On 2020/2/17 13:08, 王矛 wrote: > > Hi Chao, > > Thanks for your prompt reply. > 1. Is it possible that when the issue occurred, there is indeed all free > segment > used up?
No, f2fs reserved some sections for foreground GC, no one should use up those reserved sections, though foreground GC may use them temporarily to migrate blocks. > (End user used up all free segments, system core service(uid root) used up > all reserved segments) > > If this maybe true, it would lead such crash issue, is this right? > I mean if we can add more detailed check to see if there is free segment > and > return ENOSPC in time before hit crash. > 2. In data/node write path, it will call has_not_enough_free_secs(sbi, 0, 0) > to > make sure there is enough segments. That is for reclaim path in where wbc->for_reclaim should be true. But anyway, in data/node write path, we may call f2fs_balance_fs() to check foreground GC water line. > But in this crash callstack, it crashed in > do_write_page()->f2fs_allocate_data_block(). > Is there any corner case that has_not_enough_free_secs() is missed during this > write procedure? I've no idea, one way is to check whether there is a determinate method to reproduce this issue, if there is, maybe I can check this issue in more details. Thanks, > > Thanks, > Mao > > > At 2020-02-17 11:50:56, "Chao Yu" <yuch...@huawei.com> wrote: >>Hi 王矛， >> >>On 2020/2/16 19:39, 王矛 wrote: >>> *So the problem is:* >>> >>> 1. in new_curseg(), if the segno allocated is invalid(no free segment, max >>> segno >>> is returned). >>> >>> F2fs should do something to indicate this exception. >>> >>> 2. otherwise, we may hit the f2fs panic(se invalid). >>> >>> Maybe we should do sanity check in update_sit_entry() to see if segno is >>> really >>> out of range and caused this panic. >> >>I'm afraid it's too late to handle such error in update_sit_entry(), since we >>expect all procedures in do_write_page() will be successful, it's a little >>hard >>to handle such error in that context. >> >>So the problem here is why we can not find any free segments w/ LFS >>allocation, >>because in case of lack of free segments (check via >>has_not_enough_free_secs()), >>f2fs will force to trigger f2fs_gc() to recycle free sections. >> >>I doubt there may be some corner case we haven't considered, result all free >>segments (including reserved free segments) was exhausted by data/node writes >>when last checkpoint is triggered during umount. >> >>If this issue can be reproduced (during umount, free_segments() < >>reserved_segments()), we can add some logs to see why f2fs_balance_fs() fail >>to >>recycle enough free segments previously. >> >>Thanks, >> >>> >>> >>> Thanks, >>> Mao >>> >>> >>> >>> >>> > > > > > _______________________________________________ Linux-f2fs-devel mailing list Linux-f2fs-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
