- developers - lists.mariadb.org

[Maria-developers] Updated (by Psergey): Add DS-MRR support for clustered primary keys (121)
by worklog-noreply＠askmonty.org 12 Jun '10

12 Jun '10

----------------------------------------------------------------------- WORKLOG TASK -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- TASK...........: Add DS-MRR support for clustered primary keys CREATION DATE..: Sat, 12 Jun 2010, 08:23 SUPERVISOR.....: Igor IMPLEMENTOR....: COPIES TO......: CATEGORY.......: Client-BackLog TASK ID........: 121 (http://askmonty.org/worklog/?tid=121) VERSION........: Benchmarks-3.0 STATUS.........: Un-Assigned PRIORITY.......: 60 WORKED HOURS...: 0 ESTIMATE.......: 0 (hours remain) ORIG. ESTIMATE.: 0 PROGRESS NOTES: -=-=(Psergey - Sat, 12 Jun 2010, 08:39)=-=- High-Level Specification modified. --- /tmp/wklog.121.old.538 2010-06-12 08:39:46.000000000 +0000 +++ /tmp/wklog.121.new.538 2010-06-12 08:39:46.000000000 +0000 @@ -1 +1,16 @@ +Basic idea: DS-MRR scan should be done as follows: +1. Sort incoming keys +2. Use the sorted keys to do a disk-ordered retrieval + +Unresolved questions: + +* The "sort incoming keys" part is trivial when we have only equality ranges. + If we allow ranges of arbitrary form ( ncluding ranges with one endpoint + being infinity or ranges overlapping with one another), sorting becomes + non-trival. Do we need to support this case or support only equality ranges? + +* Can/should we use the fact rowid=={clustered PK value} for InnoDB's clustered + PKs? + +* Do we support scanning on a prefix of clustered PK? (seems to be yes?) DESCRIPTION: Currently, DS-MRR doesn't support operation over clustered primary keys. The reason for this was that - Clustered primary keys are stored in disk order and so, if the sequence of ranges is ordered, the reads will already go in disk order (and so DS-MRR's step of re-ordering reads is not necessary). - Within DS-MRR implementation, the "get rowids from keys" step is not necessary when using clustered primary key, because in InnoDB/XtraDB clustered primary key values are the rowids. However, with BKA making the MRR calls, there are cases where DS-MRR over clustered primary key is beneficial: * BKA may provide lookup keys that have duplicates and/or are in arbitrary order. In that case, DS-MRR implementation may sort the key values and order them, so that it hits the disk in key order. * When running multi-table join with high @@join_cache_level value (and so, linked join buffers), lack of MRR implementation causes the chain of linked join buffers to break. (TODO and so what? Is that really a problem?) * TODO anything else? HIGH-LEVEL SPECIFICATION: Basic idea: DS-MRR scan should be done as follows: 1. Sort incoming keys 2. Use the sorted keys to do a disk-ordered retrieval Unresolved questions: * The "sort incoming keys" part is trivial when we have only equality ranges. If we allow ranges of arbitrary form ( ncluding ranges with one endpoint being infinity or ranges overlapping with one another), sorting becomes non-trival. Do we need to support this case or support only equality ranges? * Can/should we use the fact rowid=={clustered PK value} for InnoDB's clustered PKs? * Do we support scanning on a prefix of clustered PK? (seems to be yes?) ESTIMATED WORK TIME ESTIMATED COMPLETION DATE ----------------------------------------------------------------------- WorkLog (v3.5.9)

1 0

[Maria-developers] Updated (by Psergey): Add DS-MRR support for clustered primary keys (121)
by worklog-noreply＠askmonty.org 12 Jun '10

