel@gmail.com>
Fixes: f95eec9bed ("KVM: x86/mmu: Don't put invalid SPs back on the list of active pages")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
free_nested() frees the shadow VMCS while vmcs01 still points to it. But
because it is asynchronous with respect to loaded_vmcs_clear(), the vCPU
might migrate before the pointer is cleared and __loaded_vmcs_clear()
may then execute VMCLEAR.
The VMCS needs to stay attached until its explicit VMCLEAR completes, but
then it can be hidden and the page safely freed.
Fixes: 355f4fb140 ("kvm: nVMX: VMCLEAR an active shadow VMCS after last use")
Cc: stable@vger.kernel.org
Signed-off-by: Hyunwoo Kim <imv4bel@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
airoha_npu_send_msg() always maps the mailbox buffer with DMA_TO_DEVICE,
but some callers expect the NPU to write response data back into the
same buffer:
- airoha_npu_wlan_msg_get() (NPU_OP_GET): NPU writes response into
the buffer, then the caller reads it via memcpy()
- airoha_npu_ppe_stats_setup() (NPU_OP_SET): NPU writes back
npu_stats_addr field in the response
On non-cache-coherent architectures like EN7581 (Cortex-A53 without
hardware cache coherency for NPU DMA), DMA_TO_DEVICE unmap is a no-op
— it does not invalidate the CPU cache. If the NPU-written cache line
is still present in the CPU cache when the caller reads the buffer,
the CPU observes stale data instead of the NPU response.
This is a timing-sensitive bug: small mailbox buffers (~24 bytes)
typically fit in a single cache line and may survive in the cache
until the caller reads them, producing silent data corruption rather
than a crash. The bug is more likely to trigger when the caller reads
the response immediately after dma_unmap_single() without intervening
cache-evicting operations.
Fix by using DMA_BIDIRECTIONAL for both map and unmap, which ensures
dma_unmap_single() invalidates the CPU cache on non-coherent systems.
The mailbox buffers are small so there is no performance concern.
Fixes: c52918744e ("net: airoha: npu: Move memory allocation in airoha_npu_send_msg() caller")
Signed-off-by: Wayen Yan <win847@gmail.com>
Acked-by: Lorenzo Bianconi <lorenzo@kernel.org>
Link: https://patch.msgid.link/178351055214.98729.11403147818632027428@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
fsl_mc_get_endpoint() returns the MAC endpoint device with a reference
taken through device_find_child(). The Ethernet connect path stores that
device in mac->mc_dev and keeps it for the lifetime of the connected MAC
object.
However, the disconnect path only disconnects and closes the MAC before
freeing the dpaa2_mac object. It does not drop the endpoint device
reference stored in mac->mc_dev, so every successful connect leaks that
device reference when the MAC is later disconnected.
Drop the endpoint device reference after closing the MAC and before
freeing the dpaa2_mac object.
Fixes: 7194792308 ("dpaa2-eth: add MAC/PHY support through phylink")
Signed-off-by: Guangshuo Li <lgs201920130244@gmail.com>
Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Reviewed-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://patch.msgid.link/20260708111738.750391-1-lgs201920130244@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
In the nested state, the physical interrupt has already been
deactivated through the HW bit in the LR. The extra deactivation
would be harmless but can hit an errata case on AmpereOne, so
avoid it here.
On AmpereOne, deactivating a physical interrupt through
ICC_DIR_EL1 or ICC_EOIR1_EL1 (depending on EOImode) which is not
active, but is the highest priority pending interrupt causes the
cpu to lose the interrupt pending state and also prevents the
delivery of future interrupts.
