On Fri, 15 Apr 2016 13:40:41 -0700 Cedric BAIL <[email protected]> said:
> On Wed, Apr 13, 2016 at 4:37 PM, Carsten Haitzler <[email protected]> > wrote: > > On Wed, 13 Apr 2016 11:16:58 -0700 Cedric BAIL <[email protected]> said: > >> On Wed, Apr 13, 2016 at 3:47 AM, Carsten Haitzler <[email protected]> > >> wrote: > >> > i'm wondering... shouldn't we make eo and eo base "threadsafe". ie a > >> > basic eo object should have locks around things like: > >> > >> I am not to sure of the intended idea of making eo threadsafe for all > >> of those function. If you are using the same object from two > >> differents thread, it is pretty certain in my opinion that you will > >> need to have lock to manage its use yourself or you will end up with > >> crash. I don't think it is possible to build a threadsafe and safe to > >> use API that won't require the user of that API to actually think > >> about where to put lock. So if we advertise our API as threadsafe and > >> the developer don't understand the limit of it (Like most Java > >> developers for example), he will manage to have race condition and > >> blame Eo for not being properly threadsafe. I believe it will result > >> in a more complex use of eo, when can I rely on it being threadsafe, > >> when can't I ? Which is why I don't think this is a good idea. > > > > you CANNOT build a threadsafe object ever if the base class isn't > > threadsafe. i can NEVER do it. it's impossible to do. > > I do not believe it is possible to have a generic threadsafe object in > C anyway. You need external synchronisation primitive or you will have > race condition when passing parameter and returning value. Some > primitive can be threadsafe, but the concept of threadsafety is > unlikely to be broad. example: thread1: obj = eo_add(..); eo_data_key_set(obj, "stuff", data); send_msg(thread2, obj); // send msg to thread 2 to use obj // some time later for (;;) { eo_data_key_set(obj, "otherstuff", data2); sleep(1); datax = eo_data_key_get(obj, "otherstuff"); eo_data_key_del(obj, "otherstuff"); } thread2: wait_msg(); for (;;) { data3 = eo_data_get_get(obj, "stuff"); process(data3); } this above SHOULD work if obj was intended to be threadsafe. reality is we are psetered day in and out to make efl threadsafe. all the time. reality is that for gui studff this is nigh on impossible because of the model of keeping state unrendered until going idle so changes need to be done within a lock of some sort within the loop that owns the ui. most of efl is ui. but let's talk about things like making the main loop an object so different threads can have different loops. so you can do efl_loop_begin/end(mainloop); we have talked about having inter-loop messaging similarly etc. this requires threadsafe objects. if eo_base and below is not threadsafe then we are screwed in ever being able to do this. you cannot have the above code which has no race conditions at all because thread1 has guaranteed "stuff" is stored before thread2 ever gets the message to use it and it doesnt ever mess with it again. because keys are implemented using a list and a key lookup brings the key to the front of the list (LRU) then 2 threads messing with the same list is going to end up with a crash despite the programmer doing all the right things. forcing them to create locks for every object separately to the object is just poor poor poor design. that's another bit of data to remember, store, delete/clean up when object is cleaned up and so on. > > simple stuff like i am setting data keys in one thread and getting their > > value in another... > > Will always be a race condition if you don't have any synchronisation > primitive between both side. The parameter of your function call and > the returned value are always, by construction, outside of any > potential lock, making it impossible to be threadsafe. We tried that > with ecore and we all know it failed for that exact reason. no you don't. see above. > > i'm not saying to make all of efl threadsafe. but base class and core eo > > stuff like the object table have to be to allow someone to build threadsafe > > objects on top. > > > > we already want to make the main loop "thread safe" in that one thread will > > > > mainloop.begin(); > > or > > efl_loop_beging(mainloop); > > This call has no synchronisation issue as long as eo internal code to > resolve function is threadsafe, because it doesn't influence the > parameter, nor the return value and that's all we need and can do ! > Oh, and eo_ref/unref should already be threadsafe, but anyway in the > example above, the mainloop should be refcounted when a thread get > created and unref when it get destroyed. Otherwise the above code will > always be race. no - see above. if we dont lock simple