Re: [PATCH] cifs: add support for fallocate mode 0 for non-sparse files

[ronnie sahlberg <ronniesahlberg@gmail.com>]

On Fri, Jan 17, 2020 at 10:27 AM Pavel Shilovsky <piastryyy@gmail.com> wrote:
>

Thanks.

> чт, 16 янв. 2020 г. в 01:05, ronnie sahlberg <ronniesahlberg@gmail.com>:
> >
> > The bug is basically that if we extend a file by fallocate mode==0
> > and immediately afterwards mmap() the file we will only mmap() as much
> > as end-of-file was
> > prior to the truncate  and then if we try to touch any
> > address in this extended region userspace dies with bus error.
> >
> > The patch added "extend a file with fallocate mode==0 for NON-Sparse
> > files" and caused xfstest to fail.
> > The fix is to force us to revalidate the file attributes (the size is
> > the important one) when we extend the file so
> > mmap() will work properly.
> > I have fixed this in the patch and will resend tomorrow after some more testing.
> >
> > Looking for other SMB2_set_eof() callsites I see we already had the
> > same bug for the case of extending a SPARSE
>
> I agree that regardless of file being sparse or not, we should somehow
> update a size in the VFS after calling SMB2_set_eof().
>
> > file using fallocate mode==0. I fixed that too and will audit all
> > other plases where we use SMB2_set_eof()
> > to see if they are safe or not before resending.
>
> One of those places where SMB2_set_eof() is called is
> cifs_set_file_size() which does call the following after getting a
> successful response from the server:
>
> 2250 >-------if (rc == 0) {
> 2251 >------->-------cifsInode->server_eof = attrs->ia_size;
> 2252 >------->-------cifs_setsize(inode, attrs->ia_size);
> 2253 >------->-------cifs_truncate_page(inode->i_mapping, inode->i_size);
> 2254 >-------}
>
> This is called by cifs_setattr_[no]unix() which does the following afterwards:
>
> 2569 >-------if ((attrs->ia_valid & ATTR_SIZE) &&
> 2570 >-------    attrs->ia_size != i_size_read(inode))
> 2571 >------->-------truncate_setsize(inode, attrs->ia_size);
>
> truncate_setsize() does  similar things as cifs_setsize() besides
> setting cinode->time to 0. This code path probably needs to be
> refactored. But putting this aside, for the current fallocate fix I
> think we should use the same existing mechanism to update a file size
> in the VFS without revalidating the inode.

I can do that change. Note however that since we are still setting
cinode->time to 0
we are still going to trigger a revalidation at some stage.

>
> --
> Best regards,
> Pavel Shilovsky