Fixes: 6dd333c894 ("KVM: arm64: GICv3: nv: Plug L1 LR sync into deactivation primitive")
Signed-off-by: D Scott Phillips <scott@os.amperecomputing.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/linux-arm-kernel/20260710222128.416581-1-scott@os.amperecomputing.com/
Link: https://patch.msgid.link/20260714231158.496808-1-scott@os.amperecomputing.com
Signed-off-by: Marc Zyngier <maz@kernel.org>
airoha_ppe_deinit() replaces the NPU pointer with NULL via
rcu_replace_pointer() but does not wait for existing RCU readers
to exit before calling ppe_deinit() and airoha_npu_put(). This can
cause a use-after-free if a reader in an RCU read-side critical
section still holds a reference to the NPU when it is freed.
The init path (airoha_ppe_init) already calls synchronize_rcu()
after rcu_assign_pointer(), but the deinit path introduced in
commit 6abcf751bc ("net: airoha: Fix schedule while atomic in
airoha_ppe_deinit()") omitted the matching barrier when switching
from rcu_read_lock()/rcu_dereference() to rcu_replace_pointer().
Add synchronize_rcu() before ppe_deinit() to ensure all existing
RCU readers have completed before the NPU resources are released.
Fixes: 6abcf751bc ("net: airoha: Fix schedule while atomic in airoha_ppe_deinit()")
Signed-off-by: Wayen Yan <win847@gmail.com>
Acked-by: Lorenzo Bianconi <lorenzo@kernel.org>
Link: https://patch.msgid.link/178351022574.97989.6880403520276841703@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
fsl_mc_get_endpoint() returns the MAC endpoint device with a reference
taken through device_find_child(). The switch port connect path stores
that device in mac->mc_dev and keeps it for the lifetime of the connected
MAC object.
However, the disconnect path only closes the MAC and frees the dpaa2_mac
object. It does not drop the endpoint device reference stored in
mac->mc_dev, so every successful connect leaks that device reference when
the MAC is later disconnected.
Drop the endpoint device reference before freeing the dpaa2_mac object.
Fixes: 84cba72956 ("dpaa2-switch: integrate the MAC endpoint support")
Signed-off-by: Guangshuo Li <lgs201920130244@gmail.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260708111025.749311-1-lgs201920130244@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Stefano Garzarella says:
====================
vsock/virtio: collapse receive queue under memory pressure
This series contains a patch (the first one) that is part of work I'm
doing to improve the tracking of memory used by AF_VSOCK sockets.
The second patch is a test for our suite that highlights the issue.
Since Brien reported an issue with his environment (based on Linux 6.12.y)
related to the work I’m doing, I extracted this patch and tried to make it
as easy as possible to backport. Brien tested it by backporting it to
6.12.y, which now contains the backport of the 059b7dbd20
("vsock/virtio: fix potential unbounded skb queue").
This patch primarily fixes STREAM sockets, but also partially fixes
SEQPACKET (with the exception of EOMs, which are kept in separate skbs to
avoid overcomplicating the code).
The rest of the work, I feel, is more net-next material and still needs
some work to be completed.
v1: https://lore.kernel.org/netdev/20260626134823.206676-1-sgarzare@redhat.com/
====================
Link: https://patch.msgid.link/20260708102904.50732-1-sgarzare@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add a test that sends 2 MB of data using randomly sized small packets
(129-512 bytes) over a SOCK_STREAM connection. Packets above
GOOD_COPY_LEN (128) bypass the in-place coalescing in recv_enqueue(),
forcing each one into its own skb.
Without receive queue collapsing, the per-skb overhead eventually
exceeds buf_alloc and the connection is reset. The test verifies
that all data arrives and that content integrity is preserved.
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Bobby Eshleman <bobbyeshleman@meta.com>
Link: https://patch.msgid.link/20260708102904.50732-3-sgarzare@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When many small packets accumulate in the receive queue, the skb overhead
can exceed buf_alloc even while the payload is within bounds. This causes
virtio_transport_inc_rx_pkt() to reject packets, leading to connection
resets during large transfers under backpressure.
The issue was reported by Brien, who has a reproducer, but it is also
easily reproducible with iperf-vsock [1] using a small packet size:
iperf3 --vsock -c $CID -l 129
which fails immediately without this patch but with commit 059b7dbd20
("vsock/virtio: fix potential unbounded skb queue").
