Re: [Qemu-devel] [PATCH v4 3/5] 9p: Added virtfs option "remap_inodes"

[Christian Schoenebeck]

On Freitag, 28. Juni 2019 12:09:31 CEST Greg Kurz wrote:
> On Wed, 26 Jun 2019 20:42:13 +0200
> 
> Christian Schoenebeck via Qemu-devel <address@hidden> wrote:
> > To support multiple devices on the 9p share, and avoid
> > qid path collisions we take the device id as input
> > to generate a unique QID path. The lowest 48 bits of
> > the path will be set equal to the file inode, and the
> > top bits will be uniquely assigned based on the top
> > 16 bits of the inode and the device id.
> > 
> > Signed-off-by: Antonios Motakis <address@hidden>
> 
> Same remark about changes to the original patch.

ack_once();   :)

> BTW, I had a concern with the way v9fs_do_readdir() open-codes QID
> generation without calling stat_to_qid().
> 
> See discussion here:
> 
> https://lists.gnu.org/archive/html/qemu-devel/2018-02/msg02724.html
> 
> I guess you should ensure in a preliminary patch that QIDs only
> come out of stat_to_qid().

Mja, actually I first omitted your suggestion consciously, because I first 
thought it was an overkill pure visibility issue lmited to the default case 
remap_inodes==false, but now that I look at it again, it is actually an issue 
even when remap_inodes==true since dirent would expose wrong inode numbers on 
guest as well.

I will see what to do about it. However about your other concern here, quote:

        "Also, if we hit a collision while reading the directory, I'm
         afraid the remaining entries won't be read at all. I'm not
         sure this is really what we want."

That's however still a concern here that I would consider overkill to address. 
I mean if a user gets into that situation then because of a configuration error 
that must be corrected by user; the point of this patch set is to prevent 
undefined behaviour and to make the user aware about the root cause of the 
overall issue; the purpose is not to address all possible issues while there 
is still a configuration error.

> > +static int qid_path_prefixmap(V9fsPDU *pdu, const struct stat *stbuf,
> > +                                uint64_t *path)
> > +{
> > +    QppEntry lookup = {
> > +        .dev = stbuf->st_dev,
> > +        .ino_prefix = (uint16_t) (stbuf->st_ino >> 48)
> > +    }, *val;
> > +    uint32_t hash = qpp_hash(lookup);
> > +
> > +    val = qht_lookup(&pdu->s->qpp_table, &lookup, hash);
> > +
> > +    if (!val) {
> > +        if (pdu->s->qp_prefix_next == 0) {
> > +            /* we ran out of prefixes */
> 
> And we won't ever be able to allocate a new one. Maybe worth
> adding an error_report_once() to inform the user ?

Yeah, I thought about that as well. Will do.

> >  static int stat_to_qid(V9fsPDU *pdu, const struct stat *stbuf, V9fsQID
> >  *qidp) {
> > 
> > -    size_t size;
> > +    int err;
> > 
> > -    if (pdu->s->dev_id == 0) {
> > -        pdu->s->dev_id = stbuf->st_dev;
> > -    } else if (pdu->s->dev_id != stbuf->st_dev) {
> > -        error_report_once(
> > -            "9p: Multiple devices detected in same VirtFS export. "
> > -            "You must use a separate export for each device."
> > -        );
> > -        return -ENOSYS;
> > +    if (pdu->s->ctx.export_flags & V9FS_REMAP_INODES) {
> > +        /* map inode+device to qid path (fast path) */
> > +        err = qid_path_prefixmap(pdu, stbuf, &qidp->path);
> > +        if (err) {
> > +            return err;
> > +        }
> > +    } else {
> > +        if (pdu->s->dev_id == 0) {
> > +            pdu->s->dev_id = stbuf->st_dev;
> > +        } else if (pdu->s->dev_id != stbuf->st_dev) {
> > +            error_report_once(
> > +                "9p: Multiple devices detected in same VirtFS export. "
> > +                "You must either use a separate export for each device "
> > +                "shared from host or enable virtfs option
> > 'remap_inodes'."
> > +            );
> > +            return -ENOSYS;
> > +        }
> > +        size_t size;
> 
> From CODING_STYLE:
> 
> 5. Declarations
> 
> Mixed declarations (interleaving statements and declarations within
> blocks) are generally not allowed; declarations should be at the beginning
> of blocks.
> 
> Please do so for "size" and add an extra blank line.

Ok.

> > +#define QPATH_INO_MASK        (((unsigned long)1 << 48) - 1)
> 
> This won't give the expected result on a 32-bit host. Since this
> is a mask for 64-bit entities, it should rather be:
> 
> #define QPATH_INO_MASK        ((1ULL << 48) - 1)

Correct, will fix it.

> > diff --git a/qemu-options.hx b/qemu-options.hx
> > index 0d8beb4afd..e7ea136da1 100644
> > --- a/qemu-options.hx
> > +++ b/qemu-options.hx
> > @@ -1334,7 +1334,7 @@ ETEXI
> > 
> >  DEF("virtfs", HAS_ARG, QEMU_OPTION_virtfs,
> >  
> >      "-virtfs
> >      local,path=path,mount_tag=tag,security_model=mapped-xattr|mapped-fil
> >      e|passthrough|none\n"> 
> > -    "       
> > [,id=id][,writeout=immediate][,readonly][,fmode=fmode][,dmode=dmode]\n" +
> >    "       
> > [,id=id][,writeout=immediate][,readonly][,fmode=fmode][,dmode=dmode][,rem
> > ap_inodes]\n"
> This feature applies to all backends IIUC. We don't really care for the
> synth backend since it generates non-colliding inode numbers by design,
> but the proxy backend has the same issue as local. So...

Yeah, I was not sure about these, because I did not even know what these two 
were for exactly. :)  [ lazyness disclaimer end]

Will do for the other manual locations you mentioned as well.

Best regards,
Christian Schoenebeck