- Boost - lists.preview.boost.org

Boost.Threads: RW mutexes and locks.
by Batov, Vladimir 28 Jun '04

28 Jun '04

RW mutexes/locks are important to a threading library and appear simple to implement. However, Boost.Threads lacks it. I wrote something simple that appear to do the job using the functionality already available in Boost.Threads. If the idea is considered good/robust/efficient enough and worth having in the library, I can re-write it to conform with the rest of Boost.Threads. class rw_mutex : private boost::noncopyable { typedef boost::condition Condition; public: friend class writer_lock; friend class reader_lock; rw_mutex() : num_readers_(0), num_writers_(0) {} private: boost::mutex mutex_; Condition reader_condition_; Condition writer_condition_; int num_readers_; int num_writers_; }; class writer_lock : public boost::mutex::scoped_lock { typedef boost::mutex::scoped_lock super; public: writer_lock(rw_mutex& mutex) : super(mutex.mutex_), mutex_(mutex) { ++mutex_.num_writers_; // Make new readers wait. // If there are readers in-progress, wait for them to finish. while (mutex_.num_readers_) mutex_.writer_condition_.wait(*this); } ~writer_lock() { --mutex_.num_writers_; // Last writer notifies waiting readers. // If not last, wake up the next writer. if (!mutex_.num_writers_) mutex_.reader_condition_.notify_all(); else mutex_.writer_condition_.notify_one(); } private: rw_mutex& mutex_; }; class reader_lock : public boost::mutex::scoped_lock { typedef boost::mutex::scoped_lock super; public: reader_lock(rw_mutex& mutex) : super(mutex.mutex_), mutex_(mutex) { // If there are writers waiting, let them in first. while (mutex_.num_writers_) mutex_.reader_condition_.wait(*this); ++mutex_.num_readers_; // Make new writes wait. unlock(); } ~reader_lock() { lock(); --mutex_.num_readers_; // Last reader notifies if (!mutex_.num_readers_) mutex_.writer_condition_.notify_one(); } private: rw_mutex& mutex_; };

3 2

Re:[boost] Re: 1.32 release preparation
by Johannes Brunen 28 Jun '04

28 Jun '04

On Sun, 27 Jun 2004 19:05:45 -0700, Robert Ramey wrote > Here is the current scenario. > > a) The announcement gets posted - release 1.32 scheduled for date x. > b) Each developer decides - uh oh I better get in my changes before date x. > c) Of course these changes take longer > d) and break some other things > e) which takes longer still IMHO, point d) touches a weak point: There are two kind of libraries in a system like boost. At first, there are core libraries which are used by many other libraries. Changes at these libraries are potentially dangerous and must be well-considered since they can broke a lot of stuff. They should early checked into CVS to get time for stabelizing the base. Secondly, there are libraries which 'only' provide a special feature like the boost graph library the circular_buffer library. They shouldn't be expected to lead to regressions of other boost libraries. I think that they could be handled differently than the core libraries. > a) the release manager or the core influential developers keep an eye on the > current state of the CVS tree and tests. > b) When things look "pretty good", branch for release without warning. None > of this loading my latest feature in to make the schedule. > c) the Main tree goes on as usual > d) pending issues in the release branch are addressed by nagging the person > responsible. > e) the new version is releases with a number like 1.31.1 or maybe even 28 > July 2004. > f) any changes between branch and release should be just bug fixes which can > then be merged back into the development tree. > > Under this system, releases would occur more frequently - on the order of 6 > weeks. If you're ready to check in your latest feature and the new release > branch catches you with your pants down - well too bad. But it's no big deal > because the next release will be in 6? weeks anyhow which is close to what > its going to take to shake out the repercussions of your check in anyway. We should seriouly consider such a release process. Howerver, then the CVS main tree must always be in a nearly 'Release-Ready' state. But IMO that is anyway a good idea. > But I > think it's really hard to get such a large number of people working > independently on exactly the same schedule. Yes ... > Almost all libraries are accepted subject to some changes being made. Many > such changes seem simple but end up rippling through the whole library. This > takes time. This is IMHO an indication of a flaw in the current boost review process. A library should be only accepted after the list of requested changes has been made. Additionally, I think a library should be accepted only if it is compatible to the current CVS main tree. > Once a library is accepted, then one can get serious about making it pass > with a larger number of compilers. This also takes time - a lot of time Passing compilers which weren't supported during the review process under no circumstances shouldn't delay the integration of the library into the boost main pool. Supporting a special compiler is only a nice to have feature. This means not that it not worth spending the effort supporting important platforms. However, in the first iteration I just expect a library to be written in ANSI C++ and to be usable on platforms which support it. > These changes end up having a ripple effect and can result it changes like > moving things in namespaces and subdirectories. While this is going on, the > system is broken and really not suitable to be subject to the daily testing > routine. This is IMHO an indication of a flaw ... (s.a) > So it seems attractive to deal with all this stuff before one does > the initial check in to CVS. This way one can start with a relatively clean > start. I second this. With best regards Johannes

