Re: [syzbot] [mm?] kernel BUG in sg_init_one

[Barry Song]

On Tue, Mar 19, 2024 at 10:10 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
>
> On Mon, Mar 18, 2024 at 01:17:19PM -0700, Yosry Ahmed wrote:
> > On Mon, Mar 18, 2024 at 11:00 AM Nhat Pham <nphamcs@xxxxxxxxx> wrote:
> > >
> > > On Mon, Mar 18, 2024 at 9:58 AM syzbot
> > > <syzbot+adbc983a1588b7805de3@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > Hello,
> > > >
> > > > syzbot found the following issue on:
> > > >
> > > > HEAD commit:    e5eb28f6d1af Merge tag 'mm-nonmm-stable-2024-03-14-09-36' ..
> > > > git tree:       upstream
> > > > console output: https://syzkaller.appspot.com/x/log.txt?x=13043abe180000
> > > > kernel config:  https://syzkaller.appspot.com/x/.config?x=19bb57c23dffc38e
> > > > dashboard link: https://syzkaller.appspot.com/bug?extid=adbc983a1588b7805de3
> > > > compiler:       arm-linux-gnueabi-gcc (Debian 12.2.0-14) 12.2.0, GNU ld (GNU Binutils for Debian) 2.40
> > > > userspace arch: arm
> > > > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=1706d231180000
> > > > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=13ba7959180000
> > > >
> > > > Downloadable assets:
> > > > disk image (non-bootable): https://storage.googleapis.com/syzbot-assets/8ead8862021c/non_bootable_disk-e5eb28f6.raw.xz
> > > > vmlinux: https://storage.googleapis.com/syzbot-assets/0a7371c63ff2/vmlinux-e5eb28f6.xz
> > > > kernel image: https://storage.googleapis.com/syzbot-assets/7539441b4add/zImage-e5eb28f6.xz
> > > >
> > > > IMPORTANT: if you fix the issue, please add the following tag to the commit:
> > > > Reported-by: syzbot+adbc983a1588b7805de3@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > >
> > > > ------------[ cut here ]------------
> > > > kernel BUG at include/linux/scatterlist.h:187!
> > >
> > > Looks like the provided buffer is invalid:
> > >
> > > #ifdef CONFIG_DEBUG_SG
> > > BUG_ON(!virt_addr_valid(buf));
> > > #endif
> > >
> > > which is "src" from:
> > >
> > > sg_init_one(&input, src, entry->length);
> > >
> > > Looking at the surrounding code and recent history, there's this
> > > commit that stands out:
> > >
> > > mm/zswap: remove the memcpy if acomp is not sleepable
> > > (sha: 270700dd06ca41a4779c19eb46608f076bb7d40e)
> > >
> > > which has the effect of, IIUC, using the zpool mapped memory directly
> > > as src, instead of acomp_ctx->buffer (which was previously the case,
> > > as zsmalloc was not sleepable).
> > >
> > > This might not necessarily be a bug with that commit itself, but might
> > > have revealed another bug elsewhere.
> > >
> > > Anyway, cc-ing the author, Barry Song, to fact check me :) Will take a
> > > closer look later.
> >
> > I am not a highmem expert, but the reproducer has CONFIG_HIGHMEM=y,
> > and it seems like zs_map_object() may return a highmem address if the
> > compressed object is entirely in a single page to avoid copying to a
> > buffer:
> >
> > if (off + class->size <= PAGE_SIZE) {
> >    /* this object is contained entirely within a page */
> >    area->vm_addr = kmap_atomic(page);
> >    ret = area->vm_addr + off;
> >    goto out;
> > }
> >
> > The virt_addr_valid() check seems to indicate that we expect a direct
> > map address in sg_init_one(), right?
>
> If the page is highmem, kmap_atomic() establishes a temporary mapping
> to it in the direct map, such that we have a legit kernel pointer to
> the memory. Otherwise the memcpy() in zswap also wouldn't work... Am I
> missing something?

Right, we built a map but it is not a linear mapping. so we can't use
virt_to_page
on this kind of non-linear mapping.
kmap_to_page can handle both linear and non-linear, but  Ira's commit
added a WARN_ON_ONCE in it for non-linear mapping case.