tty: n_gsm: fix gsm_queue() UAF and add a base regression test
Subsystems (3)
v3 cover tty: n_gsm: fix gsm_queue() UAF and add a base regression test
The receive worker walks gsm->dlci[] without gsm->mutex while a concurrent GSMIOC_SETCONF -> gsm_cleanup_mux() frees the DLCIs, so the control handlers can dereference a freed gsm_dlci. v1's NULL check only narrowed the window; this series fixes the use-after-free itself by pinning each DLCI the dispatch dereferences with its existing tty_port reference (dlci_get/put), so the data path stays lock-free.
v3:
- Drop the cmpxchg() in gsm_dlci_free(); the pin alone fixes the UAF (Greg).
- Keep the addr bounds check at each call site, not in the helper (Daniel).
- Correct the Fixes: tag to the refcount commit (Daniel).
- Add the AI-assist tag to patch 1 and trim the comments (Daniel).
Verification (KASAN, panic_on_warn=1): a reproducer targeting the MSC handler crashes the unpatched kernel and survives 200+ race rounds on this series. The selftest passes on both the clean and patched kernel (pass:3 fail:0 skip:0).
Weiming Shi (2):
tty: n_gsm: fix use-after-free in gsm_queue() control frame dispatch
selftests: tty: add base regression test for n_gsm line discipline
drivers/tty/n_gsm.c | 73 +++- tools/testing/selftests/tty/.gitignore | 1 + tools/testing/selftests/tty/Makefile | 2 +- tools/testing/selftests/tty/config | 1 + tools/testing/selftests/tty/tty_n_gsm_test.c | 344 +++++++++++++++++++ 5 files changed, 414 insertions(+), 7 deletions(-) create mode 100644 tools/testing/selftests/tty/tty_n_gsm_test.c
Signature · 1 line
2.43.0
PATCH v3 1/2 tty: n_gsm: fix use-after-free in gsm_queue() control frame dispatch
The receive worker (flush_to_ldisc -> gsmld_receive_buf -> gsm0_receive/ gsm1_receive -> gsm_queue) reads gsm->dlci[address] and dispatches the frame via dlci->data() without holding gsm->mutex. The control handlers reached through dlci->data() then re-read gsm->dlci[]: gsm_control_reply() re-reads gsm->dlci[0], while gsm_control_modem() (MSC), gsm_control_rls() (RLS) and gsm_control_negotiation() (PN) re-read gsm->dlci[addr] for the DLCI named in the command - a different channel from the one the frame was addressed to.
Concurrently GSMIOC_SETCONF -> gsm_config() -> gsm_cleanup_mux() takes gsm->mutex and releases every DLCI via gsm_dlci_release() -> dlci_put(). When the last reference is dropped the destructor gsm_dlci_free() clears gsm->dlci[addr] and frees the object. If the worker dereferences one of those DLCIs while it is being freed, it touches freed memory.
A peer that drives DLCI 0 control frames (e.g. CMD_TEST) while the mux owner reconfigures the line discipline with GSMIOC_SETCONF can therefore trigger a use-after-free:
BUG: KASAN: slab-use-after-free in gsm_control_reply.isra.0 Read of size 8 at addr ffff888029ae9000 by task kworker/u16:2/46 Workqueue: events_unbound flush_to_ldisc Call Trace: gsm_control_reply.isra.0 (drivers/tty/n_gsm.c:1494) gsm_dlci_command (drivers/tty/n_gsm.c:2477) gsmld_receive_buf (drivers/tty/n_gsm.c:3616) tty_ldisc_receive_buf (drivers/tty/tty_buffer.c:398) tty_port_default_receive_buf (drivers/tty/tty_port.c:37) flush_to_ldisc (drivers/tty/tty_buffer.c:502) process_one_work worker_thread kthread Freed by task 5110: kfree gsm_cleanup_mux (drivers/tty/n_gsm.c:3161) gsmld_ioctl (drivers/tty/n_gsm.c:3415) tty_ioctl
Pin each DLCI across the dereference with its existing tty_port reference. gsm_dlci_open_get() looks gsm->dlci[addr] up under gsm->mutex and, if present, takes a dlci_get() reference before dropping the mutex; the caller releases it with gsm_dlci_unget() once it is done. While the reference is held the kref cannot reach zero, so gsm_dlci_free() cannot run and the object stays live. gsm_queue() pins the addressed DLCI for the UI/UIH dispatch, and gsm_control_modem(), gsm_control_rls() and gsm_control_negotiation() each pin the DLCI they operate on; the addr range check stays at the call site so a malformed frame cannot index gsm->dlci[] out of bounds.
The reference is taken only under the mutex, around the lookup; the mutex is released before dlci->data() and before the data-path work (gsm_process_modem(), tty_flip_buffer_push(), gsm_data_queue(), ...), so the receive/transmit path is not serialised by gsm->mutex and its timing is unaffected.
