Re: [PATCH] block: make sure last_lookup set as NULL after part deleted

[Ming Lei <ming.lei@redhat.com>]

On Fri, Jan 03, 2020 at 11:06:25AM +0800, Hou Tao wrote:
> Hi,
> 
> On 2020/1/2 9:23, Ming Lei wrote:
> > On Tue, Dec 31, 2019 at 10:55:47PM +0800, Hou Tao wrote:
> >> Hi,
> >>
> >> On 2019/12/31 19:09, Yufen Yu wrote:
> >>> When delete partition executes concurrently with IOs issue,
> >>> it may cause use-after-free on part in disk_map_sector_rcu()
> >>> as following:
> >> snip
> >>
> >>>
> >>> diff --git a/block/genhd.c b/block/genhd.c
> >>> index ff6268970ddc..39fa8999905f 100644
> >>> --- a/block/genhd.c
> >>> +++ b/block/genhd.c
> >>> @@ -293,7 +293,23 @@ struct hd_struct *disk_map_sector_rcu(struct gendisk *disk, sector_t sector)
> >>>  		part = rcu_dereference(ptbl->part[i]);
> >>>  
> >>>  		if (part && sector_in_part(part, sector)) {
> >> snip
> >>
> >>>  			rcu_assign_pointer(ptbl->last_lookup, part);
> >>> +			part = rcu_dereference(ptbl->part[i]);
> >>> +			if (part == NULL) {
> >>> +				rcu_assign_pointer(ptbl->last_lookup, NULL);
> >>> +				break;
> >>> +			}
> >>>  			return part;
> >>>  		}
> >>>  	}
> >>
> >> Not ensure whether the re-read can handle the following case or not:
> >>
> We have written a similar test case for the following case and found out that
> process C still may got the freed hd_struct pointer from process A. So
> the re-read will not resolve the problem.
> 
> >> process A                                 process B                          process C
> >>
> >> disk_map_sector_rcu():                    delete_partition():               disk_map_sector_rcu():
> >>
> >> rcu_read_lock
> >>
> >>   // need to iterate partition table
> >>   part[i] != NULL   (1)                   part[i] = NULL (2)
> >>                                           smp_mb()
> >>                                           last_lookup = NULL (3)
> >>                                           call_rcu()  (4)
> >>     last_lookup = part[i] (5)
> >>
> >>
> >>                                                                              rcu_read_lock()
> >>                                                                              read last_lookup return part[i] (6)
> >>                                                                              sector_in_part() is OK (7)
> >>                                                                              return part[i] (8)
> >>
> >>   part[i] == NULL (9)
> >>       last_lookup = NULL (10)
> >>   rcu_read_unlock() (11)
> >>                                            one RCU grace period completes
> >>                                            __delete_partition() (12)
> >>                                            free hd_partition (13)
> >>                                                                              // use-after-free
> >>                                                                              hd_struct_try_get(part[i])  (14)
> >>
> >> * the number in the parenthesis is the sequence of events.
> >>
> 
> 
> 
> >> Maybe RCU experts can shed some light on this problem, so cc +paulmck@kernel.org, +joel@joelfernandes.org and +RCU maillist.
> >>
> >> If the above case is possible, maybe we can fix the problem by pinning last_lookup through increasing its ref-count
> >> (the following patch is only compile tested):
> >>
> >> diff --git a/block/genhd.c b/block/genhd.c
> >> index 6e8543ca6912..179e0056fae1 100644
> >> --- a/block/genhd.c
> >> +++ b/block/genhd.c
> >> @@ -279,7 +279,14 @@ struct hd_struct *disk_map_sector_rcu(struct gendisk *disk, sector_t sector)
> >>  		part = rcu_dereference(ptbl->part[i]);
> >>
> >>  		if (part && sector_in_part(part, sector)) {
> >> -			rcu_assign_pointer(ptbl->last_lookup, part);
> >> +			struct hd_struct *old;
> >> +
> >> +			if (!hd_struct_try_get(part))
> >> +				break;
> >> +
> >> +			old = xchg(&ptbl->last_lookup, part);
> >> +			if (old)
> >> +				hd_struct_put(old);
> >>  			return part;
> >>  		}
> >>  	}
> >> @@ -1231,7 +1238,11 @@ static void disk_replace_part_tbl(struct gendisk *disk,
> >>  	rcu_assign_pointer(disk->part_tbl, new_ptbl);
> >>
> >>  	if (old_ptbl) {
> >> -		rcu_assign_pointer(old_ptbl->last_lookup, NULL);
> >> +		struct hd_struct *part;
> >> +
> >> +		part = xchg(&old_ptbl->last_lookup, NULL);
> >> +		if (part)
> >> +			hd_struct_put(part);
> >>  		kfree_rcu(old_ptbl, rcu_head);
> >>  	}
> >>  }
> >> diff --git a/block/partition-generic.c b/block/partition-generic.c
> >> index 98d60a59b843..441c1c591c04 100644
> >> --- a/block/partition-generic.c
> >> +++ b/block/partition-generic.c
> >> @@ -285,7 +285,8 @@ void delete_partition(struct gendisk *disk, int partno)
> >>  		return;
> >>
> >>  	rcu_assign_pointer(ptbl->part[partno], NULL);
> >> -	rcu_assign_pointer(ptbl->last_lookup, NULL);
> >> +	if (cmpxchg(&ptbl->last_lookup, part, NULL) == part)
> >> +		hd_struct_put(part);
> >>  	kobject_put(part->holder_dir);
> >>  	device_del(part_to_dev(part));
> > 
> > IMO this approach looks good.
> >
> Not sure about the overhead when there are concurrent IOs on different partitions,
> we will measure that.
> 
> We have got a seemingly better solution: caching the index of last_lookup in tbl->part[]
> instead of caching the pointer itself, so we can ensure the validity of returned pointer
> by ensuring it's not NULL in tbl->part[] as does when last_lookup is NULL or 0.

