| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| The websocket client's HTTP header callback on_header_value() in subsys/net/lib/websocket/websocket.c receives the Sec-WebSocket-Accept header value as a pointer/length pair from the HTTP parser. The referenced bytes live inside the application-supplied receive buffer and are not NUL terminated. When the received value did not match the expected accept key, the pointer was passed directly to a %s conversion in a NET_DBG() call, so the logging code measured the string with strlen() and read on past the end of the header value looking for a terminator.
The value is supplied by the peer during the WebSocket handshake, before any application-level authentication, so a malicious or man-in-the-middle server reached over a ws:// (or otherwise attacker-terminated) connection can trigger the mismatch branch at will simply by answering with a wrong accept key. By sizing the response so the header value ends at the tail of the buffer passed to websocket_connect() as wreq->tmp_buf, the peer can make the scan run off the end of that buffer and into adjacent memory. The defective log statement is compiled in only when websocket debug logging is enabled (CONFIG_NET_WEBSOCKET_LOG_LEVEL_DBG=y, or CONFIG_LOG_DEFAULT_LEVEL=4); it is absent from a default build, which bounds exposure to diagnostic configurations.
The consequence is an out-of-bounds read whose contents are emitted to the log stream: memory adjacent to the receive buffer is disclosed wherever the device's logs are collected or displayed. On targets where the memory following the buffer is not mapped, the scan can instead fault and terminate the connection or thread. No out-of-bounds write occurs, so there is no integrity impact.
The fix copies at most sizeof(got) - 1 bytes of the parser-provided slice into a bounded stack buffer, NUL terminates the copy, and logs that instead, so the formatter never reads beyond the bytes the parser actually offered. |
| The Bluetooth Classic SDP client built its outgoing request PDUs with bt_sdp_create_pdu(), which allocates from the sdp_pool net-buf pool with a K_FOREVER timeout and no failure handling, in sdp_client_ss_search(), sdp_client_sa_search() and sdp_client_ssa_search() in subsys/bluetooth/host/classic/sdp.c. All of these run on the Bluetooth workqueue bt_workq (via sdp_client_work_handler(), the L2CAP sdp_client_connected() callback, or the continuation-state follow-up from a received response), which is the same single thread that dispatches incoming HCI ACL and event traffic, drives L2CAP, and performs connection and channel teardown. sdp_pool holds only CONFIG_BT_MAX_CONN buffers (one by default) and is shared with the SDP server role.
An unauthenticated peer within BR/EDR range can drive both halves of the trigger. The SDP L2CAP server is registered at BT_SECURITY_L0, so any inbound SDP request makes the device allocate the pool's buffer for a response, which is held until the controller has transmitted it. The same peer can then cause a client request to be issued while the pool is empty: in-tree profiles start SDP discovery from inbound connection callbacks (for example hfp_hf_accept() calls bt_sdp_discover() through hfp_hf_create(), and hfp_ag.c mirrors this), and a peer acting as an SDP server can force repeated client requests by answering with partial responses carrying continuation state.
With the pool empty, the allocation blocks uninterruptibly and freezes the entire Bluetooth host. The stall becomes permanent when the release of the outstanding buffer itself depends on the blocked workqueue: if the link carrying the queued PDU is dropped or stalls, hci_disconn_complete() and the L2CAP channel cleanup that purges the channel TX queue are themselves bt_workq work items, so the buffer is never returned. The result is an unrecoverable denial of service of all Bluetooth functionality until reboot; there is no memory-safety, confidentiality or integrity impact.
The fix allocates client PDUs with K_NO_WAIT via a new sdp_client_create_pdu(), propagates -ENOBUFS, parks the session on a new SDP_CLIENT_FLAG_WAIT_FOR_TX_BUF flag, and adds an sdp_pool destructor that re-submits the SDP client worker when a buffer becomes available. Note that the SDP server role still allocated from the same pool with K_FOREVER at this commit; that path was addressed separately. |
| The Virtual L2 receive handler virtual_recv() in subsys/net/l2/virtual/virtual.c ends with a fall-through path meant for a virtual interface that has nothing attached below it: it takes net_if_get_device(iface)->api, treats it as a struct virtual_interface_api, and calls api->recv(iface, pkt). That handler is however also invoked with a non-virtual interface: ethernet_recv() in subsys/net/l2/ethernet/ethernet.c deliberately passes the Ethernet interface so that a tagged frame can be dispatched over the VLAN interfaces attached to it. For an Ethernet interface dev->api is a struct ethernet_api, whose layout only shares the leading struct net_if_api, so the slot read as recv is actually get_config() (default) or get_capabilities() (when CONFIG_NET_STATISTICS_ETHERNET is enabled).
The fall-through is reached whenever no attached virtual interface claims the frame - most simply when the VLAN interface owning the frame's tag exists but is administratively down, since net_eth_get_vlan_iface() resolves the tag without checking link state while virtual_recv() skips interfaces that are not up. Any host able to place an 802.1Q-tagged Ethernet frame on the same link can then trigger the mistyped indirect call; no authentication, IP reachability or user interaction is required, because the frame is consumed at L2 before any address check.