12 Jun '10

----------------------------------------------------------------------- WORKLOG TASK -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- TASK...........: Add DS-MRR support for clustered primary keys CREATION DATE..: Sat, 12 Jun 2010, 08:23 SUPERVISOR.....: Igor IMPLEMENTOR....: COPIES TO......: CATEGORY.......: Client-BackLog TASK ID........: 121 (http://askmonty.org/worklog/?tid=121) VERSION........: Benchmarks-3.0 STATUS.........: Un-Assigned PRIORITY.......: 60 WORKED HOURS...: 0 ESTIMATE.......: 0 (hours remain) ORIG. ESTIMATE.: 0 PROGRESS NOTES: -=-=(Psergey - Sat, 12 Jun 2010, 08:39)=-=- High-Level Specification modified. --- /tmp/wklog.121.old.538 2010-06-12 08:39:46.000000000 +0000 +++ /tmp/wklog.121.new.538 2010-06-12 08:39:46.000000000 +0000 @@ -1 +1,16 @@ +Basic idea: DS-MRR scan should be done as follows: +1. Sort incoming keys +2. Use the sorted keys to do a disk-ordered retrieval + +Unresolved questions: + +* The "sort incoming keys" part is trivial when we have only equality ranges. + If we allow ranges of arbitrary form ( ncluding ranges with one endpoint + being infinity or ranges overlapping with one another), sorting becomes + non-trival. Do we need to support this case or support only equality ranges? + +* Can/should we use the fact rowid=={clustered PK value} for InnoDB's clustered + PKs? + +* Do we support scanning on a prefix of clustered PK? (seems to be yes?) DESCRIPTION: Currently, DS-MRR doesn't support operation over clustered primary keys. The reason for this was that - Clustered primary keys are stored in disk order and so, if the sequence of ranges is ordered, the reads will already go in disk order (and so DS-MRR's step of re-ordering reads is not necessary). - Within DS-MRR implementation, the "get rowids from keys" step is not necessary when using clustered primary key, because in InnoDB/XtraDB clustered primary key values are the rowids. However, with BKA making the MRR calls, there are cases where DS-MRR over clustered primary key is beneficial: * BKA may provide lookup keys that have duplicates and/or are in arbitrary order. In that case, DS-MRR implementation may sort the key values and order them, so that it hits the disk in key order. * When running multi-table join with high @@join_cache_level value (and so, linked join buffers), lack of MRR implementation causes the chain of linked join buffers to break. (TODO and so what? Is that really a problem?) * TODO anything else? HIGH-LEVEL SPECIFICATION: Basic idea: DS-MRR scan should be done as follows: 1. Sort incoming keys 2. Use the sorted keys to do a disk-ordered retrieval Unresolved questions: * The "sort incoming keys" part is trivial when we have only equality ranges. If we allow ranges of arbitrary form ( ncluding ranges with one endpoint being infinity or ranges overlapping with one another), sorting becomes non-trival. Do we need to support this case or support only equality ranges? * Can/should we use the fact rowid=={clustered PK value} for InnoDB's clustered PKs? * Do we support scanning on a prefix of clustered PK? (seems to be yes?) ESTIMATED WORK TIME ESTIMATED COMPLETION DATE ----------------------------------------------------------------------- WorkLog (v3.5.9)

1 0

[Maria-developers] New (by Psergey): Add DS-MRR support for clustered primary keys (121)
by worklog-noreply＠askmonty.org 12 Jun '10

12 Jun '10

----------------------------------------------------------------------- WORKLOG TASK -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- TASK...........: Add DS-MRR support for clustered primary keys CREATION DATE..: Sat, 12 Jun 2010, 08:23 SUPERVISOR.....: Igor IMPLEMENTOR....: COPIES TO......: CATEGORY.......: Client-BackLog TASK ID........: 121 (http://askmonty.org/worklog/?tid=121) VERSION........: Benchmarks-3.0 STATUS.........: Un-Assigned PRIORITY.......: 60 WORKED HOURS...: 0 ESTIMATE.......: 0 (hours remain) ORIG. ESTIMATE.: 0 PROGRESS NOTES: DESCRIPTION: Currently, DS-MRR doesn't support operation over clustered primary keys. The reason for this was that - Clustered primary keys are stored in disk order and so, if the sequence of ranges is ordered, the reads will already go in disk order (and so DS-MRR's step of re-ordering reads is not necessary). - Within DS-MRR implementation, the "get rowids from keys" step is not necessary when using clustered primary key, because in InnoDB/XtraDB clustered primary key values are the rowids. However, with BKA making the MRR calls, there are cases where DS-MRR over clustered primary key is beneficial: * BKA may provide lookup keys that have duplicates and/or are in arbitrary order. In that case, DS-MRR implementation may sort the key values and order them, so that it hits the disk in key order. * When running multi-table join with high @@join_cache_level value (and so, linked join buffers), lack of MRR implementation causes the chain of linked join buffers to break. (TODO and so what? Is that really a problem?) * TODO anything else? ESTIMATED WORK TIME ESTIMATED COMPLETION DATE ----------------------------------------------------------------------- WorkLog (v3.5.9)

