We did actually discuss doing something like this a long time ago. The 
worry that comes to mind is whether it should actually be forbidden (an 
invariant) or merely strong frowned upon. It is possible to arrange a 
situation where you know (based on other knowledge) that a put on a channel 
will succeed. Should that be allowed? I'd have to hammock on that one for a 
while.

On Tuesday, August 29, 2017 at 3:19:51 PM UTC-5, Aaron Iba wrote:
>
> Ahh that makes a lot of sense.  Indeed, I'm guilty of doing a blocking >!! 
> inside a go-block.  I was so careful to avoid other kinds of blocking calls 
> (like IO) that I forgot that blocking variants of core.async calls 
> themselves were forbidden.
>
> Thank you for pointing this out!  I will rewire things to not do this.
>
> Per Gary's suggestion, I also think it'd be useful if core.async blocking 
> ops checked a dynamic var (or a property of the thread itself) and at least 
> warned if they are being called from a forbidden context.  To resolve my 
> original issue, I'm considering doing this in my dev environment:
>
> (doseq [v '[<!! >!!]]
>   (alter-var-root (ns-resolve 'clojure.core.async v)
>                   (fn [f]
>                     (fn [& args]
>                       (if (.startsWith (.getName (Thread/currentThread))
>                                        "async-dispatch-")
>                         (throw (Exception. (str v " called inside 
> async-dispatch")))
>                         (apply f args))))))
>
>
>
>
> On Tuesday, August 29, 2017 at 1:43:53 PM UTC-4, Gary Trakhman wrote:
>>
>> Hm, I came across a similar ordering invariant (No code called by a go 
>> block should ever call the blocking variants of core.async functions) while 
>> wrapping an imperative API, and I thought it might be useful to use 
>> vars/binding to enforce it.  Has this or other approaches been considered 
>> in core.async?  I could see a *fixed-thread-pool* var being set and >!! 
>> checking for false.
>>
>> An analogy in existing clojure.core would be the STM commute's 'must be 
>> running in a transaction' check that uses a threadlocal. 
>> https://github.com/clojure/clojure/blob/master/src/jvm/clojure/lang/LockingTransaction.java#L205
>>
>> On Tue, Aug 29, 2017 at 1:30 PM Timothy Baldridge <[email protected]> 
>> wrote:
>>
>>> To add to what Alex said, look at this trace: 
>>> https://gist.github.com/anonymous/65049ffdd37d43df8f23630928e8fed0#file-thread-dump-out-L1337-L1372
>>>
>>> Here we see a go block calling mapcat, and inside the inner map 
>>> something is calling >!!. As Alex mentioned this can be a source of 
>>> deadlocks. No code called by a go block should ever call the blocking 
>>> variants of core.async functions (<!!, >!!, alts!!, etc.). So I'd start at 
>>> the code redacted in those lines and go from there. 
>>>
>>>
>>>
>>> On Tue, Aug 29, 2017 at 11:09 AM, Alex Miller <[email protected]> 
>>> wrote:
>>>
>>>> go blocks are multiplexed over a thread pool which has (by default) 8 
>>>> threads. You should never perform any kind of blocking activity inside a 
>>>> go 
>>>> block, because if every go block in work happens to end up blocked, you 
>>>> will prevent all go blocks from making any further progress. It sounds to 
>>>> me like that's what has happened here. The go block threads are named 
>>>> "async-dispatch-<n>" and it looks like there are 8 blocked ones in your 
>>>> thread dump.
>>>>
>>>> It also looks like they are all blocking on a >!!, which is a blocking 
>>>> call. So I would look for a go block that contains a >!! and convert that 
>>>> to a >! or do something else to avoid blocking there.
>>>>
>>>>
>>>> On Tuesday, August 29, 2017 at 11:48:25 AM UTC-5, Aaron Iba wrote:
>>>>>
>>>>> My company has a production system that uses core.async extensively. 
>>>>> We've been running it 24/7 for over a year with occasional restarts to 
>>>>> update things and add features, and so far core.async has been working 
>>>>> great.