Inspired by TCP's tcp_collapse() which solves a similar problem, add
virtio_transport_collapse_rx_queue() that walks the receive queue and
re-copies data into compact linear skbs to reduce the overhead.
The collapse is triggered proactively from when the number of skb queued
is close to exceeding the overhead budget.
A pre-scan counts the eligible bytes to size each allocation precisely,
avoiding waste for isolated small packets. Partially consumed skbs are
kept as-is to preserve buf_used/fwd_cnt accounting, EOM-marked skbs to
maintain SEQPACKET message boundaries, and skbs already larger than the
collapse target because they already have a good data-to-overhead ratio.
Walking a large queue may take a significant amount of time and cache
misses, causing traffic burstiness. To limit this, the collapse stops
once enough room is freed for this packet and the next one, but may
opportunistically free more to fill each collapsed skb to capacity.
[1] https://github.com/stefano-garzarella/iperf-vsock
Fixes: 059b7dbd20 ("vsock/virtio: fix potential unbounded skb queue")
Cc: stable@vger.kernel.org
Reported-by: Brien Oberstein <brienpub@gmail.com>
Closes: https://lore.kernel.org/netdev/618701dd023e$063de350$12b9a9f0$@gmail.com/
Tested-by: Brien Oberstein <brienpub@gmail.com>
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Bobby Eshleman <bobbyeshleman@meta.com>
Link: https://patch.msgid.link/20260708102904.50732-2-sgarzare@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The dedicated "ProArt" key on the ASUS ProArt PX13 (HN7306) emits ASUS
WMI event code 0x8b. This code is absent from the asus-nb-wmi sparse
keymap, so pressing the key only produces a "Unknown key code 0x8b"
message and does nothing in userspace.
On this model the key is exposed solely through the WMI interface and has
no HID scancode, so it can only be handled inside asus-nb-wmi. Map 0x8b
to KEY_PROG3 so the key emits a bindable event, consistent with the
existing KEY_PROG3 mapping used for the ASUS programmable/Armoury key.
Signed-off-by: Nizar Al-Kabbani <nizar.m.kabbani@gmail.com>
Reviewed-by: Denis Benato <denis.benato@linux.dev>
Link: https://patch.msgid.link/20260719090843.66777-1-nizar.m.kabbani@gmail.com
Reviewed-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
rxrpc_wake_up_io_thread() checks local->io_thread before waking it, but
then reloads the pointer for wake_up_process().
local->io_thread is cleared with WRITE_ONCE() when the I/O thread exits, so
the second load can see NULL even if the first load did not.
Take a READ_ONCE() snapshot and use it for both the NULL check and the
wake_up_process() call, as rxrpc_encap_rcv() already does.
Fixes: 5800b1cf3f ("rxrpc: Allow CHALLENGEs to the passed to the app for a RESPONSE")
Signed-off-by: Xuanqiang Luo <luoxuanqiang@kylinos.cn>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260708093534.53486-1-xuanqiang.luo@linux.dev
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
ntfs_non_resident_attr_shrink() shrinks attribute sizes but fails to
trim the page cache. This leaves orphaned dirty folios beyond the new
end of the attribute, leading to writeback failures (-ENOENT), data
loss, and $EA chain corruption.
Fix this by truncating the page cache to the new size immediately after
updating the sizes, preventing writeback from flushing out-of-range folios.
Fixes: 495e90fa33 ("ntfs: update attrib operations")
Signed-off-by: Namjae Jeon <linkinjeon@kernel.org>
Add a shared helper to safely convert runlist element counts to byte sizes
using overflow checks, and use it in both ntfs_rl_realloc() and
ntfs_rl_realloc_nofail().