1 0

[Maria-developers] New (by Psergey): Add DS-MRR support for clustered primary keys (121)
by worklog-noreply＠askmonty.org 12 Jun '10

12 Jun '10

----------------------------------------------------------------------- WORKLOG TASK -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- TASK...........: Add DS-MRR support for clustered primary keys CREATION DATE..: Sat, 12 Jun 2010, 08:23 SUPERVISOR.....: Igor IMPLEMENTOR....: COPIES TO......: CATEGORY.......: Client-BackLog TASK ID........: 121 (http://askmonty.org/worklog/?tid=121) VERSION........: Benchmarks-3.0 STATUS.........: Un-Assigned PRIORITY.......: 60 WORKED HOURS...: 0 ESTIMATE.......: 0 (hours remain) ORIG. ESTIMATE.: 0 PROGRESS NOTES: DESCRIPTION: Currently, DS-MRR doesn't support operation over clustered primary keys. The reason for this was that - Clustered primary keys are stored in disk order and so, if the sequence of ranges is ordered, the reads will already go in disk order (and so DS-MRR's step of re-ordering reads is not necessary). - Within DS-MRR implementation, the "get rowids from keys" step is not necessary when using clustered primary key, because in InnoDB/XtraDB clustered primary key values are the rowids. However, with BKA making the MRR calls, there are cases where DS-MRR over clustered primary key is beneficial: * BKA may provide lookup keys that have duplicates and/or are in arbitrary order. In that case, DS-MRR implementation may sort the key values and order them, so that it hits the disk in key order. * When running multi-table join with high @@join_cache_level value (and so, linked join buffers), lack of MRR implementation causes the chain of linked join buffers to break. (TODO and so what? Is that really a problem?) * TODO anything else? ESTIMATED WORK TIME ESTIMATED COMPLETION DATE ----------------------------------------------------------------------- WorkLog (v3.5.9)

1 0

[Maria-developers] Question about License...
by Mateusz Matan 09 Jun '10

09 Jun '10

Welcome, We designed a custom Storage Engine (CLDB) for MySQL/MariaDB which now pass all premature tests. I wonder, how can we solve license problem... We would not share our sources, but we will just use shared library (*.so)... Which licenses should our customer who bought MariaDB with CLDB use to stay legally... If it is possible, can we use in corporation use MariaDB on GPL and just sell CLDB (storage engine) on other license (which one) ?? Thank You for Your quick reply... -- ___________________________________________________________ Mateusz Matan IT Security R&D Department, C/C++ programmer ComArch S.A., Al. Jana Pawła II 41d, 31-864 Kraków tel: (+48 12) 684 8411 e-mail: Mateusz.Matan(a)comarch.pl

3 3

[Maria-developers] per-partition attributes in the CREATE TABLE
by Sergei Golubchik 09 Jun '10

