The btrfs writepages function collects a large range of pages flagged
for delayed allocation, and then sends them down through the COW code
for processing.  When compression is on, we allocate one async_chunk
structure for every 512K, and then run those pages through the
compression code for IO submission.

writepages starts all of this off with a single page, locked by the
original call to extent_write_cache_pages(), and it's important to keep
track of this page because it has already been through
clear_page_dirty_for_io().

The btrfs async_chunk struct has a pointer to the locked_page, and when
we're redirtying the page because compression had to fallback to
uncompressed IO, we use page->index to decide if a given async_chunk
struct really owns that page.

But, this is racey.  If a given delalloc range is broken up into two
async_chunks (chunkA and chunkB), we can end up with something like
this:

 compress_file_range(chunkA)
 submit_compress_extents(chunkA)
 submit compressed bios(chunkA)
 put_page(locked_page)

				 compress_file_range(chunkB)
				 ...

Or:

 async_cow_submit
  submit_compressed_extents <--- falls back to buffered writeout
   cow_file_range
    extent_clear_unlock_delalloc
     __process_pages_contig
       put_page(locked_pages)

					    async_cow_submit

The end result is that chunkA is completed and cleaned up before chunkB
even starts processing.  This means we can free locked_page() and reuse
it elsewhere.  If we get really lucky, it'll have the same page->index
in its new home as it did before.

While we're processing chunkB, we might decide we need to fall back to
uncompressed IO, and so compress_file_range() will call
__set_page_dirty_nobufers() on chunkB->locked_page.

Without cgroups in use, this creates as a phantom dirty page, which
isn't great but isn't the end of the world. What can happen, it can go
through the fixup worker and the whole COW machinery again:

in submit_compressed_extents():
  while (async extents) {
  ...
    cow_file_range
    if (!page_started ...)
      extent_write_locked_range
    else if (...)
      unlock_page
    continue;

This hasn't been observed in practice but is still possible.

With cgroups in use, we might crash in the accounting code because
page->mapping->i_wb isn't set.

  BUG: unable to handle kernel NULL pointer dereference at 00000000000000d0
  IP: percpu_counter_add_batch+0x11/0x70
  PGD 66534e067 P4D 66534e067 PUD 66534f067 PMD 0
  Oops: 0000 [#1] SMP DEBUG_PAGEALLOC
  CPU: 16 PID: 2172 Comm: rm Not tainted
  RIP: 0010:percpu_counter_add_batch+0x11/0x70
  RSP: 0018:ffffc9000a97bbe0 EFLAGS: 00010286
  RAX: 0000000000000005 RBX: 0000000000000090 RCX: 0000000000026115
  RDX: 0000000000000030 RSI: ffffffffffffffff RDI: 0000000000000090
  RBP: 0000000000000000 R08: fffffffffffffff5 R09: 0000000000000000
  R10: 00000000000260c0 R11: ffff881037fc26c0 R12: ffffffffffffffff
  R13: ffff880fe4111548 R14: ffffc9000a97bc90 R15: 0000000000000001
  FS:  00007f5503ced480(0000) GS:ffff880ff7200000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00000000000000d0 CR3: 00000001e0459005 CR4: 0000000000360ee0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   account_page_cleaned+0x15b/0x1f0
   __cancel_dirty_page+0x146/0x200
   truncate_cleanup_page+0x92/0xb0
   truncate_inode_pages_range+0x202/0x7d0
   btrfs_evict_inode+0x92/0x5a0
   evict+0xc1/0x190
   do_unlinkat+0x176/0x280
   do_syscall_64+0x63/0x1a0
   entry_SYSCALL_64_after_hwframe+0x42/0xb7

The fix here is to make asyc_chunk->locked_page NULL everywhere but the
one async_chunk struct that's allowed to do things to the locked page.

Link: https://lore.kernel.org/linux-btrfs/c2419d01-5c84-3fb4-189e-4db519d08796@suse.com/
Fixes: 771ed689 ("Btrfs: Optimize compressed writeback and reads")
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Chris Mason <clm@fb.com>
[ update changelog from mail thread discussion ]
Signed-off-by: David Sterba <dsterba@suse.com>