things like key_data_set/get/del then the object data structure internals get messed up. i'm not talking about the input params or returns. i'm talking about the object internal state is entirely thead-unsafe as 2 threads doing 2 key data ops at the same time on the same obj will screw eachother up as they mess with the same data struct. it is IMPOSSIBLE to build threadsafe objects on top of eo no matter how hard you try if eo_base itself doesnt even do the simplest attempts to be threadsafe. we need to make ref++ and ref-- atomic (threadsafe) in addition to other things like key data, and so on to even make this possible > > to do the current ecore_main_loop_begin(); stuff - this involves having > > calls safe to call from another loop. making the object already threadsafe > > for these reasons. we are not going to do ALL objects - you can't sensibly > > go mess with anything ui related from another loop and expect things to > > work well because a ui tree is in an unknown state when the owning loop > > comes to render (you need the above begin/end locks). > > I just don't believe it is doable in any sensible way. Let's take a > few example : > eo_event_callback_call > > Where should now the event be triggered ? It can be in the thread > where we are calling the function, but then it force the code in that > event handler to be thread aware or it will crash. Most people will > get that wrong and that will lead to bug. If we force it in the thread > owning the Eo object, how can we do that in a sensible way ? Should > now Eo always rely on Ecore ? Also this will not work with > Ecore_Thread object when we do them, as they will likely have event in > both thread... So this is a first example that show how confusing this > is going to be. no. i'm talking simple stuff like above. just a simple spinlock around data key stuff, or atomics for ref-- and ref++. of course callback_call calls within the current stack (the current thread). i am not saying that callback_call should magically call on the owning thread or anything like this. simple stuff like above that we have not done. > But let's continue with : eo_key_data_set/get/del. So you have one > side who does set stuff and the other get stuff... Doesn't work at > all, you need a way to be notified of a change. You can't sensibly > implement anything by doing a polling. So how do we get a see above. assume there is a message pipe between the reads. it's still broken because of no locks. > notification, an event every time we set some data ? That's going to > be extreme and slow down the default use case which is for non thread > safe case. But can we actually do the internal in a thread safe way > for a specific object and add that event in that case ? Plus maybe a > wait operation. I think it is doable by inheriting the > data_set/get/del function and adding the lock + an event emit in the > correct thread. i am saying that the data set/get/del should just have a lock by default. > That's why I am really not convinced at all that we should increase > the cost of our base object for pseudo thread safety as at the end it > is doable for specific case in the object themself and in a less > confusing way. then people are left to add locks and do super every base call and they CANNOt do this to ref++ and ref--. btw. even then for things like callback_call you CANNOt because you have to unlock when you call the cb func as you lose scope control so overriding cant do it without a complete re-implementation. > -- > Cedric BAIL > > ------------------------------------------------------------------------------ > Find and fix application performance issues faster with Applications Manager > Applications Manager provides deep performance insights into multiple tiers of > your business applications. It resolves application problems quickly and > reduces your MTTR. Get your free trial! > https://ad.doubleclick.net/ddm/clk/302982198;130105516;z > _______________________________________________ > enlightenment-devel mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/enlightenment-devel > -- ------------- Codito, ergo sum - "I code, therefore I am" -------------- The Rasterman (Carsten Haitzler) [email protected] ------------------------------------------------------------------------------ Find and fix application performance issues faster with Applications Manager Applications Manager provides deep performance insights into multiple tiers of your business applications. It resolves application problems quickly and reduces your MTTR. Get your free trial! https://ad.doubleclick.net/ddm/clk/302982198;130105516;z _______________________________________________ enlightenment-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/enlightenment-devel
