Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 1 | Kernel Memory Leak Detector |
| 2 | =========================== |
| 3 | |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 4 | Kmemleak provides a way of detecting possible kernel memory leaks in a |
| 5 | way similar to a tracing garbage collector |
Masanari Iida | ae13c65 | 2015-06-18 00:12:02 +0900 | [diff] [blame] | 6 | (https://en.wikipedia.org/wiki/Garbage_collection_%28computer_science%29#Tracing_garbage_collectors), |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 7 | with the difference that the orphan objects are not freed but only |
| 8 | reported via /sys/kernel/debug/kmemleak. A similar method is used by the |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 9 | Valgrind tool (``memcheck --leak-check``) to detect the memory leaks in |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 10 | user-space applications. |
Wang YanQing | 4762c98 | 2014-04-03 14:50:38 -0700 | [diff] [blame] | 11 | Kmemleak is supported on x86, arm, powerpc, sparc, sh, microblaze, ppc, mips, s390, metag and tile. |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 12 | |
| 13 | Usage |
| 14 | ----- |
| 15 | |
| 16 | CONFIG_DEBUG_KMEMLEAK in "Kernel hacking" has to be enabled. A kernel |
Catalin Marinas | bab4a34 | 2009-06-26 17:38:26 +0100 | [diff] [blame] | 17 | thread scans the memory every 10 minutes (by default) and prints the |
Catalin Marinas | 4698c1f | 2009-06-26 17:38:27 +0100 | [diff] [blame] | 18 | number of new unreferenced objects found. To display the details of all |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 19 | the possible memory leaks:: |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 20 | |
| 21 | # mount -t debugfs nodev /sys/kernel/debug/ |
| 22 | # cat /sys/kernel/debug/kmemleak |
| 23 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 24 | To trigger an intermediate memory scan:: |
Catalin Marinas | 4698c1f | 2009-06-26 17:38:27 +0100 | [diff] [blame] | 25 | |
| 26 | # echo scan > /sys/kernel/debug/kmemleak |
| 27 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 28 | To clear the list of all current possible memory leaks:: |
Luis R. Rodriguez | 30b3710 | 2009-09-04 17:44:51 -0700 | [diff] [blame] | 29 | |
| 30 | # echo clear > /sys/kernel/debug/kmemleak |
| 31 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 32 | New leaks will then come up upon reading ``/sys/kernel/debug/kmemleak`` |
Luis R. Rodriguez | 30b3710 | 2009-09-04 17:44:51 -0700 | [diff] [blame] | 33 | again. |
| 34 | |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 35 | Note that the orphan objects are listed in the order they were allocated |
| 36 | and one object at the beginning of the list may cause other subsequent |
| 37 | objects to be reported as orphan. |
| 38 | |
| 39 | Memory scanning parameters can be modified at run-time by writing to the |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 40 | ``/sys/kernel/debug/kmemleak`` file. The following parameters are supported: |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 41 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 42 | - off |
| 43 | disable kmemleak (irreversible) |
| 44 | - stack=on |
| 45 | enable the task stacks scanning (default) |
| 46 | - stack=off |
| 47 | disable the tasks stacks scanning |
| 48 | - scan=on |
| 49 | start the automatic memory scanning thread (default) |
| 50 | - scan=off |
| 51 | stop the automatic memory scanning thread |
| 52 | - scan=<secs> |
| 53 | set the automatic memory scanning period in seconds |
| 54 | (default 600, 0 to stop the automatic scanning) |
| 55 | - scan |
| 56 | trigger a memory scan |
| 57 | - clear |
| 58 | clear list of current memory leak suspects, done by |
| 59 | marking all current reported unreferenced objects grey, |
| 60 | or free all kmemleak objects if kmemleak has been disabled. |
| 61 | - dump=<addr> |
| 62 | dump information about the object found at <addr> |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 63 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 64 | Kmemleak can also be disabled at boot-time by passing ``kmemleak=off`` on |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 65 | the kernel command line. |
| 66 | |
Catalin Marinas | a9d9058 | 2009-06-25 10:16:11 +0100 | [diff] [blame] | 67 | Memory may be allocated or freed before kmemleak is initialised and |
| 68 | these actions are stored in an early log buffer. The size of this buffer |
| 69 | is configured via the CONFIG_DEBUG_KMEMLEAK_EARLY_LOG_SIZE option. |
| 70 | |
Masanari Iida | 6808a40 | 2014-10-24 21:25:00 +0900 | [diff] [blame] | 71 | If CONFIG_DEBUG_KMEMLEAK_DEFAULT_OFF are enabled, the kmemleak is |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 72 | disabled by default. Passing ``kmemleak=on`` on the kernel command |
