[PATCH 0/5] Speedups and check_setget_bounds reporting updates

[David Sterba <dsterba@suse.com>]

This is a followup to [1] disabling asserts by default. There's some
performance hit involved because of check_setget_bounds that's called
for any extent buffer item read/write. Trading correctness for speed is
should be considered but in general correctness should be favored.

I've optimized the helper so it should have negligible impact (one
inlined branch instead of function call and 2 branches), I haven't
measured that yet though.

How the checks and reporting works:

- any function generated by BTRFS_SETGET_FUNCS calls check_setget_bounds
  and verifies if the reguested bytes are within the extent buffer

- if the checks fail nothing is read or written to the buffer
  - set a bit in fs_info::fs_state so we know something happened
  - store information about the failed range into fs_info for later

- as the setget helpers aren't checked for errors, there are a few
  points where the transaction is ended/aborted once the bit is set

I've tested that in fstests with inserted failure [2] after each 100K
operations, this allows tests to proceed but fail eventually. This makes
i up to btrfs/020 and hits assert in warn_about_uncommitted_trans.

Possible improvements:

- there's a similar check/report for read_extent_buffer so that could
  also set the bit and fail fast

- can be used as base for the shutdown logic

[1] https://lore.kernel.org/linux-btrfs/20200724164147.39925-1-josef@toxicpanda.com

[2]
--- a/fs/btrfs/struct-funcs.c
+++ b/fs/btrfs/struct-funcs.c
@@ -33,7 +33,15 @@ static void report_setget_bounds(const struct extent_buffer *eb,
 static inline bool check_setget_bounds(const struct extent_buffer *eb,
                                       const void *ptr, unsigned off, int size)
 {
-       const unsigned long member_offset = (unsigned long)ptr + off;
+       unsigned long member_offset = (unsigned long)ptr + off;
+       static unsigned long failure = 1;
+       const unsigned long fail_after = 102400;
+
+       if (test_bit(BTRFS_FS_OPEN, &eb->fs_info->flags) &&
+           (failure++ % fail_after) == 0) {
+               printk(KERN_ERR "SETGET: trigger setget failure at %lu\n", failure);
+               member_offset = 128 * 1024;
+       }
---

David Sterba (5):
  btrfs: remove redundant check in up check_setget_bounds
  btrfs: factor out reporting when check_setget_bounds fails
  btrfs: store details about first setget bounds check failure
  btrfs: fail transaction when a setget bounds check failure is detected
  btrfs: move check_setget_bounds out of ASSERT

 fs/btrfs/ctree.h        | 16 +++++++++++--
 fs/btrfs/struct-funcs.c | 51 ++++++++++++++++++++++++++++-------------
 fs/btrfs/transaction.c  | 11 +++++++++
 3 files changed, 60 insertions(+), 18 deletions(-)

-- 
2.34.1