Hello Kent Overstreet,
The patch ff35f7b3d0cf: "bcachefs: Better journal tracepoints" from
Jan 15, 2024 (linux-next), leads to the following Smatch static
checker warning:
fs/bcachefs/journal.c:236 __journal_entry_close()
warn: sleeping in atomic context
fs/bcachefs/journal.c
200 static void __journal_entry_close(struct journal *j, unsigned
closed_val, bool trace)
201 {
202 struct bch_fs *c = container_of(j, struct bch_fs, journal);
203 struct journal_buf *buf = journal_cur_buf(j);
204 union journal_res_state old, new;
205 u64 v = atomic64_read(&j->reservations.counter);
206 unsigned sectors;
207
208 BUG_ON(closed_val != JOURNAL_ENTRY_CLOSED_VAL &&
209 closed_val != JOURNAL_ENTRY_ERROR_VAL);
210
211 lockdep_assert_held(&j->lock);
212
213 do {
214 old.v = new.v = v;
215 new.cur_entry_offset = closed_val;
216
217 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL ||
218 old.cur_entry_offset == new.cur_entry_offset)
219 return;
220 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
221 old.v, new.v)) != old.v);
222
223 if (!__journal_entry_is_open(old))
224 return;
225
226 /* Close out old buffer: */
227 buf->data->u64s =
cpu_to_le32(old.cur_entry_offset);
228
229 if (trace_journal_entry_close_enabled() && trace) {
230 struct printbuf pbuf = PRINTBUF;
231 pbuf.atomic++;
232
233 prt_str(&pbuf, "entry size: ");
234 prt_human_readable_u64(&pbuf, vstruct_bytes(buf->data));
235 prt_newline(&pbuf);
--> 236 bch2_prt_task_backtrace(&pbuf, current, 1);
^^^^^^^^^^^^^^^^^^^^^^^
It looks like bch2_prt_task_backtrace() calls kvmalloc() in
__bch2_darray_resize() which is a sleeping function. A lot of of the
callers are holding spinlocks.
bch2_journal_halt() <- disables preempt
bch2_journal_entry_close() <- disables preempt
bch2_journal_flush_seq_async() <- disables preempt
-> journal_entry_want_write()
journal_write_work() <- disables preempt
__journal_res_get() <- disables preempt
bch2_journal_entry_res_resize() <- disables preempt
bch2_journal_flush_seq_async() <- disables preempt <duplicate>
__bch2_next_write_buffer_flush_journal_buf() <- disables preempt
-> __journal_entry_close()
regards,
dan carpenter
