Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 1 | Kernel Memory Leak Detector |
| 2 | =========================== |
| 3 | |
| 4 | Introduction |
| 5 | ------------ |
| 6 | |
| 7 | Kmemleak provides a way of detecting possible kernel memory leaks in a |
| 8 | way similar to a tracing garbage collector |
| 9 | (http://en.wikipedia.org/wiki/Garbage_collection_%28computer_science%29#Tracing_garbage_collectors), |
| 10 | with the difference that the orphan objects are not freed but only |
| 11 | reported via /sys/kernel/debug/kmemleak. A similar method is used by the |
| 12 | Valgrind tool (memcheck --leak-check) to detect the memory leaks in |
| 13 | user-space applications. |
| 14 | |
| 15 | Usage |
| 16 | ----- |
| 17 | |
| 18 | CONFIG_DEBUG_KMEMLEAK in "Kernel hacking" has to be enabled. A kernel |
Catalin Marinas | bab4a34 | 2009-06-26 17:38:26 +0100 | [diff] [blame] | 19 | thread scans the memory every 10 minutes (by default) and prints the |
Catalin Marinas | 4698c1f | 2009-06-26 17:38:27 +0100 | [diff] [blame] | 20 | number of new unreferenced objects found. To display the details of all |
| 21 | the possible memory leaks: |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 22 | |
| 23 | # mount -t debugfs nodev /sys/kernel/debug/ |
| 24 | # cat /sys/kernel/debug/kmemleak |
| 25 | |
Catalin Marinas | 4698c1f | 2009-06-26 17:38:27 +0100 | [diff] [blame] | 26 | To trigger an intermediate memory scan: |
| 27 | |
| 28 | # echo scan > /sys/kernel/debug/kmemleak |
| 29 | |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 30 | Note that the orphan objects are listed in the order they were allocated |
| 31 | and one object at the beginning of the list may cause other subsequent |
| 32 | objects to be reported as orphan. |
| 33 | |
| 34 | Memory scanning parameters can be modified at run-time by writing to the |
| 35 | /sys/kernel/debug/kmemleak file. The following parameters are supported: |
| 36 | |
| 37 | off - disable kmemleak (irreversible) |
Catalin Marinas | e0a2a16 | 2009-06-26 17:38:25 +0100 | [diff] [blame] | 38 | stack=on - enable the task stacks scanning (default) |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 39 | stack=off - disable the tasks stacks scanning |
Catalin Marinas | e0a2a16 | 2009-06-26 17:38:25 +0100 | [diff] [blame] | 40 | scan=on - start the automatic memory scanning thread (default) |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 41 | scan=off - stop the automatic memory scanning thread |
Catalin Marinas | e0a2a16 | 2009-06-26 17:38:25 +0100 | [diff] [blame] | 42 | scan=<secs> - set the automatic memory scanning period in seconds |
| 43 | (default 600, 0 to stop the automatic scanning) |
Catalin Marinas | 4698c1f | 2009-06-26 17:38:27 +0100 | [diff] [blame] | 44 | scan - trigger a memory scan |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 45 | |
| 46 | Kmemleak can also be disabled at boot-time by passing "kmemleak=off" on |
| 47 | the kernel command line. |
| 48 | |
Catalin Marinas | a9d9058 | 2009-06-25 10:16:11 +0100 | [diff] [blame] | 49 | Memory may be allocated or freed before kmemleak is initialised and |
| 50 | these actions are stored in an early log buffer. The size of this buffer |
| 51 | is configured via the CONFIG_DEBUG_KMEMLEAK_EARLY_LOG_SIZE option. |
| 52 | |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 53 | Basic Algorithm |
| 54 | --------------- |
| 55 | |
| 56 | The memory allocations via kmalloc, vmalloc, kmem_cache_alloc and |
| 57 | friends are traced and the pointers, together with additional |
| 58 | information like size and stack trace, are stored in a prio search tree. |
| 59 | The corresponding freeing function calls are tracked and the pointers |
| 60 | removed from the kmemleak data structures. |
| 61 | |
| 62 | An allocated block of memory is considered orphan if no pointer to its |
| 63 | start address or to any location inside the block can be found by |
| 64 | scanning the memory (including saved registers). This means that there |
| 65 | might be no way for the kernel to pass the address of the allocated |
| 66 | block to a freeing function and therefore the block is considered a |
| 67 | memory leak. |
| 68 | |
| 69 | The scanning algorithm steps: |
| 70 | |
| 71 | 1. mark all objects as white (remaining white objects will later be |
| 72 | considered orphan) |
| 73 | 2. scan the memory starting with the data section and stacks, checking |