The resulting call passes a struct net_if where the driver expects a struct device and a struct net_pkt where it expects a struct net_if , with the callee's remaining parameters left as whatever the call site happens to leave in the argument registers. Driver implementations dereference these values: get_capabilities() typically reads dev->config/dev->data (a wild read through a non-pointer field of struct net_if), and get_config() implementations write through the uninitialized struct ethernet_config * argument. The practical impact is undefined behaviour in the network RX path - crashes and uncontrolled memory corruption - plus mis-handling of the packet when a capability bitmask is returned as a net_verdict. The attacker controls neither the corrupted address nor the value written, so this is a denial-of-service and memory-corruption issue rather than a controllable write primitive.
The same defective fall-through is reachable through any_recv() in subsys/net/l2/dummy/any/any.c, which passes a dummy interface whose struct dummy_api is shorter than the offset read, producing an out-of-bounds read of the driver's API object followed by an indirect call. The fix drops any packet whose interface is not a Virtual L2 interface, closing both entry points. |
| The Bluetooth Classic SDP server in subsys/bluetooth/host/classic/sdp.c allocated every response PDU through bt_sdp_create_pdu(), which expands to bt_l2cap_create_pdu() and therefore uses a K_FOREVER timeout. The allocation is made from sdp_pool, a fixed pool holding only CONFIG_BT_MAX_CONN buffers (one in the default configuration) that is shared with the SDP client role. When the pool is empty the call blocks without a bound.
The allocation is reached from bt_sdp_recv() on the Bluetooth host's RX context: the bt_workq "BT RX WQ" workqueue in the default configuration since v3.1.0, and the dedicated "BT RX" thread before that. This context processes all normal-priority HCI events and all received ACL data. The SDP L2CAP server is registered on PSM 0x0001 with no security requirement (BT_SECURITY_L0), so an unpaired, unauthenticated BR/EDR peer within radio range of a connectable device can drive it. Every malformed or unrecognised SDP PDU reaches send_err_rsp(), and the sdp_svc_search_req(), sdp_svc_att_req() and sdp_svc_search_att_req() handlers allocate as well, so a short burst of pipelined requests drains the pool and parks the RX context in the blocking allocation.
A pool buffer is released only once its response has been sent, and a peer that withholds baseband acknowledgements keeps the controller's buffers for the link full, so the response is never sent. While the RX context is blocked, the entire Bluetooth host, for all connections and both LE and BR/EDR, stops processing normal-priority HCI events and received data. In most affected releases the stall is permanent. Queued PDUs are freed only by connection or channel teardown, which waits for the normal-priority Disconnection Complete handler, and that handler runs on the blocked RX context, so the host does not recover without a reset. In v3.7.0 through v4.1.x the TX processor discards the queued response once the link supervision timeout drops the connection, so each stall ends at that timeout, but the peer can reconnect and repeat it for as long as it stays in range. The impact is limited to availability; no memory corruption or information disclosure is involved.
The fix adds sdp_svc_create_pdu() using K_NO_WAIT, pre-allocates the response buffer once in bt_sdp_recv(), stashes the request in sdp->recv_pending when no buffer is available, and retries it from a work handler once a buffer is freed by the pool destroy callback. |
| bt_att_find_req_by_user_data() in subsys/bluetooth/host/att.c walks every ATT channel attached to a connection and dereferences chan->req->user_data without checking chan->req. An ATT channel with no client request in flight holds chan->req == NULL — that is its normal idle state, set by att_handle_rsp() when a response resolves the transaction and tested by att_req_send_process() to pick a free channel — so the lookup reads through a NULL pointer as soon as it visits an idle channel. Every other chan->req user in the file already guards the pointer.
The lookup is reached from the public bt_gatt_cancel() API and, more importantly, from the host itself: bt_gatt_unsubscribe() calls bt_gatt_cancel() while a subscribe CCC write is still outstanding (BT_GATT_SUBSCRIBE_FLAG_WRITE_PENDING), and call_notify_cb_and_maybe_unsubscribe() in subsys/bluetooth/host/gatt.c invokes bt_gatt_unsubscribe() from the RX path whenever an application notify callback returns BT_GATT_ITER_STOP. A malicious or malfunctioning peer can hold the trigger window open by withholding the ATT Write Response to the CCC write and then sending a notification for the subscribed handle. With CONFIG_BT_EATT the traversal is very likely to meet an idle sibling channel, since ATT channels are prepended to the list and a channel opened after the pending request was sent is visited first.
The dereference is a read of a small constant offset from address zero, so the impact is availability only: no memory corruption, no attacker-chosen address and no information disclosure. On platforms and configurations where a near-NULL read traps — TrustZone non-secure images, MMU targets, native_sim, or builds with CONFIG_NULL_POINTER_EXCEPTION_DETECTION_MPU/_DWT — the fault occurs in the Bluetooth RX context and results in a fatal error and device reset. On a plain Cortex-M build with the vector table mapped at address 0 and null-pointer detection disabled (the default) the read succeeds and the defect is inert.
The fix adds an explicit chan->req != NULL test to the channel traversal and additionally rejects a NULL user_data argument, which would otherwise have matched, and cancelled, an unrelated request. |
| The ITE IT8xxx2 SHI host-command backend in subsys/mgmt/ec_host_cmd/backends/ec_host_cmd_backend_shi_ite.c arms the SPI peripheral before its receive context exists. shi_ite_init_registers() sets the state machine to SHI_STATE_READY_TO_RECV and calls irq_enable() while the device initializes at POST_KERNEL priority CONFIG_KERNEL_INIT_PRIORITY_DEFAULT (40), but data->rx_ctx and data->tx are only assigned later by shi_ite_backend_init(), which ec_host_cmd_init() invokes at CONFIG_EC_HOST_CMD_INIT_PRIORITY (default 80).
