On Thu, Jan 26, 2023 at 12:08:33AM +0100, Philippe Mathieu-Daudé wrote:
> On 25/1/23 23:40, Peter Xu wrote:
> > Teach QEMU to use /dev/userfaultfd when it existed and fallback to the
> > system call if either it's not there or doesn't have enough permission.
> >
> > Firstly, as long as the app has permission to access /dev/userfaultfd, it
> > always have the ability to trap kernel faults which QEMU mostly wants.
> > Meanwhile, in some context (e.g. containers) the userfaultfd syscall can be
> > forbidden, so it can be the major way to use postcopy in a restricted
> > environment with strict seccomp setup.
> >
> > Signed-off-by: Peter Xu <pet...@redhat.com>
> > ---
> > util/trace-events | 1 +
> > util/userfaultfd.c | 36 ++++++++++++++++++++++++++++++++++++
> > 2 files changed, 37 insertions(+)
> >
> > diff --git a/util/trace-events b/util/trace-events
> > index c8f53d7d9f..16f78d8fe5 100644
> > --- a/util/trace-events
> > +++ b/util/trace-events
> > @@ -93,6 +93,7 @@ qemu_vfio_region_info(const char *desc, uint64_t
> > region_ofs, uint64_t region_siz
> > qemu_vfio_pci_map_bar(int index, uint64_t region_ofs, uint64_t
> > region_size, int ofs, void *host) "map region bar#%d addr 0x%"PRIx64" size
> > 0x%"PRIx64" ofs 0x%x host %p"
> > #userfaultfd.c
> > +uffd_detect_open_mode(int mode) "%d"
> > uffd_query_features_nosys(int err) "errno: %i"
> > uffd_query_features_api_failed(int err) "errno: %i"
> > uffd_create_fd_nosys(int err) "errno: %i"
> > diff --git a/util/userfaultfd.c b/util/userfaultfd.c
> > index 9845a2ec81..360ecf8084 100644
> > --- a/util/userfaultfd.c
> > +++ b/util/userfaultfd.c
> > @@ -18,10 +18,46 @@
> > #include <poll.h>
> > #include <sys/syscall.h>
> > #include <sys/ioctl.h>
> > +#include <fcntl.h>
> > +
> > +typedef enum {
> > + UFFD_UNINITIALIZED = 0,
> > + UFFD_USE_DEV_PATH,
> > + UFFD_USE_SYSCALL,
> > +} uffd_open_mode;
> > +
> > +static uffd_open_mode open_mode;
>
> 'open_mode' could be reduced to uffd_detect_open_mode()'s
> scope.
Yes, will do.
>
> > +static int uffd_dev;
> > +
> > +static uffd_open_mode uffd_detect_open_mode(void)
> > +{
> > + if (open_mode == UFFD_UNINITIALIZED) {
> > + /*
> > + * Make /dev/userfaultfd the default approach because it has better
> > + * permission controls, meanwhile allows kernel faults without any
> > + * privilege requirement (e.g. SYS_CAP_PTRACE).
> > + */
> > + uffd_dev = open("/dev/userfaultfd", O_RDWR | O_CLOEXEC);
> > + if (uffd_dev >= 0) {
> > + open_mode = UFFD_USE_DEV_PATH;
> > + } else {
> > + /* Fallback to the system call */
> > + open_mode = UFFD_USE_SYSCALL;
> > + }
> > + trace_uffd_detect_open_mode(open_mode);
> > + }
> > +
> > + return open_mode;
>
> If 'open_mode' isn't relevant, this function could return uffd_dev/-1
> instead. Not really an improvement :)
Logically I think the two variables can be squashed into one. I kept that
for clearance just to easily identify e.g. uffd_dev is not chosen to be
used, or uffd_dev open failed.
>
> Reviewed-by: Philippe Mathieu-Daudé <phi...@linaro.org>
Thanks, Phil.
--
Peter Xu
