rfc:tls

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revisionPrevious revision
Next revision
Previous revision
rfc:tls [2008/08/24 20:27] – typo lbarnaudrfc:tls [2017/09/22 13:28] (current) – external edit 127.0.0.1
Line 5: Line 5:
   * Status: Under Discussion   * Status: Under Discussion
   * First Published at: http://marc.info/?l=php-internals&m=121893972814818&w=2   * First Published at: http://marc.info/?l=php-internals&m=121893972814818&w=2
-  * Initial patch: http://arnaud.lb.s3.amazonaws.com/__thread-tls.patch +  * Initial patch: http://gist.github.com/659731 
-  * Current patch: http://arnaud.lb.s3.amazonaws.com/__thread-tls-2.patch+  * Current patch: http://gist.github.com/659724
  
 Currently ZTS builds are slower than non-ZTS builds. This RFC is about avoiding some of the major overhead of ZTS builds by using native thread local storage. Currently ZTS builds are slower than non-ZTS builds. This RFC is about avoiding some of the major overhead of ZTS builds by using native thread local storage.
Line 20: Line 20:
 Results for bench.php: Results for bench.php:
  
-^ non-PIC                           ^^ +^ non-PIC                         ^ ^^ 
-|non-ZTS                       |3.7s| +|non-ZTS                       |3.7s
-|ZTS unpatched                 |5.2s| +|ZTS unpatched                 |5.2s
-|ZTS patched                   |4.0s| +|ZTS patched                   |4.0s
-|ZTS patched and static globals|3.8s| +|ZTS patched and static globals|3.8s
-^ PIC                                         ^^ +^ PIC                                    ^IA-32^X86_64^^ 
-|non-ZTS                                 |4.7s| +|non-ZTS                                 |4.7s |3.7s  
-|ZTS                                     |6.4s| +|ZTS                                     |6.4s |4.7s  
-|ZTS patched, static globals, dynamic TLS|4.8s|+|ZTS patched, static globals, dynamic TLS|4.8s |3.8s  |
  
 So the patch made ZTS builds mostly as fast as non-ZTS builds. So the patch made ZTS builds mostly as fast as non-ZTS builds.
Line 34: Line 34:
 Unfortunately the TLS model used in these tests was the static model, which is restrictive, and in particular does not allow to use it in shared libraries which will be loaded dynamically. (c.f [[tls#tls_internals|TLS internals]] bellow). Unfortunately the TLS model used in these tests was the static model, which is restrictive, and in particular does not allow to use it in shared libraries which will be loaded dynamically. (c.f [[tls#tls_internals|TLS internals]] bellow).
  
-Even if this patch can still be used with the dynamic model (it actually greatly improves the performance of ZTS-PIC builds), it can not be used as-is when building non-PIC code (the dynamic model cannot be used with non-PIC code).+For the PHP module to be able to be loaded at runtime in Apache or an other server with this patch enabled, it has to be built with the dynamic TLS model (on gcc, -ftls-model=global-dynamic), which also requires to build position independent code. While the last makes a big difference on IA-32, this is the default on x86_64 and the results on bench.php on an unpatched PHP are the same as IA-32/PIC builds. 
 + 
 +Native TLS can be enabled with %%--with-tsrm-__thread-tls%% or %%--with-tsrm-full__thread-tls%%. The last declares globals statically instead of making them pointers.
  
 ===== Second patch ===== ===== Second patch =====
Line 74: Line 76:
 Results: Results:
  
-^ Results                               ^^ +^ Results                         ^IA-32/non-PIC^X86_64/PIC^^ 
-|ZTS                              |5.2s| +|non-ZTS                          |3.7s         |3.7s      | 
-|ZTS-patched, native TLS disabled |5.0s| +|ZTS                              |5.2s         |4.7s      
-|ZTS-patched, native TLS enabled  |4.2s|+|ZTS-patched, native TLS disabled |5.0s         |4.4s      
 +|ZTS-patched, native TLS enabled  |4.2s         |4.0s      |
  
-Native TLS can be enabled with %%--with-tsrm-__thread-tls%% or %%--with-tsrm-full__thread-tls%% in the first patch.+Native TLS can be enabled with %%--with-tsrm-native-tls%%.
  
