On Tue, 20 Aug 2024 23:16:50 -0400, Lizhi Xu wrote: > > On Tue, 20 Aug 2024 23:00:05 -0400, Kent Overstreet wrote: > > > > When the nr value of a signle entry or their sum overflows, it will > > > > cause the value of ja->nr to be incorrect, this will result in the > > > > allocated memory to ja->buckets being too small, leading to out of > > > > bounds access in bch2_dev_journal_init. > > > > > > > > Reported-by: [email protected] > > > > Signed-off-by: Lizhi Xu <[email protected]> > > > > --- > > > > fs/bcachefs/journal_sb.c | 16 +++++++++++++++- > > > > 1 file changed, 15 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/fs/bcachefs/journal_sb.c b/fs/bcachefs/journal_sb.c > > > > index db80e506e3ab..230ed99130e4 100644 > > > > --- a/fs/bcachefs/journal_sb.c > > > > +++ b/fs/bcachefs/journal_sb.c > > > > @@ -107,6 +107,7 @@ static int bch2_sb_journal_v2_validate(struct > > > > bch_sb *sb, struct bch_sb_field *f > > > > unsigned nr; > > > > unsigned i; > > > > struct u64_range *b; > > > > + u64 total_nr = 0, entry_nr; > > > > > > > > nr = bch2_sb_field_journal_v2_nr_entries(journal); > > > > if (!nr) > > > > @@ -117,8 +118,21 @@ static int bch2_sb_journal_v2_validate(struct > > > > bch_sb *sb, struct bch_sb_field *f > > > > return -BCH_ERR_ENOMEM_sb_journal_v2_validate; > > > > > > > > for (i = 0; i < nr; i++) { > > > > + entry_nr = le64_to_cpu(journal->d[i].nr); > > > > + if (entry_nr > UINT_MAX) { > > > > + prt_printf(err, "Journal v2 entry d[%u] nr %llu > > > > overflow\n", > > > > + i, entry_nr); > > > > + goto err; > > > > + } > > > > > > This check is unnecessary; we know the sum can't overflow a u64 because > > > we're also checking that the entries are nonoverlapping. > > You didn't read my previous email carefully. > > In this issue, journal->d[0] is 7, journal->d[1] is 18446744073709551615, > > so the sum of their u64 type values will definitely overflow. > > It doesn't matter. We're already checking that the entries are > nonoverlapping, and within the range of [1, nbuckets), so total_nr can't > overflow nbuckets, much less an s64 (not that that matters). Are you sure? Or did I not express myself clearly? See the actual running results below: le64_to_cpu(journal->d[1].nr) + le64_to_cpu(journal->d[0].nr) = 7 + 18446744073709551615 will overflow.
When a u64 overflow occurs, total_nr will not be greater than UINT_MAX, so it is not enough to only calculate the total nr of entry to determine. Note: u64 contains 0 ~ 18446744073709551616. BR, Lizhi
