Are all Ruby built-in objects thread safe?

Are all built-in objects thread safe? For example, if I have an array
and one thread is constant appending to it while another thread is
shifting
elements off of it and there’s no synchronization going on, can the
array
object ever get corrupted? What about a similar scenario for hashes?
These
are surely complicated objects with internal state that must be
maintained.
Are they implemented to be thread safe?
Thank you…

Hi,

In message “Re: Are all Ruby built-in objects thread safe?”
on Tue, 23 Dec 2008 18:21:27 +0900, “Just Another Victim of the
Ambient M.” [email protected] writes:

| Are all built-in objects thread safe?

For MRI (1.8.x) and YARV (1.9.x), every C implemented methods are
protected by GIL (Global Interpreter Lock), so that you don’t have to
worry about. But it might depend on each implementation.

          matz.

On 23.12.2008 10:24, Just Another Victim of the Ambient M. wrote:

Are all built-in objects thread safe?  For example, if I have an array 

and one thread is constant appending to it while another thread is shifting
elements off of it and there’s no synchronization going on, can the array
object ever get corrupted? What about a similar scenario for hashes? These
are surely complicated objects with internal state that must be maintained.
Are they implemented to be thread safe?

The answer will sound a bit odd: they are not built to be thread safe
but it may well be that they are. In fact, it may well depend whether
they are thread safe in practical terms depending on the Ruby version
you are using. Since the classic interpreter uses green threads, i.e.
no preemption is happening, it may be that some or all operations are
atomic and thus thread safe.

The bottom line is: you should not rely on this and assume they are not
thread safe!

In your example, you should be using class Queue which is thread safe
and can be imported by requiring “thread”

$ irb -r thread
irb(main):001:0> q = Queue.new
=> #Queue:0x7ff6afc0
irb(main):002:0> t = Thread.new(q) {|qq| while (o = qq.deq) != qq; p o;
end}
=> #<Thread:0x7ff61c68 sleep>
irb(main):003:0> 5.times {|i| q.enq i}; q.enq q; t.join
0
1
2
3
4
=> #<Thread:0x7ff61c68 dead>

Then there are also classes Mutex and Monitor plus module MonitorMixin.
The difference is reentrance:

irb(main):009:0> require ‘monitor’
=> true
irb(main):010:0> m=Mutex.new
=> #Mutex:0x7ff3e100
irb(main):011:0> m.synchronize { m.synchronize { 1 } }
ThreadError: stopping only thread
note: use sleep to stop forever
from (irb):11:in synchronize' from (irb):11 from (irb):11:insynchronize’
from (irb):11
from (null):0
irb(main):012:0> m=Monitor.new
=> #<Monitor:0x7ff329f4 @mon_count=0, @mon_owner=nil,
@mon_waiting_queue=[], @mon_entering_queue=[]>
irb(main):013:0> m.synchronize { m.synchronize { 1 } }
=> 1
irb(main):014:0>

Kind regards

robert

Just Another Victim of the Ambient M. wrote:

Are all built-in objects thread safe?  For example, if I have an array 

and one thread is constant appending to it while another thread is shifting
elements off of it and there’s no synchronization going on, can the array
object ever get corrupted? What about a similar scenario for hashes? These
are surely complicated objects with internal state that must be maintained.
Are they implemented to be thread safe?

This is a very interesting question! And it is a question that can
ultimately only be answered by a formal Ruby Specification or more
specifically a formal Ruby Memory Model.

Until we have such a specification, the C source code of MRI or YARV
is considered to be the “specfication”. However, there is a problem:
that source code can actually be interpreted several different ways.

If you look at the implementations of Hash, Array and friends, you
will see that they are not thread-safe. Ergo: the specification says
that the user is responsible for locking Arrays and Hashes.

If, however, you look at the implementation of threads, you will see
that both MRI and YARV are actually incapable of running more than one
thread at a time – even on a 1000-core machine MRI and YARV will only
ever use one core. So, since two threads can never access an Array at
the same time, there is no need for locking. Ergo: the specification
says that the user is not responsible for locking Arrays and Hashes.

There is a conflict here – on the one hand, Arrays aren’t
thread-safe, on the other hand, MRI’s broken threading implementation
accidentally makes them thread-safe. Which do you depend on? As it
turns out, different people interpret this differently.

