Pike-devel

pike-devel@lists.lysator.liu.se

1 participants
3045 discussions

Should atexit really store a weak reference?
by ceder (-) Per Cederqvist ＠ Pike (-) developers forum 23 Nov '18

23 Nov '18

The atexit() function internally stores the exit handlers in "atexit_functions" (see src/signal_handler.c). For some reason I don't understand, f_atexit() includes this line: atexit_functions->flags |= ARRAY_WEAK_FLAG; This means that if you pass a lambda to atexit() and don't hold any other references to the lambda, it may be garbage collected. This is not documented, and I belive it is bad. At the very least, it is not what I expected. Here is a sample program. Note that is doesn't print "atexit lambda 0.200 running" unless you remove the gc call. --- cut here --- #!/usr/bin/env pike void do_exit() { werror("do_exit function running\n"); exit(0); } void setup_lambda(float x) { atexit(lambda() { werror("atexit lambda %f running\n", x); } ); } int main() { call_out(setup_lambda, 0.2, 0.2); call_out(gc, 0.4); call_out(setup_lambda, 0.6, 0.6); call_out(do_exit, 0.8); return -1; }

2 1

Pike sites downtime tomorrow, Saturday Nov 17
by Peter Bortas 16 Nov '18

16 Nov '18

All Pike sites will go down while the machines are moved a few racks over. The downtime will be somewhere between 09:00 and 18:00 and is estimated to last less than 60min. This is done together with a lot of other machines, so a more exact timeline is not possible. Regards, -- Peter Bortas

1 0

Pike 8.0.702 release candidate
by Peter Bortas 12 Oct '18

12 Oct '18

Pike 8.0.702 release candidate: https://pike.lysator.liu.se/pub/pike/beta/8.0.702/Pike-v8.0.702.tar.gz Other builds: https://pike.lysator.liu.se/pub/pike/beta/8.0.702/Pike-v8.0.702-Darwin-15.4… https://pike.lysator.liu.se/pub/pike/beta/8.0.702/Pike-v8.0.702-win32-oldli… Changes since Pike 8.0.610 (release 12) New Features ------------ o Sql.mysql In unicode decode mode, characters outside the BMP are now encoded as UTF-8 encoded UTF-16. This works around that MySQL/MariaDB don't allow characters outside the BMP in utf8 strings (and instead require the character set to be declared as utf8mb4, which has quite a few side effects). Bug fixes --------- o Runtime - Fixed NULL dereference (SIGSEGV) in _disable_threads() if another thread is holding the compiler lock. - Fixed loss of Pike_interpreter_pointer (race, --with-debug only). The Pike_interpreter_pointer must not be altered by threads not holding the interpreter lock... o Calendar - Updated tzdata to 2018e. o Compiler - The machine code generator for ppc64 now generates correct code under ABI v2 (ppc64le) - Fixed an incompatibility between the machine code generator on ppc and GCC 7. - Support inherit of CompilerEnvironment. Fix some bugs triggered in the supporter subsystem when executing in an inherited CompilerEnviroment. o Filesystem.Monitor - Fix NULL-dereference in bump(). Survive st being zero in bump(). o Fuse - Switched the Fuse module to use a different libfuse API. The way we used to use libfuse got first deprecated and support for it is now removed from the libfuse git repository entirely. In addition, our previous use of said libfuse API triggered a bug resulting in a segmentation fault in at least libfuse 2.9.2 and 2.9.7. We now use libfuse in a less elegant but more usual way, and do not trigger the bug anymore. - Handle errors in Fuse operations (i.e. Pike code implementing FUSE filesystems) gently by reporting ENOSYS instead of exiting the process. o Gmp.mpq Fix numbers between -1.0 and 1.0 sometimes missing a leading 0. o Pike.identify_cycle Fix various issues with LFUNs throwing errors. LFUNs used by mapping operations may throw errors; this caused identify_loop_visit_leave() to also throw errors. The rest of the identify_cycle code was not happy about this and - Forgot to unlock the mc_mutex. This caused all following calls to Pike.identify_cycle() and Pike.count_memory() to hang. - Leaked memory. The above issues are now avoided by instead using the addresses of objects as indices in the affected mapping. It also fixes the issue with hangs if there are objects with lfuns calling Pike.identify_cycle() or Pike.count_memory(). o Protocols.HTTP.Server Fix an issue with accounting of sent bytes that caused responses to requests over HTTPS to not finish on success, but instead the whole connection to time out. Data for the first request was usually fully written, but the connection neither was properly closed nor re-used. o Protocols.WebSocket - Added low_connect() and low_websocket_accept(). These make it easier to test the error handling in the module, and to replace the HTTP-layer. - Improved compliance with RFC 6455 section 4.1. o The nonce is now actually random. o The http version, method and headers are validated. o Header case-sensitivity issues have been fixed. - The module now has a testsuite. - Improved HTTP-header camel-case consistency. o Sass - Errors thrown by handle_sass_import() can now be caught and handled by the caller of compile_file() and compile_string(). - Added the option "sass_syntax". This option will parse the input as having indented Sass syntax. This option only has effect in compile_string(). - Added method sass2scss() to convert Sass syntax to SCSS syntax. o Sql.pgsql - Fix regression bug since rel 12 that hangs queries returning over 1024 records. - Fix all rare-deadlock-causes logged in the last two years. o Stdio.Buffer Eliminate race condition in read_cstring(). o Stdio.UDP Fixed wait() failing with ENOTSOCK on NT. o Thread.Farm - Fixed lost error in provide_error() exception handling. - Avoid delayed deallocation of jobs by clearing references immediately after execution. o Tools.Standalone - extract_autodoc: Terminate on failure. Reenable termination on extraction failure. Building & Tools ---------------- o Java - The location of libjvm is now detected on Ubuntu 18.04. - Detect and support OpenJDK 10. o Postgres Support Postgresql 10 and later. o Sass Improved detection of C++ runtime library on MacOS X. MVH, -- Peter Bortas

