Hi,
On 2020/1/6 16:11, Ming Lei wrote:
> On Mon, Jan 06, 2020 at 03:39:07PM +0800, Yufen Yu wrote:
>> Hi, Ming
>>
>> On 2020/1/3 23:16, Ming Lei wrote:
>>> Hello Yufen,
>>>
>>> OK, we still can move clearing .last_lookup into __delete_partition(),
>>> at that time all IO path can observe the partition percpu-refcount killed.
>>>
>>> Also the rcu work fn is run after one RCU grace period, at that time,
>>> the NULL .last_lookup becomes visible in all IO path too.
>>>
>>> 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..e3dec90b1f43 100644
>>> --- a/block/genhd.c
>>> +++ b/block/genhd.c
>>> @@ -286,17 +286,21 @@ 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) && hd_struct_try_get(part))
>>> 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);
>>> diff --git a/block/partition-generic.c b/block/partition-generic.c
>>> index 1d20c9cf213f..1739f750dbf2 100644
>>> --- a/block/partition-generic.c
>>> +++ b/block/partition-generic.c
>>> @@ -262,6 +262,12 @@ static void delete_partition_work_fn(struct work_struct *work)
>>> void __delete_partition(struct percpu_ref *ref)
>>> {
>>> struct hd_struct *part = container_of(ref, struct hd_struct, ref);
>>> + struct disk_part_tbl *ptbl =
>>> + rcu_dereference_protected(part->disk->part_tbl, 1);
>>> +
>>> + rcu_assign_pointer(ptbl->last_lookup, NULL);
>>> + put_device(disk_to_dev(part->disk));
>>> +
>>> INIT_RCU_WORK(&part->rcu_work, delete_partition_work_fn);
>>> queue_rcu_work(system_wq, &part->rcu_work);
>>> }
>>> @@ -283,8 +289,9 @@ void delete_partition(struct gendisk *disk, int partno)
>>> if (!part)
>>> return;
>>> + get_device(disk_to_dev(disk));
>>> rcu_assign_pointer(ptbl->part[partno], NULL);
>>> - rcu_assign_pointer(ptbl->last_lookup, NULL);
>>> +
>>> kobject_put(part->holder_dir);
>>> device_del(part_to_dev(part));
>>> @@ -349,6 +356,7 @@ struct hd_struct *add_partition(struct gendisk *disk, int partno,
>>> p->nr_sects = len;
>>> p->partno = partno;
>>> p->policy = get_disk_ro(disk);
>>> + p->disk = disk;
>>> if (info) {
>>> struct partition_meta_info *pinfo = alloc_part_info(disk);
>>> diff --git a/include/linux/genhd.h b/include/linux/genhd.h
>>> index 8bb63027e4d6..66660ec5e8ee 100644
>>> --- a/include/linux/genhd.h
>>> +++ b/include/linux/genhd.h
>>> @@ -129,6 +129,7 @@ struct hd_struct {
>>> #else
>>> struct disk_stats dkstats;
>>> #endif
>>> + struct gendisk *disk;
>>> struct percpu_ref ref;
>>> struct rcu_work rcu_work;
>>> };
>>
>>
>> IMO, this change can solve the problem. But, __delete_partition will
>> depend on the implementation of disk_release(). If disk .release modify
>> as blocked in the future, then __delete_partition will also be blocked,
>> which is not expected in rcu callback function.
>
> __delete_partition() won't be blocked because it just calls queue_rcu_work() to
> release the partition instance in wq context.
>
>>
>> We may cache index of part[] instead of part[i] itself to fix the use-after-free bug.
>> https://patchwork.kernel.org/patch/11318767/
>
> That approach can fix the issue too, but extra overhead is added in the
> fast path because partition retrieval is changed to the following way:
>
> + last_lookup = READ_ONCE(ptbl->last_lookup);
> + if (last_lookup > 0 && last_lookup < ptbl->len) {
> + part = rcu_dereference(ptbl->part[last_lookup]);
> + if (part && sector_in_part(part, sector))
> + return part;
> + }
>
> from
> part = rcu_dereference(ptbl->last_lookup);
>
> So ptbl->part[] has to be fetched, it is fine if the ->part[] array
> shares same cacheline with ptbl->last_lookup, but one disk may have
> too many partitions, then your approach may introduce one extra cache
> miss every time.
>
Yes. The solution you proposed also adds an invocation of percpu_ref_tryget_live()
in the fast path. Not sure which one will have a better performance. However the
reason we prefer the index caching is the simplicity instead of performance.
> READ_ONCE() may imply one read barrier too.
>
According to the manual, it only matters on Alpha machines and READ_ONCE() also
exists in rcu_dereference().
Regards,
Tao
>
> Thanks,
> Ming
>
>
> .
>
