<div dir="ltr">Hi Levente, Hi Chris,<br><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Apr 28, 2015 at 3:41 PM, Levente Uzonyi <span dir="ltr"><<a href="mailto:leves@elte.hu" target="_blank">leves@elte.hu</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">There's no need to store preferences in a data structure at all. We already have "pragma" preferences (since 4.1), which store the preference values independently. Since the 4.1 release it's a "permanent" goal to rewrite all preferences to "pragma" preferences.<br>
We should just make it happen.</blockquote><div><br></div><div>This seems like a lot of work, and is work that can be done over time. But right now we're suffering lock ups due to the Mutex in Preferences. For example, the Notifier/Debugger accesses the scrollBarsOnRight preference and I've often seen lock ups caused by this. So I propose that I fix the access to be as I described it. There be no access lock except for adding/updating preferences. So reading is done without synchronisation, and setting and/or adding is done by copying and assigning. I also propose to compile preferences without creating a block, so</div><div><br></div><div><div>autoIndent</div><div><span class="Apple-tab-span" style="white-space:pre"> </span>^ self</div><div><span class="Apple-tab-span" style="white-space:pre"> </span>valueOfFlag: #autoIndent</div><div><span class="Apple-tab-span" style="white-space:pre"> </span>ifAbsent: true</div></div><div><br></div><div>instead of</div><div><br></div><div><div>autoIndent</div><div><span class="Apple-tab-span" style="white-space:pre"> </span>^ self</div><div><span class="Apple-tab-span" style="white-space:pre"> </span>valueOfFlag: #autoIndent</div><div><span class="Apple-tab-span" style="white-space:pre"> </span>ifAbsent: [true]</div></div><div><br></div><div>which is well-supported by both the Interpreter and the Cog VMs, given Object>>value ^self. This to save space and time.</div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="HOEnZb"><font color="#888888">Levente<br>
</font></span><br>
P.S.: Reverting that method will solve the concurrency issue.<div class="HOEnZb"><div class="h5"><br>
<br>
On Tue, 28 Apr 2015, Eliot Miranda wrote:<br>
<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
<br>
On Tue, Apr 28, 2015 at 12:47 PM, Chris Muller <<a href="mailto:asqueaker@gmail.com" target="_blank">asqueaker@gmail.com</a>> wrote:<br>
Wait, the newer one has a non-local return in it, but<br>
Mutex>>#critical: has an ensure: in it anyway, so maybe I don't see<br>
the problem..?<br>
<br>
<br>
If one hits ctrl-period when the system is in the critical section then the debugger can't open because it interrupts the critical section, preventing the ensure block from running, attempts to access e.g.<br>
scroll bar preferences when it tries to open, and the system deadlocks. So preferences either need to be *not* protected by a critical section, or the Debugger needs not to access preferences.<br>
<br>
IMO, we should try and write preferences so that they don't require a lock. Writing them as a lock-free data structure would be a really good idea. First that critical section is slow and clunky. Second, I<br>
presume it is there only for the rare case of a write to preferences, not to protect reads.<br>
<br>
IMO, a simple implementation which copied and replaced the entire preferences dictionary on write would be sufficient. Sure there's a danger that some client would get a stale value if it read concurrently<br>
while there was a write, but then so what? A preference is a preference, not a hard-and-fast value, and code should work accessing a preference no matter its value, so momentarily getting a stale value<br>
shouldn't matter. So the implementation could be as simple as<br>
<br>
addPreference: aName categories: categoryList default: aValue balloonHelp: helpString projectLocal: localBoolean changeInformee: informeeSymbol changeSelector: aChangeSelector type: aType<br>
"Add or replace a preference as indicated. Reuses the preexisting Preference object for this symbol, if there is one, so that UI artifacts that interact with it will remain valid."<br>
<br>
| aPreference aPrefSymbol |<br>
aPrefSymbol := aName asSymbol.<br>
aPreference := DictionaryOfPreferences<br>
at: aPrefSymbol<br>
ifAbsent:<br>
[| newPreference |<br>
newPreference := aPreference <br>
name:aPrefSymbol<br>
defaultValue:aValue<br>
helpString:helpString<br>
localToProject:localBoolean<br>
categoryList:categoryList<br>
changeInformee:informeeSymbol<br>
changeSelector:aChangeSelector<br>
type: aType.<br>
AccessLock critical:<br>
[| newDict |<br>
newDict := DictionaryOfPreferences copy.<br>
newDict at: aPrefSymbol put: newPreference].<br>
self compileAccessMethodForPreference:aPreference.<br>
newPreference]<br>
<br>
<br>
On Tue, Apr 28, 2015 at 2:43 PM, Chris Muller <<a href="mailto:asqueaker@gmail.com" target="_blank">asqueaker@gmail.com</a>> wrote:<br>
>> The above change restores the old behavior of locking up the image, so it<br>
>> should be reverted. An additional comment explaininng why aBlock must not be<br>
>> evaluated inside the argument of #accessDictionaryOfPreferencesIn: would be<br>
>> helpful.<br>
><br>
> Ahh, because aBlock might have a non-local return in it, leaving the<br>
> Mutex unsignaled (and critical unenterable), is that right?<br>
><br>
> Took me a minute to see that problem.<br>
><br>
> Okay, I'll revert that method if no one else does by my next commit..<br>
><br>
>> It would be even better to finally get rid of DictionaryOfPreferences.<br>
>><br>
>><br>
>> Levente<br>
>><br>
<br>
<br>
<br>
<br>
--<br>
best,Eliot<br>
<br>
</blockquote>
</div></div><br><br>
<br></blockquote></div><br><br clear="all"><div><br></div>-- <br><div class="gmail_signature">best,<div>Eliot</div></div>
</div></div>