A couple of months ago, this actually became an issue. Originally, the
JRuby developers had implemented Arrays to be not safe. One of the big
selling points of JRuby was and still is the promise of true
concurrency and better scalability. So, naturally, people wanted to
take advantage of this feature and started running their concurrent
programs on JRuby. And those programs crashed left and right, because
they didn’t lock their Arrays properly. So, the JRuby team decided to
implement thread-safe data structures on their end, so that code that
didn’t crash on MRI could be run unmodified on JRuby.

However, they didn’t have to do that. They could just as well have
concluded that those programs were broken and they needed to become
thread-safe. That would have been perfectly acceptable. And there is
no guarantee that all Ruby Implementations will do it that way (and
there’s lots of them, something around 14 or so at the moment). Well,
unless of course, there is a specification which tells them to.

So, in short: when in doubt, lock.

jwm

Yukihiro M. wrote:

Hi,

In message “Re: Are all Ruby built-in objects thread safe?”
on Tue, 23 Dec 2008 18:21:27 +0900, “Just Another Victim of the Ambient M.” [email protected] writes:

| Are all built-in objects thread safe?

For MRI (1.8.x) and YARV (1.9.x), every C implemented methods are
protected by GIL (Global Interpreter Lock), so that you don’t have to
worry about. But it might depend on each implementation.

In JRuby we’ve made a best attempt to keep the core data structures like
Array, Hash, and String as thread-safe as possible without introducing
locks (which would slow down single-threaded cases). There are cases
where when using Array or Hash you may get ConcurrencyErrors raised.
This is the trade-off for having fast lock-free collections.

I would very much welcome a set of guaranteed-thread-safe wrapper
collections folks could use if they’re concerned about concurrency,
since it would be unreasonable to penalize all code with locking. For
now, learn and love Mutex, or don’t share collections across threads.

  • Charlie

“Jörg W Mittag” [email protected] wrote in message
news:[email protected]

Are they implemented to be thread safe?
will see that they are not thread-safe. Ergo: the specification says
that the user is responsible for locking Arrays and Hashes.

If, however, you look at the implementation of threads, you will see
that both MRI and YARV are actually incapable of running more than one
thread at a time – even on a 1000-core machine MRI and YARV will only
ever use one core. So, since two threads can never access an Array at
the same time, there is no need for locking. Ergo: the specification
says that the user is not responsible for locking Arrays and Hashes.

I don't think this is relevant.  Concurrency isn't about how many

processors you use. Multitasking systems existed long before SMP
hardware
existed. Concurrency is about doing tasks concurrently. If you have
one
method running and it may be preempted by another method then they are
running concurrently. If the two methods share data then they may
corrupt
that data for each other. This is true regardless of how these
concurrencies, or threads, are implemented. It doesn’t matter if
they’re
hardware supported system threads or if they’re Ruby green threads…

So, in short: when in doubt, lock.

This is the popular wisdom so this is what I will do.  Better safe 

than
not thread-safe!

Just Another Victim wrote:

I don't think this is relevant.  Concurrency isn't about how many

processors you use. Multitasking systems existed long before SMP
hardware
existed. Concurrency is about doing tasks concurrently. If you have
one
method running and it may be preempted by another method then they are
running concurrently. If the two methods share data then they may
corrupt
that data for each other. This is true regardless of how these
concurrencies, or threads, are implemented. It doesn’t matter if
they’re
hardware supported system threads or if they’re Ruby green threads…

Not exactly. Ruby green threads aren’t fully pre-emptive; they will only
pre-empt at the boundaries between execution steps, not within an
execution step.

Hence the single operations @array.pop and @array.push(…) are safe
against preemption. You could consider each method call to be wrapped
implicitly by Thread.critical { … }

Jörg W Mittag wrote:

A couple of months ago, this actually became an issue. Originally, the
JRuby developers had implemented Arrays to be not safe. One of the big
selling points of JRuby was and still is the promise of true
concurrency and better scalability. So, naturally, people wanted to
take advantage of this feature and started running their concurrent
programs on JRuby. And those programs crashed left and right, because
they didn’t lock their Arrays properly. So, the JRuby team decided to
implement thread-safe data structures on their end, so that code that
didn’t crash on MRI could be run unmodified on JRuby.

