lockdep-splat.rst 5.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115
  1. .. SPDX-License-Identifier: GPL-2.0
  2. =================
  3. Lockdep-RCU Splat
  4. =================
  5. Lockdep-RCU was added to the Linux kernel in early 2010
  6. (http://lwn.net/Articles/371986/). This facility checks for some common
  7. misuses of the RCU API, most notably using one of the rcu_dereference()
  8. family to access an RCU-protected pointer without the proper protection.
  9. When such misuse is detected, an lockdep-RCU splat is emitted.
  10. The usual cause of a lockdep-RCU slat is someone accessing an
  11. RCU-protected data structure without either (1) being in the right kind of
  12. RCU read-side critical section or (2) holding the right update-side lock.
  13. This problem can therefore be serious: it might result in random memory
  14. overwriting or worse. There can of course be false positives, this
  15. being the real world and all that.
  16. So let's look at an example RCU lockdep splat from 3.0-rc5, one that
  17. has long since been fixed::
  18. =============================
  19. WARNING: suspicious RCU usage
  20. -----------------------------
  21. block/cfq-iosched.c:2776 suspicious rcu_dereference_protected() usage!
  22. other info that might help us debug this::
  23. rcu_scheduler_active = 1, debug_locks = 0
  24. 3 locks held by scsi_scan_6/1552:
  25. #0: (&shost->scan_mutex){+.+.}, at: [<ffffffff8145efca>]
  26. scsi_scan_host_selected+0x5a/0x150
  27. #1: (&eq->sysfs_lock){+.+.}, at: [<ffffffff812a5032>]
  28. elevator_exit+0x22/0x60
  29. #2: (&(&q->__queue_lock)->rlock){-.-.}, at: [<ffffffff812b6233>]
  30. cfq_exit_queue+0x43/0x190
  31. stack backtrace:
  32. Pid: 1552, comm: scsi_scan_6 Not tainted 3.0.0-rc5 #17
  33. Call Trace:
  34. [<ffffffff810abb9b>] lockdep_rcu_dereference+0xbb/0xc0
  35. [<ffffffff812b6139>] __cfq_exit_single_io_context+0xe9/0x120
  36. [<ffffffff812b626c>] cfq_exit_queue+0x7c/0x190
  37. [<ffffffff812a5046>] elevator_exit+0x36/0x60
  38. [<ffffffff812a802a>] blk_cleanup_queue+0x4a/0x60
  39. [<ffffffff8145cc09>] scsi_free_queue+0x9/0x10
  40. [<ffffffff81460944>] __scsi_remove_device+0x84/0xd0
  41. [<ffffffff8145dca3>] scsi_probe_and_add_lun+0x353/0xb10
  42. [<ffffffff817da069>] ? error_exit+0x29/0xb0
  43. [<ffffffff817d98ed>] ? _raw_spin_unlock_irqrestore+0x3d/0x80
  44. [<ffffffff8145e722>] __scsi_scan_target+0x112/0x680
  45. [<ffffffff812c690d>] ? trace_hardirqs_off_thunk+0x3a/0x3c
  46. [<ffffffff817da069>] ? error_exit+0x29/0xb0
  47. [<ffffffff812bcc60>] ? kobject_del+0x40/0x40
  48. [<ffffffff8145ed16>] scsi_scan_channel+0x86/0xb0
  49. [<ffffffff8145f0b0>] scsi_scan_host_selected+0x140/0x150
  50. [<ffffffff8145f149>] do_scsi_scan_host+0x89/0x90
  51. [<ffffffff8145f170>] do_scan_async+0x20/0x160
  52. [<ffffffff8145f150>] ? do_scsi_scan_host+0x90/0x90
  53. [<ffffffff810975b6>] kthread+0xa6/0xb0
  54. [<ffffffff817db154>] kernel_thread_helper+0x4/0x10
  55. [<ffffffff81066430>] ? finish_task_switch+0x80/0x110
  56. [<ffffffff817d9c04>] ? retint_restore_args+0xe/0xe
  57. [<ffffffff81097510>] ? __kthread_init_worker+0x70/0x70
  58. [<ffffffff817db150>] ? gs_change+0xb/0xb
  59. Line 2776 of block/cfq-iosched.c in v3.0-rc5 is as follows::
  60. if (rcu_dereference(ioc->ioc_data) == cic) {
  61. This form says that it must be in a plain vanilla RCU read-side critical
  62. section, but the "other info" list above shows that this is not the
  63. case. Instead, we hold three locks, one of which might be RCU related.
  64. And maybe that lock really does protect this reference. If so, the fix
  65. is to inform RCU, perhaps by changing __cfq_exit_single_io_context() to
  66. take the struct request_queue "q" from cfq_exit_queue() as an argument,
  67. which would permit us to invoke rcu_dereference_protected as follows::
  68. if (rcu_dereference_protected(ioc->ioc_data,
  69. lockdep_is_held(&q->queue_lock)) == cic) {
  70. With this change, there would be no lockdep-RCU splat emitted if this
  71. code was invoked either from within an RCU read-side critical section
  72. or with the ->queue_lock held. In particular, this would have suppressed
  73. the above lockdep-RCU splat because ->queue_lock is held (see #2 in the
  74. list above).
  75. On the other hand, perhaps we really do need an RCU read-side critical
  76. section. In this case, the critical section must span the use of the
  77. return value from rcu_dereference(), or at least until there is some
  78. reference count incremented or some such. One way to handle this is to
  79. add rcu_read_lock() and rcu_read_unlock() as follows::
  80. rcu_read_lock();
  81. if (rcu_dereference(ioc->ioc_data) == cic) {
  82. spin_lock(&ioc->lock);
  83. rcu_assign_pointer(ioc->ioc_data, NULL);
  84. spin_unlock(&ioc->lock);
  85. }
  86. rcu_read_unlock();
  87. With this change, the rcu_dereference() is always within an RCU
  88. read-side critical section, which again would have suppressed the
  89. above lockdep-RCU splat.
  90. But in this particular case, we don't actually dereference the pointer
  91. returned from rcu_dereference(). Instead, that pointer is just compared
  92. to the cic pointer, which means that the rcu_dereference() can be replaced
  93. by rcu_access_pointer() as follows::
  94. if (rcu_access_pointer(ioc->ioc_data) == cic) {
  95. Because it is legal to invoke rcu_access_pointer() without protection,
  96. this change would also suppress the above lockdep-RCU splat.