If the attached host processor starts a SPI transaction inside that window, the Rx-valid-length interrupt fires, shi_ite_int_handler() passes its state check and calls shi_ite_parse_header(), which stores the host-derived request length through the still-NULL data->rx_ctx. On the IT8xxx2 low addresses are not backed by RAM, and with CONFIG_NULL_POINTER_EXCEPTION_DETECTION_PMP the first bytes of the address space are explicitly made inaccessible, so the store raises a fatal access fault in interrupt context. The window widens further when CONFIG_EC_HOST_CMD_INITIALIZE_AT_BOOT is disabled and the application calls ec_host_cmd_init() itself.
The result is a kernel panic on the embedded controller: with Zephyr's default fatal handler the CPU halts, so the controller stops servicing host commands until a hardware reset. Because the trigger is traffic from the attached host processor -- an untrusted peer from the controller firmware's point of view -- software on that host can provoke the fault by forcing a controller reset and issuing transactions during the reboot, turning a recoverable reset into a persistent outage. Impact is limited to availability: the faulting address is fixed and not attacker-chosen, so no memory-corruption primitive follows.
The fix leaves the interface at SHI_STATE_DISABLED with the interrupt masked during device init, moves irq_enable() into shi_ite_backend_init() after the buffers are published, and makes shi_ite_int_handler() reject a NULL data->rx_ctx while write-clearing IT83XX_SPI_ISR and IT83XX_SPI_RX_VLISR so that an unexpected early interrupt cannot storm. |
| The Synopsys DesignWare I3C controller driver (drivers/i3c/i3c_dw.c) tracks free Device Address Table (DAT) slots with get_free_pos(), which signals an exhausted table by returning -1. The helper declared its return type as uint8_t, so the sentinel was truncated to 255, and both consumers (dw_i3c_attach_device() and dw_i3c_i2c_attach_device()) also stored it in a uint8_t before testing if (pos < 0) — a comparison that is never true for an unsigned value. The exhaustion guard was therefore inert in every build.
With the guard defeated, an attach attempted on a full DAT proceeds with pos == 255: it writes data->dw_i3c_i2c_priv_data[255].id, more than 200 bytes past the end of the 32-element dw_i3c_i2c_priv_data array and past the end of struct dw_i3c_data; it stores the resulting out-of-bounds address in desc->controller_priv, from which later detach, reattach and IBI operations read the slot index back; and it performs a 32-bit MMIO store 1020 bytes beyond the hardware DAT base, outside the table the controller reports. The attach still returns success, so the upper layer believes the device is bound.
The DAT is smaller than the software array (its depth is read from DEVICE_ADDR_TABLE_POINTER, commonly 8 or 16 entries), so exhaustion is reachable. In a dual-role build (CONFIG_I3C_DUAL_ROLE with CONFIG_I3C_USE_IBI) acting as a secondary controller, the attach count is supplied by the bus's active controller: i3c_sec_handoffed() attaches one device per entry of the received DEFTGTS list, so a hostile or malfunctioning controller on the local I3C bus that declares more targets than the local DAT depth triggers the condition once per surplus entry. Exhaustion can also be reached from a devicetree that declares more devices than the table holds, or by repeated i3c attach shell commands.
Impact is bounded: the out-of-bounds store is a single byte of fixed value at a fixed displacement, with no attacker-chosen data or offset, so the realistic outcomes are corruption of neighbouring static driver state, writes into reserved controller registers, and inconsistent driver bookkeeping (a crash or a wedged I3C bus) rather than a controllable memory-write primitive. The attack vector is the local chip-to-chip I3C bus; there is no network-reachable path. |
| The Zephyr WebSocket client validates the server's handshake response in on_header_value() in subsys/net/lib/websocket/websocket.c. The comparison was performed as strncmp(at, str, length), where length is the length of the value the server sent rather than the length of the expected value, so the check only asked whether the received value was a prefix of the expected base64(SHA-1(key || GUID)). A server that returned a handful of correct characters passed, and a response carrying an empty Sec-WebSocket-Accept: header caused strncmp() to be called with a count of zero, which returns 0 and set ctx->sec_accept_ok with no comparison performed at all. The header-value callback also cleared ctx->sec_accept_present after the first chunk, so a value split across receive-buffer boundaries was only ever examined in part.
ctx->sec_accept_ok is the sole WebSocket-specific gate in websocket_connect(); the client does not check the HTTP status code or the Upgrade:/Connection: headers. Any HTTP response containing an empty or prefix-matching Sec-WebSocket-Accept header therefore completed the handshake, after which the socket is finalized as a WebSocket and all further peer bytes are parsed as frames. The affected path is driven directly by remote input through subsys/net/lib/http/http_parser.c, and is reached by every user of websocket_connect(), including MQTT over WebSocket (subsys/net/lib/mqtt/mqtt_transport_websocket.c) and subsys/net/lib/ocpp/ocpp.c.
