On Thu, Oct 11, 2018 at 2:47 PM, Amir Goldstein <amir73il@xxxxxxxxx> wrote:
> On Thu, Oct 11, 2018 at 3:25 PM Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>>
>> On Thu, Oct 11, 2018 at 1:24 PM, Amir Goldstein <amir73il@xxxxxxxxx> wrote:
>> > Implement the FIGETBSZ query with the statfs(2) filesystem
>> > interface. Reading inode->i_sb->s_blocksize directly is incorrect
>> > on stacked filesystems (i.e. overlayfs).
>> >
>> > This fixes a Floating point exception in e2fsprogs utility filefrag,
>> > which divides the size of the file with the value returned by FIGETBSZ.
>> >
>> > Fixes: d1d04ef8572b ("ovl: stack file ops")
>> > Signed-off-by: Amir Goldstein <amir73il@xxxxxxxxx>
>> > ---
>> >
>> > Miklos,
>> >
>> > Consider this for-rc8?
>> > There is another trivial kfree(ERR_PTR()) fix I posted yesterday.
>> >
>> > xfstests actually has coverage for FIGETBSZ with fiemap-tester
>> > program. However, the tests that use fiemap-tester (generic/094
>> > generic/225) did not fail on v4.19-rc1, but rather thier runtime
>> > dropped from > 10s to 0s, so I'll need to go fix those tests.
>> >
>> > Thanks,
>> > Amir.
>> >
>> > fs/ioctl.c | 15 ++++++++++++++-
>> > 1 file changed, 14 insertions(+), 1 deletion(-)
>> >
>> > diff --git a/fs/ioctl.c b/fs/ioctl.c
>> > index 2005529af560..a0e55a8e94fd 100644
>> > --- a/fs/ioctl.c
>> > +++ b/fs/ioctl.c
>> > @@ -17,6 +17,7 @@
>> > #include <linux/buffer_head.h>
>> > #include <linux/falloc.h>
>> > #include <linux/sched/signal.h>
>> > +#include <linux/statfs.h>
>> >
>> > #include "internal.h"
>> >
>> > @@ -219,6 +220,18 @@ static int ioctl_fiemap(struct file *filp, unsigned long arg)
>> > return error;
>> > }
>> >
>> > +static int ioctl_getbsize(struct file *file, int __user *argp)
>> > +{
>> > + struct kstatfs buf;
>> > + int err;
>> > +
>> > + err = vfs_statfs(&file->f_path, &buf);
>> > + if (err)
>> > + return err;
>> > +
>> > + return put_user(buf.f_bsize, argp);
>>
>> We must be careful not to change the interface for non-overlayfs while
>> trying to fix overlayfs, and I don't see any guarantees, that for
>> every other fs f_bsize would be the same as s_blocksize. There are
>> other, minor side effects, like possibly blocking and/or generating
>> network traffic, unlike the previous version.
>>
>> But first let's understand what this interface is for. Haven't found
>> any other documentation than this:
>>
>> #define FIGETBSZ _IO(0x00,2) /* get the block size used for bmap */
>>
>> So it's for the deprecated bmap interface, and we no longer support
>> that. Hooray! Then just let's just put some dummy value into
>> overlayfs's s_blocksize and s_blocksize_bits (e.g.
>> PAGE_SIZE/PAGE_SHIFT) and be done with that.
>>
>
> OK. That's fine by me, but notice that filefrag will use the value
> it gets from FIGETBSZ even in the FIEMAP implementation,
> which is the default implementation.
>
> I think we can set s_blocksize to upper s_blocksize and cover some
> common cases. If s_blocksize doesn't match lower s_blocksize we
> can warn and set s_blocksize back to 0.
>
> Then, we can return -EINVAL from FIGETBSZ in case s_blocksize
> is not set, which is the equivalent of what we do with FIBMAP.
>
> It that too much?
Does any other filesystem leave s_blocksize unset? Apparently ceph
does, for example, so that's not without precedent.
Just returning -EINVAL from FIGETBSZ if s_blocksize is zero would fix
filefrag, since it would gracefully fall back to f_bsize.
Thanks,
Miklos
