On Sat, Nov 13, 2004 at 02:58:55AM +1000, Adam Conrad wrote: > Mark: Can you think of any reason why forcing MMFILE might lead to the > symptoms described in the two bug reports CCd?
Well, the MMFILE implementation is a mmapped file. This will cause it to allocate backing store for the shared memory segment when it is used. If this is only showing up on UML systems my first guess would be that sending the process a signal (as a configuration reload does) causes the UML kernel to allocate backing store for the entire memory mapping. Otherwise the reload probably just causes PHP to walk over the entire shared memoy segment. I'd guess that the old build would've been using System V shared memory which wouldn't have the backing store - it could be figured out readily although the mirror I use doesn't seem to have the old stable mm binary package and I managed to loose my archive a while ago. If it had ended up doing that then there would be no such issues with backing store. MMFILE was chosen because it's the most conservative mechanism in terms of kernel compatibility and when System V shared memory did get used it caused problems for PHP since the default system limit on shared memory is 32Mb with 2.6 kernels - PHP tries to allocate 64Mb and fails. -- "You grabbed my hand and we fell into it, like a daydream - or a fever."
