On 12/01/2021 15:36, David Sterba wrote:

On Fri, Jan 08, 2021 at 12:02:48AM +0000, Martin Raiber wrote:

quoted

When reading from btrfs file via io_uring I get following
call traces:

Is there a way to reproduce by common tools (fio) or is a specialized
one needed?

I'm not familiar with this particular issue, but:
should _probably_ be reproducible with fio with io_uring engine
or fio/t/io_uring tool.

quoted

[<0>] wait_on_page_bit+0x12b/0x270
[<0>] read_extent_buffer_pages+0x2ad/0x360
[<0>] btree_read_extent_buffer_pages+0x97/0x110
[<0>] read_tree_block+0x36/0x60
[<0>] read_block_for_search.isra.0+0x1a9/0x360
[<0>] btrfs_search_slot+0x23d/0x9f0
[<0>] btrfs_lookup_csum+0x75/0x170
[<0>] btrfs_lookup_bio_sums+0x23d/0x630
[<0>] btrfs_submit_data_bio+0x109/0x180
[<0>] submit_one_bio+0x44/0x70
[<0>] extent_readahead+0x37a/0x3a0
[<0>] read_pages+0x8e/0x1f0
[<0>] page_cache_ra_unbounded+0x1aa/0x1f0
[<0>] generic_file_buffered_read+0x3eb/0x830
[<0>] io_iter_do_read+0x1a/0x40
[<0>] io_read+0xde/0x350
[<0>] io_issue_sqe+0x5cd/0xed0
[<0>] __io_queue_sqe+0xf9/0x370
[<0>] io_submit_sqes+0x637/0x910
[<0>] __x64_sys_io_uring_enter+0x22e/0x390
[<0>] do_syscall_64+0x33/0x80
[<0>] entry_SYSCALL_64_after_hwframe+0x44/0xa9

Prevent those by setting IOCB_NOIO before calling
generic_file_buffered_read.

Async read has the same problem. So disable that by removing
FMODE_BUF_RASYNC. This was added with commit
8730f12b7962b21ea9ad2756abce1e205d22db84 ("btrfs: flag files as

Oh yeah that's the commit that went to btrfs code out-of-band. I am not
familiar with the io_uring support and have no good idea what the new
flag was supposed to do.

iirc, Jens did make buffered IO asynchronous by waiting on a page
with wait_page_queue, but don't remember well enough.

quoted

supporting buffered async reads") with 5.9. Io_uring will read
the data via worker threads if it can't be read without sync IO
this way.

What are the implications of that? Like more context switching (due to
the worker threads) or other potential performance related problems?

io_uring splits submission and completion steps and usually expect
submissions to happen quick and not block (at least for long),
otherwise it can't submit other requests, that reduces QD and so
forth. In the worst case it can serialise it to QD1. I guess the
same can be applied to AIO.

-- 
Pavel Begunkov