Attaching the n_gsm line discipline requires CAP_NET_ADMIN (gsmld_open() uses capable(), not ns_capable()), so this is a local denial of service for a privileged mux owner whose control channel is driven by an untrusted peer on the serial link while it reconfigures; harden the receive path regardless.
--- drivers/tty/n_gsm.c | 73 +++++++++++++++++++++++++++++++++++++++++---- 1 file changed, 67 insertions(+), 6 deletions(-)
drivers/tty/n_gsm.c +67 -6
@@ -453,6 +453,8 @@ static const u8 gsm_fcs8[256] = {#define GOOD_FCS 0xCFstatic void gsm_dlci_close (struct gsm_dlci *dlci); +static struct gsm_dlci *gsm_dlci_open_get (struct gsm_mux *gsm,unsigned int addr); +static void gsm_dlci_unget (struct gsm_dlci *dlci);static int gsmld_output (struct gsm_mux *gsm,u8 *data,int len);static int gsm_modem_update (struct gsm_dlci *dlci,u8 brk); static struct gsm_msg *gsm_data_alloc(struct gsm_mux *gsm, u8 addr, int len,
@@ -1695,9 +1697,8 @@ static void gsm_control_modem(struct gsm_mux *gsm, const u8 *data, int clen) addr >>=1 ;/* Closed port, or invalid ? */ -if (addr ==0 || addr >=NUM_DLCI || gsm->dlci[addr] ==NULL ) +if (addr ==0 || addr >=NUM_DLCI )return ; - dlci = gsm->dlci[addr];/* Must be at least one byte following the EA */ if ((cl - len) <1 )
@@ -1711,12 +1712,18 @@ static void gsm_control_modem(struct gsm_mux *gsm, const u8 *data, int clen)if (len <1 )return ; +/* Hold the DLCI until we are done; see gsm_dlci_open_get(). */ + dlci =gsm_dlci_open_get (gsm, addr); +if (dlci ==NULL ) +return ; + tty =tty_port_tty_get (&dlci->port);gsm_process_modem (tty, dlci, modem, cl);if (tty) {tty_wakeup (tty);tty_kref_put (tty); } +gsm_dlci_unget (dlci);gsm_control_reply (gsm,CMD_MSC , data, clen); }
@@ -1746,15 +1753,22 @@ static void gsm_control_negotiation(struct gsm_mux *gsm, unsigned int cr,/* Invalid DLCI? */ params = (struct gsm_dlci_param_bits *)data; addr =FIELD_GET (PN_D_FIELD_DLCI , params->d_bits); -if (addr ==0 || addr >=NUM_DLCI || !gsm->dlci[addr]) { +if (addr ==0 || addr >=NUM_DLCI ) { + gsm->open_error++; +return ; + } + +/* Hold the DLCI until we are done; see gsm_dlci_open_get(). */ + dlci =gsm_dlci_open_get (gsm, addr); +if (dlci ==NULL ) { gsm->open_error++;return ; } - dlci = gsm->dlci[addr];/* Too late for parameter negotiation? */ if ((!cr && dlci->state ==DLCI_OPENING ) || dlci->state ==DLCI_OPEN ) { gsm->open_error++; +gsm_dlci_unget (dlci);return ; }
@@ -1765,6 +1779,7 @@ static void gsm_control_negotiation(struct gsm_mux *gsm, unsigned int cr,pr_info ("%s PN failed\n" , __func__); gsm->open_error++;gsm_dlci_close (dlci); +gsm_dlci_unget (dlci);return ; }
@@ -1785,6 +1800,7 @@ static void gsm_control_negotiation(struct gsm_mux *gsm, unsigned int cr,pr_info ("%s PN in invalid state\n" , __func__); gsm->open_error++; } +gsm_dlci_unget (dlci); }/**
@@ -1800,6 +1816,7 @@ static void gsm_control_negotiation(struct gsm_mux *gsm, unsigned int cr,static void gsm_control_rls (struct gsm_mux *gsm,const u8 *data,int clen) { +struct gsm_dlci *dlci;struct tty_port *port;unsigned int addr =0 ;u8 bits;
@@ -1817,14 +1834,18 @@ static void gsm_control_rls(struct gsm_mux *gsm, const u8 *data, int clen)return ; addr >>=1 ;/* Closed port, or invalid ? */ -if (addr ==0 || addr >=NUM_DLCI || gsm->dlci[addr] ==NULL ) +if (addr ==0 || addr >=NUM_DLCI )return ;/* No error ? */ bits = *dp;if ((bits &1 ) ==0 )return ; - port = &gsm->dlci[addr]->port; +/* Hold the DLCI until we are done; see gsm_dlci_open_get(). */ + dlci =gsm_dlci_open_get (gsm, addr); +if (dlci ==NULL ) +return ; + port = &dlci->port;if (bits &2 )tty_insert_flip_char (port,0 ,TTY_OVERRUN );
@@ -1835,6 +1856,7 @@ static void gsm_control_rls(struct gsm_mux *gsm, const u8 *data, int clen)tty_flip_buffer_push (port); +gsm_dlci_unget (dlci);gsm_control_reply (gsm,CMD_RLS , data, clen); }
