[BUG] tracing: trace_pipe reads out of bounds when it races with the ring buffer mmap reader

From: Kyuwon Cho

Date: Mon Nov 16 2026 - 12:56:37 EST


Hi Steve, Masami, Vincent,

While testing v7.4-rc3 with an LLM-guided test generator that we are
developing as a research prototype, we found that reading a CPU's
trace_pipe while another task consumes the same CPU buffer through the
mmap interface (TRACE_MMAP_IOCTL_GET_READER) makes trace_pipe print
entries that have already been overwritten. With KASAN this ends in an
out-of-bounds read in the event output code:

BUG: KASAN: use-after-free in string (lib/vsprintf.c:650 lib/vsprintf.c:732)
Read of size 1 at addr ff11000011e2d8ed by task test/89
Call Trace:
<TASK>
string (lib/vsprintf.c:650 lib/vsprintf.c:732)
vsnprintf (lib/vsprintf.c:2948)
seq_buf_vprintf (lib/seq_buf.c:74)
trace_seq_vprintf (kernel/trace/trace_seq.c:175)
trace_event_printf (kernel/trace/trace_output.c:360)
trace_raw_output_workqueue_rescued (include/trace/events/workqueue.h:210)
print_trace_line (kernel/trace/trace.c:3082 kernel/trace/trace.c:3248)
tracing_read_pipe (kernel/trace/trace.c:5475)
vfs_read (fs/read_write.c:552)
ksys_read (fs/read_write.c:696)
do_syscall_64 (arch/x86/entry/syscall_64.c:61 arch/x86/entry/syscall_64.c:89)
entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:121)
</TASK>
The buggy address belongs to the physical page:
page: refcount:0 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x11e2d
page_type: f0(buddy)

and the type check in trace_assign_type() fires as well:

WARNING: kernel/trace/trace_output.c:1798 at trace_print_print+0xeb/0x130, CPU#1: test/89
Call Trace:
<TASK>
print_trace_line (kernel/trace/trace.c:3082 kernel/trace/trace.c:3248)
tracing_read_pipe (kernel/trace/trace.c:5475)
vfs_read (fs/read_write.c:552)
ksys_read (fs/read_write.c:696)
do_syscall_64 (arch/x86/entry/syscall_64.c:61 arch/x86/entry/syscall_64.c:89)
entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:121)
</TASK>

The stale entry carries another event's type, so that event's output
function decodes trace_marker data as its fields, including __data_loc
offsets and lengths that point outside the ring buffer page. Other runs
reported slab-out-of-bounds and global-out-of-bounds reads from the same
path; the ring buffer pages themselves are not freed.

tracing_buffers_ioctl() calls ring_buffer_map_get_reader() without
trace_access_lock():

if (cmd == TRACE_MMAP_IOCTL_GET_READER) {
...
return ring_buffer_map_get_reader(iter->array_buffer->buffer,
iter->cpu_file);
}

The other consuming readers of a CPU buffer, tracing_read_pipe(),
tracing_buffers_read() and tracing_buffers_splice_read(), all take
trace_access_lock(iter->cpu_file). tracing_read_pipe() peeks an event,
prints it with reader_lock dropped and only then consumes it, so a
GET_READER in between can hand the reader page back to the writers. This
has been the case since the ioctl was added in cf9f0f7c4c5b ("tracing:
Allow user-space mapping of the ring-buffer").

Taking trace_access_lock(iter->cpu_file) around
ring_buffer_map_get_reader() in tracing_buffers_ioctl(), as the other
readers do, makes the problem go away here: the reproducer below ran
three times without a report, and the ring-buffer map_test selftest
still passes.

The reproducer needs root, for tracefs. One thread writes to trace_marker
on CPU 0; on CPU 1, one thread loops on TRACE_MMAP_IOCTL_GET_READER on
the mmapped per_cpu/cpu0/trace_pipe_raw and another reads
per_cpu/cpu0/trace_pipe. On an x86_64 QEMU guest with 2 vCPUs, KASAN
and lockdep (gcc 13.3, config attached), the WARNING appears within
seconds in every run and KASAN reports in 2 of 3 runs. Neither reader
alone triggers it.

Checked against v7.4-rc3 (aebb72744764); tracing_buffers_ioctl() is the
same in trace/for-next. We found no earlier report on syzbot or on
linux-trace-kernel.

// Run as root with tracefs available and at least 2 CPUs.
// CPU 0: write to trace_marker. CPU 1: one thread polls
// TRACE_MMAP_IOCTL_GET_READER on the mmapped per_cpu/cpu0/trace_pipe_raw,
// another reads per_cpu/cpu0/trace_pipe.
#define _GNU_SOURCE
#include <fcntl.h>
#include <pthread.h>
#include <sched.h>
#include <stdio.h>
#include <string.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/mount.h>
#include <sys/stat.h>
#include <unistd.h>

#define T "/sys/kernel/tracing"
#define TRACE_MMAP_IOCTL_GET_READER _IO('R', 0x20)

static volatile int stop;

static void pin(int cpu)
{
cpu_set_t set;

CPU_ZERO(&set);
CPU_SET(cpu, &set);
sched_setaffinity(0, sizeof(set), &set);
}

static void put(const char *file, const char *val)
{
int fd = open(file, O_WRONLY | O_TRUNC);

if (fd >= 0) {
write(fd, val, strlen(val));
close(fd);
}
}

static void *writer(void *arg)
{
char buf[1000];
int fd = open(T "/trace_marker", O_WRONLY);

pin(0);
memset(buf, 'a', sizeof(buf));
for (unsigned long i = 0; !stop; i++)
write(fd, buf, 1 + i % sizeof(buf));
return NULL;
}

static void *mapper(void *arg)
{
int fd = open(T "/per_cpu/cpu0/trace_pipe_raw", O_RDONLY | O_NONBLOCK);

pin(1);
if (mmap(NULL, getpagesize(), PROT_READ, MAP_SHARED, fd, 0) == MAP_FAILED) {
perror("mmap");
return NULL;
}
while (!stop)
ioctl(fd, TRACE_MMAP_IOCTL_GET_READER);
return NULL;
}

static void *cooked(void *arg)
{
char buf[4096];
int fd = open(T "/per_cpu/cpu0/trace_pipe", O_RDONLY | O_NONBLOCK);

pin(1);
while (!stop)
read(fd, buf, sizeof(buf));
return NULL;
}

int main(void)
{
pthread_t th[3];

mount("tracefs", T, "tracefs", 0, NULL);
put(T "/buffer_size_kb", "16");
put(T "/tracing_on", "1");
pthread_create(&th[0], NULL, writer, NULL);
pthread_create(&th[1], NULL, mapper, NULL);
pthread_create(&th[2], NULL, cooked, NULL);
sleep(60);
stop = 1;
for (int i = 0; i < 3; i++)
pthread_join(th[i], NULL);
return 0;
}

Thanks,
Kyuwon Cho

Attachment: config.gz
Description: application/gzip