Actually, we made a minimal attempt to ensure that concurrent operations
against Array were usually safe across threads but also would raise a
Ruby-land “ConcurrencyError” when concurrent changes could not be
reconciled. It’s a trade-off; adding locks to all the core collections
would severely penalize performance for what’s generally the rare case
of concurrent access. And really there should be a separate set of
classes with guaranteed concurrency that people can use if the
performance considerations of locking are acceptabe for safe concurrent
access.

So in short, you’re absolutely right; nobody should ever rely on the
core collections to be thread-safe, even if they happen to be
thread-safe by accident in the C implementations right now. That won’t
be the case on all implementations, and may not even be the case on
future versions of the C impls. The safe answer is to ensure you’re
watching your own back and properly synchronizing access to shared data
structures.

  • Charlie

On Dec 24, 2008, at 1:56 AM, Just Another Victim of the Ambient
Morality wrote:

they’re
hardware supported system threads or if they’re Ruby green threads…

I think the key here is the granularity of Ruby’s atomicity. You’re
assuming that preemption can occur on the granularity of machine
instructions. Were that the case, two simultaneous threads on a single
core could, potentially, cause problems. I think what Matz was saying
is that, because of the GIL, simultaneous threads will only preempt at
a much higher granularity.

So I have a question for Matz and Charles: Would it be reasonable to
specify that YARV instructions should be atomic? Charles, how does
this work with JVM ops? Last I heard, JRuby was still skipping YARV
and going straight to Java bytecodes, which could make this a
difficult proposition. My completely uneducated guess, though, is that
unless we specify that certain implementation provided data structures
must be thread safe (at the very least Mutex), then there would have
to be a minimum level at which everything is atomic to be able to
write implementation independent thread-safe libraries.

  • Josh

Joshua B. wrote:

I think the key here is the granularity of Ruby’s atomicity. You’re
assuming that preemption can occur on the granularity of machine
instructions. Were that the case, two simultaneous threads on a single
core could, potentially, cause problems. I think what Matz was saying is
that, because of the GIL, simultaneous threads will only preempt at a
much higher granularity.

And rarely within C-based code, unless that code explicitly yields
control to the thread scheduler. This also means that calls out to C
libraries have to be written to use asynchronous calls or they just
plain block all threads.

In JRuby, threads may preempt at any time, and indeed can and will run
“really” in parallel at any given time if the hardware supports it.

So I have a question for Matz and Charles: Would it be reasonable to
specify that YARV instructions should be atomic? Charles, how does this
work with JVM ops? Last I heard, JRuby was still skipping YARV and going
straight to Java bytecodes, which could make this a difficult
proposition. My completely uneducated guess, though, is that unless we
specify that certain implementation provided data structures must be
thread safe (at the very least Mutex), then there would have to be a
minimum level at which everything is atomic to be able to write
implementation independent thread-safe libraries.

Everything in the thread(.rb) library obviously has thread-safety as
part of its contract, so you don’t have to worry about that. The core
collections (Array, String, Hash) do not have such guarantees explicitly
in their contract, and I believe they should stay that way since the
vast majority of uses are single-threaded. We (JRuby) have additionally
added locking (as smartly as possible) to ensure that method and
instance variable tables are thread-safe, since they’re crucial to
Ruby’s operation.

I don’t think YARV instructions are good atomic units. In JRuby we can’t
even (and won’t even) guarantee individual Java bytecodes are atomic.
Nothing can be atomic unless you lock around it, and in most cases you
can’t have atomicity and still allow code to execute in parallel. Plus
YARV instructions cover a wide range of things, some of which are
obviously not atomic like arbitrary method calls or test-and-set (||=,
&&=) logic. Atomicity and thread-safety should be specified on a
per-mutator basis for all the mutable structures in Ruby, rather than as
a blanket assertion.

  • Charlie

On 26.12.2008 10:43, Robert D. wrote:

I would very much welcome a set of guaranteed-thread-safe wrapper

collections folks could use if they’re concerned about concurrency, since it
would be unreasonable to penalize all code with locking. For now, learn and
love Mutex, or don’t share collections across threads.

That sounds like a fun job. The issue goes well with a question which
burns on the top of my tongue:
How to test thread issues? I do not have any experience in this field
but consider it somehow critical before “promising” functional thread
safe wrappers.

