1
2.. _volatile_considered_harmful:
3
4Why the "volatile" type class should not be used
5------------------------------------------------
6
7C programmers have often taken volatile to mean that the variable could be
8changed outside of the current thread of execution; as a result, they are
9sometimes tempted to use it in kernel code when shared data structures are
10being used.  In other words, they have been known to treat volatile types
11as a sort of easy atomic variable, which they are not.  The use of volatile in
12kernel code is almost never correct; this document describes why.
13
14The key point to understand with regard to volatile is that its purpose is
15to suppress optimization, which is almost never what one really wants to
16do.  In the kernel, one must protect shared data structures against
17unwanted concurrent access, which is very much a different task.  The
18process of protecting against unwanted concurrency will also avoid almost
19all optimization-related problems in a more efficient way.
20
21Like volatile, the kernel primitives which make concurrent access to data
22safe (spinlocks, mutexes, memory barriers, etc.) are designed to prevent
23unwanted optimization.  If they are being used properly, there will be no
24need to use volatile as well.  If volatile is still necessary, there is
25almost certainly a bug in the code somewhere.  In properly-written kernel
26code, volatile can only serve to slow things down.
27
28Consider a typical block of kernel code::
29
30    spin_lock(&the_lock);
31    do_something_on(&shared_data);
32    do_something_else_with(&shared_data);
33    spin_unlock(&the_lock);
34
35If all the code follows the locking rules, the value of shared_data cannot
36change unexpectedly while the_lock is held.  Any other code which might
37want to play with that data will be waiting on the lock.  The spinlock
38primitives act as memory barriers - they are explicitly written to do so -
39meaning that data accesses will not be optimized across them.  So the
40compiler might think it knows what will be in shared_data, but the
41spin_lock() call, since it acts as a memory barrier, will force it to
42forget anything it knows.  There will be no optimization problems with
43accesses to that data.
44
45If shared_data were declared volatile, the locking would still be
46necessary.  But the compiler would also be prevented from optimizing access
47to shared_data _within_ the critical section, when we know that nobody else
48can be working with it.  While the lock is held, shared_data is not
49volatile.  When dealing with shared data, proper locking makes volatile
50unnecessary - and potentially harmful.
51
52The volatile storage class was originally meant for memory-mapped I/O
53registers.  Within the kernel, register accesses, too, should be protected
54by locks, but one also does not want the compiler "optimizing" register
55accesses within a critical section.  But, within the kernel, I/O memory
56accesses are always done through accessor functions; accessing I/O memory
57directly through pointers is frowned upon and does not work on all
58architectures.  Those accessors are written to prevent unwanted
59optimization, so, once again, volatile is unnecessary.
60
61Another situation where one might be tempted to use volatile is
62when the processor is busy-waiting on the value of a variable.  The right
63way to perform a busy wait is::
64
65    while (my_variable != what_i_want)
66        cpu_relax();
67
68The cpu_relax() call can lower CPU power consumption or yield to a
69hyperthreaded twin processor; it also happens to serve as a compiler
70barrier, so, once again, volatile is unnecessary.  Of course, busy-
71waiting is generally an anti-social act to begin with.
72
73There are still a few rare situations where volatile makes sense in the
74kernel:
75
76  - The above-mentioned accessor functions might use volatile on
77    architectures where direct I/O memory access does work.  Essentially,
78    each accessor call becomes a little critical section on its own and
79    ensures that the access happens as expected by the programmer.
80
81  - Inline assembly code which changes memory, but which has no other
82    visible side effects, risks being deleted by GCC.  Adding the volatile
83    keyword to asm statements will prevent this removal.
84
85  - The jiffies variable is special in that it can have a different value
86    every time it is referenced, but it can be read without any special
87    locking.  So jiffies can be volatile, but the addition of other
88    variables of this type is strongly frowned upon.  Jiffies is considered
89    to be a "stupid legacy" issue (Linus's words) in this regard; fixing it
90    would be more trouble than it is worth.
91
92  - Pointers to data structures in coherent memory which might be modified
93    by I/O devices can, sometimes, legitimately be volatile.  A ring buffer
94    used by a network adapter, where that adapter changes pointers to
95    indicate which descriptors have been processed, is an example of this
96    type of situation.
97
98For most code, none of the above justifications for volatile apply.  As a
99result, the use of volatile is likely to be seen as a bug and will bring
100additional scrutiny to the code.  Developers who are tempted to use
101volatile should take a step back and think about what they are truly trying
102to accomplish.
103
104Patches to remove volatile variables are generally welcome - as long as
105they come with a justification which shows that the concurrency issues have
106been properly thought through.
107
108
109References
110==========
111
112[1] https://lwn.net/Articles/233481/
113
114[2] https://lwn.net/Articles/233482/
115
116Credits
117=======
118
119Original impetus and research by Randy Dunlap
120
121Written by Jonathan Corbet
122
123Improvements via comments from Satyam Sharma, Johannes Stezenbach, Jesper
124Juhl, Heikki Orsila, H. Peter Anvin, Philipp Hahn, and Stefan
125Richter.
126