Performance is slow on Lion 64-bit

[krblock]

I recently tried using mach_override.c on Lion 64-bit and the performance was a lot slower than in previous OS versions on 32-bit. Based on some performance measurements I determine that the performance problem was due to a large number of calls to vm_allocate in the following snippet of code from allocateBrandIsland. In previous versions vm_allocate was only called once, but on Lion 64-bit is was called around 64K times.

        while( !err && !allocated && page != last ) {

            err = vm_allocate( task_self, &page, pageSize, 0 );
            if( err == err_none )
                allocated = 1;
            else if( err == KERN_NO_SPACE ) {

if defined(x86_64)

                page -= pageSize;

else

                page += pageSize;

endif

                err = err_none;
            }
        }

Not sure what the idea solution would be, however, I made two changes that greatly improved performance for me:

1. I cached the last call, if the same "first" address is used, I started at the next available page based on the last call.
2. When not using a cached call, make larger jumps through memory to find an available page quicker.

The changes where the following where COVERITY macro enables the change:

--- 1.15/build/capture/mach-override.c 2012-01-28 12:54:07 -08:00 +++ 1.16/build/capture/mach-override.c 2012-01-28 13:50:08 -08:00 @@ -361,6 +361,11 @@ mach_override_ptr(

    ************************************************************************

***/

+#if COVERITY +static vm_address_t lastFunctionAddr = 0; +static vm_address_t cacheAddrFirst = 0; +#endif + mach_error_t allocateBranchIsland( BranchIsland **island, @@ -389,7 +394,26 @@ allocateBranchIsland( vm_address_t last = 0xfffe0000;

endif

+#if COVERITY

• /* If we are starting from the same address again, skip ahead to where we
• found the address last time. */
• int using_cache = FALSE;
• vm_address_t page;
• if (lastFunctionAddr == first) {
• using_cache = TRUE; +#if defined(x86_64)
• page = cacheAddrFirst - pageSize; +#else
• page = cacheAddrFirst + pageSize; +#endif
• }
• else {
• page = first;
• } +#else vm_address_t page = first; +#endif + int allocated = 0; vm_map_t task_self = mach_task_self();

@@ -400,15 +424,30 @@ allocateBranchIsland( allocated = 1; else if( err == KERN_NO_SPACE ) {

if defined(x86_64)

+#if COVERITY

• /* If we are not using a cached value, the memory we want could be fa r away.
• Take bigger jumps */
• page -= (using_cache ? pageSize : pageSi ze*64); +#else page -= pageSize; +#endif

  else

                                  page += pageSize;

  endif

                                  err = err_none;
                          }
                  }

  +

+#if COVERITY

• if( allocated ) {
• lastFunctionAddr = first;
• cacheAddrFirst = page;
• island = (BranchIsland) page;
• } +#else if( allocated ) island = (BranchIsland) page; +#endif