The Sec-WebSocket-Accept exchange is not a secret-based authentication -- an endpoint or on-path attacker that sees the request can compute the expected value -- so the impact is bounded to parties that never observed the handshake request: a caching or transparent intermediary replaying a canned response, an off-path attacker injecting a spoofed response, or a non-WebSocket endpoint the client was induced to talk to. For those, the per-connection binding that RFC 6455 requires the client to enforce provided no protection at all, letting a single crafted response convince the device it holds a live WebSocket session and causing it to send application data to, and accept frames from, an unverified peer. There is no memory-safety consequence.
The fix compares each delivered chunk against the expected value at the position reached so far, refuses a chunk that would overrun it, and accepts only once every byte of the expected value -- and no more -- has been matched; on_header_field() now also clears the flag for any other header so a partial value cannot be continued across an intervening header. |
| bt_hci_le_per_adv_report_common() in subsys/bluetooth/host/scan.c handles the HCI_LE_Periodic_Advertising_Report (and _v2) event. It verified that the fixed event header was present and that the controller-declared Data_Length fits the tailroom of the periodic-sync reassembly buffer, but never checked that the event actually carried as many data bytes as evt->length claims. Both reassembly paths — DATA_STATUS_PARTIAL, and DATA_STATUS_COMPLETE while a chain is already in progress — then called net_buf_simple_add_mem(&per_adv_sync->reassembly, buf->data, evt->length), copying up to 255 bytes starting past the end of the event's real payload.
The Data_Length field originates in the HCI controller, so the flaw is reachable in host-only/split builds where the controller is a separate part reached over H4 UART, SPI, USB or an inter-core IPC transport: a malicious, compromised or buggy controller (or an attacker able to inject on that link) can send a report declaring a length larger than the event payload. The generic event dispatcher in subsys/bluetooth/host/hci_core.c only enforces a per-event minimum length and never cross-checks the subevent's internal length field, so nothing else catches it. Builds using Zephyr's own in-image controller are not affected, because the controller always emits a length matching the bytes it appends.
Exploitation additionally requires CONFIG_BT_PER_ADV_SYNC with CONFIG_BT_PER_ADV_SYNC_BUF_SIZE > 0 (the option defaults to 0, which compiles the reassembly block out; LE Audio broadcast-sink and Public Broadcast Profile applications that reassemble fragmented periodic advertising data set it), and an established periodic advertising sync.
The consequence is an out-of-bounds read of adjacent HCI RX-pool memory — including stale contents of previously received HCI events — which is copied into the reassembly buffer and then delivered to every registered bt_le_per_adv_sync_cb.recv listener as if it were received advertising data, so the disclosed memory becomes application-visible and may be forwarded off-device. The write side remains in bounds because the tailroom check uses the same length value, so there is no memory corruption; impact is limited to information disclosure and undefined behaviour on the over-read. The fix drops any report whose declared length exceeds the remaining event payload, matching the guard already present in the extended and legacy advertising report handlers. |
| The simulated EEPROM driver validated access ranges in eeprom_range_is_valid() (drivers/eeprom/eeprom_simulator.c) with a single test, (offset + len) <= config->size. Because offset is a signed off_t and len an unsigned size_t of the same rank, the addition is evaluated in unsigned arithmetic: a negative offset combined with a compensating len wraps the sum back into the accepted range. eeprom_sim_read() and eeprom_sim_write() then memcpy() through the EEPROM(addr) macro, which is plain pointer arithmetic on the driver's static mock_eeprom array, so the copy addresses memory located before that array.
eeprom_read() and eeprom_write() are __syscall entry points, and their verification handlers in drivers/eeprom/eeprom_handlers.c check only the device object and the caller's own data buffer (K_SYSCALL_MEMORY_READ/K_SYSCALL_MEMORY_WRITE) — the offset argument is passed to the driver unvalidated. With CONFIG_USERSPACE=y, a CONFIG_EEPROM_SIMULATOR device instance, and the device object granted to a user thread (the usage the in-tree tests/drivers/eeprom/api suite exercises from ZTEST_USER threads), an unprivileged thread could supply a negative offset and have the kernel perform the copy on its behalf.
The result is an out-of-bounds read of kernel .bss returned into the caller's buffer and, via the write path, an out-of-bounds write of caller-supplied bytes over kernel memory preceding mock_eeprom. Both the start address and the length of the affected span are attacker-selected, so the primitive is sufficient to corrupt adjacent kernel state and escape the user-mode sandbox. Exposure is limited to builds that instantiate the simulated EEPROM (status = "okay" by default on qemu_x86 and native_sim, otherwise added by a devicetree overlay) together with userspace isolation.
The fix rejects negative offsets and performs the remaining bound check without overflow. |
| The user-mode syscall verifiers for the 16-bit ("wide data") UART API in drivers/serial/uart_handlers.c validated the wrong driver operation. z_vrfy_uart_poll_out_u16(), z_vrfy_uart_tx_u16() and z_vrfy_uart_rx_enable_u16() each ran K_SYSCALL_DRIVER_UART() against the 8-bit op (poll_out, tx, rx_enable) rather than the 16-bit op that the corresponding implementation dereferences (poll_out_u16, tx_u16, rx_enable_u16). z_impl_uart_poll_out_u16(), z_impl_uart_tx_u16() and z_impl_uart_rx_enable_u16() in include/zephyr/drivers/uart/uart_internal.h call their op unconditionally, with no NULL guard of their own, because the verifier is supposed to have established that the op exists.
