P.S.
On 16-08-19 11:01, Hans de Goede wrote:
On 16-08-19 09:56, Christoph Hellwig wrote:
<snip>
+ * Ideally we would wrap generic_file_read_iter with a function which also
+ * does this check, to reduce the chance of us missing writes happening on the
+ * host side after open(). But the vboxsf stat call to the host only works on
+ * filenames, so that would require caching the filename in our
+ * file->private_data and there is no guarantee that filename will still
+ * be valid at read_iter time. So this would be in no way bulletproof.
Well, you can usually generate a file name from file->f_path.dentry.
The only odd case is opened by unliked files. NFS has a special hack
for those called sillyrename (you can grep for that).
Right, so since the unlink or a normal rename could happen on the host side
and there is no notification of that, those will be 2 areas where a stat
call to verify will fail, which leaves us with 3 options:
1) Make stat calls before read() calls, if they fail purge the cache to be safe
2) Make stat calls before read(), on failure ignore the stat call
3) Treat read() calls like other page-cache reads such as sendfile or mmap
and only check if the cache is stale at open() time.
I just realized there is a 4th option, which is to make vboxsf read_iter
simply always do:
4) Always evict non dirt pages from page-cache from read_iter, to ensure we
get the latest version from the host:
So something like this:
/*
* Evict non dirt pages from page-cache, so that we reget them from
* the host in case they have been changed.
* /
invalidate_mapping_pages(inode->i_mapping, pos >> PAGE_SHIFT,
(pos + len) >> PAGE_SHIFT);
generic_file_read_iter(...)
This is in essence what the out of tree driver is doing on read(), except
that it does not go through the page-cache at all.
I think this might be the best option, perhaps controlled by a mount
flag ?
Regards,
Hans