This is tricky. There are few things I’d like to say to this: first,
there is an easy way to provide basic thread safety by wrapping all
method calls in a synchronized block - for this even a
SynchronizedDelegator would be sufficient which could be used for all
classes - not just collections. In this case, because of the way the
wrapping is done there is no need to test thread safety because - as
long as the delegator ensures that all methods are wrapped in this way
there is no chance of corrupting internal state.

But, in the general case thread safety cannot be achieved on the class
level. My typical example is this

if hash.contains_key? k
dat = hash[k]
else
dat = create_dat(k)
hash[k] = dat
end

The basic property of this bit of code which makes it impossible to
ensure thread safety on level of class Hash is that there are two
methods invoked on the same instance and there must not be any state
change between them because then you either end up with garbage (nil) in
“dat” or you invoke create_dat(k) more than once per key value and this
leads to different threads having a different idea of what hash[k] is.

So in this case you need a lock around the complete block of code.
(Note, the method level lock would work if using a Hash with a default
block, but this solves only a small portion of the cases.) This is also
the reason why a general per method locking is only of limited use. It
only ensures consistency of the internal state of an object but can
never ensure overall concurrency correctness of an application.

Testing thread safety is difficult to impossible for several reasons.
One of the reasons is that for proper control of the test case you would
need to ensure exact timing of thread execution. While you could do
that I have never seen people actually doing this - maybe because this
will make test suites run much slower. Another reason is that you
vastly depend on the underlying machine (i.e. hardware, OS and VM). I
have seen Java programs break as soon as they were executed on a multi
core machine with more than n cores.

Things aren’t made easier by the fact that - concluding from postings I
see on comp.lang.java.* newsgroups for example - few people seem to have
a thorough understanding of concurrency and all the issues involved.
Another item on the “makes it hard” list is the fact that most OO and
procedural languages only have a very basic toolkit for concurrency
control; while Java started out pretty good with built in “synchronized”
it took until version 5 that they incorporated Doug Lea’s concurrency
utilities into the language’s standard library and also into the EJB
spec.

It’s different in other programming languages: functional languages are
by definition thread safe because they are free of side effects. (At
least in theory. :-)) Also, other languages built for concurrent
applications which have a different programming model (e.g. Occam) of
course have much better support for concurrency.

Kind regards

robert

I would very much welcome a set of guaranteed-thread-safe wrapper

collections folks could use if they’re concerned about concurrency, since it
would be unreasonable to penalize all code with locking. For now, learn and
love Mutex, or don’t share collections across threads.

That sounds like a fun job. The issue goes well with a question which
burns on the top of my tongue:
How to test thread issues? I do not have any experience in this field
but consider it somehow critical before “promising” functional thread
safe wrappers.

Cheers
R.


Il computer non è una macchina intelligente che aiuta le persone
stupide, anzi, è una macchina stupida che funziona solo nelle mani
delle persone intelligenti.
Computers are not smart to help stupid people, rather they are stupid
and will work only if taken care of by smart people.

Umberto Eco

Just Another Victim of the Ambient M. wrote:

I don't think this is relevant.  Concurrency isn't about how many 

processors you use. Multitasking systems existed long before SMP hardware
existed. Concurrency is about doing tasks concurrently. If you have one
method running and it may be preempted by another method then they are
running concurrently. If the two methods share data then they may corrupt
that data for each other. This is true regardless of how these
concurrencies, or threads, are implemented. It doesn’t matter if they’re
hardware supported system threads or if they’re Ruby green threads…

Keep in mind though that on smp systems, most instructions are not
atomic. On a single processor, they are.

if hash.contains_key? k
dat = hash[k]
else
dat = create_dat(k)
hash[k] = dat
end

[…]

So in this case you need a lock around the complete block of code.
(Note, the method level lock would work if using a Hash with a default
block, but this solves only a small portion of the cases.)

If Hash had a method get_with_default that took a default value or a
block like initialize you could solve these cases.

mfg, simon … l

On Sat, Dec 27, 2008 at 10:04 AM, Simon K. [email protected] wrote:

So in this case you need a lock around the complete block of code.
(Note, the method level lock would work if using a Hash with a default
block, but this solves only a small portion of the cases.)

If Hash had a method get_with_default that took a default value or a
I believe you mean
“get and assign with default”
because IIRC there is a get_with_default, it is simply called
Hash#fetch
R.

On 27.12.2008 09:55, Simon K. wrote:

So in this case you need a lock around the complete block of code.
(Note, the method level lock would work if using a Hash with a default
block, but this solves only a small portion of the cases.)