>>>>>
>>>>> The other day, during a particularly high workload, the whole system 
>>>>> got locked up. All the channels seemed blocked at once.  I was able to 
>>>>> connect with a REPL and poke around, and noticed strange behavior of 
>>>>> core.async. Specifically, the following code, when evaluated in the REPL, 
>>>>> blocked on the put (third expression):
>>>>>
>>>>> (def c (async/chan))
>>>>> (go-loop []
>>>>>   (when-some [x (<! c)]
>>>>>     (println x)
>>>>>     (recur)))
>>>>> (>!! c true)
>>>>>
>>>>> Whereas on any fresh system, the above expressions obviously succeed.
>>>>>
>>>>> Puts succeeded if they went onto the channel's buffer, but not when 
>>>>> they should go through to a consumer. For example with the following 
>>>>> expressions, evaluated in the REPL, the first put succeeded (presumably 
>>>>> because it went on the buffer), but subsequent puts blocked:
>>>>>
>>>>> (def c (async/chan 1))
>>>>> (def m (async/mult c))
>>>>> (def out (async/chan (async/sliding-buffer 3)))
>>>>> (async/tap m out)
>>>>> (>!! c true) ;; succeeds
>>>>> (>!! c true) ;; blocks forever
>>>>>
>>>>> This leads me to wonder if core.async itself somehow got into a bad 
>>>>> state. It's entirely possible I caused this by misusing the API somewhere 
>>>>> in the codebase, but we use core.async so extensively that I wouldn't 
>>>>> know 
>>>>> where to begin looking.
>>>>>
>>>>> I'm wondering if someone more familiar with core.async internals has 
>>>>> an idea about what could cause the above situation. Or if we notice it 
>>>>> happening again, what could I do to gather more helpful information.
>>>>>
>>>>> I also have a redacted thread dump, in case it's useful:
>>>>>
>>>>> https://gist.github.com/anonymous/65049ffdd37d43df8f23630928e8fed0
>>>>>
>>>>> Any help would be much appreciated,
>>>>>
>>>>> Aaron
>>>>>
>>>>> P.S. core.async has been a godsend in terms of helping us structure 
>>>>> and modularize our large system.  Thank you to all those who contributed 
>>>>> to 
>>>>> this wonderful library!
>>>>>
>>>>> -- 
>>>> You received this message because you are subscribed to the Google
>>>> Groups "Clojure" group.
>>>> To post to this group, send email to [email protected]
>>>> Note that posts from new members are moderated - please be patient with 
>>>> your first post.
>>>> To unsubscribe from this group, send email to
>>>> [email protected]
>>>> For more options, visit this group at
>>>> http://groups.google.com/group/clojure?hl=en
>>>> --- 
>>>> You received this message because you are subscribed to the Google 
>>>> Groups "Clojure" group.
>>>> To unsubscribe from this group and stop receiving emails from it, send 
>>>> an email to [email protected].
>>>> For more options, visit https://groups.google.com/d/optout.
>>>>
>>>
>>>
>>>
>>> -- 
>>> “One of the main causes of the fall of the Roman Empire was that–lacking 
>>> zero–they had no way to indicate successful termination of their C 
>>> programs.”
>>> (Robert Firth) 
>>>
>>> -- 
>>> You received this message because you are subscribed to the Google
>>> Groups "Clojure" group.
>>> To post to this group, send email to [email protected]
>>> Note that posts from new members are moderated - please be patient with 
>>> your first post.
>>> To unsubscribe from this group, send email to
>>> [email protected]
>>> For more options, visit this group at
>>> http://groups.google.com/group/clojure?hl=en
>>> --- 
>>> You received this message because you are subscribed to the Google 
>>> Groups "Clojure" group.
>>> To unsubscribe from this group and stop receiving emails from it, send 
>>> an email to [email protected].
>>> For more options, visit https://groups.google.com/d/optout.
>>>
>>

-- 
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to [email protected]
Note that posts from new members are moderated - please be patient with your 
first post.
To unsubscribe from this group, send email to
[email protected]
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
--- 
You received this message because you are subscribed to the Google Groups 
"Clojure" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to