Fixes: 11ccc9107d ("ntfs: update runlist handling and cluster allocator")
Co-developed-by: Alper Mudar <kommandant_alper@proton.me>
Signed-off-by: Alper Mudar <kommandant_alper@proton.me>
Tested-by: Alper Mudar <kommandant_alper@proton.me>
Signed-off-by: Namjae Jeon <linkinjeon@kernel.org>
rt_flush_dev() can replace rt->dst.dev with blackhole_netdev while RCU
readers are running. ip_rt_send_redirect() and ip_rt_get_source() both
read rt->dst.dev more than once and use the results in one operation.
If rt->dst.dev changes between those reads, the operation can use values
from two devices. For example, ip_rt_send_redirect() can use in_dev from
the old device and the L3 master ifindex from blackhole_netdev.
Read rt->dst.dev once in these two functions and use the snapshot for the
later device accesses.
Signed-off-by: Xuanqiang Luo <luoxuanqiang@kylinos.cn>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20260708060537.17188-3-xuanqiang.luo@linux.dev
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
rt_flush_dev() replaces rt->dst.dev with blackhole_netdev on uncached
routes. The field is also exposed as dst.dev_rcu, and existing readers
use dst_dev_rcu().
Use rcu_assign_pointer() for the replacement, as dst_dev_put() already
does for the same field.
Signed-off-by: Xuanqiang Luo <luoxuanqiang@kylinos.cn>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20260708060537.17188-2-xuanqiang.luo@linux.dev
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Commit 98d0912e9f ("net: skbuff: fix missing zerocopy reference in
pskb_carve helpers") introduced two calls of net_zcopy_get(skb_zcopy(skb)).
In fact, skb_zcopy() has already been executed once before. When calling
net_zcopy_get(), skb_zcopy() always returns skb_uarg(skb), which results
in adding some unnecessary instructions in skb_zcopy. So, change these
two calls to directly use skb_uarg(skb) instead of skb_zcopy.
In addition, also use net_zcopy_get() instead of refcount_inc() in
pskb_expand_head() for code consistency.
No functional change intended.
Signed-off-by: Yun Lu <luyun@kylinos.cn>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260708055454.9167-1-luyun_611@163.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Sparse does not know __builtin_infer_alloc_token() and complains:
sparse: sparse: undefined identifier '__builtin_infer_alloc_token'
Fix it by using a dummy variant of __kmalloc_token() if __CHECKER__ is
defined.
Fixes: feb662d916 ("slab: support for compiler-assisted type-based slab cache partitioning")
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202607110912.nZTqfCrH-lkp@intel.com/
Signed-off-by: Marco Elver <elver@google.com>
Link: https://patch.msgid.link/20260721092005.1986693-1-elver@google.com
Acked-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
Improve PAT index selection logic in xe_migrate.c to avoid unnecessary
coherency overhead when host-side memory is uncached. Previously,
we defaulted to XE_CACHE_WB, which enforces 2-way coherency and may
trigger cacheline pulls from CPU even when host-side memory is never
dirty.
This change introduces xe_migrate_pat_index() to choose the appropriate
PAT index based on the actual TTM caching mode of the buffer object
being mapped. For iGPUs with WC host mappings, we now prefer
XE_CACHE_NONE to skip coherency snoops. For compressed PTEs on newer
platforms, we select XE_CACHE_NONE_COMPRESSION.
This avoids unnecessary cache traffic for uncached host mappings.
v6: (sashiko)
- Only apply the BO's host-side caching for system-memory PTEs.
v5: (Matt A)
- Simplify emit_pte() to derive caching from res->bo directly, removing
the separate bo parameter
- Leave changes in __xe_migrate_update_pgtables() and
build_pt_update_batch_sram()
- Fix comment about page-walker coherency in xe_migrate_pat_index()
v4:
- Keep xe_migrate_prepare_vm() on XE_CACHE_WB since page tables require
page-walker coherency.
- Pass BO into emit_pte() and select PAT attributes from the BO's TTM
caching mode.
