Le mercredi 27 juillet 2011 à 22:56 +0200, Ingo Weinhold a écrit : > François Revol wrote: > > Le mercredi 27 juillet 2011 à 17:44 +0200, Stephan Assmus a écrit : > > > That being said, I believe the point of implementing PAE support (yes it > > > exists) was to be able to address more than 4 GB. Each individual > > > process in Haiku can only address 4GB, but the kernel should be able to > > > use more than 4GB for the mapping of virtual to physical memory. That is > > > > Yep, the PAE paging code uses 64bit types to hold the physical > > addresses, which seems to be 40bit according to the bitmasks in paging.h > > The actually supported physical address width is processor dependent. The > maximum possible width is 52 bit (4 PiB). That's the same for x86_64. Oh right, the 40bits are the page addresses, +12bits for offsets in the page. > > > of course only interesting when more than one process is running at a > > > time, and affects the necessity for swapping out pages. For example, > > > when one process indeed uses all of it's 4GB address space, the Haiku > > > kernel does not necessarily need to swap out virtual memory, when there > > > are more then 4GB available in physical RAM, since the address space of > > > another process could be mapped to another range of the physical memory. > > > Hope I have remembered these details correctly. > > More interestingly even, the kernel directly uses physical memory for the > file cache (unfortunately not for the block cache yet). So one can still > benefit from more installed RAM even when not using memory hungry > applications. And RAM disks, ... François.