On 08/03, Chao Yu wrote: > On 2021/8/3 9:44, Jaegeuk Kim wrote: > > On 08/03, Chao Yu wrote: > > > On 2021/8/3 1:59, Jaegeuk Kim wrote: > > > > On 08/01, Chao Yu wrote: > > > > > On 2021/7/31 6:18, Jaegeuk Kim wrote: > > > > > > On 07/20, Chao Yu wrote: > > > > > > > On 2021/7/20 2:25, Jaegeuk Kim wrote: > > > > > > > > On 07/19, Chao Yu wrote: > > > > > > > > > On 2021/4/27 20:37, Chao Yu wrote: > > > > > > > > > > I think just reverting dirty/free bitmap is not enough if > > > > > > > > > > checkpoint fails, > > > > > > > > > > due to we have updated sbi->cur_cp_pack and nat/sit bitmap, > > > > > > > > > > next CP tries > > > > > > > > > > to overwrite last valid meta/node/data, then filesystem > > > > > > > > > > will be corrupted. > > > > > > > > > > > > > > > > > > > > So I suggest to set cp_error if do_checkpoint() fails until > > > > > > > > > > we can handle > > > > > > > > > > all cases, which is not so easy. > > > > > > > > > > > > > > > > > > > > How do you think? > > > > > > > > > > > > > > > > > > Let's add below patch first before you figure out the patch > > > > > > > > > which covers all > > > > > > > > > things. > > > > > > > > > > > > > > > > > > From 3af957c98e9e04259f8bb93ca0b74ba164f3f27e Mon Sep 17 > > > > > > > > > 00:00:00 2001 > > > > > > > > > From: Chao Yu <[email protected]> > > > > > > > > > Date: Mon, 19 Jul 2021 16:37:44 +0800 > > > > > > > > > Subject: [PATCH] f2fs: fix to stop filesystem update once CP > > > > > > > > > failed > > > > > > > > > > > > > > > > > > During f2fs_write_checkpoint(), once we failed in > > > > > > > > > f2fs_flush_nat_entries() or do_checkpoint(), metadata of > > > > > > > > > filesystem > > > > > > > > > such as prefree bitmap, nat/sit version bitmap won't be > > > > > > > > > recovered, > > > > > > > > > it may cause f2fs image to be inconsistent, let's just set CP > > > > > > > > > error > > > > > > > > > flag to avoid further updates until we figure out a scheme to > > > > > > > > > rollback > > > > > > > > > all metadatas in such condition. > > > > > > > > > > > > > > > > > > Reported-by: Yangtao Li <[email protected]> > > > > > > > > > Signed-off-by: Yangtao Li <[email protected]> > > > > > > > > > Signed-off-by: Chao Yu <[email protected]> > > > > > > > > > --- > > > > > > > > > fs/f2fs/checkpoint.c | 10 +++++++--- > > > > > > > > > 1 file changed, 7 insertions(+), 3 deletions(-) > > > > > > > > > > > > > > > > > > diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c > > > > > > > > > index 6c208108d69c..096c85022f62 100644 > > > > > > > > > --- a/fs/f2fs/checkpoint.c > > > > > > > > > +++ b/fs/f2fs/checkpoint.c > > > > > > > > > @@ -1639,8 +1639,10 @@ int f2fs_write_checkpoint(struct > > > > > > > > > f2fs_sb_info *sbi, struct cp_control *cpc) > > > > > > > > > > > > > > > > > > /* write cached NAT/SIT entries to NAT/SIT area */ > > > > > > > > > err = f2fs_flush_nat_entries(sbi, cpc); > > > > > > > > > - if (err) > > > > > > > > > + if (err) { > > > > > > > > > + f2fs_stop_checkpoint(sbi, false); > > > > > > > > > > > > > > > > I think we should abuse this, since we can get any known ENOMEM > > > > > > > > as well. > > > > > > > > > > > > > > Yup, but one critical issue here is it can break A/B update of > > > > > > > NAT area, > > > > > > > so, in order to fix this hole, how about using NOFAIL memory > > > > > > > allocation > > > > > > > in f2fs_flush_nat_entries() first until we figure out the finial > > > > > > > scheme? > > > > > > > > > > > > NOFAIL is risky, so how about adding a retry logic on ENOMEM with a > > > > > > message > > > > > > and then giving up if we can't get the memory? BTW, what about EIO > > > > > > or other > > > > > > family? > > > > > > > > > > How about this? > > > > > > > > Hmm, it seems we won't get ENOMEM. > > > > > > > > __flush_nat_entry_set > > > > -> get_next_nat_page > > > > -> ... > > > > -> __get_meta_page > > > > -> repeat on ENOMEM, but