On May 11, 2021 / 14:36, Ming Lei wrote: > Hello Shinichiro, > > Thanks for your test! > > On Tue, May 11, 2021 at 05:05:52AM +0000, Shinichiro Kawasaki wrote: > > On May 07, 2021 / 22:42, Ming Lei wrote: > > > Hi Jens, > > > > > > This patchset fixes the request UAF issue by one simple approach, > > > without clearing ->rqs[] in fast path, please consider it for 5.13. > > > > > > 1) grab request's ref before calling ->fn in blk_mq_tagset_busy_iter, > > > and release it after calling ->fn, so ->fn won't be called for one > > > request if its queue is frozen, done in 2st patch > > > > > > 2) clearing any stale request referred in ->rqs[] before freeing the > > > request pool, one per-tags spinlock is added for protecting > > > grabbing request ref vs. clearing ->rqs[tag], so UAF by refcount_inc_not_zero > > > in bt_tags_iter() is avoided, done in 3rd patch. > > > > Ming, thank you for your effort to fix the UAF issue. I applied the V6 series to > > the kernel v5.13-rc1, and confirmed that the series avoids the UAF I had been > > observing with blktests block/005 and HDD behind HBA. This is good. However, I > > found that the series triggered block/029 hang. Let me share the kernel message > > below, which was printed at the hang. KASAN reported null-ptr-deref. > > > > [ 2124.489023] run blktests block/029 at 2021-05-11 13:42:22 > > [ 2124.561386] null_blk: module loaded > > [ 2125.201166] general protection fault, probably for non-canonical address 0xdffffc0000000012: 0000 [#1] SMP KASAN PTI > > [ 2125.212387] KASAN: null-ptr-deref in range [0x0000000000000090-0x0000000000000097] > > It is because this hw queue isn't mapped yet and new added hw queue is > mapped in blk_mq_map_swqueue(), and the following change can fix it, and > I will post V7 after careful test. > > > diff --git a/block/blk-mq.c b/block/blk-mq.c > index fcd5ed79011f..691b555c26fa 100644 > --- a/block/blk-mq.c > +++ b/block/blk-mq.c > @@ -2652,6 +2652,10 @@ static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, > int i; > unsigned long flags; > > + /* return if hw queue isn't mapped */ > + if (!tags) > + return; > + > WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0); > > for (i = 0; i < queue_depth; i++) Thank you Ming. I confirmed that this change avoids the hang at block/029. -- Best Regards, Shin'ichiro Kawasaki