If Hash had a method get_with_default that took a default value or a
block like initialize you could solve these cases.

Please read my note: this was just an example for the class of
synchronization problems that cannot be solved within the class. You
can invent arbitrary more complex scenarios that would not make any
sense to be covered by a standard method in class Hash.

My point is totally independent of the functionality of class Hash.
This is about sequences of method invocations which do not tolerate any
state changes by other threads between method calls.

Put it differently: in concurrent applications there are many scenarios
where lock granularity “method” is insufficient to guarantee correct
code. Instead you need explicit locking for larger portions of client
code.

Kind regards

robert

On Fri, Dec 26, 2008 at 11:24 AM, Robert K.
[email protected] wrote:

That sounds like a fun job. The issue goes well with a question which
thread safety because - as long as the delegator ensures that all methods
end
this solves only a small portion of the cases.) This is also the reason why
break as soon as they were executed on a multi core machine with more than n


remember.guy do |as, often| as.you_can - without end

Robert, IIUC we do not want either the one, nor the other.
For what I am concerned one would need Read/Write Locks.
The built in methods like [] and []= would obtain read and write
locks, while the read lock automatically obtains a write lock and only
releases it when its count is to zero, the write lock would inhibit
the read lock to be obtained.

I do not know if one should allow user access to these locks? My first
thought would be no, all we want to do is to have thread-safe Hash,
Array, String but not to provide advanced synchronization mechanisms.
OTOH one could expose the write_lock and read_lock of each object
which would allow for the following

hash = TS_Hash::new

dat = hash.read_synchronize do
if hash.has_key? key then
hash[ key ]
else
hash.write_synchronize do
hash[key] = compute_data( key )
end
end
end
N.B. We have to solve the interlock problem here as normally the
write_synchronize does not obtain the write_lock
as the read_synchronize did get a read_lock. We would need to grant
the write lock if the read_lock is obtained by the current thread
only, but imagine two threads waiting for the write_lock while
containing the read_lock, the good old
philosophers-forks-spoon-spaghetti interlock. [ That is why we eat
spaghetti with a fork only :wink: ]

Therefore I guess the RW-locking in a ThreadSafe Container class shall
rather not be exposed as we avoiding interlocking is not that
complicated IIRC

And if we expose read_synchronize &blk, and write_synchronize &blk we
should probably raise something like an IllegalMonitorState exception
if a thread tries to call the one inside the other.

More thoughts please.
Cheers
Robert


Il computer non è una macchina intelligente che aiuta le persone
stupide, anzi, è una macchina stupida che funziona solo nelle mani
delle persone intelligenti.
Computers are not smart to help stupid people, rather they are stupid
and will work only if taken care of by smart people.

Umberto Eco

On 26.12.2008 19:17, Robert D. wrote:

On Fri, Dec 26, 2008 at 11:24 AM, Robert K.
[email protected] wrote:

full quote

Please invest a bit more time in quoting and trimming. I may speak for
me only but I find it unnecessary hard to follow a discussion when there
is a full quote and no clear indication of references in your reply.

Robert, IIUC we do not want either the one, nor the other.

What are you referring to with “one” and “other”?

For what I am concerned one would need Read/Write Locks.

Why?

The built in methods like [] and []= would obtain read and write
locks, while the read lock automatically obtains a write lock and only
releases it when its count is to zero, the write lock would inhibit
the read lock to be obtained.

Let’s first talk about semantics not jump into implementation. How the
read write locking is implemented is rather unimportant for the question
what kind of additional locking we want to propose for default
collections (if any). As far as I can see three variants are lying on
the table with the lock free collection always be present as well:

  1. none at all, locking needs to be explicitly done in client code

  2. exclusive locking, no two methods can execute concurrently on the
    same instance

  3. read write locking with the usual semantics, allowing multiple
    methods with read semantic to be executed concurrently or at most one
    “write” method at a time. “write” lock excludes any “read” locks.