09 Jun '10

Hi, we've talked about engine attributes in the CREATE TABLE, and that one should be able to specify them per partition as well. Now, thinking about it, I'm not quite sure what the semantics shuld be. What is your use case ? How do you want them to work ? I see different possibilities. Say, there is create table ... (.....) XXX=1 partition by list (a) ( partition p0 values in (1) YYY=2, partition p1 values in (2) ); 1. We can say that XXX should be listed in the engine's hton->table_options, and YYY - in the hton->partition_options. this works fine because the engine can use table_share->option_struct and it will contain correct values, independent from whether a table is partitioned or not. but it will break when partitioning will support different engines in different partitions. 2. We can say that XXX is partition engine options, and a pluggable engine can only see YYY. YYY should come from hton->partition_options, and the engine's table level attributes are not applicable. The drawback - every engine needs to have special code to take care of the partitioned case, and to duplicate all table-level options in the partition-level options. 3. Same as 2, but YYY can come from either table level or partition level arrays. Every engine still needs the special code for partitioned case, but does not need to duplicate the table_options array. Regards, Sergei

5 6

[Maria-developers] Ever wanted to save/restore your GDB breakpoints? Here's a solution.
by Timour Katchaounov 09 Jun '10

09 Jun '10

Hi, Today I got really annoyed while debugging a server crash with GDB. I never managed to make the same GDB reattach to a new process after a crash. This means that after each crash one has to restart GDB, and reattach to the new process. As a result all breakpoints are lost, and need to be set each time, which is pretty annoying. Today I thought that others might have been equally annoyed, and might have figured a solution. Indeed, Google is our friend, and I found the following short snippet that should be added to .gdbinit: define bsave shell rm -f brestore.txt set logging file brestore.txt set logging on info break set logging off # reformat on-the-fly to a valid gdb command file shell perl -n -e 'print "break $1\n" if /^\d+.+?(\S+)$/g' brestore.txt > brestore.gdb end document bsave store actual breakpoints end define brestore source brestore.gdb end document brestore restore breakpoints saved by bsave end The credits go to this page: http://stackoverflow.com/questions/501486/getting-gdb-to-save-a-list-of-bre… Happy debugging, Timour

1 1

[Maria-developers] bzr commit into MariaDB 5.1, with Maria 1.5:maria branch (knielsen:2850)
by knielsen＠knielsen-hq.org 09 Jun '10