3 3

Pike 8.0.690 beta/release candidate
by Peter Bortas 12 Sep '18

12 Sep '18

Pike 8.0.690 beta/release candidate: https://pike.lysator.liu.se/pub/pike/beta/8.0.690/Pike-v8.0.690.tar.gz Other builds: https://pike.lysator.liu.se/pub/pike/beta/8.0.690/Pike-v8.0.690-Darwin-15.4… https://pike.lysator.liu.se/pub/pike/beta/8.0.690/Pike-v8.0.690-win32-oldli… Changes since Pike 8.0.610 (release 12) New Features ------------ o Sql.mysql In unicode decode mode, characters outside the BMP are now encoded as UTF-8 encoded UTF-16. This works around that MySQL/MariaDB don't allow characters outside the BMP in utf8 strings (and instead require the character set to be declared as utf8mb4, which has quite a few side effects). Bug fixes --------- o Runtime - Fixed loss of Pike_interpreter_pointer (race, --with-debug only). The Pike_interpreter_pointer must not be altered by threads not holding the interpreter lock... o Calendar - Updated tzdata to 2018e. o Compiler - The machine code generator for ppc64 now generates correct code under ABI v2 (ppc64le) - Fixed an incompatibility between the machine code generator on ppc and GCC 7. - Support inherit of CompilerEnvironment. Fix some bugs triggered in the supporter subsystem when executing in an inherited CompilerEnviroment. o Filesystem.Monitor - Fix NULL-dereference in bump(). Survive st being zero in bump(). o Fuse - Switched the Fuse module to use a different libfuse API. The way we used to use libfuse got first deprecated and support for it is now removed from the libfuse git repository entirely. In addition, our previous use of said libfuse API triggered a bug resulting in a segmentation fault in at least libfuse 2.9.2 and 2.9.7. We now use libfuse in a less elegant but more usual way, and do not trigger the bug anymore. - Handle errors in Fuse operations (i.e. Pike code implementing FUSE filesystems) gently by reporting ENOSYS instead of exiting the process. o Gmp.mpq Fix numbers between -1.0 and 1.0 sometimes missing a leading 0. o Pike.identify_cycle Fix various issues with LFUNs throwing errors. LFUNs used by mapping operations may throw errors; this caused identify_loop_visit_leave() to also throw errors. The rest of the identify_cycle code was not happy about this and - Forgot to unlock the mc_mutex. This caused all following calls to Pike.identify_cycle() and Pike.count_memory() to hang. - Leaked memory. The above issues are now avoided by instead using the addresses of objects as indices in the affected mapping. It also fixes the issue with hangs if there are objects with lfuns calling Pike.identify_cycle() or Pike.count_memory(). o Protocols.HTTP.Server Fix an issue with accounting of sent bytes that caused responses to requests over HTTPS to not finish on success, but instead the whole connection to time out. Data for the first request was usually fully written, but the connection neither was properly closed nor re-used. o Sass - Errors thrown by handle_sass_import() can now be caught and handled by the caller of compile_file() and compile_string(). - Added the option "sass_syntax". This option will parse the input as having indented Sass syntax. This option only has effect in compile_string(). - Added method sass2scss() to convert Sass syntax to SCSS syntax. o Sql.pgsql - Fix regression bug since rel 12 that hangs queries returning over 1024 records. - Fix all rare-deadlock-causes logged in the last two years. o Stdio.Buffer Eliminate race condition in read_cstring(). o Thread.Farm - Fixed lost error in provide_error() exception handling. - Avoid delayed deallocation of jobs by clearing references immediately after execution. o Tools.Standalone - extract_autodoc: Terminate on failure. Reenable termination on extraction failure. Building & Tools ---------------- o Java - The location of libjvm is now detected on Ubuntu 18.04. - Detect and support OpenJDK 10. o Postgres Support Postgresql 10 and later. Regards, -- Peter Bortas

