On Wed, Mar 27, 2013 at 12:03 AM, Hugh Dickins <hughd@xxxxxxxxxx> wrote: > On Tue, 26 Mar 2013, Toralf Foerster wrote: >> On 03/25/2013 11:53 PM, Andrew Morton wrote: >> > On Fri, 22 Mar 2013 18:28:36 +0100 Toralf Foerster <toralf.foerster@xxxxxx> wrote: >> > >> >> > Using trinity I often trigger under a user mode linux image with host kernel 3.8.4 >> >> > and guest kernel linux-v3.9-rc3-244-g9217cbb the following : >> >> > (The UML guest is a 32bit stable Gentoo Linux) >> > I assume 3.8 is OK? >> > >> With UML kernel 3.7.10 (host kernel still 3.8.4) I can trigger this >> issue too. >> Just to clarify it - here the bug appears in the UML kernel - the host >> kernel is ok (I can of course crash a host kernel too by trinity'ing an >> UML guest, but that's another thread - see [1]) >> >> >> FWIW he trinity command is just a test of 1 syscall: >> >> $> trinity --children 1 --victims /mnt/nfs/n22/victims -c mremap >> >> >> >> [1] https://lkml.org/lkml/2013/3/24/174 > > I should think it's been like this for five years, or even more: maybe > you are the first person to try unmapping user address 0x100000 on UML; > though it's odd that you find it using mremap than the more common munmap. Sounds sane. I fear trinity will find more "you are the first person to try"-bugs in UML. I'll look at it. > uml_setup_stubs() sets up the special vma with install_special_mapping(), > but instead of then faulting in the two pages concerned, it has preset > the ptes with init_stub_pte(), which did not increment page mapcount. > > munmap() that area (or set up another mapping in that place), and > zap_pte_range() will decrement page mapcount negative, hence the > "Bad page" errors. Whereas UML uses an arch_exit_mmap() hook to > clear the ptes at exit time, to avoid encountering such errors. > > I think that adding VM_PFNMAP to those install_special_mapping() flags > would be enough to fix it (and avoid the need for the arch_exit_mmap(), > and let vm_insert_pfn() do the work of init_stub_pte()); but I'm not > certain that would be the approved way, and I may have missed problems > doing it like this (which would disallow get_user_pages(), e.g. ptrace, > on that area: which might or might not be a good thing, I don't know). > > I'm saying this just by examination, I've not tried any of it at all. > Over to Richard. :-) -- Thanks, //richard -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>