[Devel] [RFC PATCH vz9 v6 00/62] ploop optimistations and scalling

[Pavel Tikhomirov]

On 1/10/25 15:55, Pavel Tikhomirov wrote:
> 
> 
> On 1/10/25 15:52, Alexander Atanasov wrote:
>> On 10.01.25 9:43, Pavel Tikhomirov wrote:
>>> I didn't find specific patch for this comment so I'll leave it to 
>>> patch -00.
>>>
>>> What if ploop_dispatch_pios and do_ploop_run_work(do_ploop_work) run 
>>> concurrently for one ploop on different cpus? Would not it be bad if
>>> ploop_dispatch_pios only dispatched part of pios and 
>>> do_ploop_run_work already starts processing them, or if 
>>> do_ploop_run_work yet does not "see" all pios from already fully 
>>> dispatched ploop_dispatch_pios? That can lead to arbitrary pio 
>>> reordering.
>>
>> This is not a problem if we reorder pios - i had the same worries and 
>> even tried to make the to be executed in sequence but it is not 
>> required. Filesystems are receiving and handling out of order requests 
>> and the callers must wait for completion before issuing further 
>> requests. Inside ploop completion is organized so when all split pios 
>> are done then the completion is performed.
>>
>> Said in another way - if a caller issues two writes or read and write 
>> to the same place and he did not wait completion it is not our problem.
>>
> 
> If a caller issues write and sync in this order, should the caller 
> expect that on receiving both completions sync was done after write?
> 

For history, message from Alexander in slack:

 > if you issue a write and a sync before write is completed you have no 
warranty that write is completed before sync

that is reasonable so reordering here between do_ploop_run_work and 
ploop_dispatch_pios is completely fine.

>>
>>
>>>
>>> For instance imagine:
>>>
>>> ploop_dispatch_pios () {
>>>                            do_ploop_run_work () {
>>>                              llcow_pios = llist_del_all(...);
>>>                              ...
>>>
>>>    add data pio #1
>>>    add data pio #2
>>>    add flush pio #3
>>>                              llflush_pios = llist_del_all(...);
>>>                              process_ploop_fsync_work(llflush_pios);
>>>                            }
>>> }
>>>
>>> This way a flush pio #3 will be processed before data pios #1 and #2.
>>>
>>
> 

-- 
Best regards, Tikhomirov Pavel
Senior Software Developer, Virtuozzo.