On 2025/01/10 17:33, Michael S. Tsirkin wrote:
On Fri, Jan 10, 2025 at 01:38:06PM +0900, Akihiko Odaki wrote:
On 2025/01/09 21:46, Willem de Bruijn wrote:
Akihiko Odaki wrote:
On 2025/01/09 16:31, Michael S. Tsirkin wrote:
On Thu, Jan 09, 2025 at 03:58:44PM +0900, Akihiko Odaki wrote:
tun used to simply advance iov_iter when it needs to pad virtio header,
which leaves the garbage in the buffer as is. This is especially
problematic when tun starts to allow enabling the hash reporting
feature; even if the feature is enabled, the packet may lack a hash
value and may contain a hole in the virtio header because the packet
arrived before the feature gets enabled or does not contain the
header fields to be hashed. If the hole is not filled with zero, it is
impossible to tell if the packet lacks a hash value.
Zero is a valid hash value, so cannot be used as an indication that
hashing is inactive.
Zeroing will initialize the hash_report field to
VIRTIO_NET_HASH_REPORT_NONE, which tells it does not have a hash value.
In theory, a user of tun can fill the buffer with zero before calling
read() to avoid such a problem, but leaving the garbage in the buffer is
awkward anyway so fill the buffer in tun.
Signed-off-by: Akihiko Odaki <akihiko.odaki@xxxxxxxxxx>
But if the user did it, you have just overwritten his value,
did you not?
Yes. but that means the user expects some part of buffer is not filled
after read() or recvmsg(). I'm a bit worried that not filling the buffer
may break assumptions others (especially the filesystem and socket
infrastructures in the kernel) may have.
If this is user memory that is ignored by the kernel, just reflected
back, then there is no need in general to zero it. There are many such
instances, also in msg_control.
More specifically, is there any instance of recvmsg() implementation which
returns N and does not fill the complete N bytes of msg_iter?
The one in tun. It was a silly idea but it has been here for years now.
Except tun. If there is such an example of recvmsg() implementation and
it is not accidental and people have agreed to keep it functioning, we
can confidently say this construct is safe without fearing pushback from
people maintaining filesystem/networking infrastructure. Ultimately I
want those people decide if this can be supported for the future or not.
If not zeroing leads to ambiguity with the new feature, that would be
a reason to add it -- it is always safe to do so.
If we are really confident that it will not cause problems, this
behavior can be opt-in based on a flag or we can just write some
documentation warning userspace programmers to initialize the buffer.