09 Jun '10

#At lp:maria 2850 knielsen(a)knielsen-hq.org 2010-06-09 MWL#116: Fix a couple of races in group commit. modified: include/atomic/gcc_builtins.h include/atomic/x86-gcc.h sql/handler.cc === modified file 'include/atomic/gcc_builtins.h' --- a/include/atomic/gcc_builtins.h 2008-02-06 16:55:04 +0000 +++ b/include/atomic/gcc_builtins.h 2010-06-09 11:17:39 +0000 @@ -19,8 +19,9 @@ v= __sync_lock_test_and_set(a, v); #define make_atomic_cas_body(S) \ int ## S sav; \ - sav= __sync_val_compare_and_swap(a, *cmp, set); \ - if (!(ret= (sav == *cmp))) *cmp= sav; + int ## S cmp_val= *cmp; \ + sav= __sync_val_compare_and_swap(a, cmp_val, set);\ + if (!(ret= (sav == cmp_val))) *cmp= sav #ifdef MY_ATOMIC_MODE_DUMMY #define make_atomic_load_body(S) ret= *a === modified file 'include/atomic/x86-gcc.h' --- a/include/atomic/x86-gcc.h 2007-02-28 16:50:51 +0000 +++ b/include/atomic/x86-gcc.h 2010-06-09 11:17:39 +0000 @@ -38,15 +38,33 @@ #define asm __asm__ #endif +/* + The atomic operations imply a memory barrier for the CPU, to ensure that all + prior writes are flushed from cache, and all subsequent reads reloaded into + cache. + + We need to imply a similar memory barrier for the compiler, so that all + pending stores (to memory that may be aliased in other parts of the code) + will be flushed to memory before the operation, and all reads from such + memory be re-loaded. This is achieved by adding the "memory" pseudo-register + to the clobber list, see GCC documentation for more explanation. + + The compiler and CPU memory barriers are needed to make sure changes in one + thread are made visible in another by the atomic operation. +*/ #ifndef MY_ATOMIC_NO_XADD #define make_atomic_add_body(S) \ - asm volatile (LOCK_prefix "; xadd %0, %1;" : "+r" (v) , "+m" (*a)) + asm volatile (LOCK_prefix "; xadd %0, %1;" : "+r" (v) , "+m" (*a): : "memory") #endif #define make_atomic_fas_body(S) \ - asm volatile ("xchg %0, %1;" : "+r" (v) , "+m" (*a)) + asm volatile ("xchg %0, %1;" : "+r" (v) , "+m" (*a) : : "memory") #define make_atomic_cas_body(S) \ + int ## S sav; \ asm volatile (LOCK_prefix "; cmpxchg %3, %0; setz %2;" \ - : "+m" (*a), "+a" (*cmp), "=q" (ret): "r" (set)) + : "+m" (*a), "=a" (sav), "=q" (ret) \ + : "r" (set), "a" (*cmp) : "memory"); \ + if (!ret) \ + *cmp= sav #ifdef MY_ATOMIC_MODE_DUMMY #define make_atomic_load_body(S) ret=*a @@ -59,9 +77,9 @@ #define make_atomic_load_body(S) \ ret=0; \ asm volatile (LOCK_prefix "; cmpxchg %2, %0" \ - : "+m" (*a), "+a" (ret): "r" (ret)) + : "+m" (*a), "+a" (ret) : "r" (ret) : "memory") #define make_atomic_store_body(S) \ - asm volatile ("; xchg %0, %1;" : "+m" (*a), "+r" (v)) + asm volatile ("; xchg %0, %1;" : "+m" (*a), "+r" (v) : : "memory") #endif /* TODO test on intel whether the below helps. on AMD it makes no difference */ === modified file 'sql/handler.cc' --- a/sql/handler.cc 2010-05-26 08:16:18 +0000 +++ b/sql/handler.cc 2010-06-09 11:17:39 +0000 @@ -1103,14 +1103,30 @@ ha_check_and_coalesce_trx_read_only(THD static THD * enqueue_atomic(THD *thd) { - my_atomic_rwlock_wrlock(&LOCK_group_commit_queue); + THD *orig_queue; + thd->next_commit_ordered= group_commit_queue; + + my_atomic_rwlock_wrlock(&LOCK_group_commit_queue); + do + { + /* + Save the read value of group_commit_queue in each iteration of the loop. + When my_atomic_casptr() returns TRUE, we know that orig_queue is equal + to the value of group_commit_queue when we enqueued. + + However, as soon as we enqueue, thd->next_commit_ordered may be + invalidated by another thread (the group commit leader). So we need to + save the old queue value in a local variable orig_queue like this. + */ + orig_queue= thd->next_commit_ordered; + } while (!my_atomic_casptr((void **)(&group_commit_queue), (void **)(&thd->next_commit_ordered), - thd)) - ; + thd)); my_atomic_rwlock_wrunlock(&LOCK_group_commit_queue); - return thd->next_commit_ordered; + + return orig_queue; } static THD * @@ -1399,6 +1415,9 @@ int ha_commit_trans(THD *thd, bool all) int cookie; if (tc_log->use_group_log_xid) { + // ToDo: if xid==NULL here, we may use is_group_commit_leader uninitialised. + // ToDo: Same for cookie below when xid==NULL. + // Seems we generally need to check the case xid==NULL. if (is_group_commit_leader) { pthread_mutex_lock(&LOCK_group_commit); @@ -1434,9 +1453,18 @@ int ha_commit_trans(THD *thd, bool all) } pthread_mutex_unlock(&LOCK_group_commit); - /* Wake up everyone except ourself. */ - while ((queue= queue->next_commit_ordered) != NULL) - group_commit_wakeup_other(queue); + /* Wake up everyone except ourself. */ + THD *current= queue->next_commit_ordered; + while (current != NULL) + { + /* + Careful not to access current->next_commit_ordered after waking up + the other thread! As it may change immediately after wakeup. + */ + THD *next= current->next_commit_ordered; + group_commit_wakeup_other(current); + current= next; + } } else {

