[RFC,0/2] Block and buffer invalidation under a filesystem

Message

Jan Kara Aug. 25, 2020, 12:05 p.m. UTC

Hello,

Recently Ye Bin has reported an ext4 crash which he tracked tracked down to a
problem that several places (block_write_full_page(), fallocate(2) on blkdev,
etc.) can invalidate buffers under a live filesystem - block_invalidatepage()
will clear (among other) BH_Mapped flag and following lookup of the buffer_head
will reinitialize it (init_page_buffers()) which among other things clears
bh->b_private fields which then makes jbd2 crash.

I was thinking how to best fix this. block_write_full_page() is easy to deal
with as the invalidation there is just a relict from the past and we don't need
to invalidate pages there at all (patch 1/2). Other cases are more
questionable. In patch 2/2, I have made fallocate(2) on the block device and
discard ioctls bail with EBUSY if there's filesystem mounted because it seems
very weird and problematic to mess with a block device like that under a
filesystem. What do people think? Is anyone aware of a user that would be
broken by this?

There are also other possibilities of fixing this like making
block_invalidatepage() (or rather new ->invalidatepage callback for the
block device) less aggressive so that it does not discard that much state from
buffer_heads. But details of that are not yet clear to me.

Or other possibilities people see to fix this?

								Honza