1 0

Pike conference 2018
by Tobias S. Josefowitz 07 Sep '18

07 Sep '18

Hi, we invite you to this year's Pike developer conference on the weekend of the 3rd and 4th of November 2018. Opera Software will host the conference in its offices in Linköping. On Saturday and Sunday we meet at the Opera offices (Repslagaregatan 21) starting from 10:00. Just call "Opera Software" using the fancy door-bell-replacement technology at the entrance (or follow instructions as posted on the door). If you intend to join, please consider dropping me an email to let me know. You are also most welcome to send me an email in case you have any other questions. In case you want to schedule an agenda item for the conference, please also just send me an email. On behalf of the 2018 conference committee, Tobi

2 1

wrong exit code of extract_autodoc on errors
by ceder (-) Per Cederqvist ＠ Pike (-) developers forum 13 Aug '18

13 Aug '18

I think extract_autodoc should exit with a non-0 exitcode whenever it produces an error message. Right now, it doesn't, which can cause documentation markup errors to go unnoticed. When running "make documentation" on the current tip of the 8.1 branch, (commit 73945f040279bfe66d), you get this error message: > /home/cederp/rgit/pike/src/cpp.c:5918..5921: DocParser error: @expr{ without matching @} > > ERROR: <Invalid error container: Tools.AutoDoc.AutoDocError(SourcePosition(File: /home/cederp/rgit/pike/src/cpp.c, lines: 5918..5921), "DocParser", "@expr{ without matching @}")> The resulting ./build/linux-3.13.0-153-generic-x86_64/doc_build/src/cpp.c.xml file contains a single newline. However, the extract_autodoc command used exit code 0, so it didn't fail. (In this case, a later make command fails, but I've seen cases where the entire build succeeds -- but maybe that wasn't in the Pike tree.) The patch below is one way of making sure that extract_autodoc fails. Maybe there are better ways. diff --git a/lib/modules/Tools.pmod/Standalone.pmod/extract_autodoc.pike b/lib/modules/Tools.pmod/Standalone.pmod/extract_autodoc.pike index a3fb82645c..a346f9a132 100644 --- a/lib/modules/Tools.pmod/Standalone.pmod/extract_autodoc.pike +++ b/lib/modules/Tools.pmod/Standalone.pmod/extract_autodoc.pike @@ -717,6 +717,8 @@ string extract(string filename, string imgdest, if (!verbosity) ; werror("\nERROR: %s\n", describe_error(err)); + if (!(flags & Tools.AutoDoc.FLAG_KEEP_GOING)) + exit(1); // return 0; } And the actual documentation markup error is trivial to fix: diff --git a/src/cpp.cmod b/src/cpp.cmod index 3b98d89746..440dceb3c7 100644 --- a/src/cpp.cmod +++ b/src/cpp.cmod @@ -5018,7 +5018,7 @@ static void insert_callback_define_no_args(struct object *cpp_obj, *! it will default to @expr{"-"@}. *! *! @param charset - *! Turn on automatic character set detection if @expr{1}, otherwise + *! Turn on automatic character set detection if @expr{1@}, otherwise *! specifies the character set used by the input. *! Defaults to @expr{"ISO-10646"@}. *!