2 1

Re: 1.32 release preparation
by Johannes Brunen 28 Jun '04

28 Jun '04

On Sun, 27 Jun 2004 16:35:15 -0400, David Abrahams wrote > Should we have accepted libraries checked in immediately (on a branch, > maybe?), Maybe we should reconsider the meaning of 'acceptance' in the review process. Apart of that, I think that this is a very good idea. > and then only added to the libs page when the authors > deem them ready? We should list them on the lib page but in a special section. > If they're checked in on the trunk they can even get the benefit of > regression testing while the authors whip them into shape. Yes, that would be fine. With best regards Johannes

2 1

Re: [boost] Re: 1.32 release preparation
by Johannes Brunen 28 Jun '04

28 Jun '04

On Sun, 27 Jun 2004 08:30:20 -0700, Jeff Garland wrote: > As much as I'd like to see circular_buffer, serialization, etc in this release I'd rather > see us pull off a release that doesn't take 2-3 months to accomplish. I > believe we should hold the line on the schedule now and anything that isn't > ready should simply go in the next release. I disagree on that. We don't have that many Boost releases. Waiting for the next one could be a long time. Therefore, I propose to wait until all (or as many as possibly considering a reasonable time schedule) accepeted libraries will be checked into CVS and be part of the release. > And if there are a large > number of new things left out by that approach then we should plan another > release sooner rather than waiting 4 months. However, then you need new volunteers which have to spend the time for the whole procedure again. > One is that we have been reviewing and > accepting a large number of libraries recently. So, many of these have been > accepted in the last couple months. The other thing that seems to be a > pattern is that libraries get accepted, then authors get a list of changes to > make. I could think of a review process which accepts a library only after the list of requested changes has beem made. > If they get busy it often takes months to get these done and then > finally they check into CVS. So a release tends to trigger the evaluation of > anything that is in that multi-month pipeline. That isn't good. I would prefere a review process which leads to final acceptance only after a library is 'CVS' compatible i.e. does not yield to any (known) regressions of the existing pool libraries and is equipted with a working testing facility. With best regards Johannes

2 1

Re:[boost] Re: 1.32 release preparation
by Robert Ramey 28 Jun '04

28 Jun '04

Dave Abrahams wrote: >I'd like to know the reasons for "all the sudden rush". We did >announce a schedule long ago, and it has already been delayed by three >weeks. I'm not asking in order to point fingers; I just want to know >how to avoid the "sudden rush" next time. Here is the current scenario. a) The announcement gets posted - release 1.32 scheduled for date x. b) Each developer decides - uh oh I better get in my changes before date x. c) Of course these changes take longer d) and break some other things e) which takes longer still Personally I think the concept of scheduling a release fundamentally flawed. I would suggest another idea. a) the release manager or the core influential developers keep an eye on the current state of the CVS tree and tests. b) When things look "pretty good", branch for release without warning. None of this loading my latest feature in to make the schedule. c) the Main tree goes on as usual d) pending issues in the release branch are addressed by nagging the person responsible. e) the new version is releases with a number like 1.31.1 or maybe even 28 July 2004. f) any changes between branch and release should be just bug fixes which can then be merged back into the development tree. Under this system, releases would occur more frequently - on the order of 6 weeks. If you're ready to check in your latest feature and the new release branch catches you with your pants down - well too bad. But it's no big deal because the next release will be in 6? weeks anyhow which is close to what its going to take to shake out the repercussions of your check in anyway. I realize that this is not the most orthodox way of doing things. But I think it's really hard to get such a large number of people working independently on exactly the same schedule. >How come we have so many >accepted libraries that have not even been put in the CVS? Almost all libraries are accepted subject to some changes being made. Many such changes seem simple but end up rippling through the whole library. This takes time. Once a library is accepted, then one can get serious about making it pass with a larger number of compilers. This also takes time - a lot of time These changes end up having a ripple effect and can result it changes like moving things in namespaces and subdirectories. While this is going on, the system is broken and really not suitable to be subject to the daily testing routine. So it seems attractive to deal with all this stuff before one does the initial check in to CVS. This way one can start with a relatively clean start. Robert Ramey

6 6

Problems compiling regex with gcc3 on aix and hpux
by Goetz Isenmann 28 Jun '04

28 Jun '04

Hi! I get ugly warnings (they look dangerous to me) when I compile boost/regex on AIX and HP-UX with gcc. On Linux and FreeBSD everything is fine. For example /usr/local/gcc-3.3.3/bin/g++ -c -I. libs/regex/src/cregex.cpp compiles without any complains. On aix43 and hpux11 this gives me the attached warning messages. The order of the messages is slightly different on both platforms, but after sorting the output is identical. I use boost-1.31.0 and see no differences with and without boost-1.31.0-regex-patch-20040503. -- Goetz Isenmann