CONFIG_UART_WIDE_DATA is an application-selectable option that no driver selects, and only a handful of the serial drivers in the tree implement any _u16 operation; every other driver leaves those struct uart_driver_api members NULL while still providing the 8-bit ops the old check inspected. In a build with CONFIG_USERSPACE and CONFIG_UART_WIDE_DATA (plus CONFIG_UART_ASYNC_API for the uart_tx_u16() and uart_rx_enable_u16() variants), an unprivileged user-mode thread that legitimately holds access to such a UART device object can invoke the _u16 syscall: the object, permission and op checks all pass, and the kernel-side implementation then performs an indirect call through a NULL function pointer while running at supervisor privilege.
The result is a CPU exception taken in kernel context instead of the intended K_OOPS or -ENOSYS return. With the default k_sys_fatal_error_handler() this halts the system (or warm-reboots it under CONFIG_RESET_ON_FATAL_ERROR), so a thread confined by the userspace sandbox can deny service to the whole device — a capability the user-mode privilege boundary is specifically meant to withhold. The call target is a fixed NULL, so there is no attacker-controlled code execution and no memory disclosure; the separate K_SYSCALL_MEMORY_ARRAY_READ()/K_SYSCALL_MEMORY_ARRAY_WRITE() bounds checks on the caller's buffer were and remain correct. The fix points each verifier at the matching _u16 op, using a plain K_SYSCALL_OBJ() object check where the op does not exist or where the implementation already tests it for NULL. |
| The IEEE 1588 (PTP) port implementation in subsys/net/lib/ptp/port.c adopted the logMessageInterval field of a received Delay_Resp message without any validation. port_delay_resp_msg_process() stored the raw wire value (a full-range int8_t) into port_ds.log_min_delay_req_interval, and that value is later used as a shift count by port_timer_set_timeout() when the Delay_Req timer is armed. Values below about -30 truncate the computed timeout to zero, and values with a magnitude of 64 or more perform an undefined shift of a uint64_t; the standards-defined "not applicable" encoding 0x7F fell into the latter case as well.
PTP messages are received over unauthenticated link-local traffic (UDP multicast on 224.0.1.129, IPv6 multicast, or raw IEEE 802.3), and the handler accepts a Delay_Resp from any sender as long as the port is a time receiver using the end-to-end delay mechanism and the message echoes the port identity and a pending sequence id — both of which the victim itself broadcasts in its Delay_Req. Any node on the same link can therefore inject a single crafted Delay_Resp, or simply become the time transmitter and advertise the bogus interval.
Once the delay timer is armed with a zero timeout, the cooperative PTP thread re-arms and retransmits a Delay_Req on every tick. Each transmission takes a buffer from the fixed CONFIG_PTP_MSG_POLL_SIZE message slab (default 10) and holds it on port->delay_req_list for three seconds, so the slab is emptied within milliseconds and ptp_msg_alloc() then blocks with K_FOREVER inside the only thread that frees those buffers — permanently stalling time synchronization until reboot. Where a transmit-timestamp callback on another thread keeps recycling buffers, the result is instead a sustained multicast Delay_Req flood and continuous CPU consumption. There is no memory corruption and no disclosure of data; the impact is availability of the PTP subsystem and of the local network segment.
The fix rejects intervals outside a representable range and treats 0x7F as "not applicable" as required by IEEE 1588-2019 Table 42. The equivalent unchecked adoption of logMessageInterval from Sync messages in port_sync_msg_process(), a weaker variant of the same issue, is fixed by the companion change 334bf57f416f (#118150), which applies the same range check, ignores the field for unicast Sync messages and treats 0x7F as not applicable. |
| The HTTP/2 front-end of the Zephyr HTTP server aliased connection receive state as a scratch output. In subsys/net/lib/http/http_server_http2.c, handle_http2_static_fs_resource() passed &client->data_len as the file_size out-parameter of http_server_find_file(). client->data_len is the count of unprocessed bytes remaining in the fixed-size per-connection receive buffer client->buffer[CONFIG_HTTP_SERVER_CLIENT_BUFFER_SIZE] (256 bytes by default), an invariant every frame handler and the socket receive path rely on. Serving any static file therefore replaced that bounded value with the file's on-disk size, and nothing resets it afterwards.
The defect is reached by an unauthenticated HTTP/2 GET for any file under the web root, in applications that register a static-filesystem resource (HTTP_RESOURCE_TYPE_STATIC_FS) with CONFIG_FILE_SYSTEM=y; CONFIG_HTTP_SERVER_VERSION_2 is enabled by default and no TLS or credentials are required on this path. Once client->data_len exceeds the buffer size, the frame dispatch loop in subsys/net/lib/http/http_server_core.c keeps parsing past the end of the received data, and the peer controls the bytes at client->cursor, so it can choose which fabricated frame the state machine consumes next by appending bytes after its request.
