projects / linux-block.git / commitdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: a5806cd)
ring-buffer: Move cpus_read_lock() outside of buffer->mutex
authorSteven Rostedt <rostedt@goodmis.org>
Tue, 27 May 2025 14:58:20 +0000 (10:58 -0400)
committerSteven Rostedt (Google) <rostedt@goodmis.org>
Tue, 27 May 2025 19:46:39 +0000 (15:46 -0400)
Running a modified trace-cmd record --nosplice where it does a mmap of the
ring buffer when '--nosplice' is set, caused the following lockdep splat:

 ======================================================
 WARNING: possible circular locking dependency detected
 6.15.0-rc7-test-00002-gfb7d03d8a82f #551 Not tainted
 ------------------------------------------------------
 trace-cmd/1113 is trying to acquire lock:
 ffff888100062888 (&buffer->mutex){+.+.}-{4:4}, at: ring_buffer_map+0x11c/0xe70

 but task is already holding lock:
 ffff888100a5f9f8 (&cpu_buffer->mapping_lock){+.+.}-{4:4}, at: ring_buffer_map+0xcf/0xe70

 which lock already depends on the new lock.

 the existing dependency chain (in reverse order) is:

 -> #5 (&cpu_buffer->mapping_lock){+.+.}-{4:4}:
        __mutex_lock+0x192/0x18c0
        ring_buffer_map+0xcf/0xe70
        tracing_buffers_mmap+0x1c4/0x3b0
        __mmap_region+0xd8d/0x1f70
        do_mmap+0x9d7/0x1010
        vm_mmap_pgoff+0x20b/0x390
        ksys_mmap_pgoff+0x2e9/0x440
        do_syscall_64+0x79/0x1c0
        entry_SYSCALL_64_after_hwframe+0x76/0x7e

 -> #4 (&mm->mmap_lock){++++}-{4:4}:
        __might_fault+0xa5/0x110
        _copy_to_user+0x22/0x80
        _perf_ioctl+0x61b/0x1b70
        perf_ioctl+0x62/0x90
        __x64_sys_ioctl+0x134/0x190
        do_syscall_64+0x79/0x1c0
        entry_SYSCALL_64_after_hwframe+0x76/0x7e

 -> #3 (&cpuctx_mutex){+.+.}-{4:4}:
        __mutex_lock+0x192/0x18c0
        perf_event_init_cpu+0x325/0x7c0
        perf_event_init+0x52a/0x5b0
        start_kernel+0x263/0x3e0
        x86_64_start_reservations+0x24/0x30
        x86_64_start_kernel+0x95/0xa0
        common_startup_64+0x13e/0x141

 -> #2 (pmus_lock){+.+.}-{4:4}:
        __mutex_lock+0x192/0x18c0
        perf_event_init_cpu+0xb7/0x7c0
        cpuhp_invoke_callback+0x2c0/0x1030
        __cpuhp_invoke_callback_range+0xbf/0x1f0
        _cpu_up+0x2e7/0x690
        cpu_up+0x117/0x170
        cpuhp_bringup_mask+0xd5/0x120
        bringup_nonboot_cpus+0x13d/0x170
        smp_init+0x2b/0xf0
        kernel_init_freeable+0x441/0x6d0
        kernel_init+0x1e/0x160
        ret_from_fork+0x34/0x70
        ret_from_fork_asm+0x1a/0x30

 -> #1 (cpu_hotplug_lock){++++}-{0:0}:
        cpus_read_lock+0x2a/0xd0
        ring_buffer_resize+0x610/0x14e0
        __tracing_resize_ring_buffer.part.0+0x42/0x120
        tracing_set_tracer+0x7bd/0xa80
        tracing_set_trace_write+0x132/0x1e0
        vfs_write+0x21c/0xe80
        ksys_write+0xf9/0x1c0
        do_syscall_64+0x79/0x1c0
        entry_SYSCALL_64_after_hwframe+0x76/0x7e

 -> #0 (&buffer->mutex){+.+.}-{4:4}:
        __lock_acquire+0x1405/0x2210
        lock_acquire+0x174/0x310
        __mutex_lock+0x192/0x18c0
        ring_buffer_map+0x11c/0xe70
        tracing_buffers_mmap+0x1c4/0x3b0
        __mmap_region+0xd8d/0x1f70
        do_mmap+0x9d7/0x1010
        vm_mmap_pgoff+0x20b/0x390
        ksys_mmap_pgoff+0x2e9/0x440
        do_syscall_64+0x79/0x1c0
        entry_SYSCALL_64_after_hwframe+0x76/0x7e

 other info that might help us debug this:

 Chain exists of:
   &buffer->mutex --> &mm->mmap_lock --> &cpu_buffer->mapping_lock

  Possible unsafe locking scenario:

        CPU0                    CPU1
        ----                    ----
   lock(&cpu_buffer->mapping_lock);
                                lock(&mm->mmap_lock);
                                lock(&cpu_buffer->mapping_lock);
   lock(&buffer->mutex);