2 2

Sql.Sql()->query() returns double values with MySQL backend
by Tomasz Jamroszczak 11 Jun '18

11 Jun '18

When I run db->query("select columna,columnb from table") on SQLite, the query() code calls master_sql->query() and I get array of two-element mappings: ({ ([columna: valuea1, columnb: valueb1)], ([columna: valuea2, columnb: valueb2)], ([columna: valuea3, columnb: valueb3)] }). When I call the same Sql.Sql()->query() on MySQL, the query() code sees there's no master_sql->query() and it calls res_obj_to_array(master_sql->big_query()). This in turn returns array of four-element mappings: ({ ([columna: valuea1, table.columna: valuea1, columnb: valueb1, table.columnb: valueb1)], ([columna: valuea2, table.columna: valuea2, columnb: valueb2, table.columnb: valueb2)], ([columna: valuea3, table.columna: valuea3, columnb: valueb3, table.columnb: valueb3)] }). That's because of 19 year old https://github.com/pikelang/Pike/commit/1e1a5b4a89bad32c3be7d0554d7727b60e1… commit (the code is intact in Pike 8.1 version, although moved to https://github.com/pikelang/Pike/blob/8.1/lib/modules/__builtin.pmod/Sql.pm…). The same res_obj_to_array() is called for example in Sql.Sql()->list_tables(), but then the extra values are filtered out. That's quite unexpected to have double values when changing SQL backends. What's the rationale behind the commit? -- Best Regards Tomasz Jamroszczak

2 1

Recent change in Gmp.mpq
by Marcus Comstedt (ACROSS) (Hail Ilpalazzo!) ＠ Pike (-) developers forum 30 May '18

30 May '18

Before: > sprintf("%3.3f", Gmp.mpq(1,3)); (1) Result: "0.333" > sprintf("%3.3f", Gmp.mpq(2,3)); (2) Result: ".666" > Ok, I can see that seems a bit wrong. After: > sprintf("%3.3f", Gmp.mpq(1,3)); (1) Result: "0.333" > sprintf("%3.3f", Gmp.mpq(2,3)); (2) Result: "066\0\0" > Uh, but that's not right either, is it? Could we please make it a habit to test our code at least once before pushing it? K thx bye...

2 1

Re: calling methods
by Marcus Comstedt (ACROSS) (Hail Ilpalazzo!) ＠ Pike (-) developers forum 29 May '18

29 May '18

invoke_signal_func calls Pike code (via apply_svalue). To run Pike code, you need to hold the interpreter lock. But that is not all. You also need to have a Pike stack and a frame pointer, which are thread local. These are kept in a structure called a "Pike_interpreter_struct", which in turn is part of something called a "thread_state". All threads created with thread_create() already has such a state, stored in a lookup table indexed on the thread id. However, a thread created by (say) C-code in an external library will not have such a state created for it. call_with_interpreter is simply a convenience function used to be able to call Pike code from an unknown context: * If the running thread already has an entry in the lookup table, then it just takes the interpreter lock, if it wasn't already taken, and calls the function. Then it unlocks the interpreter lock again if it was unlocked before. * If the running thread does _not_ have an entry in the lookup table, then a new thread_state is created, and made the current interpreter (after taking the lock). The thread_state is also stored in the lookup table, in case the pike code calls the C library back, which then again makes a call into Pike and we reenter call_with_interpreter within the same context. After the Pike code has been executed, the thread_state is removed from the table and discarded.

2 1

Re: calling methods
by Henrik Grubbstr�m (Lysator) ＠ Pike (-) developers forum 29 May '18

29 May '18

> Could I trouble you for some insight into your GI branch code? > I was looking at > >static void signal_func_wrapper(struct signal_data *d, gpointer go, guint n_params, > const GValue *param_values, GValue *return_value) { > struct signal_wrapper_context ctx = { > d, n_params, param_values, return_value > }; > call_with_interpreter(invoke_signal_func, &ctx); >} > > > Why call_with_interpreter instead of just invoke_signal_func(ctx)? > I see it has something to do with having the interpreter lock, but > what does this give you over a direct function call? call_with_interpreter() ensures that there is a valid Pike context for the current thread. This is used to handle the case that the signal might have been triggered from some strange non-pike thread. /grubba

1 0

Jump to page:

2026

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

2005

2004

2003

2002

Pike-devel