On Wed, 8 Jun 2022 at 10:20, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote: > > On Tue, Jun 07, 2022 at 09:15:09AM +0200, Dmitry Vyukov wrote: > > > On Mon, 6 Jun 2022 at 14:39, Dan Carpenter <dan.carpenter@xxxxxxxxxx> wrote: > > > > > > > > On Sat, Jun 04, 2022 at 10:32:46AM +0200, 'Dmitry Vyukov' via syzkaller-bugs wrote: > > > > > On Fri, 3 Jun 2022 at 18:12, Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote: > > > > > > > > > > > > But again, is this a "real and able to be triggered from userspace" > > > > > > problem, or just fault-injection-induced? > > > > > > > > > > Then this is something to fix in the fault injection subsystem. > > > > > Testing systems shouldn't be reporting false positives. > > > > > What allocations cannot fail in real life? Is it <=page_size? > > > > > > > > > > > > > Apparently in 2014, anything less than *EIGHT?!!* pages succeeded! > > > > > > > > https://lwn.net/Articles/627419/ > > > > > > > > I have been on the look out since that article and never seen anyone > > > > mention it changing. I think we should ignore that and say that > > > > anything over PAGE_SIZE can fail. Possibly we could go smaller than > > > > PAGE_SIZE... > > > > > > +linux-mm for GFP expertise re what allocations cannot possibly fail > > > and should be excluded from fault injection. > > > > > > Interesting, thanks for the link. > > > > > > PAGE_SIZE looks like a good start. Once we have the predicate in > > > place, we can refine it later when/if we have more inputs. > > > > > > But I wonder about GFP flags. They definitely have some impact on allocations. > > > If GFP_ACCOUNT is set, all allocations can fail, right? > > > If GFP_DMA/DMA32 is set, allocations can fail, right? What about other zones? > > > If GFP_NORETRY is set, allocations can fail? > > > What about GFP_NOMEMALLOC and GFP_ATOMIC? > > > What about GFP_IO/GFP_FS/GFP_DIRECT_RECLAIM/GFP_KSWAPD_RECLAIM? At > > > least some of these need to be set for allocations to not fail? Which > > > ones? > > > Any other flags are required to be set/unset for allocations to not fail? > > > > I'm not the expert on page allocation, but ... > > > > I don't think GFP_ACCOUNT makes allocations fail. It might make reclaim > > happen from within that cgroup, and it might cause an OOM kill for > > something in that cgroup. But I don't think it makes a (low order) > > allocation more likely to fail. > > Interesting. > I was thinking of some malicious specifically crafted configurations > with very low limit and particular pattern of allocations. Also what > if there is just 1 process (current)? Is it possible to kill and > reclaim the current process when a thread is stuck in the middle of > the kernel on a kmalloc? > Also I see e.g.: > Tasks with the OOM protection (oom_score_adj set to -1000) > are treated as an exception and are never killed. > > I am not an expert on this either, but I think it may be hard to fight > with a specifically crafted attack. > > > > There's usually less memory avilable in DMA/DMA32 zones, but we have > > so few allocations from those zones, I question the utility of focusing > > testing on those allocations. > > > > GFP_ATOMIC allows access to emergency pools, so I would say _less_ likely > > to fail. KSWAPD_RECLAIM has no effect on whether _this_ allocation > > succeeds or fails; it kicks kswapd to do reclaim, rather than doing > > reclaim directly. DIRECT_RECLAIM definitely makes allocations more likely > > to succeed. GFP_FS allows (direct) reclaim to happen from filesystems. > > GFP_IO allows IO to start (ie writeback can start) in order to clean > > dirty memory. > > > > Anyway, I hope somebody who knows the page allocator better than I do > > can say smarter things than this. Even better if they can put it into > > Documentation/ somewhere ;-) > > Even better to put this into code as a predicate function that fault > injection will use. It will also serve as precise up-to-date > documentation. Also at the end of kmalloc as: WARN_ON(!ret && !cant_fail(size, gfp)); ! > > https://www.kernel.org/doc/html/latest/core-api/memory-allocation.html > > exists but isn't quite enough to answer this question.
