On 4/28/15 7:24 AM, Lukáš Czerner wrote:
> On Tue, 28 Apr 2015, Jan Kara wrote:
>
>> Date: Tue, 28 Apr 2015 14:21:02 +0200
>> From: Jan Kara <jack@xxxxxxx>
>> To: Eric Sandeen <sandeen@xxxxxxxxxx>
>> Cc: Jan Kara <jack@xxxxxxx>, Andreas Dilger <adilger@xxxxxxxxx>,
>> Lukas Czerner <lczerner@xxxxxxxxxx>, linux-ext4@xxxxxxxxxxxxxxx
>> Subject: Re: [PATCH v2] e2fsprogs: Limit number of reserved gdt blocks on
>> small fs
>>
>> On Mon 27-04-15 11:23:19, Eric Sandeen wrote:
>>> On 4/27/15 11:14 AM, Jan Kara wrote:
>>>> On Fri 24-04-15 22:25:06, Andreas Dilger wrote:
>>>>> On Apr 24, 2015, at 3:51 PM, Eric Sandeen <sandeen@xxxxxxxxxx> wrote:
>>>>>> On 3/25/15 5:46 AM, Lukas Czerner wrote:
>>>>>>> Currently we're unable to online resize very small (smaller than 32 MB)
>>>>>>> file systems with 1k block size because there is not enough space in the
>>>>>>> journal to put all the reserved gdt blocks.
>>>>>>
>>>>>> So, I'll get to the patch review if I need to, but this all seemed a little
>>>>>> odd; this is a regression, so do we really need to restrict things at mkfs
>>>>>> time?
>>>>>>
>>>>>> On the userspace side, things were ok until:
>>>>>>
>>>>>> 9f6ba88 resize2fs: add support for new in-kernel online resize ioctl
>>>>>>
>>>>>> and even with that, on the kernelspace side, things were ok until:
>>>>>>
>>>>>> 8f7d89f jbd2: transaction reservation support
>>>>>>
>>>>>> I guess I'm trying to understand why that jbd2 commit regressed this.
>>>>>> I've not been paying enough attention to ext4 lately. ;)
>>>>>>
>>>>>> I mean, the threshold got chopped in half:
>>>>>>
>>>>>> - if (nblocks > journal->j_max_transaction_buffers) {
>>>>>> + /*
>>>>>> + * 1/2 of transaction can be reserved so we can practically handle
>>>>>> + * only 1/2 of maximum transaction size per operation
>>>>>> + */
>>>>>> + if (WARN_ON(blocks > journal->j_max_transaction_buffers / 2)) {
>>>>>> printk(KERN_ERR "JBD2: %s wants too many credits (%d > %d)\n",
>>>>>> - current->comm, nblocks,
>>>>>> - journal->j_max_transaction_buffers);
>>>>>> + current->comm, blocks,
>>>>>> + journal->j_max_transaction_buffers / 2);
>>>>>> return -ENOSPC;
>>>>>> }
>>>>>>
>>>>>> so it's clear why the behavior changed, I guess, but it feels like I
>>>>>> must be missing something here.
>>>>>
>>>>> Is there some way to reserve these journal blocks only in the case of
>>>>> delalloc usage? This has caused a performance regression with Lustre
>>>>> servers on 3.10 kernels because the journal commits twice as often.
>>>>> We've worked around this for now by doubling the journal size, but it
>>>>> seems a bit of a hack since we can never use the whole journal anymore.
>>>> Hum, so the above hunk only limits maximum number of credits used by a
>>>> single handle. Multiple handles can still consume upto maximum transaction
>>>> size buffers (at least that's the intention :). So I don't see how that can
>>>> cause the problem you describe. What can happen though is that there are
>>>> quite a few outstanding reserved handles and so we have to reserve space
>>>> for them in the running transaction. Do you use dioread_nolock option? That
>>>> enables the use of reserved handles in ext4 for conversion of unwritten
>>>> extents...
>>>
>>> You're probably asking Andreas, but just in case, for my testcase, it's
>>> all defaults & standard options.
>>>
>>> i.e. just this fails, after the above commit, whereas it worked before.
>>>
>>> mkfs.ext4 /dev/sda 20M
>>> mount /dev/sda /mnt/test
>>> resize2fs /dev/sda 200M
>> Yeah, I understand your failure - transaction reservation has reduced
>> max transaction size to a half. After that your fs resize exceeds max
>> transaction size and we are in trouble. I'd prefer solution for that to be
>> in resize code though because it's really a corner case and I wouldn't like
>> to slow down the common transaction start path for it...
>
> Hi Jan,
>
> if you have not already, please see the patch which started the
> discussion.
It just doesn't feel right to me to place limits on fs geometry for this
reason. We could, but it seems like a stopgap. Can we modify the
resize code so that it doesn't need such a large transaction?
In the "old" resize world, I think we played tricks with restarting
transactions, because until the resize was complete and superblocks were
updated, it was ok if we lost updates to new block groups (or something
like that...) i.e. we didn't need one giant atomic update of the filesystem.
Maybe we can do something similar here? I've kind of lost track
of how resize is working now, TBH.
-Eric
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
