On Tue, 03/07 11:58, Paolo Bonzini wrote:
>
>
> On 07/03/2017 03:17, Fam Zheng wrote:
> > Linux exposes a separate limit, /sys/block/.../queue/max_segments, which
> > in the worst case can be more restrictive than BLKSECTGET (as they are
> > two different things). Similar to the BLKSECTGET story, guests don't see
> > this limit and send big requests will get -EINVAL error on SG_IO.
> >
> > Lean on the safer side to clamp max_transfer according to max_segments
> > and page size, because in the end what host HBA gets is the mapped host
> > pages rather than a guest buffer.
> >
> > Signed-off-by: Fam Zheng <f...@redhat.com>
> > ---
> > block/file-posix.c | 58
> > ++++++++++++++++++++++++++++++++++++++++++++++++++++++
> > 1 file changed, 58 insertions(+)
> >
> > diff --git a/block/file-posix.c b/block/file-posix.c
> > index 4de1abd..b615262 100644
> > --- a/block/file-posix.c
> > +++ b/block/file-posix.c
> > @@ -668,6 +668,59 @@ static int
> > hdev_get_max_transfer_length(BlockDriverState *bs, int fd)
> > #endif
> > }
> >
> > +static int hdev_get_max_segments(BlockDriverState *bs)
> > +{
> > +#ifdef CONFIG_LINUX
> > + char buf[32];
> > + const char *end;
> > + char *sysfspath, *fullpath;
> > + int ret;
> > + int fd = -1;
> > + long max_segments;
> > +
> > + fullpath = realpath(bs->exact_filename, NULL);
> > + if (!fullpath) {
> > + return -errno;
> > + }
> > + if (strncmp(fullpath, "/dev/", 5) || !fullpath[5]) {
> > + ret = -ENOTSUP;
> > + goto out;
> > + }
> > + sysfspath = g_strdup_printf("/sys/block/%s/queue/max_segments",
> > + &fullpath[5]);
>
> I think you cannot rely on the /dev/... path. Luckily, there is an
> alternative path to "queue" via /sys/dev/block/MAJOR:MINOR/queue, where
> MAJOR:MINOR can be retrieved via fstat. This also avoids any possible
> TOC-TOU races.
Sounds good, I'll send another version.
Fam