Masanari Iida | 6808a40 | 2014-10-24 21:25:00 +0900 | [diff] [blame] | 73 | line enables the function. |
| 74 | |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 75 | Basic Algorithm |
| 76 | --------------- |
| 77 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 78 | The memory allocations via :c:func:`kmalloc`, :c:func:`vmalloc`, |
| 79 | :c:func:`kmem_cache_alloc` and |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 80 | friends are traced and the pointers, together with additional |
Wang YanQing | 4762c98 | 2014-04-03 14:50:38 -0700 | [diff] [blame] | 81 | information like size and stack trace, are stored in a rbtree. |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 82 | The corresponding freeing function calls are tracked and the pointers |
| 83 | removed from the kmemleak data structures. |
| 84 | |
| 85 | An allocated block of memory is considered orphan if no pointer to its |
| 86 | start address or to any location inside the block can be found by |
| 87 | scanning the memory (including saved registers). This means that there |
| 88 | might be no way for the kernel to pass the address of the allocated |
| 89 | block to a freeing function and therefore the block is considered a |
| 90 | memory leak. |
| 91 | |
| 92 | The scanning algorithm steps: |
| 93 | |
| 94 | 1. mark all objects as white (remaining white objects will later be |
| 95 | considered orphan) |
| 96 | 2. scan the memory starting with the data section and stacks, checking |
Wang YanQing | 4762c98 | 2014-04-03 14:50:38 -0700 | [diff] [blame] | 97 | the values against the addresses stored in the rbtree. If |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 98 | a pointer to a white object is found, the object is added to the |
| 99 | gray list |
| 100 | 3. scan the gray objects for matching addresses (some white objects |
| 101 | can become gray and added at the end of the gray list) until the |
| 102 | gray set is finished |
| 103 | 4. the remaining white objects are considered orphan and reported via |
| 104 | /sys/kernel/debug/kmemleak |
| 105 | |
| 106 | Some allocated memory blocks have pointers stored in the kernel's |
| 107 | internal data structures and they cannot be detected as orphans. To |
| 108 | avoid this, kmemleak can also store the number of values pointing to an |
| 109 | address inside the block address range that need to be found so that the |
| 110 | block is not considered a leak. One example is __vmalloc(). |
| 111 | |
Luis R. Rodriguez | 30b3710 | 2009-09-04 17:44:51 -0700 | [diff] [blame] | 112 | Testing specific sections with kmemleak |
| 113 | --------------------------------------- |
| 114 | |
| 115 | Upon initial bootup your /sys/kernel/debug/kmemleak output page may be |
| 116 | quite extensive. This can also be the case if you have very buggy code |
| 117 | when doing development. To work around these situations you can use the |
| 118 | 'clear' command to clear all reported unreferenced objects from the |
| 119 | /sys/kernel/debug/kmemleak output. By issuing a 'scan' after a 'clear' |
| 120 | you can find new unreferenced objects; this should help with testing |
| 121 | specific sections of code. |
| 122 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 123 | To test a critical section on demand with a clean kmemleak do:: |
Luis R. Rodriguez | 30b3710 | 2009-09-04 17:44:51 -0700 | [diff] [blame] | 124 | |
| 125 | # echo clear > /sys/kernel/debug/kmemleak |
| 126 | ... test your kernel or modules ... |
| 127 | # echo scan > /sys/kernel/debug/kmemleak |
| 128 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 129 | Then as usual to get your report with:: |
Luis R. Rodriguez | 30b3710 | 2009-09-04 17:44:51 -0700 | [diff] [blame] | 130 | |
| 131 | # cat /sys/kernel/debug/kmemleak |
| 132 | |
Li Zefan | c89da70 | 2014-04-03 14:46:27 -0700 | [diff] [blame] | 133 | Freeing kmemleak internal objects |
| 134 | --------------------------------- |
| 135 | |
Rahul Bedarkar | abb3b1f | 2014-07-31 23:50:19 +0530 | [diff] [blame] | 136 | To allow access to previously found memory leaks after kmemleak has been |
Li Zefan | c89da70 | 2014-04-03 14:46:27 -0700 | [diff] [blame] | 137 | disabled by the user or due to an fatal error, internal kmemleak objects |
| 138 | won't be freed when kmemleak is disabled, and those objects may occupy |
| 139 | a large part of physical memory. |
| 140 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 141 | In this situation, you may reclaim memory with:: |
Li Zefan | c89da70 | 2014-04-03 14:46:27 -0700 | [diff] [blame] | 142 | |
| 143 | # echo clear > /sys/kernel/debug/kmemleak |
| 144 | |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 145 | Kmemleak API |