I’ll follow up with discussion at the end.

  hash[ key ]

else
hash.write_synchronize do
hash[key] = compute_data( key )
end
end
end

This is a bad idiom because it is prone to deadlocking. When having
multiple levels of locks you must only go from stronger to weaker locks,
not the other way round. Otherwise this can happen

t1: read lock
t2: read lock
t1: write lock (deadlock, blocked by t2’s read lock)

See also
http://java.sun.com/j2se/1.5.0/docs/api/java/util/concurrent/locks/ReentrantReadWriteLock.html

N.B. We have to solve the interlock problem here as normally the
write_synchronize does not obtain the write_lock
as the read_synchronize did get a read_lock. We would need to grant
the write lock if the read_lock is obtained by the current thread
only,

With this definition you force code to obtain a read lock before the
write lock can be held. I assume you mean that the rule should rather
be “grant the write lock only if no other thread holds the read lock” -
which is what rw locking usually does.

but imagine two threads waiting for the write_lock while
containing the read_lock, the good old
philosophers-forks-spoon-spaghetti interlock. [ That is why we eat
spaghetti with a fork only :wink: ]

Exactly, as show above.

Therefore I guess the RW-locking in a ThreadSafe Container class shall
rather not be exposed as we avoiding interlocking is not that
complicated IIRC

And if we expose read_synchronize &blk, and write_synchronize &blk we
should probably raise something like an IllegalMonitorState exception
if a thread tries to call the one inside the other.

Not necessarily: read_synchronize inside write_synchronize is perfectly
ok although it has no noticeable effect. But since it can occur in
another method it is reasonable to allow it.

Some more remarks about the complexity of read write locking can be
found here

http://java.sun.com/j2se/1.5.0/docs/api/java/util/concurrent/locks/ReadWriteLock.html

For the interested Wikipedia also has a number of pages on the matter
“locking”:

Now, what would be reasonable to provide in a thread safe standard
collection? IMHO the discussion above and the complexity and wide
variance in implementation choices for read write locking makes it next
to impossible to come up with an optimal read write lock implementation
in standard collections for most application scenarios.

This yields a bad cost benefit ratio for option 3 above. This leaves us
with options 1 and 2. I lean toward option 1 because even the dead
simple exclusive lock approach is of limited use only. Even in light of
default values or the default_proc of a Hash there are some things to
note.

Consider

assuming Synchronize() wraps an instance with a delegator

which synchronizes all method calls on a single lock

$all_values = Synchronize(Hash.new {|h,k| h[k] = Synchronize([])})

in multiple threads:

$all_values[some_key] << item

While our basic locking ensures internal state of all collections is
consistent the code has the disadvantage that every thread needs to
obtain two locks and the number of locks is not limited (1 + no of
keys). A single lock would be sufficient here.

Another disadvantage of built in locking is that it might trick some
people in believing that this is all they need to use to get thread safe
code. If there would be no such thing as a default thread safe Array
and Hash people are forced to think about how they want to go about
locking and I believe that way they will get better (i.e. more correct
and more efficient) code.

As a compromise I suggest to provide something like the Synchronize()
method I showed above which does two things: 1. it extends the object
passed with MonitorMixin and 2. it wraps the instance with another
instance which synchronizes all methods like this:

class SynchWrapper
def initialize(o)
@obj = o.extend(::MonitorMixin)
end

 # remove all Object methods

 def method_missing(*a,&b)
    @obj.synchronize { @obj.__send__(*a,&b) }
 end

end

Advantage is that there is a single mechanism that uniformly works for
all classes not just collections. And, the mechanism is made explicit
while not requiring too much manual intervention or additional typing.
For all cases where the logic requires other locking schemes explicit
locking needs to be employed anyway.

Kind regards

robert

PS: I just recognize we should probably move this discussion over to
ruby-core…

On Sat, Dec 27, 2008 at 12:59 PM, Robert K.
[email protected] wrote:

On 26.12.2008 19:17, Robert D. wrote:

On Fri, Dec 26, 2008 at 11:24 AM, Robert K.
[email protected] wrote:

full quote

Please invest a bit more time in quoting and trimming. I may speak for me
only but I find it unnecessary hard to follow a discussion when there is a
full quote and no clear indication of references in your reply.
Oh this was bad work, I messed up, my apologies.

Robert, IIUC we do not want either the one, nor the other.

What are you referring to with “one” and “other”?
IIRC, because I indeed lost context, sorry again, I meant that your
observations, however useful and learned, did not have a direct impact
to what Charles wants here. Look e.g. at the synchronized containers
in Java. They are thread safe and yet all your concerns are still
valid. In other words you are already talking about problems of the
next level of abstraction while the level below still has problems. So
all I wanted to say is that we still should tackle the low level
thread-safety.