1 0

[Maria-developers] Fixes for an issue with my_atomic
by Kristian Nielsen 09 Jun '10

09 Jun '10

Hi, I had a problem with my group commit patch, and tracked it down to a problem in my_atomic. The issue is that my_atomic_cas*(val, cmp, new) accesses *cmp after successful CAS operation (in one place reading it, in another place writing it). Here is the fix: Index: work-5.1-groupcommit/include/atomic/gcc_builtins.h =================================================================== --- work-5.1-groupcommit.orig/include/atomic/gcc_builtins.h 2010-06-09 11:53:59.000000000 +0200 +++ work-5.1-groupcommit/include/atomic/gcc_builtins.h 2010-06-09 11:54:06.000000000 +0200 @@ -19,8 +19,9 @@ v= __sync_lock_test_and_set(a, v); #define make_atomic_cas_body(S) \ int ## S sav; \ - sav= __sync_val_compare_and_swap(a, *cmp, set); \ - if (!(ret= (sav == *cmp))) *cmp= sav; + int ## S cmp_val= *cmp; \ + sav= __sync_val_compare_and_swap(a, cmp_val, set);\ + if (!(ret= (sav == cmp_val))) *cmp= sav #ifdef MY_ATOMIC_MODE_DUMMY #define make_atomic_load_body(S) ret= *a Index: work-5.1-groupcommit/include/atomic/x86-gcc.h =================================================================== --- work-5.1-groupcommit.orig/include/atomic/x86-gcc.h 2010-06-09 11:53:59.000000000 +0200 +++ work-5.1-groupcommit/include/atomic/x86-gcc.h 2010-06-09 11:54:06.000000000 +0200 @@ -45,8 +45,12 @@ #define make_atomic_fas_body(S) \ asm volatile ("xchg %0, %1;" : "+r" (v) , "+m" (*a)) #define make_atomic_cas_body(S) \ + int ## S sav; \ asm volatile (LOCK_prefix "; cmpxchg %3, %0; setz %2;" \ - : "+m" (*a), "+a" (*cmp), "=q" (ret): "r" (set)) + : "+m" (*a), "=a" (sav), "=q" (ret) \ + : "r" (set), "a" (*cmp)); \ + if (!ret) \ + *cmp= sav #ifdef MY_ATOMIC_MODE_DUMMY #define make_atomic_load_body(S) ret=*a This makes the behaviour consistent with the other implementations, see for example generic-msvc.h. (It is also pretty important for correct operation. In my code, I use my_atomic_casptr() to atomically enqueue a struct from one thread, and as soon as it is enqueued another thread may grab the struct and change it, including the field pointed to by cmp. So it is essential that my_atomic_casptr() neither reads nor writes *cmp after successful CAS, as the above patch ensures.) It was btw. a bit funny how I tracked this down. After digging for a few hours in my own code without success I got the idea to check the CAS implementation, and spotted the problem in gcc_builtins.h. After fixing I was baffled as to why my code stil failed, until I realised I was not using gcc_builtins.h but x86-gcc.h, and found and fixed the similar problem there ;-) Now the question is, where should I push this (if at all)? Any opinions? ----------------------------------------------------------------------- While I was there, I also noticed another potential problem in gcc_builtins.h, suggesting this patch: Index: work-5.1-groupcommit/include/atomic/x86-gcc.h =================================================================== --- work-5.1-groupcommit.orig/include/atomic/x86-gcc.h 2010-06-09 11:37:12.000000000 +0200 +++ work-5.1-groupcommit/include/atomic/x86-gcc.h 2010-06-09 11:52:47.000000000 +0200 @@ -38,17 +38,31 @@ #define asm __asm__ #endif +/* + The atomic operations imply a memory barrier for the CPU, to ensure that all + prior writes are flushed from cache, and all subsequent reads reloaded into + cache. + + We need to imply a similar memory barrier for the compiler, so that all + pending stores (to memory that may be aliased in other parts of the code) + will be flushed to memory before the operation, and all reads from such + memory be re-loaded. This is achieved by adding the "memory" pseudo-register + to the clobber list, see GCC documentation for more explanation. + + The compiler and CPU memory barriers are needed to make sure changes in one + thread are made visible in another by the atomic operation. +*/ #ifndef MY_ATOMIC_NO_XADD #define make_atomic_add_body(S) \ - asm volatile (LOCK_prefix "; xadd %0, %1;" : "+r" (v) , "+m" (*a)) + asm volatile (LOCK_prefix "; xadd %0, %1;" : "+r" (v) , "+m" (*a): : "memory") #endif #define make_atomic_fas_body(S) \ - asm volatile ("xchg %0, %1;" : "+r" (v) , "+m" (*a)) + asm volatile ("xchg %0, %1;" : "+r" (v) , "+m" (*a) : : "memory") #define make_atomic_cas_body(S) \ int ## S sav; \ asm volatile (LOCK_prefix "; cmpxchg %3, %0; setz %2;" \ : "+m" (*a), "=a" (sav), "=q" (ret) \ - : "r" (set), "a" (*cmp)); \ + : "r" (set), "a" (*cmp) : "memory"); \ if (!ret) \ *cmp= sav @@ -63,9 +77,9 @@ #define make_atomic_load_body(S) \ ret=0; \ asm volatile (LOCK_prefix "; cmpxchg %2, %0" \ - : "+m" (*a), "+a" (ret): "r" (ret)) + : "+m" (*a), "+a" (ret) : "r" (ret) : "memory") #define make_atomic_store_body(S) \ - asm volatile ("; xchg %0, %1;" : "+m" (*a), "+r" (v)) + asm volatile ("; xchg %0, %1;" : "+m" (*a), "+r" (v) : : "memory") #endif /* TODO test on intel whether the below helps. on AMD it makes no difference */ The comment in the patch explains the idea I think. Basically, these memory barrier operations need to be a compiler barrier also. Otherwise there is nothing that prevents GCC from moving unrelated stores across the memory barrier operation. This is a problem for example when filling in a structure and then atomically linking it into a shared list. The CPU memory barrier ensures that the fields in the structure will be visible to other threads on the CPU level, but it is also necessary to tell GCC to keep the stores into the struct fields prior to the store linking into the shared list. (It also makes the volatile declaration on the updated memory location unnecessary, but maybe it is needed for other implementations (though it shouldn't be, volatile is almost always wrong)). - Kristian.