The inflated length defeats the only guards on the buffer. parse_http_frame_header() and the per-frame handlers read and skip up to a 24-bit frame length beyond the buffer (out-of-bounds read); the compaction memmove(client->buffer, client->cursor, client->data_len) at the end of request handling copies the file-size number of bytes into the 256-byte buffer, overflowing the rest of struct http_client_ctx — including the current_detail/current_stream pointers, the parser_settings function-pointer table and the inactivity_timer work handler — and the adjacent client contexts in static memory; and the subsequent zsock_recv(client->fd, client->buffer + client->data_len, sizeof(client->buffer) - client->data_len, 0) underflows its size argument, writing attacker-supplied data of attacker-chosen length out of bounds.
The impact is remote memory corruption in the server's single address space: at minimum a crash or hang of the device, and plausibly control-data overwrite leading to code execution. The HTTP/1 and HTTP/3 static-filesystem handlers already used a local variable and are unaffected; the fix makes the HTTP/2 handler use a local size_t file_size as well. |
| The RFCOMM multiplexer control-channel handler in subsys/bluetooth/host/classic/rfcomm.c computed the echo length for a BT_RFCOMM_TEST command as buf->len - 1 and passed it to rfcomm_send_test(), whose length parameter is a uint8_t. A malformed but FCS-valid UIH frame on DLCI 0 can leave buf->len == 0 after rfcomm_handle_msg() pulls the two-byte MCC header, so buf->len - 1 underflows and narrows to 255.
rfcomm_send_test() then copies 255 bytes out of a receive buffer that holds no remaining valid data, and appends them to an outgoing L2CAP PDU allocated with only CONFIG_BT_L2CAP_TX_MTU (253 by default for CONFIG_BT_CLASSIC) bytes of tailroom. The source read runs past the end of the ACL receive buffer into adjacent buffers of the same net_buf pool, and the destination copy overruns the transmit buffer by about eight bytes; net_buf_add_mem() performs no runtime bounds check in ordinary builds, where the underlying __ASSERT_NO_MSG() is compiled out.
The trigger is a single five-byte RFCOMM frame on the multiplexer control channel, which is processed as soon as the L2CAP channel for PSM 3 is established — before any DLC exists and independent of RFCOMM session state. bt_rfcomm_init() registers that L2CAP server with .sec_level = BT_SECURITY_L1, which l2cap_br_conn_security() accepts without any link key from a peer that does not claim host Secure Simple Pairing support, so an unpaired BR/EDR device in radio range of a connectable target with CONFIG_BT_RFCOMM enabled (SPP, HFP, HSP, GOEP and similar profiles all enable it) can reach the defect.
The resulting 255-byte TEST response is transmitted back to the attacker, disclosing roughly 255 bytes of adjacent stack/pool memory — including the contents of neighbouring receive buffers, which an attacker can prime with earlier packets — and the accompanying out-of-bounds write corrupts memory following the transmit buffer, which can crash the Bluetooth host. The request is trivially repeatable, so both the disclosure and the corruption can be driven at will.
Two fixes are needed, and the first on its own does not close the out-of-bounds write. Fix 1 (0040ef781e1d, #115768) strips the FCS byte in rfcomm_recv() and echoes buf->len, rejecting anything over UINT8_MAX, which removes the underflow and the over-read. But rfcomm_make_uih_msg() still allocates the response with bt_l2cap_create_pdu() and never compares the requested length with the buffer's tailroom, so a well-formed TEST command carrying a pattern of about 248 bytes or more still builds a response (3-byte frame header, 2-byte message header, pattern, FCS) larger than the 253 bytes available. A pattern that long can only be received when the RFCOMM L2CAP receive MTU (CONFIG_BT_RFCOMM_L2CAP_MTU, which defaults to CONFIG_BT_BUF_ACL_RX_SIZE) exceeds CONFIG_BT_L2CAP_TX_MTU. The plain BR/EDR defaults (200 against 253) cannot reach it, but from v4.3.0 enabling CONFIG_BT_GOEP raises CONFIG_BT_BUF_ACL_RX_SIZE to 264 (265 from v4.4.0), so GOEP builds with only fix 1 applied still overrun the transmit buffer by up to eight bytes. This includes v4.3.2, the last v4.3 release, which carries fix 1 but will not receive fix 2. Fix 2 (214530895dd3, #118293) checks the tailroom on every RFCOMM PDU allocation and closes this residual write. |
| The AVRCP browsing channel's L2CAP receive-buffer callback, browsing_avrcp_l2cap_alloc_buf() in subsys/bluetooth/host/classic/avrcp.c, allocated from the shared avctp_browsing_rx_pool with K_FOREVER. L2CAP invokes that callback from the Bluetooth RX work queue thread (bt_workq, "BT RX WQ"), both for the head buffer of a new SDU and — through l2cap_br_alloc_frag() and net_buf_append_bytes() — for every further fragment of a reassembled SDU. That same thread is the only context that ever releases these buffers, after bt_l2cap_br_recv_seg() delivers a completed SDU and calls net_buf_unref(), so a wait taken there can never be satisfied.
The pool holds CONFIG_BT_MAX_CONN + 1 buffers of CONFIG_BT_AVRCP_BROWSING_L2CAP_MTU bytes (two buffers of 339 bytes in the default configuration). A connected browsing peer can drain it because bt_l2cap_br_recv_seg() in subsys/bluetooth/host/classic/l2cap_br.c accepts the SDU length declared in an SAR_START segment without bounding it by br_chan->rx.mtu: a peer that announces a large SDU and then streams continuation segments forces one pool buffer per segment-worth of data. Once the pool is empty the next allocation blocks forever inside the callback.
