Re: [bugfix ping2] Re: [PATCH v2 0/2] fix qcow2_can_store_new_dirty_bitmap

[John Snow]

On 12/10/19 8:24 AM, Max Reitz wrote:
> On 10.12.19 09:11, Max Reitz wrote:
>> On 09.12.19 23:03, Eric Blake wrote:
>>> On 12/9/19 11:58 AM, Max Reitz wrote:
>>>> On 09.12.19 17:30, Max Reitz wrote:
>>>>> On 02.12.19 15:09, Vladimir Sementsov-Ogievskiy wrote:
>>>>>> Hi again!
>>>>>>
>>>>>> Still forgotten bug-fix :(
>>>>>>
>>>>>> Is it too late for 4.2?
>>>>>
>>>>> Sorry. :-/
>>>>>
>>>>> Yes, I think I just forgot it.  I don’t think it’s too important for
>>>>> 4.2, so, well, it isn’t too bad, but...  Sorry.
>>>>>
>>>>>> I can't imagine better test, and it tests exactly what written in
>>>>>> https://bugzilla.redhat.com/show_bug.cgi?id=1712636
>>>>>>
>>>>>> (Hmm, actually, I doubt that it is real use-case, more probably it's
>>>>>> a bug in management layer)
>>>>>>
>>>>>> So, take this with test or without test, to 4.2 or 5.0.
>>>>>
>>>>> I was thinking of seeing whether I could write a quicker test, but of
>>>>> course we should take the patch either way.
>>>>
>>>> OK, I give up.  It’s very much possible to create an image with 65535
>>>> bitmaps very quickly (like, under a second) outside of qemu, but just
>>>> opening it takes 2:30 min (because of the quadratic complexity of
>>>> checking whether a bitmap of the same name already exists).
>>>
>>> Can we fix that to use a hash table for amortized O(1) lookup rather
>>> than the current O(n) lookup?
>>
>> Not unreasonable, considering that this is probably what we would’ve
>> done from the start in any language where hash tables are built in.
>>
>> But OTOH when you have 66k bitmaps, you probably have other problems.
>> Like, writes being incredibly slow, because all those bitmaps have to be
>> updated.
>>
>> (Well, you can technically have 99 % of them disabled, but who’d do such
>> a thing?)
>>
>> ((Maybe I’ll look into it.))
> 
> Hmm, now I did.  This gets the test down to 24 s.  Still not sure
> whether it’s worth it, though...
> 
> Max
> 

I agree we very likely have other problems once we reach resource usage
of this level.


Still, if we want to make this blazing fast for the love of doing so:

(1) Read in the directory *once*, and cache it. We have avoided doing
this largely to feel more confident that the code is correct and is
never working on an "outdated" version of the directory.

[On cache invalidation, we can write the directory back out to the
bitmap, and delete our cache. The next time we need the list, we can
reload it. This should alleviate consistency concerns.]


(2) Store the entries in an rbtree! 65536 entries is only ~16 lookups
maximum in the worst case. I took a look at the linux rbtree
implementation and did a very quick back-of-the-envelope benchmarking
for inserting strings (len=32) into a tree:

name generation 53151 usec
insert [0-10] 5 usec
insert [10-100] 14 usec
insert [100-1000] 195 usec
insert [1000-10000] 2919 usec
insert [10000-65536] 41485 usec

This seems fast enough that we're likely going to be eclipsed just by
other string handling concerns.

--js