1 0

[Maria-developers] Patch to build xtradb and federatedx statically on Windows
by Bo Thorsen 09 Jun '10

09 Jun '10

Hi everyone, Recently, we found out that xtradb and federatedx were compiled as plugins on Windows. This is not correct because they are core storage engines and should be linked statically instead. On Unix, it's possible to build both types, but the CMake files doesn't really handle this. They just look in plug.in, and if there is a dynamic line, the storage engine is built dynamically. I'm hesitant to fix this properly, as it's probably already done in MySQL 5.5, or at least it will be. I see no reason to come up with a good solution for a temporary problem - there are many other issues on Windows that are more pressing. FederatedX couldn't compile as a statically linked plugin at all on Windows. I have fixed this now, but I *really* don't like the way we do these hacks on Windows. It's absurdly difficult to keep track of when a variable says xtradb and when it says innobase. And debugging CMakeLists.txt files is not too much fun. Anyway, I did the same thing on FederatedX as is done in XtraDB, and it works now. Finally, I removed a subdir that didn't have a CMakeLists.txt. Removed a warning during the cmake run. With this patch, the Windows zip file release build should work correctly and produce a MariaDB with both XtraDB and FederatedX in it. Should I check this into ~maria-captains/maria/5.1, or is there a branch more suitable at the moment? Bo Thorsen. Monty Program AB. -- MariaDB: MySQL replacement Community developed. Feature enhanced. Backward compatible.

2 2