1===========================================================================
  2Proper Locking Under a Preemptible Kernel: Keeping Kernel Code Preempt-Safe
  3===========================================================================
  4
  5:Author: Robert Love <rml@tech9.net>
  6:Last Updated: 28 Aug 2002
  7
  8
  9Introduction
 10============
 11
 12
 13A preemptible kernel creates new locking issues.  The issues are the same as
 14those under SMP: concurrency and reentrancy.  Thankfully, the Linux preemptible
 15kernel model leverages existing SMP locking mechanisms.  Thus, the kernel
 16requires explicit additional locking for very few additional situations.
 17
 18This document is for all kernel hackers.  Developing code in the kernel
 19requires protecting these situations.
 20 
 21
 22RULE #1: Per-CPU data structures need explicit protection
 23^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 24
 25
 26Two similar problems arise. An example code snippet::
 27
 28	struct this_needs_locking tux[NR_CPUS];
 29	tux[smp_processor_id()] = some_value;
 30	/* task is preempted here... */
 31	something = tux[smp_processor_id()];
 32
 33First, since the data is per-CPU, it may not have explicit SMP locking, but
 34require it otherwise.  Second, when a preempted task is finally rescheduled,
 35the previous value of smp_processor_id may not equal the current.  You must
 36protect these situations by disabling preemption around them.
 37
 38You can also use put_cpu() and get_cpu(), which will disable preemption.
 39
 40
 41RULE #2: CPU state must be protected.
 42^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 43
 44
 45Under preemption, the state of the CPU must be protected.  This is arch-
 46dependent, but includes CPU structures and state not preserved over a context
 47switch.  For example, on x86, entering and exiting FPU mode is now a critical
 48section that must occur while preemption is disabled.  Think what would happen
 49if the kernel is executing a floating-point instruction and is then preempted.
 50Remember, the kernel does not save FPU state except for user tasks.  Therefore,
 51upon preemption, the FPU registers will be sold to the lowest bidder.  Thus,
 52preemption must be disabled around such regions.
 53
 54Note, some FPU functions are already explicitly preempt safe.  For example,
 55kernel_fpu_begin and kernel_fpu_end will disable and enable preemption.
 56However, fpu__restore() must be called with preemption disabled.
 57
 58
 59RULE #3: Lock acquire and release must be performed by same task
 60^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 61
 62
 63A lock acquired in one task must be released by the same task.  This
 64means you can't do oddball things like acquire a lock and go off to
 65play while another task releases it.  If you want to do something
 66like this, acquire and release the task in the same code path and
 67have the caller wait on an event by the other task.
 68
 69
 70Solution
 71========
 72
 73
 74Data protection under preemption is achieved by disabling preemption for the
 75duration of the critical region.
 76
 77::
 78
 79  preempt_enable()		decrement the preempt counter
 80  preempt_disable()		increment the preempt counter
 81  preempt_enable_no_resched()	decrement, but do not immediately preempt
 82  preempt_check_resched()	if needed, reschedule
 83  preempt_count()		return the preempt counter
 84
 85The functions are nestable.  In other words, you can call preempt_disable
 86n-times in a code path, and preemption will not be reenabled until the n-th
 87call to preempt_enable.  The preempt statements define to nothing if
 88preemption is not enabled.
 89
 90Note that you do not need to explicitly prevent preemption if you are holding
 91any locks or interrupts are disabled, since preemption is implicitly disabled
 92in those cases.
 93
 94But keep in mind that 'irqs disabled' is a fundamentally unsafe way of
 95disabling preemption - any spin_unlock() decreasing the preemption count
 96to 0 might trigger a reschedule. A simple printk() might trigger a reschedule.
 97So use this implicit preemption-disabling property only if you know that the
 98affected codepath does not do any of this. Best policy is to use this only for
 99small, atomic code that you wrote and which calls no complex functions.
100
101Example::
102
103	cpucache_t *cc; /* this is per-CPU */
104	preempt_disable();
105	cc = cc_data(searchp);
106	if (cc && cc->avail) {
107		__free_block(searchp, cc_entry(cc), cc->avail);
108		cc->avail = 0;
109	}
110	preempt_enable();
111	return 0;
112
113Notice how the preemption statements must encompass every reference of the
114critical variables.  Another example::
115
116	int buf[NR_CPUS];
117	set_cpu_val(buf);
118	if (buf[smp_processor_id()] == -1) printf(KERN_INFO "wee!\n");
119	spin_lock(&buf_lock);
120	/* ... */
121
122This code is not preempt-safe, but see how easily we can fix it by simply
123moving the spin_lock up two lines.
124
125
126Preventing preemption using interrupt disabling
127===============================================
128
129
130It is possible to prevent a preemption event using local_irq_disable and
131local_irq_save.  Note, when doing so, you must be very careful to not cause
132an event that would set need_resched and result in a preemption check.  When
133in doubt, rely on locking or explicit preemption disabling.
134
135Note in 2.5 interrupt disabling is now only per-CPU (e.g. local).
136
137An additional concern is proper usage of local_irq_disable and local_irq_save.
138These may be used to protect from preemption, however, on exit, if preemption
139may be enabled, a test to see if preemption is required should be done.  If
140these are called from the spin_lock and read/write lock macros, the right thing
141is done.  They may also be called within a spin-lock protected region, however,
142if they are ever called outside of this context, a test for preemption should
143be made. Do note that calls from interrupt context or bottom half/ tasklets
144are also protected by preemption locks and so may use the versions which do
145not check preemption.