On 2019/8/16 下午6:03, Filipe Manana wrote:
> On Fri, Aug 16, 2019 at 10:53 AM Qu Wenruo <quwenruo.btrfs@xxxxxxx> wrote:
>>
>>
>>
>> On 2019/8/16 下午5:33, Filipe Manana wrote:
>>> On Thu, Aug 15, 2019 at 9:36 AM Qu Wenruo <wqu@xxxxxxxx> wrote:
>>>>
>>>> Btrfs has btrfs_end_transaction_throttle() which could try to commit
>>>> transaction when needed.
>>>>
>>>> However under most cases btrfs_end_transaction_throttle() won't really
>>>> commit transaction, due to the hard timing requirement.
>>>>
>>>> Now introduce a new error injection point, btrfs_need_trans_pressure(),
>>>> to allow btrfs_should_end_transaction() to return 1 and
>>>> btrfs_end_transaction_throttle() to fallback to
>>>> btrfs_commit_transaction().
>>>>
>>>> With such more aggressive transaction commit, we can dig deeper into
>>>> cases like snapshot drop.
>>>> Now each reference drop of btrfs_drop_snapshot() will lead to a
>>>> transaction commit, allowing dm-logwrites to catch more details, other
>>>> than one big transaction dropping everything.
>>>>
>>>> Signed-off-by: Qu Wenruo <wqu@xxxxxxxx>
>>>> ---
>>>> Changelog:
>>>> v2:
>>>> - Add comment to explain why this function is needed
>>>> ---
>>>> fs/btrfs/transaction.c | 18 ++++++++++++++++++
>>>> 1 file changed, 18 insertions(+)
>>>>
>>>> diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
>>>> index 3f6811cdf803..8c5471b01d03 100644
>>>> --- a/fs/btrfs/transaction.c
>>>> +++ b/fs/btrfs/transaction.c
>>>> @@ -10,6 +10,7 @@
>>>> #include <linux/pagemap.h>
>>>> #include <linux/blkdev.h>
>>>> #include <linux/uuid.h>
>>>> +#include <linux/error-injection.h>
>>>> #include "ctree.h"
>>>> #include "disk-io.h"
>>>> #include "transaction.h"
>>>> @@ -749,10 +750,25 @@ void btrfs_throttle(struct btrfs_fs_info *fs_info)
>>>> wait_current_trans(fs_info);
>>>> }
>>>>
>>>> +/*
>>>> + * This function is to allow BPF to override the return value so that we can
>>>> + * make btrfs to commit transaction more aggressively.
>>>> + *
>>>> + * It's a debug only feature, mainly used with dm-log-writes to catch more details
>>>> + * of transient operations like balance and subvolume drop.
>>>
>>> Transient? I think you mean long running operations that can span
>>> multiple transactions.
>>
>> Nope, really transient details.
>>
>> E.g catching subvolume dropping for each drop_progress update. While
>> under most one transaction can contain multiple drop_progress update or
>> even the whole tree just get dropped in one transaction.
>>
>>>
>>>> + */
>>>> +static noinline bool btrfs_need_trans_pressure(struct btrfs_trans_handle *trans)
>>>> +{
>>>> + return false;
>>>> +}
>>>> +ALLOW_ERROR_INJECTION(btrfs_need_trans_pressure, TRUE);
>>>
>>> So, I'm not sure if it's really a good idea to have such specific
>>> things like this.
>>> Has this proven useful already? I.e., have you already found any bug using this?
>>
>> Not exactly.
>> I have observed a case where btrfs check gives false alert on missing a
>> backref of a dropped tree.
>
> Wasn't this fixed by Josef in
> https://github.com/kdave/btrfs-progs/commit/42a1aaeec47bc34ae4a923e3e8b2e55b59c01711
> ?
Oh...
> That's normal, the first phase of dropping a tree removes the
> references in the extent tree, and then only in the second phase we
> drop the leafs/nodes that pointed to the extent.
Yes, that's just a false alert from btrfs check, so it is nothing
related to kernel.
>
>>
>> Originally planned to use this feature to catch the exact update, but
>> the problem is, with this pressure, we need an extra ioctl to wait the
>> full subvolume drop to finish.
>
> That, the ioctl to wait (or better, poll) for subvolume removal to
> complete (either all subvolumes or just a specific one), would be
> useful.
OK, would work on that feature.
Thanks,
Qu
>
> Thanks.
>
>>
>> Or we will get the fs unmounted before we really go to DROP_REFERENCE
>> phase, thus dm-log-writes gets nothing interesting.
>>
>> Thanks,
>> Qu
>>
>>>
>>> I often add such similar things myself for testing and debugging, but
>>> because they are so specific, or ugly or verbose, I keep them to
>>> myself.
>>>
>>> Allowing the return value of should_end_transaction() to be
>>> overridden, using the same approach, would be more generic for
>>> example.
>>>
>>> Thanks.
>>>
>>>> +
>>>> static int should_end_transaction(struct btrfs_trans_handle *trans)
>>>> {
>>>> struct btrfs_fs_info *fs_info = trans->fs_info;
>>>>
>>>> + if (btrfs_need_trans_pressure(trans))
>>>> + return 1;
>>>> if (btrfs_check_space_for_delayed_refs(fs_info))
>>>> return 1;
>>>>
>>>> @@ -813,6 +829,8 @@ static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
>>>>
>>>> btrfs_trans_release_chunk_metadata(trans);
>>>>
>>>> + if (throttle && btrfs_need_trans_pressure(trans))
>>>> + return btrfs_commit_transaction(trans);
>>>> if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
>>>> if (throttle)
>>>> return btrfs_commit_transaction(trans);
>>>> --
>>>> 2.22.0
>>>>
>>>
>>>
>>
>
>
Attachment:
signature.asc
Description: OpenPGP digital signature