Assisted-by: Github-Copilot:claude-opus-4.8
Signed-off-by: Sanjay Yadav <sanjay.kumar.yadav@intel.com>
Suggested-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patch.msgid.link/20260710083004.1546599-2-sanjay.kumar.yadav@intel.com
GTPv1-U packets may carry a chain of extension headers before the inner
IP packet. The receive path already parses and skips these extension
headers, but it currently reads the inner protocol before doing so.
As a result, the first extension header byte is interpreted as the inner
IP version. Packets with extension headers are then dropped before PDP
lookup.
Parse the extension header chain before calling gtp_inner_proto(), so the
inner protocol is read from the actual inner IP header.
Fixes: c75fc0b9e5 ("gtp: identify tunnel via GTP device + GTP version + TEID + family")
Signed-off-by: Zhixing Chen <running910@gmail.com>
Link: https://patch.msgid.link/20260708042244.120898-1-running910@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
rds_find_bound() looks up the destination socket using a global
rhashtable keyed solely on (addr, port, scope_id). Network namespaces
are not part of the key, so a sender in netns A can deliver an incoming
message (inc) to a socket that lives in a different netns B.
When this happens, inc->i_conn points to an rds_connection whose c_net
is netns A, but the receiving rs lives in netns B. Once the child
process that created netns A exits, cleanup_net() calls
rds_loop_exit_net() -> rds_loop_kill_conns() -> rds_conn_destroy(),
freeing that connection. If the survivor socket in netns B still holds
the inc, any subsequent dereference of inc->i_conn is a use-after-free.
There are two dangerous sites in rds_clear_recv_queue():
1. inc->i_conn->c_lcong (offset 88 of freed rds_connection, size 200)
read via rds_recv_rcvbuf_delta() -- confirmed by KASAN.
2. inc->i_conn->c_trans->inc_free(inc) (function pointer at offset 80)
called via rds_inc_put() when the inc refcount reaches zero -- same
race window, potential call-through-freed-object primitive.
The bug is reachable from unprivileged user namespaces
(CLONE_NEWUSER + CLONE_NEWNET), available since Linux 3.8.
Fix this by rejecting the delivery in rds_recv_incoming() when the
socket returned by rds_find_bound() belongs to a different network
namespace than the connection that carried the message. Use the
existing rds_conn_net() / sock_net() helpers and net_eq() for the
comparison.
Fixes: c809195f55 ("rds: clean up loopback rds_connections on netns deletion")
Signed-off-by: Aldo Ariel Panzardo <qwe.aldo@gmail.com>
Reviewed-by: Allison Henderson <achender@kernel.org>
Tested-by: Allison Henderson <achender@kernel.org>
Signed-off-by: Allison Henderson <achender@kernel.org>
Link: https://patch.msgid.link/20260708024314.601139-1-achender@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Anton Danilov says:
====================
udp: fix FOU/GUE over multicast
UDP encapsulation (FOU, GUE) has never worked correctly with multicast
destination addresses. When a FOU-encapsulated packet arrives at a
multicast address, it enters __udp4_lib_mcast_deliver() /
__udp6_lib_mcast_deliver() which call consume_skb() on packets that
need resubmission to the inner protocol handler, silently dropping
them instead.
The unicast delivery paths handle this correctly by propagating the
return value up to ip[6]_protocol_deliver_rcu() for resubmission, but
the multicast paths were never updated to support UDP encapsulation
resubmit.
This causes silent packet loss for FOU/GRETAP tunnels configured with
multicast remote addresses (both IPv4 and IPv6).
