[bug report] bcachefs: sleeping in atomic

[Dan Carpenter]

Hello Kent Overstreet,

The patch 93a640e2570b: "bcachefs: GFP_NOIO -> GFP_NOFS" from May 28,
2023 (linux-next), leads to the following Smatch static checker
warning:

fs/bcachefs/journal_io.c:1461 journal_buf_realloc() warn: sleeping in atomic 
context
fs/bcachefs/journal_io.c:1662 bch2_journal_entries_postprocess() warn: sleeping 
in atomic context
fs/bcachefs/journal_io.c:1764 bch2_journal_write() warn: sleeping in atomic 
context
fs/bcachefs/journal_io.c:788 jset_validate() warn: sleeping in atomic context
fs/bcachefs/replicas.c:439 bch2_mark_replicas() warn: sleeping in atomic context

All this sleeping in atomic warnings start in journal_write_done()
where it takes spin_lock(&j->lock) so preempt is disabled.

The call tree for the first warning is:

journal_write_done() <- disables preempt
-> bch2_journal_write()
   -> journal_buf_realloc()

fs/bcachefs/journal_io.c
    1452 static void journal_buf_realloc(struct journal *j, struct journal_buf 
*buf)
    1453 {
    1454         /* we aren't holding j->lock: */
    1455         unsigned new_size = READ_ONCE(j->buf_size_want);
    1456         void *new_buf;
    1457 
    1458         if (buf->buf_size >= new_size)
    1459                 return;
    1460 
--> 1461         new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);

This sleeps.  GFP_NOWAIT and GFP_ATOMIC don't sleep.  But kvpmalloc() is
always going to have to be treated as a sleeping function because it
allocates giant amounts of memory.

    1462         if (!new_buf)
    1463                 return;
    1464 
    1465         memcpy(new_buf, buf->data, buf->buf_size);
    1466 
    1467         spin_lock(&j->lock);
    1468         swap(buf->data,                new_buf);
    1469         swap(buf->buf_size,        new_size);
    1470         spin_unlock(&j->lock);
    1471 
    1472         kvpfree(new_buf, new_size);
    1473 }

regards,
dan carpenter