On Wed, Sep 06, 2023 at 04:33:35PM +0530, Ritesh Harjani wrote: > Matthew Wilcox <willy@xxxxxxxxxxxxx> writes: > > > On Mon, Sep 04, 2023 at 02:08:19AM -0400, Theodore Ts'o wrote: > >> #regzbot introduced: 8147c4c4546f9f05ef03bb839b741473b28bb560 ^ > >> > >> OK, I've isolated the regression of generic/455 failing with ext4/1k > >> to this commit, which came in via the mm tree. Nothing seems > >> *obviously* wrong, but I'm not sure if there are any differences in > >> the semantics of the new folio functions such as kmap_local_folio, > >> offset_in_folio, set_folio_bh() which might be making a difference. > > > > Thanks for the cc, Let's see what we can do ... > > > > virt_to_folio() - For an order-0 page, there is no difference. > > offset_in_folio() - Ditto > > bh->b_page vs bh->b_folio - Ditto > > virt_to_folio() - Ditto > > folio_set_bh() - Ditto > > > > kmap_local_folio() vs kmap_atomic - Here, we have a difference. > > memcpy_from_folio() - Same difference as above. > > > > I suppose it must be this, and yet I cannot understand how it would > > make a difference. Perhaps you can help me? > > > > static inline void *kmap_atomic_prot(struct page *page, pgprot_t prot) > > { > > if (IS_ENABLED(CONFIG_PREEMPT_RT)) > > migrate_disable(); > > else > > preempt_disable(); > > > > pagefault_disable(); > > return __kmap_local_page_prot(page, prot); > > } > > > > vs > > > > static inline void *kmap_local_folio(struct folio *folio, size_t offset) > > { > > struct page *page = folio_page(folio, offset / PAGE_SIZE); > > return __kmap_local_page_prot(page, kmap_prot) + offset % PAGE_SIZE; > > } > > > > I don't believe that returning the address with the offset included > > is the problem here. It must be disabling preemption / migration. > > There's no chace this funcation accesses userspace (... is there?) so > > it can't be the pagefault_disable(). > > > > We can try splitting this up into tiny commits and figuring out which > > of them is the problem. I'll be back at work tomorrow and can look > > more deeply then. > > > >> Using kvm-xfstests[1] I bisected this via the command: > >> > >> % install-kconfig ; kbuild ; kvm-xfstests -c ext4/1k -C 10 generic/455 > >> > >> [1] https://github.com/tytso/xfstests-bld/blob/master/Documentation/kvm-quickstart.md > >> > >> > >> And the bisection pointed me at this commit: > >> > >> commit 8147c4c4546f9f05ef03bb839b741473b28bb560 (refs/bisect/bad) > >> Author: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx> > >> AuthorDate: Thu Jul 13 04:55:11 2023 +0100 > >> Commit: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> > >> CommitDate: Fri Aug 18 10:12:30 2023 -0700 > >> > >> jbd2: use a folio in jbd2_journal_write_metadata_buffer() > >> > > This is inline with my observation too. > > However, is this log expected with below diff when running with ext4/1k? > I am finding a folio with order > 0 here. > > <diff> > diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c > index 768fa05bcbed..152c08e83fa2 100644 > --- a/fs/jbd2/journal.c > +++ b/fs/jbd2/journal.c > @@ -369,6 +369,12 @@ int jbd2_journal_write_metadata_buffer(transaction_t *transaction, > new_offset = offset_in_folio(new_folio, jh2bh(jh_in)->b_data); > } > > + if (folio_size(new_folio) > PAGE_SIZE) { > + pr_crit("%s: folio_size=%lu, folio_order=%d, new_offset=%u bh_size=%lu folio_test_large=%d\n", > + __func__, folio_size(new_folio), folio_order(new_folio), new_offset, > + bh_in->b_size, folio_test_large(new_folio)); > + } > + > mapped_data = kmap_local_folio(new_folio, new_offset); > /* > * Fire data frozen trigger if data already wasn't frozen. Do this > > <dmesg log> > [ 40.419772] jbd2_journal_write_metadata_buffer: folio_size=16384, folio_order=2, new_offset=0 bh_size=1024 folio_test_large=1 > [ 40.444737] jbd2_journal_write_metadata_buffer: folio_size=16384, folio_order=2, new_offset=2048 bh_size=1024 folio_test_large=1 > [ 40.472385] jbd2_journal_write_metadata_buffer: folio_size=16384, folio_order=2, new_offset=3072 bh_size=1024 folio_test_large=1 > [ 40.560581] jbd2_journal_write_metadata_buffer: folio_size=16384, folio_order=2, new_offset=8192 bh_size=1024 folio_test_large=1 > [ 40.588512] jbd2_journal_write_metadata_buffer: folio_size=16384, folio_order=2, new_offset=10240 bh_size=1024 folio_test_large=1 > [ 40.612103] jbd2_journal_write_metadata_buffer: folio_size=16384, folio_order=2, new_offset=7168 bh_size=1024 folio_test_large=1 > [ 40.636800] jbd2_journal_write_metadata_buffer: folio_size=16384, folio_order=2, new_offset=9216 bh_size=1024 folio_test_large=1 > [ 40.661166] jbd2_journal_write_metadata_buffer: folio_size=16384, folio_order=2, new_offset=10240 bh_size=1024 folio_test_large=1 > > > Is this code path a possibility, which can cause above logs? > > ptr = jbd2_alloc() -> kmem_cache_alloc() > <..> > new_folio = virt_to_folio(ptr) > new_offset = offset_in_folio(new_folio, ptr) > > And then I am still not sure what the problem really is? > Is it because at the time of checkpointing, the path is still not fully > converted to folio? Oh yikes! I didn't know that the allocation might come from kmalloc! Yes, slab might use high-order allocations. I'll have to look through this and figure out what the problem might be. Thanks for debugging this.