| 74 | the values against the addresses stored in the prio search tree. If |
| 75 | a pointer to a white object is found, the object is added to the |
| 76 | gray list |
| 77 | 3. scan the gray objects for matching addresses (some white objects |
| 78 | can become gray and added at the end of the gray list) until the |
| 79 | gray set is finished |
| 80 | 4. the remaining white objects are considered orphan and reported via |
| 81 | /sys/kernel/debug/kmemleak |
| 82 | |
| 83 | Some allocated memory blocks have pointers stored in the kernel's |
| 84 | internal data structures and they cannot be detected as orphans. To |
| 85 | avoid this, kmemleak can also store the number of values pointing to an |
| 86 | address inside the block address range that need to be found so that the |
| 87 | block is not considered a leak. One example is __vmalloc(). |
| 88 | |
| 89 | Kmemleak API |
| 90 | ------------ |
| 91 | |
| 92 | See the include/linux/kmemleak.h header for the functions prototype. |
| 93 | |
| 94 | kmemleak_init - initialize kmemleak |
| 95 | kmemleak_alloc - notify of a memory block allocation |
| 96 | kmemleak_free - notify of a memory block freeing |
| 97 | kmemleak_not_leak - mark an object as not a leak |
| 98 | kmemleak_ignore - do not scan or report an object as leak |
| 99 | kmemleak_scan_area - add scan areas inside a memory block |
| 100 | kmemleak_no_scan - do not scan a memory block |
| 101 | kmemleak_erase - erase an old value in a pointer variable |
| 102 | kmemleak_alloc_recursive - as kmemleak_alloc but checks the recursiveness |
| 103 | kmemleak_free_recursive - as kmemleak_free but checks the recursiveness |
| 104 | |
| 105 | Dealing with false positives/negatives |
| 106 | -------------------------------------- |
| 107 | |
| 108 | The false negatives are real memory leaks (orphan objects) but not |
| 109 | reported by kmemleak because values found during the memory scanning |
| 110 | point to such objects. To reduce the number of false negatives, kmemleak |
| 111 | provides the kmemleak_ignore, kmemleak_scan_area, kmemleak_no_scan and |
| 112 | kmemleak_erase functions (see above). The task stacks also increase the |
| 113 | amount of false negatives and their scanning is not enabled by default. |
| 114 | |
| 115 | The false positives are objects wrongly reported as being memory leaks |
| 116 | (orphan). For objects known not to be leaks, kmemleak provides the |
| 117 | kmemleak_not_leak function. The kmemleak_ignore could also be used if |
| 118 | the memory block is known not to contain other pointers and it will no |
| 119 | longer be scanned. |
| 120 | |
| 121 | Some of the reported leaks are only transient, especially on SMP |
| 122 | systems, because of pointers temporarily stored in CPU registers or |
| 123 | stacks. Kmemleak defines MSECS_MIN_AGE (defaulting to 1000) representing |
| 124 | the minimum age of an object to be reported as a memory leak. |
| 125 | |
| 126 | Limitations and Drawbacks |
| 127 | ------------------------- |
| 128 | |
| 129 | The main drawback is the reduced performance of memory allocation and |
| 130 | freeing. To avoid other penalties, the memory scanning is only performed |
| 131 | when the /sys/kernel/debug/kmemleak file is read. Anyway, this tool is |
| 132 | intended for debugging purposes where the performance might not be the |
| 133 | most important requirement. |
| 134 | |
| 135 | To keep the algorithm simple, kmemleak scans for values pointing to any |
| 136 | address inside a block's address range. This may lead to an increased |
| 137 | number of false negatives. However, it is likely that a real memory leak |
| 138 | will eventually become visible. |
| 139 | |
| 140 | Another source of false negatives is the data stored in non-pointer |
| 141 | values. In a future version, kmemleak could only scan the pointer |
| 142 | members in the allocated structures. This feature would solve many of |
| 143 | the false negative cases described above. |
| 144 | |
| 145 | The tool can report false positives. These are cases where an allocated |
| 146 | block doesn't need to be freed (some cases in the init_call functions), |
| 147 | the pointer is calculated by other methods than the usual container_of |
| 148 | macro or the pointer is stored in a location not scanned by kmemleak. |
| 149 | |
| 150 | Page allocations and ioremap are not tracked. Only the ARM and x86 |
| 151 | architectures are currently supported. |