On Mon, Apr 6, 2015 at 9:42 PM, Jarod Wilson <jarod@xxxxxxxxxx> wrote:
> On Mon, Apr 06, 2015 at 12:40:12AM +0800, Ming Lei wrote:
>> On Mon, Apr 6, 2015 at 12:12 AM, Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
>> >> +/*
>> >> + * This is exported as API for block driver, can be called
>> >> + * with requiring bd_mutex or not.
>> >> + */
>> >> +int __blkdev_reread_part(struct block_device *bdev, bool lock)
>> >> {
>> >> struct gendisk *disk = bdev->bd_disk;
>> >> int res;
>> >> @@ -159,12 +163,14 @@ static int blkdev_reread_part(struct block_device *bdev)
>> >> return -EINVAL;
>> >> if (!capable(CAP_SYS_ADMIN))
>> >> return -EACCES;
>> >> - if (!mutex_trylock(&bdev->bd_mutex))
>> >> + if (lock && !mutex_trylock(&bdev->bd_mutex))
>> >> return -EBUSY;
>> >
>> > Please don't add funtions that do conditional locking, instead move
>> > all the code into blkdev_reread_part_nolock, and then wrap it:
>> >
>> > int blkdev_reread_part(struct block_device *bdev)
>> > {
>> > if (!mutex_trylock(&bdev->bd_mutex))
>> > return -EBUSY;
>> > blkdev_reread_part_nolock(bdev);
>> > mutex_unlock(&bdev->bd_mutex);
>> > }
>>
>> Yes, it is more clean, but with extra acquiring lock cost for the
>> failure cases, especially when we replace trylock with mutex_lock().
>
> I was working on a version of this myself over the past few days, I
> actually removed blkdev_reread_part() entirely, renamed
> fs/partition-generic.c::reread_partitions() to __reread_partitions(), then
> moved the locking from blkdev_reread_part() into a new reread_partitions()
> that wrapped around __reread_partitions(). Same difference, I guess.
>
>> > Please also add a lockdep_assert_held to blkdev_reread_part_nolock to
>> > ensure callers actually do hold the lock.
>>
>> Good point!
>
> Looks like fs/block_dev.c::__blkdev_get() is the only thing that would be
> calling the _nolock variant of whichever route, as it handles bd_mutex
> acquisition within __blkdev_get().
I guess you forget __blkdev_put(), :-)
>
> As an aside, there's a piece of that function that could be worth
> duplicating over into loop.c as well:
>
> if (bdev->bd_invalidated) {
> if (!ret)
> rescan_partitions(bdev);
> else if (ret == -ENOMEDIUM)
> invalidate_partitions(disk, bdev);
>
> Might this possibly be put to use to help with the problem commit
> 8761a3dc1f07b163414e2215a2cadbb4cfe2a107 was trying to solve?
I am wondering if the problem claimed in this commit exists in reality,
at least fdisk need to run reread partition first before adding partition.
- LO_FLAGS_PARTSCAN is set for both 'losetup -P' and max_parts
- if max_parts isn't set, GENHD_FL_NO_PART_SCAN is set, so user
can't reread partition successfully because of disk_part_scan_enabled().
If there is really the problem, it can be fixed by exporting
rescan_partitions or the approach in commit 8761a3dc
with not acquiring bd_mutex in release().
Thanks,
Ming Lei
>
> --
> Jarod Wilson
> jarod@xxxxxxxxxx
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
--
To unsubscribe from this list: send the line "unsubscribe linux-s390" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
