io_uring/rsrc: fix folio accounting
authorPavel Begunkov <asml.silence@gmail.com>
Thu, 16 Mar 2023 15:26:05 +0000 (15:26 +0000)
committerJens Axboe <axboe@kernel.dk>
Thu, 16 Mar 2023 15:32:18 +0000 (09:32 -0600)
| BUG: Bad page state in process kworker/u8:0  pfn:5c001
| page:00000000bfda61c8 refcount:0 mapcount:0 mapping:0000000000000000 index:0x20001 pfn:0x5c001
| head:0000000011409842 order:9 entire_mapcount:0 nr_pages_mapped:0 pincount:1
| anon flags: 0x3fffc00000b0004(uptodate|head|mappedtodisk|swapbacked|node=0|zone=0|lastcpupid=0xffff)
| raw: 03fffc0000000000 fffffc0000700001 ffffffff00700903 0000000100000000
| raw: 0000000000000200 0000000000000000 00000000ffffffff 0000000000000000
| head: 03fffc00000b0004 dead000000000100 dead000000000122 ffff00000a809dc1
| head: 0000000000020000 0000000000000000 00000000ffffffff 0000000000000000
| page dumped because: nonzero pincount
| CPU: 3 PID: 9 Comm: kworker/u8:0 Not tainted 6.3.0-rc2-00001-gc6811bf0cd87 #1
| Hardware name: linux,dummy-virt (DT)
| Workqueue: events_unbound io_ring_exit_work
| Call trace:
|  dump_backtrace+0x13c/0x208
|  show_stack+0x34/0x58
|  dump_stack_lvl+0x150/0x1a8
|  dump_stack+0x20/0x30
|  bad_page+0xec/0x238
|  free_tail_pages_check+0x280/0x350
|  free_pcp_prepare+0x60c/0x830
|  free_unref_page+0x50/0x498
|  free_compound_page+0xcc/0x100
|  free_transhuge_page+0x1f0/0x2b8
|  destroy_large_folio+0x80/0xc8
|  __folio_put+0xc4/0xf8
|  gup_put_folio+0xd0/0x250
|  unpin_user_page+0xcc/0x128
|  io_buffer_unmap+0xec/0x2c0
|  __io_sqe_buffers_unregister+0xa4/0x1e0
|  io_ring_exit_work+0x68c/0x1188
|  process_one_work+0x91c/0x1a58
|  worker_thread+0x48c/0xe30
|  kthread+0x278/0x2f0
|  ret_from_fork+0x10/0x20

Mark reports an issue with the recent patches coalescing compound pages
while registering them in io_uring. The reason is that we try to drop
excessive references with folio_put_refs(), but pages were acquired
with pin_user_pages(), which has extra accounting and so should be put
down with matching unpin_user_pages() or at least gup_put_folio().

As a fix unpin_user_pages() all but first page instead, and let's figure
out a better API after.

Fixes: 57bebf807e2abcf8 ("io_uring/rsrc: optimise registered huge pages")
Reported-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Tested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/10efd5507d6d1f05ea0f3c601830e08767e189bd.1678980230.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring/rsrc.c

index 09a16d709cb558a6381b45b6a0d23b0ba7b98960..e2bac9f89902952ea2323ee601cc592b6bedf1bc 100644 (file)
@@ -1235,7 +1235,13 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
                        }
                }
                if (folio) {
-                       folio_put_refs(folio, nr_pages - 1);
+                       /*
+                        * The pages are bound to the folio, it doesn't
+                        * actually unpin them but drops all but one reference,
+                        * which is usually put down by io_buffer_unmap().
+                        * Note, needs a better helper.
+                        */
+                       unpin_user_pages(&pages[1], nr_pages - 1);
                        nr_pages = 1;
                }
        }