stop_checkpoint on EIO > > > > > > Correct, I missed to check __get_meta_page() and > > > f2fs_get_meta_page_retry(). > > > > > > > > > > > If we have an error here, we should have stopped checkpoint. Have you > > > > seen other > > > > issue? > > > > > > Still we should fix the case from below path? > > > > > > - f2fs_write_checkpoint > > > - do_checkpoint > > > - f2fs_flush_device_cache failed > > > > What about adding a retry logic to deal with EIO in __submit_flush_wait()? > > We probably need to retry submitting FLUSH commands, and then give up > > with f2fs_stop_checkpoint(). And, then how about adding f2fs_bug_on() if > > f2fs_flush_nat_entries() returns error without f2fs_cp_error()? > > Agreed, how about this? > > From 357da99968f6a31375ea47423b691400c5f66547 Mon Sep 17 00:00:00 2001 > From: Chao Yu <[email protected]> > Date: Mon, 19 Jul 2021 16:37:44 +0800 > Subject: [PATCH v3] f2fs: fix to stop filesystem update once CP failed > > During f2fs_write_checkpoint(), once we failed in > f2fs_flush_nat_entries() or do_checkpoint(), metadata of filesystem > such as prefree bitmap, nat/sit version bitmap won't be recovered, > it may cause f2fs image to be inconsistent, let's just set CP error > flag to avoid further updates until we figure out a scheme to rollback > all metadatas in such condition. > > Reported-by: Yangtao Li <[email protected]> > Signed-off-by: Yangtao Li <[email protected]> > Signed-off-by: Chao Yu <[email protected]> > --- > v3: > - add f2fs_bug_on in error path of f2fs_flush_nat_entries() > - add retry logic in f2fs_flush_device_cache() > fs/f2fs/checkpoint.c | 12 +++++++++--- > fs/f2fs/f2fs.h | 1 + > fs/f2fs/segment.c | 8 +++++++- > 3 files changed, 17 insertions(+), 4 deletions(-) > > diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c > index 6c208108d69c..1d5a2a867282 100644 > --- a/fs/f2fs/checkpoint.c > +++ b/fs/f2fs/checkpoint.c > @@ -1639,8 +1639,11 @@ int f2fs_write_checkpoint(struct f2fs_sb_info *sbi, > struct cp_control *cpc) > > /* write cached NAT/SIT entries to NAT/SIT area */ > err = f2fs_flush_nat_entries(sbi, cpc); > - if (err) > + if (err) { > + f2fs_err(sbi, "f2fs_flush_nat_entries failed err:%d, stop > checkpoint", err); > + f2fs_bug_on(sbi, !f2fs_cp_error(sbi)); > goto stop; > + } > > f2fs_flush_sit_entries(sbi, cpc); > > @@ -1648,10 +1651,13 @@ int f2fs_write_checkpoint(struct f2fs_sb_info *sbi, > struct cp_control *cpc) > f2fs_save_inmem_curseg(sbi); > > err = do_checkpoint(sbi, cpc); > - if (err) > + if (err) { > + f2fs_err(sbi, "do_checkpoint failed err:%d, stop checkpoint", > err); > + f2fs_stop_checkpoint(sbi, false); > f2fs_release_discard_addrs(sbi); > - else > + } else { > f2fs_clear_prefree_segments(sbi, cpc); > + } > > f2fs_restore_inmem_curseg(sbi); > stop: > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h > index 1feef4cb78b6..97eae0e066ee 100644 > --- a/fs/f2fs/f2fs.h > +++ b/fs/f2fs/f2fs.h > @@ -548,6 +548,7 @@ enum { > }; > > #define DEFAULT_RETRY_IO_COUNT 8 /* maximum retry read IO count > */ > +#define DEFAULT_RETRY_FLUSH_COUNT 3 /* maximum retry flush count */ > > /* congestion wait timeout value, default: 20ms */ > #define DEFAULT_IO_TIMEOUT (msecs_to_jiffies(20)) > diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c > index 80f26158e304..37c7a05659fe 100644 > --- a/fs/f2fs/segment.c > +++ b/fs/f2fs/segment.c > @@ -776,9 +776,15 @@ int f2fs_flush_device_cache(struct f2fs_sb_info *sbi) > return 0; > > for (i = 1; i < sbi->s_ndevs; i++) { > + int count = DEFAULT_RETRY_FLUSH_COUNT; > + > if (!f2fs_test_bit(i, (char *)&sbi->dirty_device)) > continue; > - ret = __submit_flush_wait(sbi, FDEV(i).bdev); > + > + do { > + ret = __submit_flush_wait(sbi, FDEV(i).bdev); > + } while (ret && --count); > + > if (ret)
Actually, we need to stop checkpoint here? > break; > > -- > 2.22.1 > _______________________________________________ Linux-f2fs-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
