On Mon, Jul 18, 2022 at 04:51:59PM +0200, Alexander Gordeev wrote:
> On Mon, Jul 18, 2022 at 02:48:21PM +0100, Matthew Wilcox wrote:
> > On Mon, Jul 18, 2022 at 03:32:40PM +0200, Alexander Gordeev wrote:
> > > +++ b/arch/s390/kernel/crash_dump.c
> > > @@ -53,6 +53,7 @@ struct save_area {
> > > };
> > >
> > > static LIST_HEAD(dump_save_areas);
> >
> > I'd suggest you need a mutex here so that simultaneous calls to
> > copy_to_user_real() don't corrupt each others data.
>
> We stop all (but one) CPUs before calling into the capture kernel -
> one that calls these functions. Similarily to racy hsa_buf[] access
> from memcpy_hsa_iter() this should not hit.
Could you show me how that works when two processes read from
/proc/vmcore at the same time?
> As you noticed last time, it is a pre-existing race and I was
> actually going to address it in a separate fix - if the problem
> really exists.
>
> > > +static char memcpy_real_buf[PAGE_SIZE];
> > >
> > > /*
> > > * Allocate a save area
> > > @@ -179,23 +180,18 @@ int copy_oldmem_kernel(void *dst, unsigned long src, size_t count)
> > > static int copy_to_user_real(void __user *dest, unsigned long src, unsigned long count)
> > > {
> > > int offs = 0, size, rc;
> > > - char *buf;
> > >
> > > - buf = (char *)__get_free_page(GFP_KERNEL);
> > > - if (!buf)
> > > - return -ENOMEM;
> > > rc = -EFAULT;
> > > while (offs < count) {
> > > size = min(PAGE_SIZE, count - offs);
> > > - if (memcpy_real(buf, src + offs, size))
> > > + if (memcpy_real(memcpy_real_buf, src + offs, size))
> > > goto out;
> > > - if (copy_to_user(dest + offs, buf, size))
> > > + if (copy_to_user(dest + offs, memcpy_real_buf, size))
> > > goto out;
> > > offs += size;
> > > }
> > > rc = 0;
> > > out:
> > > - free_page((unsigned long)buf);
> > > return rc;
> > > }
>
> Thanks!
