Kyle Swenson | 8d8f654 | 2021-03-15 11:02:55 -0600 | [diff] [blame^] | 1 | Why the "volatile" type class should not be used |
| 2 | ------------------------------------------------ |
| 3 | |
| 4 | C programmers have often taken volatile to mean that the variable could be |
| 5 | changed outside of the current thread of execution; as a result, they are |
| 6 | sometimes tempted to use it in kernel code when shared data structures are |
| 7 | being used. In other words, they have been known to treat volatile types |
| 8 | as a sort of easy atomic variable, which they are not. The use of volatile in |
| 9 | kernel code is almost never correct; this document describes why. |
| 10 | |
| 11 | The key point to understand with regard to volatile is that its purpose is |
| 12 | to suppress optimization, which is almost never what one really wants to |
| 13 | do. In the kernel, one must protect shared data structures against |
| 14 | unwanted concurrent access, which is very much a different task. The |
| 15 | process of protecting against unwanted concurrency will also avoid almost |
| 16 | all optimization-related problems in a more efficient way. |
| 17 | |
| 18 | Like volatile, the kernel primitives which make concurrent access to data |
| 19 | safe (spinlocks, mutexes, memory barriers, etc.) are designed to prevent |
| 20 | unwanted optimization. If they are being used properly, there will be no |
| 21 | need to use volatile as well. If volatile is still necessary, there is |
| 22 | almost certainly a bug in the code somewhere. In properly-written kernel |
| 23 | code, volatile can only serve to slow things down. |
| 24 | |
| 25 | Consider a typical block of kernel code: |
| 26 | |
| 27 | spin_lock(&the_lock); |
| 28 | do_something_on(&shared_data); |
| 29 | do_something_else_with(&shared_data); |
| 30 | spin_unlock(&the_lock); |
| 31 | |
| 32 | If all the code follows the locking rules, the value of shared_data cannot |
| 33 | change unexpectedly while the_lock is held. Any other code which might |
| 34 | want to play with that data will be waiting on the lock. The spinlock |
| 35 | primitives act as memory barriers - they are explicitly written to do so - |
| 36 | meaning that data accesses will not be optimized across them. So the |
| 37 | compiler might think it knows what will be in shared_data, but the |
| 38 | spin_lock() call, since it acts as a memory barrier, will force it to |
| 39 | forget anything it knows. There will be no optimization problems with |
| 40 | accesses to that data. |
| 41 | |
| 42 | If shared_data were declared volatile, the locking would still be |
| 43 | necessary. But the compiler would also be prevented from optimizing access |
| 44 | to shared_data _within_ the critical section, when we know that nobody else |
| 45 | can be working with it. While the lock is held, shared_data is not |
| 46 | volatile. When dealing with shared data, proper locking makes volatile |
| 47 | unnecessary - and potentially harmful. |
| 48 | |
| 49 | The volatile storage class was originally meant for memory-mapped I/O |
| 50 | registers. Within the kernel, register accesses, too, should be protected |
| 51 | by locks, but one also does not want the compiler "optimizing" register |
| 52 | accesses within a critical section. But, within the kernel, I/O memory |
| 53 | accesses are always done through accessor functions; accessing I/O memory |
| 54 | directly through pointers is frowned upon and does not work on all |
| 55 | architectures. Those accessors are written to prevent unwanted |
| 56 | optimization, so, once again, volatile is unnecessary. |
| 57 | |
| 58 | Another situation where one might be tempted to use volatile is |
| 59 | when the processor is busy-waiting on the value of a variable. The right |
| 60 | way to perform a busy wait is: |
| 61 | |
| 62 | while (my_variable != what_i_want) |
| 63 | cpu_relax(); |
| 64 | |
| 65 | The cpu_relax() call can lower CPU power consumption or yield to a |
| 66 | hyperthreaded twin processor; it also happens to serve as a compiler |
| 67 | barrier, so, once again, volatile is unnecessary. Of course, busy- |
| 68 | waiting is generally an anti-social act to begin with. |
| 69 | |
| 70 | There are still a few rare situations where volatile makes sense in the |
| 71 | kernel: |
| 72 | |
| 73 | - The above-mentioned accessor functions might use volatile on |
| 74 | architectures where direct I/O memory access does work. Essentially, |
| 75 | each accessor call becomes a little critical section on its own and |
| 76 | ensures that the access happens as expected by the programmer. |
| 77 | |
| 78 | - Inline assembly code which changes memory, but which has no other |
| 79 | visible side effects, risks being deleted by GCC. Adding the volatile |
| 80 | keyword to asm statements will prevent this removal. |
| 81 | |
| 82 | - The jiffies variable is special in that it can have a different value |
| 83 | every time it is referenced, but it can be read without any special |
| 84 | locking. So jiffies can be volatile, but the addition of other |
| 85 | variables of this type is strongly frowned upon. Jiffies is considered |
| 86 | to be a "stupid legacy" issue (Linus's words) in this regard; fixing it |
| 87 | would be more trouble than it is worth. |
| 88 | |
| 89 | - Pointers to data structures in coherent memory which might be modified |
| 90 | by I/O devices can, sometimes, legitimately be volatile. A ring buffer |
| 91 | used by a network adapter, where that adapter changes pointers to |
| 92 | indicate which descriptors have been processed, is an example of this |
| 93 | type of situation. |
| 94 | |
| 95 | For most code, none of the above justifications for volatile apply. As a |
| 96 | result, the use of volatile is likely to be seen as a bug and will bring |
| 97 | additional scrutiny to the code. Developers who are tempted to use |
| 98 | volatile should take a step back and think about what they are truly trying |
| 99 | to accomplish. |
| 100 | |
| 101 | Patches to remove volatile variables are generally welcome - as long as |
| 102 | they come with a justification which shows that the concurrency issues have |
| 103 | been properly thought through. |
| 104 | |
| 105 | |
| 106 | NOTES |
| 107 | ----- |
| 108 | |
| 109 | [1] http://lwn.net/Articles/233481/ |
| 110 | [2] http://lwn.net/Articles/233482/ |
| 111 | |
| 112 | CREDITS |
| 113 | ------- |
| 114 | |
| 115 | Original impetus and research by Randy Dunlap |
| 116 | Written by Jonathan Corbet |
| 117 | Improvements via comments from Satyam Sharma, Johannes Stezenbach, Jesper |
| 118 | Juhl, Heikki Orsila, H. Peter Anvin, Philipp Hahn, and Stefan |
| 119 | Richter. |