We hit a regression while rolling out 5.2 internally where we were
hitting the following panic

  kernel BUG at mm/page-writeback.c:2659!
  RIP: 0010:clear_page_dirty_for_io+0xe6/0x1f0
  Call Trace:
   __process_pages_contig+0x25a/0x350
   ? extent_clear_unlock_delalloc+0x43/0x70
   submit_compressed_extents+0x359/0x4d0
   normal_work_helper+0x15a/0x330
   process_one_work+0x1f5/0x3f0
   worker_thread+0x2d/0x3d0
   ? rescuer_thread+0x340/0x340
   kthread+0x111/0x130
   ? kthread_create_on_node+0x60/0x60
   ret_from_fork+0x1f/0x30

This is happening because the page is not locked when doing
clear_page_dirty_for_io.  Looking at the core dump it was because our
async_extent had a ram_size of 24576 but our async_chunk range only
spanned 20480, so we had a whole extra page in our ram_size for our
async_extent.

This happened because we try not to compress pages outside of our
i_size, however a cleanup patch changed us to do

actual_end = min_t(u64, i_size_read(inode), end + 1);

which is problematic because i_size_read() can evaluate to different
values in between checking and assigning.  So either an expanding
truncate or a fallocate could increase our i_size while we're doing
writeout and actual_end would end up being past the range we have
locked.

I confirmed this was what was happening by installing a debug kernel
that had

  actual_end = min_t(u64, i_size_read(inode), end + 1);
  if (actual_end > end + 1) {
	  printk(KERN_ERR "KABOOM\n");
	  actual_end = end + 1;
  }

and installing it onto 500 boxes of the tier that had been seeing the
problem regularly.  Last night I got my debug message and no panic,
confirming what I expected.

[ dsterba: the assembly confirms a tiny race window:

    mov    0x20(%rsp),%rax
    cmp    %rax,0x48(%r15)           # read
    movl   $0x0,0x18(%rsp)
    mov    %rax,%r12
    mov    %r14,%rax
    cmovbe 0x48(%r15),%r12           # eval

  Where r15 is inode and 0x48 is offset of i_size.

  The original fix was to revert 62b37622 that would do an
  intermediate assignment and this would also avoid the doulble
  evaluation but is not future-proof, should the compiler merge the
  stores and call i_size_read anyway.

  There's a patch adding READ_ONCE to i_size_read but that's not being
  applied at the moment and we need to fix the bug. Instead, emulate
  READ_ONCE by two barrier()s that's what effectively happens. The
  assembly confirms single evaluation:

    mov    0x48(%rbp),%rax          # read once
    mov    0x20(%rsp),%rcx
    mov    $0x20,%edx
    cmp    %rax,%rcx
    cmovbe %rcx,%rax
    mov    %rax,(%rsp)
    mov    %rax,%rcx
    mov    %r14,%rax

  Where 0x48(%rbp) is inode->i_size stored to %eax.
]

Fixes: 62b37622

 ("btrfs: Remove isize local variable in compress_file_range")
CC: stable@vger.kernel.org # v5.1+
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ changelog updated ]
Signed-off-by: David Sterba <dsterba@suse.com>