-For the second patch, the switch is %%--with-tsrm-native-tls%%.+So we have two patches: 
 + 
 +  * The first one will work only with position independent codeand is the faster on targets where this is the default or when comparing only to PIC builds. At least Debian builds PHP --with-pic, and I guess this is the case on other distributions too. 
 +  * The second one does not requires to build PIC code, can not fully take profit of TLS, but is the faster at least on IA-32. 
 + 
 +===== Windows ===== 
 + 
 +Dynamically loaded DLLs can use TLS starting with Windows Vista and Server 2008. But there is a restriction: TLS variables can't be exported, which means that they can't be accessed outside of the DLL. CLI and ISAPI SAPIs works with TLS enabled, but they must be built like this is done on other platforms, with all code embeded in the executable/library (instead of a separate php5ts.dll linked by SAPIs). The same apply for extension, they must be built statically in PHP.
  
 ===== TLS internals ===== ===== TLS internals =====
Line 89: Line 99:
 ==== Static model ==== ==== Static model ====
  
-Each block is allocated at a fixed (linker-defined) offset from an address specific to each thread. As this address can be accessed very quickly, this allows very quick access to each TLS block. For instance, on Linux/IA-32, this thread-specific-address is the Thread Control Block, whose address is stored in offset 0 of the %gs segment register.+Each block is allocated at a fixed (loader-defined) offset from an address specific to each thread. As this address can be accessed very quickly, this allows very quick access to each TLS block. For instance, on Linux/IA-32, this thread-specific-address is the Thread Control Block, whose address is stored in offset 0 of the %gs segment register.
  
 The way the static model works requires that the memory needed by each TLS variable to be allocated before program startup. This means that the static model can not be used in shared libraries loaded at runtime.  The way the static model works requires that the memory needed by each TLS variable to be allocated before program startup. This means that the static model can not be used in shared libraries loaded at runtime. 
Line 97: Line 107:
 Linux, Solaris, FreeBSD, Windows. Linux, Solaris, FreeBSD, Windows.
  
-Linux, Solaris and FreeBSD implementations allocate a fixed amount of surplus memory especially to allow dynamically loaded libraries to use the static model. Linux allocates 1664 bytes, FreeBSD 64 and Solaris 512. This amount of memory is always allocated in addition of the memory allocated for TLS before program startup, and is always available (this memory can be used only by dlopen()ed modules using static TLS).+Linux, Solaris and FreeBSD implementations allocate a fixed amount of surplus memory especially to allow dynamically loaded libraries to use the static model. Linux allocates 1664 bytes, FreeBSD 64 and Solaris 512. This amount of memory is always allocated in addition of the memory allocated for TLS before program startup, and is always available (this memory can be used only by dlopen()ed modules using static TLS). These behaviors are undocumented (except by comments in Linux and FreeBSD loaders/linkers code). This has been tested with a test program and verified by reading the relevant code on Linux and FreeBSD.
  
 On GCC this model can be selected by using -ftls-model=initial-exec. On SunStudio: -xthreadvar=no%dynamic. For both, this model is the default one when building non-PIC code. On GCC this model can be selected by using -ftls-model=initial-exec. On SunStudio: -xthreadvar=no%dynamic. For both, this model is the default one when building non-PIC code.
Line 107: Line 117:
 === Implementation === === Implementation ===
  
-Linux, Solaris, FreeBSD, Windows Vista. The Windows documentation does not gives many implementation details, but it seems that Windows Vista allows DLLs to use thread local storage (not tested). Other Windows versions have no equivalent of the dynamic model.+Linux, Solaris, FreeBSD, Windows Vista/Server 2008 
 + 
 +Windows Vista and Server 2008 can use TLS in DLLs loaded using LoadLibrary(), but TLS symbols cannot be exported, which means that only the DLL where a TLS variable is declared can refer to this variable.
  
 On GCC this model can be selected by using -ftls-model=general-dynamic. On SunStudio: -xthreadvar=dynamic. For both, this is the default when building PIC code. On GCC this model can be selected by using -ftls-model=general-dynamic. On SunStudio: -xthreadvar=dynamic. For both, this is the default when building PIC code.
Line 153: Line 165:
 As the patch avoids passing tsrm_ls across function calls, #ifdef ZTS is not anymore relevant to check that. As the patch avoids passing tsrm_ls across function calls, #ifdef ZTS is not anymore relevant to check that.
 The new PASS_TSRMLS macro is now defined when tsrm_ls needs to be passed across function calls. For instance this is needed by ZEND_ATTRIBUTE_FORMAT and some other places. The new PASS_TSRMLS macro is now defined when tsrm_ls needs to be passed across function calls. For instance this is needed by ZEND_ATTRIBUTE_FORMAT and some other places.
- 
rfc/tls.1219609665.txt.gz · Last modified: 2017/09/22 13:28 (external edit)