For what I am concerned one would need Read/Write Locks.

Why?
Because we want simultaneous read access. And that was the other point
I referred to in my reply, we do not want to do exclusive reads.
lock free collection always be present as well:

  1. read write locking with the usual semantics, allowing multiple methods
    with read semantic to be executed concurrently or at most one “write” method
    at a time. “write” lock excludes any “read” locks.
    3 by all means

if hash.has_key? key then
hash[ key ]
else
hash.write_synchronize do
hash[key] = compute_data( key )
end
end
end

This is a bad idiom because it is prone to deadlocking.
That was my point to show why I am against exposure of the internal RW
lock.
I guess I did not make that clear either :frowning:
When having
multiple levels of locks you must only go from stronger to weaker locks, not
the other way round. Otherwise this can happen

t1: read lock
t2: read lock
t1: write lock (deadlock, blocked by t2’s read lock)
I disagree, write lock upgrade should throw an exception.

See also
http://java.sun.com/j2se/1.5.0/docs/api/java/util/concurrent/locks/ReentrantReadWriteLock.html
I am quite fond of Doug Lea’s Read Write Lock from Java Threads but I
will definitely have a look.
rw locking usually does.
Here I messed up again, you got it right, but this was not a stupid
error, I just got it wrong :frowning:

but imagine two threads waiting for the write_lock while
containing the read_lock, the good old
philosophers-forks-spoon-spaghetti interlock. [ That is why we eat
spaghetti with a fork only :wink: ]

Exactly, as show above.

Yup

method it is reasonable to allow it.
Hmm we have not yet defined the semantics of this, you were right when
complaining about my jump to implementation, right now I do not see
the need for this but I will try to make my home work.

Now, what would be reasonable to provide in a thread safe standard
collection? IMHO the discussion above and the complexity and wide variance
in implementation choices for read write locking makes it next to impossible
to come up with an optimal read write lock implementation in standard
collections for most application scenarios.
I agree but we should opt for a down the middle road.

This yields a bad cost benefit ratio for option 3 above. This leaves us
with options 1 and 2. I lean toward option 1 because even the dead simple
exclusive lock approach is of limited use only. Even in light of default
values or the default_proc of a Hash there are some things to note.
I fail to see why 3 is so costly, I mean why optimize. Thread
programming is complex (that’s why we all love immutable objects,
right :wink: and I fail to see why one could hope that it becomes less
complex just because we add
a small but still useful feature.
While our basic locking ensures internal state of all collections is
consistent the code has the disadvantage that every thread needs to obtain
two locks and the number of locks is not limited (1 + no of keys). A single
lock would be sufficient here.

Another disadvantage of built in locking is that it might trick some people
in believing that this is all they need to use to get thread safe code. If
there would be no such thing as a default thread safe Array and Hash people
are forced to think about how they want to go about locking and I believe
that way they will get better (i.e. more correct and more efficient) code.
No please that really could be said against any progress. Furthermore
I believe that folks asking for
thread safe collections know mostly why they do that, after all thread
safety is not a hype.

where the logic requires other locking schemes explicit locking needs to be
employed anyway.
However there is too big a prise to pay, ok we could get rid of the
metaprogramming and just follow the idea you have presented, but
exclusive read access just will make the wrapped collection unusable
for any heavy weight parallel read access.

Kind regards

   robert

PS: I just recognize we should probably move this discussion over to
ruby-core…
Hmm maybe
Anyway thanks for your time.

Robert

Il computer non è una macchina intelligente che aiuta le persone
stupide, anzi, è una macchina stupida che funziona solo nelle mani
delle persone intelligenti.
Computers are not smart to help stupid people, rather they are stupid
and will work only if taken care of by smart people.

Umberto Eco

[…]

So in this case you need a lock around the complete block of code.
(Note, the method level lock would work if using a Hash with a default
block, but this solves only a small portion of the cases.)

If Hash had a method get_with_default that took a default value or a
I believe you mean
“get and assign with default”
because IIRC there is a get_with_default, it is simply called
Hash#fetch

You are right. There is a little difference: fetch doesn’t pass the hash
to its block. But one can “abuse” both with an assignment in the block:

| dat = hash.fetch(k) { | key | hash[key] = create_dat(key) }

mfg, simon … l