Thinking of the problem further, looks we don't need to hold ref for
.last_lookup.

What we need is to make sure the partition's ref is increased just
before assigning .last_lookup, so how about something like the following?

diff --git a/block/blk-core.c b/block/blk-core.c
index 089e890ab208..79599f5fd5b7 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -1365,18 +1365,6 @@ void blk_account_io_start(struct request *rq, bool new_io)
 		part_stat_inc(part, merges[rw]);
 	} else {
 		part = disk_map_sector_rcu(rq->rq_disk, blk_rq_pos(rq));
-		if (!hd_struct_try_get(part)) {
-			/*
-			 * The partition is already being removed,
-			 * the request will be accounted on the disk only
-			 *
-			 * We take a reference on disk->part0 although that
-			 * partition will never be deleted, so we can treat
-			 * it as any other partition.
-			 */
-			part = &rq->rq_disk->part0;
-			hd_struct_get(part);
-		}
 		part_inc_in_flight(rq->q, part, rw);
 		rq->part = part;
 	}
diff --git a/block/genhd.c b/block/genhd.c
index ff6268970ddc..21f4a9b8d24d 100644
--- a/block/genhd.c
+++ b/block/genhd.c
@@ -286,17 +286,24 @@ struct hd_struct *disk_map_sector_rcu(struct gendisk *disk, sector_t sector)
 	ptbl = rcu_dereference(disk->part_tbl);
 
 	part = rcu_dereference(ptbl->last_lookup);
-	if (part && sector_in_part(part, sector))
+	if (part && sector_in_part(part, sector)) {
+		if (!hd_struct_try_get(part))
+			goto exit;
 		return part;
+	}
 
 	for (i = 1; i < ptbl->len; i++) {
 		part = rcu_dereference(ptbl->part[i]);
 
 		if (part && sector_in_part(part, sector)) {
+                       if (!hd_struct_try_get(part))
+                               goto exit;
 			rcu_assign_pointer(ptbl->last_lookup, part);
 			return part;
 		}
 	}
+ exit:
+	hd_struct_get(&disk->part0);
 	return &disk->part0;
 }
 EXPORT_SYMBOL_GPL(disk_map_sector_rcu);


> 
> > Given partition is actually protected by percpu-refcount now, I guess the
> > RCU annotation for referencing ->part[partno] and ->last_lookup may not
> > be necessary, together with the part->rcu_work.
> > 
> So we will depends on the invocation of of call_rcu() on __percpu_ref_switch_mode() to
> ensure the RCU readers will find part[i] is NULL before trying to increasing
> the atomic ref-counter of part[i], right ?

Yeah.

Thanks,
Ming