On Thu, Nov 24, 2016 at 4:08 PM, Amir Goldstein <amir73il@xxxxxxxxx> wrote:
> On Thu, Nov 24, 2016 at 3:51 PM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>> On Thu, Nov 24, 2016 at 2:12 PM, Amir Goldstein <amir73il@xxxxxxxxx> wrote:
>>> On Thu, Nov 24, 2016 at 2:03 PM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>>>> On Thu, Nov 24, 2016 at 12:52 PM, Amir Goldstein <amir73il@xxxxxxxxx> wrote:
>>>>> On Thu, Nov 24, 2016 at 12:55 PM, Miklos Szeredi <mszeredi@xxxxxxxxxx> wrote:
>>>>
>>>>>> + /*
>>>>>> + * These should be intercepted, but they are very unlikely to be
>>>>>> + * a problem in practice. Leave them alone for now.
>>>>>
>>>>> It could also be handled in vfs helpers.
>>>>> Since these ops all start with establishing that src and dest are on
>>>>> the same sb,
>>>>> then the cost of copy up of src is the cost of clone_file_range from
>>>>> lower to upper,
>>>>> so it is probably worth to copy up src and leave those fops alone.
>>>>>
>>>>>> + */
>>>>>> + ofop->fops.copy_file_range = orig->copy_file_range;
>>>>>> + ofop->fops.clone_file_range = orig->clone_file_range;
>>>>>> + ofop->fops.dedupe_file_range = orig->dedupe_file_range;
>>>>
>>>> Not sure I understand. Why should we copy up src? Copy up is the
>>>> problem not the solution.
>>>>
>>>
>>> Maybe the idea is ill conceived, but the reasoning is:
>>> To avoid the corner case of cloning from a stale lower src,
>>> call d_real() in vfs helpers to always copy up src before cloning from it
>>> and pass the correct file onwards.
>>
>> Which correct file? src is still the wrong one after calling d_real.
>> We need to clone-open src, just like we do in ovl_read_iter to get the
>> correct file. But then what's the use of copying it up beforehand?
>>
>> We could move the whole logic into the vfs, but I don't really see the point.
>>
Here is a relevant use case (creating several clones),
although not directly related to ro/rw inconsistency, which
justified putting the logic in vfs.
X is a file in lower
lower is different fs then upper
upper supports clone/dedup/copy_range
for i in `seq 1 100`; do cp --reflink=auto X X${i}; done
With current code the src and destination files are on the same
mount (test in ioctl_file_clone), but not on the same sb (test in
vfs_clone_file_range), so cp will fall back to 100 expensive data copies.
*If* instead we d_real() and clone-open src in start of vfs_clone_file_range
*after* verifying the dest file ops support clone, then we will get only one
expensive copy up and 100 cheap clones, so its a big win.
And for the case of src and dst inodes already on the same sb, we can
skip d_real() to avoid possible unneeded copy up, although a clone up
is going to be cheap anyway.
The so called worst case is that this was a one time clone (to X1),
but the cost in this case is not huge - 1 data copy up of X and 1 clone
X->X1 instead of just 1 data copy X->X1, so the difference is negligible.
Now it's true that this is heuristic, but arguably a good one.
Amir.
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