Reproducing the issue (IPv4):
ip netns add ns_a && ip netns add ns_b
ip link add veth0 netns ns_a type veth peer name veth1 netns ns_b
ip -n ns_a addr add 10.0.0.1/24 dev veth0 && ip -n ns_a link set veth0 up
ip -n ns_b addr add 10.0.0.2/24 dev veth1 && ip -n ns_b link set veth1 up
ip -n ns_a route add 239.0.0.0/8 dev veth0
ip -n ns_b route add 239.0.0.0/8 dev veth1
# Disable early demux to expose the issue (otherwise it's partially masked)
ip netns exec ns_b sysctl -w net.ipv4.ip_early_demux=0
# Join multicast group on receiver
ip -n ns_b addr add 239.0.0.1/32 dev veth1 autojoin
# Sender: GRETAP with FOU encap
ip -n ns_a link add eoudp0 type gretap \
remote 239.0.0.1 local 10.0.0.1 \
encap fou encap-sport 4797 encap-dport 4797 key 239.0.0.1
ip -n ns_a link set eoudp0 up
ip -n ns_a addr add 192.168.99.1/24 dev eoudp0
# Receiver: FOU listener + GRETAP
ip netns exec ns_b ip fou add port 4797 ipproto 47
ip -n ns_b link add eoudp0 type gretap \
remote 239.0.0.1 local 10.0.0.2 \
encap fou encap-sport 4797 encap-dport 4797 key 239.0.0.1
ip -n ns_b link set eoudp0 up
ip -n ns_b addr add 192.168.99.2/24 dev eoudp0
# Static neigh: ARP replies can't traverse unidirectional mcast tunnel
recv_mac=$(ip -n ns_b link show eoudp0 | awk '/ether/{print $2}')
ip -n ns_a neigh add 192.168.99.2 lladdr $recv_mac dev eoudp0
# Test: ping through the FOU/GRETAP tunnel
ip netns exec ns_a ping -c 100 192.168.99.2
# -> without this patch: 0 packets received on eoudp0
# -> with this patch: all packets received on eoudp0
IPv6 (using fou6 + ip6gretap) exhibits the same silent drop with a
different fix (see 1/2 for the sign-of-ret difference between
ip_protocol_deliver_rcu() and ip6_protocol_deliver_rcu()).
AI assistance (Claude, claude-opus-4-6) was used during root cause
analysis of the kernel source code (tracing the call chain from
udp[6]_queue_rcv_skb through encap_rcv to ip[6]_protocol_deliver_rcu,
comparing unicast/GSO/multicast paths) and during patch and selftest
authoring.
v5: https://lore.kernel.org/netdev/cover.1783218197.git.littlesmilingcloud@gmail.com/
v4: https://lore.kernel.org/netdev/cover.1782945956.git.littlesmilingcloud@gmail.com/
v3: https://lore.kernel.org/netdev/cover.1777934869.git.littlesmilingcloud@gmail.com/
v2: https://lore.kernel.org/netdev/ad_dal164gVmImWl@dau-home-pc/
v1 (RFC): https://lore.kernel.org/netdev/ad7MsSJOuUU6EGwS@dau-home-pc/
====================
Link: https://patch.msgid.link/cover.1783372173.git.littlesmilingcloud@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add a selftest to verify that FOU-encapsulated packets addressed to a
multicast destination are correctly resubmitted to the inner protocol
handler (GRE) via the UDP multicast delivery path. Both IPv4 and IPv6
paths are tested.
The test creates two network namespaces connected by a veth pair with
a FOU/GRETAP (IPv4) and FOU/ip6gretap (IPv6) tunnel using multicast
remote addresses (239.0.0.1 and ff0e::1). Ping is sent through each
tunnel and received packets are counted on the receiver's tunnel
interface.
The veth pair is created directly inside the namespaces to avoid
possible name collisions with devices in the root namespace.
Static neighbor entries are configured on the sender because ARP/ND
replies from the receiver cannot traverse the unidirectional multicast
tunnel back to the sender.
The early demux optimization (net.ipv4.ip_early_demux, which controls
both IPv4 and IPv6) is disabled on the receiver to force packets
through __udp4_lib_mcast_deliver() / __udp6_lib_mcast_deliver(), which
is the code path being tested.
