Messages in this thread | | | Subject | Re: Possible memory leak via inotify_add_watch | From | Eric Paris <> | Date | Mon, 06 Jul 2009 18:59:00 -0400 |
| |
On Mon, 2009-07-06 at 23:03 +0100, Catalin Marinas wrote: > Hi Eric, > > I'm getting a few kmemleak reports like the one below (it may as well > be just a false positive). All of these allocations happened during > udevd.
Ok, I'll take a look. I know it highly unlikely but possible to leak in sys_inotify_add_watch on some of the error paths. Seeing as how you object has a mask it was at some point actually attached to an inode and was removed. Since free_i_list is still unused that means it was removed by an explicit request from inotify (inotify_rm_watch I guess), not from the inode disappearing. I'll try to run it down exactly tonight.
-Eric
> unreferenced object 0xc399fd80 (size 84): > comm "udevd", pid 879, jiffies 4294897228 > backtrace: > [<c01e0c3a>] create_object+0xfa/0x250 > [<c01e1e7d>] kmemleak_alloc+0x5d/0x70 > [<c01dac1b>] kmem_cache_alloc+0x14b/0x190 > [<c0213800>] sys_inotify_add_watch+0xc0/0x2a0 > [<c010300c>] sysenter_do_call+0x12/0x38 > [<ffffffff>] 0xffffffff > > Printing this object with gdb on /proc/kcore shows: > > (gdb) print {struct inotify_inode_mark_entry}0xc399fd80 > $2 = {fsn_entry = {mask = 134250504, refcnt = {counter = 1}, inode = 0x0, > group = 0x0, i_list = {next = 0x0, pprev = 0x0}, g_list = { > next = 0xc399fd98, prev = 0xc399fd98}, lock = {raw_lock = { > slock = 1028}, magic = 3735899821, owner_cpu = 4294967295, > owner = 0xffffffff, dep_map = {key = 0xc0d3e59c, class_cache = 0x0, > name = 0xc068413d "&entry->lock"}}, free_i_list = {next = 0x6b6b6b6b, > prev = 0x6b6b6b6b}, free_g_list = {next = 0x6b6b6b6b, > prev = 0x6b6b6b6b}, free_mark = 0xc0213720 <inotify_free_mark>}, wd = 28} > > It seems that is was freed via fsnotify_destroy_mark_by_entry() since > group and inode members are NULL and it was removed from any list. The > fsn_entry.refcnt, however, is still 1. Kmemleak cannot find any > pointer to this object (though it doesn't track alloc_pages memory > blocks). > > Thanks. >
| |