Triggering the flaw requires a build with CONFIG_BT_CLASSIC and the experimental CONFIG_BT_AVRCP_BROWSING option, an application that registers AVRCP callbacks, and a BR/EDR peer that has established the AVRCP control channel and then the browsing channel, which requires an encrypted (paired) link. No malformed packet is needed — a few hundred bytes of ordinary segmented L2CAP traffic suffice.
The consequence is a permanent hang of the whole Bluetooth host, not just of AVRCP: the blocked thread drains bt_dev.rx_queue for all HCI ACL data and low-priority events and also runs the host's internal work items, so all connections stop being serviced and the state is unrecoverable without a reboot. There is no memory-corruption or information-disclosure component. The fix allocates with K_NO_WAIT, so an empty pool now yields a local-busy indication or a channel disconnect that the application can recover from. |
| The user-mode syscall verifiers for the charger driver API in drivers/charger/charger_handlers.c did not inspect the prop argument before copying union charger_propval across the user/kernel boundary. Two members of that union are function pointers — status_notification (CHARGER_PROP_STATUS_NOTIFICATION) and online_notification (CHARGER_PROP_ONLINE_NOTIFICATION), declared in include/zephyr/drivers/charger.h — and z_vrfy_charger_set_prop() forwarded the caller-supplied union verbatim to the driver's set_property callback for any property. z_vrfy_charger_get_prop() had the mirror defect and copied the stored kernel callback address back out to user memory.
With CONFIG_USERSPACE enabled, any user-mode thread that has been granted a charger device object could call charger_set_prop() with CHARGER_PROP_STATUS_NOTIFICATION or CHARGER_PROP_ONLINE_NOTIFICATION and an arbitrary pointer value. Drivers that implement these properties — drivers/charger/charger_max20335.c, drivers/charger/charger_pf1550.c and drivers/charger/charger_bq2518x.c — store the pointer in device data and later invoke it from their interrupt work handlers, which run in supervisor mode, on any charger state change such as a cable plug/unplug.
The result is a control-flow hijack at kernel privilege from an unprivileged thread, defeating the isolation that CONFIG_USERSPACE is meant to provide; the read path additionally disclosed a kernel code address to user mode, and clearing a notifier installed by a supervisor thread let user mode suppress charger notifications. Supervisor-mode callers of charger_set_prop()/charger_get_prop() are unaffected.
The fix makes both verifiers reject the two notifier properties outright — including when the supplied callback is NULL, since the notifiers are device state — and adds a BUILD_ASSERT on CHARGER_PROP_COMMON_COUNT so that new properties force a re-review of the union's user-mode safety. |
| The user-mode system call verifier z_vrfy_charger_get_prop() in drivers/charger/charger_handlers.c declares an uninitialized union charger_propval k_val on the kernel (privileged) stack, passes it to the charger driver back end via z_impl_charger_get_prop(), and then copies the full sizeof(union charger_propval) back into the calling user thread's buffer with k_usermode_to_copy(). That discloses kernel stack contents on two paths:
On error. The copy-out ignores the driver's return code. In-tree back ends return an error without writing through the value pointer for properties they do not implement: charger_gpio_get_prop() returns -ENOTSUP unconditionally, and bq24190_get_prop() has a default: return -ENOTSUP; arm that never touches val. The whole union is then stack residue.
On success. A back end writes only the one union member the requested property selects (for example sbs_charger.c writes the one-byte val->present for CHARGER_PROP_PRESENT and returns 0), yet the handler still copies the whole union, so every byte beyond that member is uninitialized.
The handler is compiled only when CONFIG_USERSPACE=y and is reachable by any user-mode thread that has been granted access to a charger device object. K_SYSCALL_DRIVER_CHARGER() validates only the device object and its API pointer, not the property identifier, so an unprivileged thread can choose prop freely: an unimplemented property forces the error path, a narrow one the success path, and the call can be repeated at will.
Each call discloses up to the size of the union (12 bytes on a 32-bit target, 16 bytes on 64-bit targets where the function-pointer members widen it) of supervisor-stack memory. The bytes are residue of earlier kernel-side frames on the thread's privileged stack and can contain kernel object addresses and saved return addresses, which an unprivileged thread cannot otherwise observe and which help defeat address-layout assumptions when exploiting a separate memory-safety defect. There is no integrity or availability impact.
Two commits fix it, and both are needed. 5054384e983c skips the copy-out when the back end reports an error. c31cac727d02 zeroes the union before the driver call, so the bytes copied out on success are driver-written data or zeros. A tree carrying only the first still leaks on every successful call. |
| The NXP ENET-QOS Ethernet driver's receive work handler eth_nxp_enet_qos_rx() in drivers/ethernet/eth_nxp_enet_qos hands each processed descriptor back to the DMA by writing desc->read.control = rx_desc_refresh_flags, which sets OWN_FLAG and BUF1_ADDR_VALID_FLAG. RDES0 of a receive descriptor is the DMA buffer address in read format, but the hardware write-back aliases that word with the VLAN tag fields (or, for a context descriptor, the timestamp low word), so it is destroyed on every completion. The success path rewrites desc->read.buf1_addr before re-arming; the four drop paths — errored or non-first descriptor, net_pkt_rx_alloc() failure, invalid descriptor length, and net_pkt_get_frag() failure — did not, re-arming the ring slot with the write-back residue as its DMA target address.