@@ -2711,6 +2733,35 @@ static inline void dlci_put(struct gsm_dlci *dlci)tty_port_put (&dlci->port); } +/** +* gsm_dlci_open_get - look up a DLCI and take a reference +* @gsm: GSM mux +* @addr: DLCI address +* +* Look up gsm->dlci[addr] under gsm->mutex and take a reference. Returns +* NULL if not present. Release with gsm_dlci_unget(). +*/ +static struct gsm_dlci *gsm_dlci_open_get (struct gsm_mux *gsm,unsigned int addr) +{ +struct gsm_dlci *dlci; + +mutex_lock (&gsm->mutex); + dlci = gsm->dlci[addr]; +if (dlci !=NULL ) +dlci_get (dlci); +mutex_unlock (&gsm->mutex); +return dlci; +} + +/** +* gsm_dlci_unget - drop a reference from gsm_dlci_open_get() +* @dlci: DLCI to release +*/ +static void gsm_dlci_unget (struct gsm_dlci *dlci) +{ +dlci_put (dlci); +} +static void gsm_destroy_network (struct gsm_dlci *dlci);/**
@@ -2839,11 +2890,21 @@ static void gsm_queue(struct gsm_mux *gsm)case UI |PF :case UIH :case UIH |PF : +/* +* Pin the DLCI so gsm_cleanup_mux() cannot free it during +* dispatch. The mutex is dropped before dlci->data(). +*/ +mutex_lock (&gsm->mutex); + dlci = gsm->dlci[address];if (dlci ==NULL || dlci->state !=DLCI_OPEN ) { +mutex_unlock (&gsm->mutex);gsm_response (gsm, address,DM |PF );return ; } +dlci_get (dlci); +mutex_unlock (&gsm->mutex); dlci->data (dlci, gsm->buf, gsm->len); +dlci_put (dlci);break ;default : goto invalid;
Signature · 1 line
2.43.0
PATCH v3 2/2 selftests: tty: add base regression test for n_gsm line discipline
In reply to Weiming Shin_gsm has no selftest coverage. Add a base functional regression test that drives the line discipline over a local pty pair, so no real serial hardware or modem is required, and exercises the GSM 07.10 / 3GPP TS 27.010 basic-option mux from userspace.
The test attaches N_GSM to the pty master, then:
- basic: brings up the mux (SETCONF, initiator side), drives the
DLCI 0 control channel SABM/UA handshake and tears it down.
- getconf: round-trips GSMIOC_GETCONF/GSMIOC_SETCONF and checks the
configuration is preserved.
- data_dlci: opens a data DLCI (DLCI 1) via the SABM/UA exchange and verifies the responder side answers, covering the control -> data DLCI path.
Frames are encoded by hand against 3GPP TS 27.010 (address EA/C-R/DLCI bits, SABM/UA/UIH control fields, the reversed CRC-8 FCS) with the clause numbers referenced in the comments, so the test doubles as a small, readable description of the on-wire format.
It is a functional/regression test, not a race reproducer: it gives the subsystem a green baseline to catch behavioural regressions, including in the gsm_queue() control-frame dispatch path.
Wire it into the tty selftest Makefile, add CONFIG_N_GSM=y to the config fragment, and ignore the built binary. The test SKIPs cleanly when N_GSM is not built, /dev/ptmx is missing, or it lacks the capability to attach the ldisc.
--- tools/testing/selftests/tty/.gitignore | 1 + tools/testing/selftests/tty/Makefile | 2 +- tools/testing/selftests/tty/config | 1 + tools/testing/selftests/tty/tty_n_gsm_test.c | 344 +++++++++++++++++++ 4 files changed, 347 insertions(+), 1 deletion(-) create mode 100644 tools/testing/selftests/tty/tty_n_gsm_test.c
tools/testing/selftests/tty/.gitignore +1 -0
@@ -1,3 +1,4 @@ # SPDX-License-Identifier: GPL-2.0-only tty_tiocsti_test tty_tstamp_update +tty_n_gsm_test
tools/testing/selftests/tty/Makefile +1 -1
@@ -1,6 +1,6 @@ # SPDX-License-Identifier: GPL-2.0 CFLAGS = -O2 -Wall -TEST_GEN_PROGS := tty_tstamp_update tty_tiocsti_test +TEST_GEN_PROGS := tty_tstamp_update tty_tiocsti_test tty_n_gsm_test LDLIBS += -lcap include ../lib.mk
tools/testing/selftests/tty/config +1 -0
@@ -1 +1,2 @@ CONFIG_LEGACY_TIOCSTI=y +CONFIG_N_GSM=y
tools/testing/selftests/tty/tty_n_gsm_test.c new file +344 -0
344 lines not on this page · raw message
Signature · 1 line
2.43.0