Messages in this thread | | | Date | Wed, 1 Jul 2009 13:04:38 +0200 | From | Ingo Molnar <> | Subject | Re: [PATCH] kmemleak: Fix scheduling-while-atomic bug |
| |
* Catalin Marinas <catalin.marinas@arm.com> wrote:
> On Wed, 2009-07-01 at 11:30 +0200, Ingo Molnar wrote: > > * Catalin Marinas <catalin.marinas@arm.com> wrote: > > > > > > The minimal fix below removes scan_yield() and adds a > > > > cond_resched() to the outmost (safe) place of the scanning > > > > thread. This solves the regression. > > > > > > With CONFIG_PREEMPT disabled it won't reschedule during the bss > > > scanning but I don't see this as a real issue (task stacks > > > scanning probably takes longer anyway). > > > > Yeah. I suspect one more cond_resched() could be added - i just > > didnt see an obvious place for it, given that scan_block() is being > > called with asymetric held-locks contexts. > > Yes, scan_block shouldn't call cond_resched(). The code is cleaner if > functions don't have too many side-effects. I can see about 1 sec of bss > scanning on an ARM board but with processor at < 500MHz and slow memory > system. On a standard x86 systems BSS scanning may not be noticeable > (and I think PREEMPT enabling is quite common these days). > > Since we are at locking, I just noticed this on my x86 laptop when > running cat /sys/kernel/debug/kmemleak (I haven't got it on an ARM > board): > > ================================================ > [ BUG: lock held when returning to user space! ] > ------------------------------------------------ > cat/3687 is leaving the kernel with locks still held! > 1 lock held by cat/3687: > #0: (scan_mutex){+.+.+.}, at: [<c01e0c5c>] kmemleak_open+0x3c/0x70 > > kmemleak_open() acquires scan_mutex and unconditionally releases > it in kmemleak_release(). The mutex seems to be released as a > subsequent acquiring works fine. > > Is this caused just because cat may have exited without closing > the file descriptor (which should be done automatically anyway)?
This lockdep warning has a 0% false positives track record so far: all previous cases it triggered showed some real (and fatal) bug in the underlying code.
The above one probably means scan_mutex is leaked out of a /proc syscall - that would be a bug in kmemleak.
Ingo
| |