The drop paths are driven entirely by received traffic and need no privilege or special configuration. With the defaults for this driver (CONFIG_NET_BUF_DATA_SIZE of 128 and 16 receive descriptors) a single full-size frame consumes twelve ring slots and twelve K_NO_WAIT fragment allocations, so a short burst of frames sent to the device exhausts the receive pools and reaches the allocation-failure paths; a frame with a receive error, or any frame whose first descriptor was already dropped, reaches the non-first-descriptor path.
Once a slot has been poisoned, the DMA writes the next frame received in that slot to an address derived from the frame's VLAN tag fields rather than to a receive buffer. On the supported MCX SoCs that address falls in flash or reserved space, so the transfer raises a fatal bus error that the driver only logs — there is no reset or recovery path — and reception stops permanently until the device is rebooted; where the write is silently discarded instead, the untouched buffer is later handed to the network stack with a length, delivering stale buffer contents as if they were a received frame. A single triggering event is enough to make the condition persistent, so an attacker able to send traffic to the device can permanently disable its network interface. |
| The NXP ENET QoS Ethernet driver in drivers/ethernet/eth_nxp_enet_qos serialises transmission with a binary semaphore, data->tx.tx_sem, taken in eth_nxp_enet_qos_tx() and returned only by the transmit-completion interrupt handler tx_dma_done(). A frame that has been handed to the transmit DMA when the Ethernet link goes down is never written back, so that completion interrupt never arrives. The PHY callback eth_nxp_enet_qos_phy_cb() reacts to carrier loss only by calling net_eth_carrier_set(); it does not stop the DMA, take the descriptors back or release the outstanding frame, and no timeout exists to do so.
The trigger is a link-down/link-up transition that lands while a frame is under the transmit DMA. It can be produced by anyone who controls the link partner or the cable — for example by administratively downing and restoring the attached switch port, or by restarting auto-negotiation from a peer device — and can be made reliable by generating traffic that keeps the device transmitting and repeating the transition. No credentials or user interaction are required, and the trigger is at the link layer rather than over IP.
Because the semaphore is never given back, every subsequent call to eth_nxp_enet_qos_tx() fails with -EBUSY for the remaining life of the boot; restoring the link does not recover the interface, since the link-up path only rewrites MAC_CONFIGURATION. The affected interface continues to receive but can no longer transmit anything — no ARP replies, no TCP acknowledgements, no DHCP renewals — until the device is reset. The net_pkt and the net_buf fragments of the abandoned frame are likewise never unreferenced and never return to their pools. The security-relevant property is persistence: a momentary link flap leaves the interface permanently unable to send, so the denial of service outlasts the attacker's access to the link. There is no confidentiality or integrity impact.
The fix adds a claim/release pair around the outstanding packet so that completion and abandonment are mutually exclusive, aborts an outstanding frame from the PHY callback on carrier loss (stopping the DMA, flushing the MTL transmit queue and reinitialising the descriptor ring), and arms a transmit watchdog, CONFIG_ETH_NXP_ENET_QOS_TX_TIMEOUT_MS, so that any non-completing DMA — not only a carrier loss — reclaims the transmit slot and the packet. |
| When CONFIG_THREAD_STACK_MEM_MAPPED is enabled, each thread stack is mapped into the kernel virtual address space with a guard page on either side, and the mapping is released during abort cleanup. k_thread_abort_cleanup() in kernel/thread.c has two paths: a deferred one for self-abort, which stashes the mapping in thread_cleanup_stack_addr/thread_cleanup_stack_sz, and an immediate one for a thread aborted by some other thread. The immediate path called do_thread_cleanup(), which ignored its thread argument and only ever unmapped the stashed address — a value populated exclusively by the self-abort path. The mapped stack of a thread aborted by another thread was therefore never unmapped.
Every k_thread_abort() of a non-current thread leaks the thread's stack region plus its two guard pages from the kernel virtual address space. Threads that end by returning from their entry function take the working self-abort path, so the leak is specific to explicit cancellation — k_thread_abort(), pthread_cancel(), osThreadTerminate(), or the kernel thread kill shell command. Both k_thread_create() and k_thread_abort() are system calls, so with CONFIG_USERSPACE an unprivileged user-mode thread that has been granted a thread object and a user-capable stack object can drive the leak from user mode, consuming a kernel-side resource it cannot otherwise touch.
With CONFIG_KERNEL_VM_SIZE defaulting to 8 MB, a few hundred create/abort cycles exhaust the runtime mapping region. k_mem_map_phys_guard() then returns NULL, and map_thread_stack() guards that only with __ASSERT_NO_MSG(), which is compiled out of production builds; thread creation proceeds to write through a NULL-derived stack base and faults in kernel mode. The impact is a persistent denial of service — a kernel fatal error requiring a reboot — with no memory disclosure or corruption, and it is limited to configurations where memory-mapped thread stacks are enabled (x86 targets with X86_MMU and demand paging disabled, where the option is implied by default). |