  *** DEADLOCK ***

 2 locks held by trace-cmd/1113:
  #0: ffff888106b847e0 (&mm->mmap_lock){++++}-{4:4}, at: vm_mmap_pgoff+0x192/0x390
  #1: ffff888100a5f9f8 (&cpu_buffer->mapping_lock){+.+.}-{4:4}, at: ring_buffer_map+0xcf/0xe70

 stack backtrace:
 CPU: 5 UID: 0 PID: 1113 Comm: trace-cmd Not tainted 6.15.0-rc7-test-00002-gfb7d03d8a82f #551 PREEMPT
 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
 Call Trace:
  <TASK>
  dump_stack_lvl+0x6e/0xa0
  print_circular_bug.cold+0x178/0x1be
  check_noncircular+0x146/0x160
  __lock_acquire+0x1405/0x2210
  lock_acquire+0x174/0x310
  ? ring_buffer_map+0x11c/0xe70
  ? ring_buffer_map+0x11c/0xe70
  ? __mutex_lock+0x169/0x18c0
  __mutex_lock+0x192/0x18c0
  ? ring_buffer_map+0x11c/0xe70
  ? ring_buffer_map+0x11c/0xe70
  ? function_trace_call+0x296/0x370
  ? __pfx___mutex_lock+0x10/0x10
  ? __pfx_function_trace_call+0x10/0x10
  ? __pfx___mutex_lock+0x10/0x10
  ? _raw_spin_unlock+0x2d/0x50
  ? ring_buffer_map+0x11c/0xe70
  ? ring_buffer_map+0x11c/0xe70
  ? __mutex_lock+0x5/0x18c0
  ring_buffer_map+0x11c/0xe70
  ? do_raw_spin_lock+0x12d/0x270
  ? find_held_lock+0x2b/0x80
  ? _raw_spin_unlock+0x2d/0x50
  ? rcu_is_watching+0x15/0xb0
  ? _raw_spin_unlock+0x2d/0x50
  ? trace_preempt_on+0xd0/0x110
  tracing_buffers_mmap+0x1c4/0x3b0
  __mmap_region+0xd8d/0x1f70
  ? ring_buffer_lock_reserve+0x99/0xff0
  ? __pfx___mmap_region+0x10/0x10
  ? ring_buffer_lock_reserve+0x99/0xff0
  ? __pfx_ring_buffer_lock_reserve+0x10/0x10
  ? __pfx_ring_buffer_lock_reserve+0x10/0x10
  ? bpf_lsm_mmap_addr+0x4/0x10
  ? security_mmap_addr+0x46/0xd0
  ? lock_is_held_type+0xd9/0x130
  do_mmap+0x9d7/0x1010
  ? 0xffffffffc0370095
  ? __pfx_do_mmap+0x10/0x10
  vm_mmap_pgoff+0x20b/0x390
  ? __pfx_vm_mmap_pgoff+0x10/0x10
  ? 0xffffffffc0370095
  ksys_mmap_pgoff+0x2e9/0x440
  do_syscall_64+0x79/0x1c0
  entry_SYSCALL_64_after_hwframe+0x76/0x7e
 RIP: 0033:0x7fb0963a7de2
 Code: 00 00 00 0f 1f 44 00 00 41 f7 c1 ff 0f 00 00 75 27 55 89 cd 53 48 89 fb 48 85 ff 74 3b 41 89 ea 48 89 df b8 09 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 76 5b 5d c3 0f 1f 00 48 8b 05 e1 9f 0d 00 64
 RSP: 002b:00007ffdcc8fb878 EFLAGS: 00000246 ORIG_RAX: 0000000000000009
 RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007fb0963a7de2
 RDX: 0000000000000001 RSI: 0000000000001000 RDI: 0000000000000000
 RBP: 0000000000000001 R08: 0000000000000006 R09: 0000000000000000
 R10: 0000000000000001 R11: 0000000000000246 R12: 0000000000000000
 R13: 00007ffdcc8fbe68 R14: 00007fb096628000 R15: 00005633e01a5c90
  </TASK>

The issue is that cpus_read_lock() is taken within buffer->mutex. The
memory mapped pages are taken with the mmap_lock held. The buffer->mutex
is taken within the cpu_buffer->mapping_lock. There's quite a chain with
all these locks, where the deadlock can be fixed by moving the
cpus_read_lock() outside the taking of the buffer->mutex.

Cc: stable@vger.kernel.org
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Vincent Donnefort <vdonnefort@google.com>
Link: https://lore.kernel.org/20250527105820.0f45d045@gandalf.local.home
Fixes: 117c39200d9d7 ("ring-buffer: Introducing ring-buffer mapping functions")
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
kernel/trace/ring_buffer.c patch | blob | blame | history

diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c
index 3f9bf562beea23eea828d09bd260d88857ef107a..ca1a8e70600487bded8114d4ed4b520eea5e0cfa 100644 (file)
--- a/kernel/trace/ring_buffer.c
+++ b/kernel/trace/ring_buffer.c
@@ -2849,6 +2849,12 @@ int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
        if (nr_pages < 2)
                nr_pages = 2;
 
+       /*
+        * Keep CPUs from coming online while resizing to synchronize
+        * with new per CPU buffers being created.
+        */
+       guard(cpus_read_lock)();
+
        /* prevent another thread from changing buffer sizes */
        mutex_lock(&buffer->mutex);
        atomic_inc(&buffer->resizing);
@@ -2893,7 +2899,6 @@ int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
                        cond_resched();
                }
 
-               cpus_read_lock();
                /*
                 * Fire off all the required work handlers
                 * We can't schedule on offline CPUs, but it's not necessary
@@ -2933,7 +2938,6 @@ int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
                        cpu_buffer->nr_pages_to_update = 0;
                }
 
-               cpus_read_unlock();
        } else {
                cpu_buffer = buffer->buffers[cpu_id];
 
@@ -2961,8 +2965,6 @@ int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
                        goto out_err;
                }
 
-               cpus_read_lock();
-
                /* Can't run something on an offline CPU. */
                if (!cpu_online(cpu_id))
                        rb_update_pages(cpu_buffer);
@@ -2981,7 +2983,6 @@ int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
                }
 
                cpu_buffer->nr_pages_to_update = 0;
-               cpus_read_unlock();
        }
 
  out:
Linux 6.x block layer and io_uring tree(s)