Re: [PATCH] block: flush the disk cache on BLKFLSBUF

[Mikulas Patocka <mpatocka@redhat.com>]

> On Mon, 26 Jun 2023, Mikulas Patocka wrote:
> 
> > The BLKFLSBUF ioctl doesn't send the flush bio to the block device, thus
> > flushed data may be lurking in the disk cache and they may not be really
> > flushed to the stable storage.
> > 
> > This patch adds the call to blkdev_issue_flush to blkdev_flushbuf.
> 
> Umm, why?  This is an ioctl no one should be using, and we certainly
> should not add new functionality to it.  Can you explain what you're
> trying to do here?

Marc Smith reported a bug where he wrote to the dm-writecache target using 
O_DIRECT, then reset the machine without proper shutdown and the freshly 
written data were lost. It turned out that he didn't use the fsync or 
fdatasync syscall (and dm-writecache makes its metadata persistent on a 
FLUSH bio).

When I was analyzing this issue, it turned out that there is no easy way 
how to send the FLUSH bio to a block device from a command line.

The sync command synchronizes only filesystems, it doesn't flush cache for 
unmounted block devices (do you think that it should flush block devices 
too?).

The "blockdev --flushbufs" command also doesn't send the FLUSH bio, but I 
would expect it to send it. Without sending the FLUSH bio, "blockdev 
--flushbufs" doesn't really guarantee anything.

Mikulas