Message ID | 1403858078-3402-1-git-send-email-dgilbert@redhat.com |
---|---|
State | New |
Headers | show |
Il 27/06/2014 10:34, Dr. David Alan Gilbert (git) ha scritto: > From: "Dr. David Alan Gilbert" <dgilbert@redhat.com> > > Commit 'virtio: validate config_len on load' restricted config_len > loaded from the wire to match the config_len that the device had. > > Unfortunately, there are cases where this isn't true, the one > we found it on was the wqe addition in virtio-blk. Indeed, the alternative here is to break migration. As a follow up, it would be nice to let the bus detect whether the config_len change is valid or not. For virtio-mmio and s390, mst said that config length must always match (luckily, these machines aren't versioned so they are not affected by the wce change). For virtio-pci, it is okay as long as the old_length + VIRTIO_PCI_REGION_SIZE(vdev) and new_length + VIRTIO_PCI_REGION_SIZE(vdev) do not cross a power of two. Paolo > Allow mismatched config-lengths: > *) If the version on the wire is shorter then ensure that the > remainder is 0xff filled (as virtio_config_read does on > out of range reads) > *) If the version on the wire is longer, load what we have space > for and skip the rest. > > Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com> > --- > hw/virtio/virtio.c | 30 ++++++++++++++++++++++++++---- > 1 file changed, 26 insertions(+), 4 deletions(-) > > diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c > index a3082d5..2b11142 100644 > --- a/hw/virtio/virtio.c > +++ b/hw/virtio/virtio.c > @@ -927,11 +927,33 @@ int virtio_load(VirtIODevice *vdev, QEMUFile *f) > } > config_len = qemu_get_be32(f); > if (config_len != vdev->config_len) { > - error_report("Unexpected config length 0x%x. Expected 0x%zx", > - config_len, vdev->config_len); > - return -1; > + /* > + * Unfortunately the reality is that there are cases where we > + * see mismatched config lengths, so we have to deal with them > + * rather than rejecting them. > + */ > + > + if (config_len < vdev->config_len) { > + /* This is normal in some devices when they add a new option */ > + memset(vdev->config, 0xff, vdev->config_len); > + qemu_get_buffer(f, vdev->config, config_len); > + } else { > + int32_t diff; > + /* config_len > vdev->config_len > + * This is rarer, but is here to allow us to fix the case above > + */ > + qemu_get_buffer(f, vdev->config, vdev->config_len); > + /* > + * Even though we expect the diff to be small, we can't use > + * qemu_file_skip because it's not safe for a large skip. > + */ > + for (diff = config_len - vdev->config_len; diff > 0; diff--) { > + qemu_get_byte(f); > + } > + } > + } else { > + qemu_get_buffer(f, vdev->config, vdev->config_len); > } > - qemu_get_buffer(f, vdev->config, vdev->config_len); > > num = qemu_get_be32(f); > >
On Fri, Jun 27, 2014 at 09:34:38AM +0100, Dr. David Alan Gilbert (git) wrote: > From: "Dr. David Alan Gilbert" <dgilbert@redhat.com> > > Commit 'virtio: validate config_len on load' restricted config_len > loaded from the wire to match the config_len that the device had. > > Unfortunately, there are cases where this isn't true, the one > we found it on was the wqe addition in virtio-blk. I think you mean wce. > > Allow mismatched config-lengths: > *) If the version on the wire is shorter then ensure that the > remainder is 0xff filled (as virtio_config_read does on > out of range reads) > *) If the version on the wire is longer, load what we have space > for and skip the rest. > > Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com> Looks good overall, but I am having thoughts about the padding with 0xff. We previously didn't do this (before virtio: validate config_len on load) so it seems safest (at least for 2.1) not to do it now either. > --- > hw/virtio/virtio.c | 30 ++++++++++++++++++++++++++---- > 1 file changed, 26 insertions(+), 4 deletions(-) > > diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c > index a3082d5..2b11142 100644 > --- a/hw/virtio/virtio.c > +++ b/hw/virtio/virtio.c > @@ -927,11 +927,33 @@ int virtio_load(VirtIODevice *vdev, QEMUFile *f) > } > config_len = qemu_get_be32(f); > if (config_len != vdev->config_len) { > - error_report("Unexpected config length 0x%x. Expected 0x%zx", > - config_len, vdev->config_len); > - return -1; > + /* > + * Unfortunately the reality is that there are cases where we > + * see mismatched config lengths, so we have to deal with them > + * rather than rejecting them. > + */ > + Drop extra line please. > + if (config_len < vdev->config_len) { > + /* This is normal in some devices when they add a new option */ > + memset(vdev->config, 0xff, vdev->config_len); > + qemu_get_buffer(f, vdev->config, config_len); > + } else { > + int32_t diff; > + /* config_len > vdev->config_len > + * This is rarer, but is here to allow us to fix the case above > + */ > + qemu_get_buffer(f, vdev->config, vdev->config_len); > + /* > + * Even though we expect the diff to be small, we can't use > + * qemu_file_skip because it's not safe for a large skip. > + */ > + for (diff = config_len - vdev->config_len; diff > 0; diff--) { > + qemu_get_byte(f); > + } > + } > + } else { > + qemu_get_buffer(f, vdev->config, vdev->config_len); > } > - qemu_get_buffer(f, vdev->config, vdev->config_len); > > num = qemu_get_be32(f); So I would say handle config_len < vdev->config_len and config_len == vdev->config_len the same: qemu_get_buffer(f, vdev->config, MIN(config_len, vdev->config_len)); and then skip the remainder if any while (config_len > vdev->config_len) { qemu_get_byte(f); config_len--; } > > -- > 1.9.3
* Michael S. Tsirkin (mst@redhat.com) wrote: > On Fri, Jun 27, 2014 at 09:34:38AM +0100, Dr. David Alan Gilbert (git) wrote: > > From: "Dr. David Alan Gilbert" <dgilbert@redhat.com> > > > > Commit 'virtio: validate config_len on load' restricted config_len > > loaded from the wire to match the config_len that the device had. > > > > Unfortunately, there are cases where this isn't true, the one > > we found it on was the wqe addition in virtio-blk. > > I think you mean wce. Oops - yes. > > Allow mismatched config-lengths: > > *) If the version on the wire is shorter then ensure that the > > remainder is 0xff filled (as virtio_config_read does on > > out of range reads) > > *) If the version on the wire is longer, load what we have space > > for and skip the rest. > > > > Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com> > > Looks good overall, but I am having thoughts about the > padding with 0xff. > We previously didn't do this (before virtio: validate config_len on > load) so it seems safest (at least for 2.1) not to do it now either. Who allocates that memory? If it's known to be a value then I agree; however if it's uninitialised then I think it's best to pick a value rather than have behaviour that depends on random junk in the memory. > > --- > > hw/virtio/virtio.c | 30 ++++++++++++++++++++++++++---- > > 1 file changed, 26 insertions(+), 4 deletions(-) > > > > diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c > > index a3082d5..2b11142 100644 > > --- a/hw/virtio/virtio.c > > +++ b/hw/virtio/virtio.c > > @@ -927,11 +927,33 @@ int virtio_load(VirtIODevice *vdev, QEMUFile *f) > > } > > config_len = qemu_get_be32(f); > > if (config_len != vdev->config_len) { > > - error_report("Unexpected config length 0x%x. Expected 0x%zx", > > - config_len, vdev->config_len); > > - return -1; > > + /* > > + * Unfortunately the reality is that there are cases where we > > + * see mismatched config lengths, so we have to deal with them > > + * rather than rejecting them. > > + */ > > + > > Drop extra line please. > > > + if (config_len < vdev->config_len) { > > + /* This is normal in some devices when they add a new option */ > > + memset(vdev->config, 0xff, vdev->config_len); > > + qemu_get_buffer(f, vdev->config, config_len); > > + } else { > > + int32_t diff; > > + /* config_len > vdev->config_len > > + * This is rarer, but is here to allow us to fix the case above > > + */ > > + qemu_get_buffer(f, vdev->config, vdev->config_len); > > + /* > > + * Even though we expect the diff to be small, we can't use > > + * qemu_file_skip because it's not safe for a large skip. > > + */ > > + for (diff = config_len - vdev->config_len; diff > 0; diff--) { > > + qemu_get_byte(f); > > + } > > + } > > + } else { > > + qemu_get_buffer(f, vdev->config, vdev->config_len); > > } > > - qemu_get_buffer(f, vdev->config, vdev->config_len); > > > > num = qemu_get_be32(f); > > > So I would say handle config_len < vdev->config_len and config_len == > vdev->config_len the same: > > qemu_get_buffer(f, vdev->config, MIN(config_len, vdev->config_len)); > > and then skip the remainder if any > while (config_len > vdev->config_len) { > qemu_get_byte(f); > config_len--; > } That probably still needs that MIN splitting out (int32_t vs size_t); but other than that I guess I can redo that. Dave -- Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
On Fri, Jun 27, 2014 at 03:42:10PM +0100, Dr. David Alan Gilbert wrote: > * Michael S. Tsirkin (mst@redhat.com) wrote: > > On Fri, Jun 27, 2014 at 09:34:38AM +0100, Dr. David Alan Gilbert (git) wrote: > > > From: "Dr. David Alan Gilbert" <dgilbert@redhat.com> > > > > > > Commit 'virtio: validate config_len on load' restricted config_len > > > loaded from the wire to match the config_len that the device had. > > > > > > Unfortunately, there are cases where this isn't true, the one > > > we found it on was the wqe addition in virtio-blk. > > > > I think you mean wce. > > Oops - yes. > > > > Allow mismatched config-lengths: > > > *) If the version on the wire is shorter then ensure that the > > > remainder is 0xff filled (as virtio_config_read does on > > > out of range reads) > > > *) If the version on the wire is longer, load what we have space > > > for and skip the rest. > > > > > > Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com> > > > > Looks good overall, but I am having thoughts about the > > padding with 0xff. > > We previously didn't do this (before virtio: validate config_len on > > load) so it seems safest (at least for 2.1) not to do it now either. > > Who allocates that memory? If it's known to be a value then I agree; however > if it's uninitialised then I think it's best to pick a value rather than > have behaviour that depends on random junk in the memory. It's initialized: e.g. for net it includes the mac, for block the wce value. > > > --- > > > hw/virtio/virtio.c | 30 ++++++++++++++++++++++++++---- > > > 1 file changed, 26 insertions(+), 4 deletions(-) > > > > > > diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c > > > index a3082d5..2b11142 100644 > > > --- a/hw/virtio/virtio.c > > > +++ b/hw/virtio/virtio.c > > > @@ -927,11 +927,33 @@ int virtio_load(VirtIODevice *vdev, QEMUFile *f) > > > } > > > config_len = qemu_get_be32(f); > > > if (config_len != vdev->config_len) { > > > - error_report("Unexpected config length 0x%x. Expected 0x%zx", > > > - config_len, vdev->config_len); > > > - return -1; > > > + /* > > > + * Unfortunately the reality is that there are cases where we > > > + * see mismatched config lengths, so we have to deal with them > > > + * rather than rejecting them. > > > + */ > > > + > > > > Drop extra line please. > > > > > + if (config_len < vdev->config_len) { > > > + /* This is normal in some devices when they add a new option */ > > > + memset(vdev->config, 0xff, vdev->config_len); > > > + qemu_get_buffer(f, vdev->config, config_len); > > > + } else { > > > + int32_t diff; > > > + /* config_len > vdev->config_len > > > + * This is rarer, but is here to allow us to fix the case above > > > + */ > > > + qemu_get_buffer(f, vdev->config, vdev->config_len); > > > + /* > > > + * Even though we expect the diff to be small, we can't