On Sat, Jan 2, 2021 at 5:53 AM Dave Chinner [off-list ref] wrote:

On Fri, Jan 01, 2021 at 05:10:49PM +0800, Yafang Shao wrote:

quoted

On Thu, Dec 31, 2020 at 10:46 AM kernel test robot
[off-list ref] wrote:

.....

quoted

quoted

[  552.905799] XFS: Assertion failed: !current->journal_info, file: fs/xfs/xfs_trans.h, line: 280
[  553.104459]  xfs_trans_reserve+0x225/0x320 [xfs]
[  553.110556]  xfs_trans_roll+0x6e/0xe0 [xfs]
[  553.116134]  xfs_defer_trans_roll+0x104/0x2a0 [xfs]
[  553.122489]  ? xfs_extent_free_create_intent+0x62/0xc0 [xfs]
[  553.129780]  xfs_defer_finish_noroll+0xb8/0x620 [xfs]
[  553.136299]  xfs_defer_finish+0x11/0xa0 [xfs]
[  553.142017]  xfs_itruncate_extents_flags+0x141/0x440 [xfs]
[  553.149053]  xfs_setattr_size+0x3da/0x480 [xfs]
[  553.154939]  ? setattr_prepare+0x6a/0x1e0
[  553.160250]  xfs_vn_setattr+0x70/0x120 [xfs]
[  553.165833]  notify_change+0x364/0x500
[  553.170820]  ? do_truncate+0x76/0xe0
[  553.175673]  do_truncate+0x76/0xe0
[  553.180184]  path_openat+0xe6c/0x10a0
[  553.184981]  do_filp_open+0x91/0x100
[  553.189707]  ? __check_object_size+0x136/0x160
[  553.195493]  do_sys_openat2+0x20d/0x2e0
[  553.200481]  do_sys_open+0x44/0x80
[  553.204926]  do_syscall_64+0x33/0x40
[  553.209588]  entry_SYSCALL_64_after_hwframe+0x44/0xa9

Thanks for the report.

At a first glance, it seems we should make a similar change as we did
in xfs_trans_context_clear().

static inline void
xfs_trans_context_set(struct xfs_trans *tp)
{
    /*
     * We have already handed over the context via xfs_trans_context_swap().
     */
    if (current->journal_info)
        return;
    current->journal_info = tp;
    tp->t_pflags = memalloc_nofs_save();
}

Ah, no.

Remember how I said "split out the wrapping of transaction
context setup in xfs_trans_reserve() from
the lifting of the context setting into xfs_trans_alloc()"?

Well, you did the former and dropped the latter out of the patch
set.

I misunderstood what you mean.

Now when a transaction rolls after xfs_trans_context_swap(), it
calls xfs_trans_reserve() and tries to do transaction context setup
work inside a transaction context that already exists.  IOWs, you
need to put the patch that lifts of the context setting up into
xfs_trans_alloc() back into the patchset before adding the
current->journal functionality patch.

Sure.

Also, you need to test XFS code with CONFIG_XFS_DEBUG=y so that
asserts are actually built into the code and exercised, because this
ASSERT should have fired on the first rolling transaction that the
kernel executes...

I really forgot to enable CONFIG_XFS_DEBUG...   -_-b


-- 
Thanks
Yafang