Signed-off-by: Anton Danilov <littlesmilingcloud@gmail.com>
Assisted-by: Claude:claude-opus-4-6
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/a5b65f092d22a12b52fc536c0565b948cd8ecae3.1783372173.git.littlesmilingcloud@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When a UDP encapsulation socket (e.g., FOU) receives a multicast
packet, __udp4_lib_mcast_deliver() and __udp6_lib_mcast_deliver()
call consume_skb() when udp_queue_rcv_skb() returns a positive value.
A positive return value from udp_queue_rcv_skb() indicates that the
encap_rcv handler (e.g., fou_udp_recv) has consumed the UDP header
and wants the packet to be resubmitted to the IP protocol handler
for further processing (e.g., as a GRE packet).
The unicast paths handle this correctly by propagating the return
value up to ip_protocol_deliver_rcu() / ip6_protocol_deliver_rcu()
for resubmission. However, the multicast paths destroy the packet
via consume_skb() instead of resubmitting it, causing silent packet
loss.
This affects any UDP encapsulation (FOU, GUE) combined with multicast
destination addresses.
Fix this by returning the value from udp_queue_rcv_skb() when it is
positive, matching the behavior of the corresponding unicast paths.
Note the sign difference between IPv4 and IPv6:
- IPv4: udp_unicast_rcv_skb() returns -ret, and
ip_protocol_deliver_rcu() resubmits when ret < 0
(using -ret as the protocol number).
- IPv6: udp6_unicast_rcv_skb() returns ret, and
ip6_protocol_deliver_rcu() resubmits when ret > 0
(using ret as the nexthdr).
Both mcast paths now follow the same convention as their respective
unicast paths.
Suggested-by: Kuniyuki Iwashima <kuniyu@google.com>
Signed-off-by: Anton Danilov <littlesmilingcloud@gmail.com>
Assisted-by: Claude:claude-opus-4-6
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/5372ccac062193147e02b991d5328a5c3fa3a85a.1783372173.git.littlesmilingcloud@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
After PCIe DPC recovery, mlx5 reloads the affected functions and
replays multiport affiliation events. In the reported failure, the
first relevant device error was:
pcieport 0000:10:01.1: DPC: containment event
pcieport 0000:10:01.1: PCIe Bus Error: severity=Uncorrected (Fatal)
pcieport 0000:10:01.1: [ 5] SDES (First)
mlx5 recovered the PCI functions and resumed 0000:11:00.1. During
that resume, RDMA multiport binding replayed
MLX5_DRIVER_EVENT_AFFILIATION_DONE and mlx5e sent
MPV_DEVCOM_MASTER_UP. The host then panicked with:
BUG: kernel NULL pointer dereference, address: 0000000000000010
RIP: mlx5_devcom_comp_set_ready+0x5/0x40 [mlx5_core]
RDI: 0000000000000000
Call trace included:
mlx5_devcom_comp_set_ready
mlx5e_devcom_event_mpv
mlx5_devcom_send_event
mlx5_ib_bind_slave_port
mlx5r_mp_probe
mlx5_pci_resume
MPV devcom registration publishes mlx5e private data to the component
peer list before mlx5e_devcom_init_mpv() stores the returned component
device in priv->devcom. A concurrent master-up event can therefore
reach a peer whose private data is visible but whose priv->devcom
backpointer is still NULL.
MPV_DEVCOM_MASTER_UP already carries the sender/master mlx5e private
data as event_data. The ready bit is stored on the shared devcom
component, not on an individual peer. Use the sender devcom when
marking the MPV component ready.
This preserves the readiness transition while avoiding a NULL
dereference of the peer devcom pointer during affiliation replay after
PCI error recovery.
Fixes: bf11485f84 ("net/mlx5: Register mlx5e priv to devcom in MPV mode")
Assisted-by: Codex:gpt-5
Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
Cc: stable@vger.kernel.org # 6.7+
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260707233911.3651139-1-manjunath.b.patil@oracle.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