2 3

Provisional schedule for Boost-1.33
by John Maddock 28 Jun '04

28 Jun '04

Booster's: It's apparent that several of us will just miss the 1.32 deadlines, and I'm sure some people will be concerned when their code will get released as a result, so provisionally I'd like to propose a schedule for 1.33 so we all know where we are. Unless someone else would like to volunteer (and please don't be shy in doing so!), then I guess that means I've just volunteered to manage the release.... Regards, John Maddock. Provisional 1.33 release schedule: ~~~~~~~~~~~~~~~~~~~~~~ Mon 30th August 2004 [1]: ~~~~~~~~~~~~~~~~ Deadline for additions and changes (other than patches to known issues, and docs) to core Boost libraries [2]. Deadline for the addition of new libraries to cvs unless they pass all tests with the core compiler set [3]. Mon 13th September 2004: ~~~~~~~~~~~~~~~~~~ Deadline for additions and changes (other than patches to regressions, and docs) to all Boost libraries. Deadline for all new library additions. Mon 27th September 2004: ~~~~~~~~~~~~~~~~~~ Branch for release, no changes to the release branch without the release managers permission (except for docs). Mon 4th October 2004: ~~~~~~~~~~~~~~~~ First release candidate published, actual release will be as soon as possible thereafter. Notes: ~~~~ 1) There should be a few reminders of this deadline in the weeks beforehand. 2) The core Boost libraries are those that other Boost libraries heavily depend upon: bind config (unless change is restricted to a single compiler, and can be tested with all effected versions of that compiler before commit). enable_if. integer. iterators (the construction framework subset only). mpl operators preprocessor smart_ptr's static_assert test type_traits. 3) The core compiler set includes those compilers that are known to be highly conforming, and for which no failures will normally be expected: gcc 3.3 or later on Linux. Visual C++ 7.1 on Win32. Intel C++ 8 on Win32 or Linux.

1 0

Re: 1.32 release prepar
by Robert Ramey 28 Jun '04

28 Jun '04

Last time we ended up making a release candidate which isolated a couple of final issues. Maybe this is a symptom and maybe it points in the direction of an interesting idea. New code is checked in and tested in the development tree just as it is now. Regression testing and everything remains exactly as it is. Code that documents all test failures and is considered ready for release is moved to the Release Candidate tree. The exact same regression testing setup used on the development tree is used here. Presumably, it would be run much less as things would only be checked in when thought to be perfect which is an infrequent occurrence. Developers develop against the Release Candidate tree. When they have something they like, the check into the Main Development tree as they do now. Traffic on the main development tree is the same as it is now. The Release Candidate Tree should be ready for "shipment" at any convenient time. This basically should consist of assignation of a version number to the tree state. It would not be or scheduled event dramatic event. It could happen as frequently as is convenient. Of course a test failure on the RC tree would be dramatic event. The check in causing such an event could be cast from the RC Tree (Heaven) back into the Main Development Tree (Purgatory) and the author would suffer the attendant embarrassment. Hopefully this would be an infrequent event. I realize that I haven't personally had a library go through the release process so feel free to ignore this idea. Robert Ramey

2 1

[config] about BOOST_NO_POINTER_TO_MEMBER_TEMPLATE_PARAMETERS
by Joaquín Mª López Muñoz 28 Jun '04

28 Jun '04

In config/intel.hpp, this defect macro is defined as follows #if defined(__ICL) && (__ICL <= 710) # define BOOST_NO_POINTER_TO_MEMBER_TEMPLATE_PARAMETERS #endif The macro should be defined *only* for Intel C++ 7.1 or older for Windows 32-bit. So my questions are: 1. Is the definition correct in light of what I said above? 2. What's the meaning of all those different macros __ICC, __ICL and __ECC? I've googled for them but found nothing conclusive. Thanks Joaquín M López Muñoz Telefónica, Investigación y Desarrollo

3 4

RE: Boost.Threads: RW mutexes and locks.
by Batov, Vladimir 28 Jun '04

28 Jun '04

John Torjo wrote: > Batov, Vladimir wrote: >> >> class writer_lock : public boost::mutex::scoped_lock >> { >> typedef boost::mutex::scoped_lock super; >> >> public: >> >> writer_lock(rw_mutex& mutex) : super(mutex.mutex_), mutex_(mutex) >> { >> ++mutex_.num_writers_; // Make new readers wait. > ^^^^ > > just curios, shouldn't this be thread-safe? I believe it is thread-safe as writer_lock currently inherits from scoped_lock and, consequently, we get inside the writer_lock constructor after its base-class scoped_lock constructor. The base-class constructor does the synchronization.

1 0