On Wed, 21 Mar 2018, Jens Axboe wrote:

On 3/21/18 10:42 AM, Mikulas Patocka wrote:

quoted

Early alpha processors cannot write a single byte or word; they read 8
bytes, modify the value in registers and write back 8 bytes.

The type blk_status_t is defined as one byte, it is often written
asynchronously by I/O completion routines, this asynchronous modification
can corrupt content of nearby bytes if these nearby bytes can be written
simultaneously by another CPU.

- one example of such corruption is the structure dm_io where
  "blk_status_t status" is written by an asynchronous completion routine
  and "atomic_t io_count" is modified synchronously
- another example is the structure dm_buffer where "unsigned hold_count"
  is modified synchronously from process context and "blk_status_t
  write_error" is modified asynchronously from bio completion routine

This patch fixes the bug by changing the type blk_status_t to 32 bits if
we are on Alpha and if we are compiling for a processor that doesn't have
the byte-word-extension.

That's nasty. Is alpha the only problematic arch here?

Yes. All the other architectures supported by Linux have byte writes.

As to the patch in question, normally I'd just say we should make it
unconditionally u32. But we pack so nicely in the bio, and I don't think
the bio itself has this issue as the rest of the members that share this
word are all set before the bio is submitted. But callers embedding
the status var in other structures don't necessarily have that
guarantee, as your dm examples show.

-- 
Jens Axboe

Keeping blk_status_t 8-bit for most architectures will save a few bytes in 
some of device mapper structures.

Mikulas