use > > > + * qemu_file_skip because it's not safe for a large skip. > > > + */ > > > + for (diff = config_len - vdev->config_len; diff > 0; diff--) { > > > + qemu_get_byte(f); > > > + } > > > + } > > > + } else { > > > + qemu_get_buffer(f, vdev->config, vdev->config_len); > > > } > > > - qemu_get_buffer(f, vdev->config, vdev->config_len); > > > > > > num = qemu_get_be32(f); > > > > > > So I would say handle config_len < vdev->config_len and config_len == > > vdev->config_len the same: > > > > qemu_get_buffer(f, vdev->config, MIN(config_len, vdev->config_len)); > > > > and then skip the remainder if any > > while (config_len > vdev->config_len) { > > qemu_get_byte(f); > > config_len--; > > } > > That probably still needs that MIN splitting out (int32_t vs size_t); but > other than that I guess I can redo that. > > Dave > > -- > Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
* Michael S. Tsirkin (mst@redhat.com) wrote: > On Fri, Jun 27, 2014 at 03:42:10PM +0100, Dr. David Alan Gilbert wrote: > > * Michael S. Tsirkin (mst@redhat.com) wrote: > > > On Fri, Jun 27, 2014 at 09:34:38AM +0100, Dr. David Alan Gilbert (git) wrote: > > > > From: "Dr. David Alan Gilbert" <dgilbert@redhat.com> > > > > > > > > Commit 'virtio: validate config_len on load' restricted config_len > > > > loaded from the wire to match the config_len that the device had. > > > > > > > > Unfortunately, there are cases where this isn't true, the one > > > > we found it on was the wqe addition in virtio-blk. > > > > > > I think you mean wce. > > > > Oops - yes. > > > > > > Allow mismatched config-lengths: > > > > *) If the version on the wire is shorter then ensure that the > > > > remainder is 0xff filled (as virtio_config_read does on > > > > out of range reads) > > > > *) If the version on the wire is longer, load what we have space > > > > for and skip the rest. > > > > > > > > Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com> > > > > > > Looks good overall, but I am having thoughts about the > > > padding with 0xff. > > > We previously didn't do this (before virtio: validate config_len on > > > load) so it seems safest (at least for 2.1) not to do it now either. > > > > Who allocates that memory? If it's known to be a value then I agree; however > > if it's uninitialised then I think it's best to pick a value rather than > > have behaviour that depends on random junk in the memory. > > It's initialized: e.g. for net it includes the mac, > for block the wce value. OK, I've just posted a v2 that is the simplified code you suggested. (Not had quite as much testing as my previous version yet). Dave -- Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c index a3082d5..2b11142 100644 --- a/hw/virtio/virtio.c +++ b/hw/virtio/virtio.c @@ -927,11 +927,33 @@ int virtio_load(VirtIODevice *vdev, QEMUFile *f) } config_len = qemu_get_be32(f); if (config_len != vdev->config_len) { - error_report("Unexpected config length 0x%x. Expected 0x%zx", - config_len, vdev->config_len); - return -1; + /* + * Unfortunately the reality is that there are cases where we + * see mismatched config lengths, so we have to deal with them + * rather than rejecting them. + */ + + if (config_len < vdev->config_len) { + /* This is normal in some devices when they add a new option */ + memset(vdev->config, 0xff, vdev->config_len); + qemu_get_buffer(f, vdev->config, config_len); + } else { + int32_t diff; + /* config_len > vdev->config_len + * This is rarer, but is here to allow us to fix the case above + */ + qemu_get_buffer(f, vdev->config, vdev->config_len); + /* + * Even though we expect the diff to be small, we can't use + * qemu_file_skip because it's not safe for a large skip. + */ + for (diff = config_len - vdev->config_len; diff > 0; diff--) { + qemu_get_byte(f); + } + } + } else { + qemu_get_buffer(f, vdev->config, vdev->config_len); } - qemu_get_buffer(f, vdev->config, vdev->config_len); num = qemu_get_be32(f);