在 2022/12/26 19:23, Dave Chinner 写道: > > Yup, but did you notice that xfs_trans_alloc() is doing GFP_KERNEL > allocation from a context that is doing filesystem work on behalf of > memory reclaim? > > The right fix is to make the inodegc workers use > memalloc_nofs_save() context, similar to what is done in > xfs_end_ioend(), as both the IO completion workers and the inodegc > workers can be doing work on behalf of memory reclaim.... > Yes, you're right, this does solve the problem. Dave, Thanks for the explanation. I will send V2 patch. > -Dave. >