| 146 | ------------ |
| 147 | |
| 148 | See the include/linux/kmemleak.h header for the functions prototype. |
| 149 | |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 150 | - ``kmemleak_init`` - initialize kmemleak |
| 151 | - ``kmemleak_alloc`` - notify of a memory block allocation |
| 152 | - ``kmemleak_alloc_percpu`` - notify of a percpu memory block allocation |
Catalin Marinas | 94f4a16 | 2017-07-06 15:40:22 -0700 | [diff] [blame] | 153 | - ``kmemleak_vmalloc`` - notify of a vmalloc() memory allocation |
Jonathan Corbet | ca90a7a3 | 2016-08-07 15:46:10 -0600 | [diff] [blame] | 154 | - ``kmemleak_free`` - notify of a memory block freeing |
| 155 | - ``kmemleak_free_part`` - notify of a partial memory block freeing |
| 156 | - ``kmemleak_free_percpu`` - notify of a percpu memory block freeing |
| 157 | - ``kmemleak_update_trace`` - update object allocation stack trace |
| 158 | - ``kmemleak_not_leak`` - mark an object as not a leak |
| 159 | - ``kmemleak_ignore`` - do not scan or report an object as leak |
| 160 | - ``kmemleak_scan_area`` - add scan areas inside a memory block |
| 161 | - ``kmemleak_no_scan`` - do not scan a memory block |
| 162 | - ``kmemleak_erase`` - erase an old value in a pointer variable |
| 163 | - ``kmemleak_alloc_recursive`` - as kmemleak_alloc but checks the recursiveness |
| 164 | - ``kmemleak_free_recursive`` - as kmemleak_free but checks the recursiveness |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 165 | |
Catalin Marinas | 9099dae | 2016-10-11 13:55:11 -0700 | [diff] [blame] | 166 | The following functions take a physical address as the object pointer |
| 167 | and only perform the corresponding action if the address has a lowmem |
| 168 | mapping: |
| 169 | |
| 170 | - ``kmemleak_alloc_phys`` |
| 171 | - ``kmemleak_free_part_phys`` |
| 172 | - ``kmemleak_not_leak_phys`` |
| 173 | - ``kmemleak_ignore_phys`` |
| 174 | |
Catalin Marinas | 04f7033 | 2009-06-11 13:22:39 +0100 | [diff] [blame] | 175 | Dealing with false positives/negatives |
| 176 | -------------------------------------- |
| 177 | |
| 178 | The false negatives are real memory leaks (orphan objects) but not |
| 179 | reported by kmemleak because values found during the memory scanning |
| 180 | point to such objects. To reduce the number of false negatives, kmemleak |
| 181 | provides the kmemleak_ignore, kmemleak_scan_area, kmemleak_no_scan and |
| 182 | kmemleak_erase functions (see above). The task stacks also increase the |
| 183 | amount of false negatives and their scanning is not enabled by default. |
| 184 | |
| 185 | The false positives are objects wrongly reported as being memory leaks |
| 186 | (orphan). For objects known not to be leaks, kmemleak provides the |
| 187 | kmemleak_not_leak function. The kmemleak_ignore could also be used if |
| 188 | the memory block is known not to contain other pointers and it will no |
| 189 | longer be scanned. |
| 190 | |
| 191 | Some of the reported leaks are only transient, especially on SMP |
| 192 | systems, because of pointers temporarily stored in CPU registers or |
| 193 | stacks. Kmemleak defines MSECS_MIN_AGE (defaulting to 1000) representing |
| 194 | the minimum age of an object to be reported as a memory leak. |
| 195 | |
| 196 | Limitations and Drawbacks |
| 197 | ------------------------- |
| 198 | |
| 199 | The main drawback is the reduced performance of memory allocation and |
| 200 | freeing. To avoid other penalties, the memory scanning is only performed |
| 201 | when the /sys/kernel/debug/kmemleak file is read. Anyway, this tool is |
| 202 | intended for debugging purposes where the performance might not be the |
| 203 | most important requirement. |
| 204 | |
| 205 | To keep the algorithm simple, kmemleak scans for values pointing to any |
| 206 | address inside a block's address range. This may lead to an increased |
| 207 | number of false negatives. However, it is likely that a real memory leak |
| 208 | will eventually become visible. |
| 209 | |
| 210 | Another source of false negatives is the data stored in non-pointer |
| 211 | values. In a future version, kmemleak could only scan the pointer |
| 212 | members in the allocated structures. This feature would solve many of |
| 213 | the false negative cases described above. |
| 214 | |
| 215 | The tool can report false positives. These are cases where an allocated |
| 216 | block doesn't need to be freed (some cases in the init_call functions), |
| 217 | the pointer is calculated by other methods than the usual container_of |
| 218 | macro or the pointer is stored in a location not scanned by kmemleak. |
| 219 | |
Daniel Baluta | 21b86bd | 2011-04-04 14:58:03 -0700 | [diff] [blame] | 220 | Page allocations and ioremap are not tracked. |