On Fri, 10 Feb 2017 01:28:47 +0900 Masami Hiramatsu <[email protected]> wrote:
> Hi, > > Here is 2nd version of the series. I just updated the > patch description to make it easier to understand > on arm and arm64, no code change. > > V1 is here: > http://www.mail-archive.com/[email protected]/msg1327856.html > > ---- > This series will fix a possible deadlock case in kretprobe > on x86, arm, arm64. Since kretprobe has been optimized on > those arch, they have similar possible deadlock issue. > > Problem > ===== > > The deadlock senario is when a user puts 2 kretprobes, > one on normal function and one on a function which can be > called from NMI or FIQ where normal interrupt disabled. > (we don't recommend it, but possible.) In this case, if > the kernel hits the 1st kretprobe on a normal function > return which calls trampoline_handler(), acquire a > spinlock on the hash table in kretprobe_hash_lock() and > disable irqs. > After that, if NMI(or FIQ on arm/arm64) is occurred and > the 2nd kretprobe is kicked, it also calls > trampoline_handler() and tries to acquire the same > spinlock (since the hash is based on current task, same > as the 1st kretprobe), it causes a deadlock on the > spinlock. > Note that this is very rare case, but theoretically happens. Please ignore this series. I found that the kretprobe already rejected probes in NMI (and FIQ) so this senario never be true (as far as in_nmi() works in nested interrupt context.) Sorry for bother you. And thank you! > > Reason and Affected Arch > ===== > > Actually, this bug has been introduced by kretprobe-booster, > which removes a kprobe from return trampoline code, but also > resets current kprobe, which can be a stopper for the nested > k(ret)probes. So, currently only x86, arm, and arm64 are > affected, because other arch have not implemented the > kretprobe-booster. > > Solution > ===== > > To fix this issue, I introduced a dummy kprobe which is set > as a current kprobe while holding the kretprobe-hash lock. > With that, if an NMI/FIQ occurred and 2nd kretprobe's kprobe > is kicked (to modify the return address, a kprobe is kicked > when the target function is called), the kprobe (and the 2nd > kretprobe also) is skipped because it detects there is > another kprobe is running. > > This reentrance detection and nested kprobe blocker had > existed when the original kretprobe was implemented by > using kprobe on trampoline code. This fix just revived it. > > Thank you, > > --- > > Masami Hiramatsu (3): > kprobes/x86: Fix a possible deadlock case in kretprobe > kprobes/arm64: Fix a possible deadlock case in kretprobe > kprobes/arm: Fix a possible deadlock case in kretprobe > > > arch/arm/probes/kprobes/core.c | 12 ++++++++++-- > arch/arm64/kernel/probes/kprobes.c | 12 ++++++++++-- > arch/x86/kernel/kprobes/core.c | 13 ++++++++++--- > 3 files changed, 30 insertions(+), 7 deletions(-) > > -- > Masami Hiramatsu -- Masami Hiramatsu <[email protected]>
