The paging design used on the x86-64 linux kernel port in 2.4.x provides: o per process virtual address space limit of 512 Gigabytes o top of userspace stack located at address 0x0000007fffffffff o start of the kernel mapping = 0x0000010000000000 o global RAM per system 508*512GB=254 Terabytes o no need of any common code change o 512GB of vmalloc/ioremap space Description: x86-64 has a 4 level page structure, similar to ia32 PSE but with some extensions. Each level consits of a 4K page with 512 64bit entries. The levels are named in Linux PML4, PGD, PMD, PTE; AMD calls them PML4E, PDPE, PDE, PTE respectively. For direct and kernel mapping only 3 levels are used with the PMD pointing to 2MB pages. Userspace is able to modify and it sees only the 3rd/2nd/1st level pagetables (pgd_offset() implicitly walks the 1st slot of the 4th level pagetable and it returns an entry into the 3rd level pagetable). This is where the per-process 512 Gigabytes limit cames from. The common code pgd is the PDPE, the pmd is the PDE, the pte is the PTE. The PML4 remains invisible to the common code. Since the per-process limit is 512 Gigabytes (due to kernel common code 3 level pagetable limitation), the higher virtual address mapped into userspace is 0x7fffffffff and it makes sense to use it as the top of the userspace stack to allow the stack to grow as much as possible. The kernel mapping and the direct memory mapping are split. Direct memory mapping starts directly after userspace after a 512GB gap, while kernel mapping is at the end of (negative) virtual address space to exploit the kernel code model. There is no support for discontig memory, this implies that kernel mapping/vmalloc/ioremap/module mapping are not represented in their "real" mapping in mem_map, but only with their direct mapped (but normally not used) alias. Future: During 2.5.x we can break the 512 Gigabytes per-process limit possibly by removing from the common code any knowledge about the architectural dependent physical layout of the virtual to physical mapping. Once the 512 Gigabytes limit will be removed the kernel stack will be moved (most probably to virtual address 0x00007fffffffffff). Nothing will break in userspace due that move, as nothing breaks in IA32 compiling the kernel with CONFIG_2G. Linus agreed on not breaking common code and to live with the 512 Gigabytes per-process limitation for the 2.4.x timeframe and he has given me and Andi some very useful hints... (thanks! :) Thanks also to H. Peter Anvin for his interesting and useful suggestions on the x86-64-discuss lists! Current PML4 Layout: Each CPU has an PML4 page that never changes. Each slot is 512GB of virtual memory. 0 user space pgd or 40MB low mapping at bootup. Changed at context switch. 1 unmapped 2 __PAGE_OFFSET - start of direct mapping of physical memory ... direct mapping in further slots as needed. 509 some io mappings (others are in a memory hole below 4gb) 510 vmalloc and ioremap space 511 kernel code mapping, fixmaps and modules. Other memory management related issues follows: PAGE_SIZE: If somebody is wondering why these days we still have a so small 4k pagesize (16 or 32 kbytes would be much better for performance of course), the PAGE_SIZE have to remain 4k for 32bit apps to provide 100% backwards compatible IA32 API (we can't allow silent fs corruption or as best a loss of coherency with the page cache by allocating MAP_SHARED areas in MAP_ANONYMOUS memory with a do_mmap_fake). I think it could be possible to have a dynamic page size between 32bit and 64bit apps but it would need extremely intrusive changes in the common code as first for page cache and we sure don't want to depend on them right now even if the hardware would support that. PAGETABLE SIZE: In turn we can't afford to have pagetables larger than 4k because we could not be able to allocate them due physical memory fragmentation, and failing to allocate the kernel stack is a minor issue compared to failing the allocation of a pagetable. If we fail the allocation of a pagetable the only thing we can do is to sched_yield polling the freelist (deadlock prone) or to segfault the task (not even the sighandler would be sure to run). KERNEL STACK: 1st stage: The kernel stack will be at first allocated with an order 2 allocation (16k) (the utilization of the stack for a 64bit platform really isn't exactly the double of a 32bit platform because the local variables may not be all 64bit wide, but not much less). This will make things even worse than they are right now on IA32 with respect of failing fork/clone due memory fragmentation. 2nd stage: We'll benchmark if reserving one register as task_struct pointer will improve performance of the kernel (instead of recalculating the task_struct pointer starting from the stack pointer each time). My guess is that recalculating will be faster but it worth a try. If reserving one register for the task_struct pointer will be faster we can as well split task_struct and kernel stack. task_struct can be a slab allocation or a PAGE_SIZEd allocation, and the kernel stack can then be allocated in a order 1 allocation. Really this is risky, since 8k on a 64bit platform is going to be less than 7k on a 32bit platform but we could try it out. This would reduce the fragmentation problem of an order of magnitude making it equal to the current IA32. We must also consider the x86-64 seems to provide in hardware a per-irq stack that could allow us to remove the irq handler footprint from the regular per-process-stack, so it could allow us to live with a smaller kernel stack compared to the other linux architectures. 3rd stage: Before going into production if we still have the order 2 allocation we can add a sysctl that allows the kernel stack to be allocated with vmalloc during memory fragmentation. This have to remain turned off during benchmarks :) but it should be ok in real life. Order of PAGE_CACHE_SIZE and other allocations: On the long run we can increase the PAGE_CACHE_SIZE to be an order 2 allocations and also the slab/buffercache etc.ec.. could be all done with order 2 allocations. To make the above to work we should change lots of common code thus it can be done only once the basic port will be in a production state. Having a working PAGE_CACHE_SIZE would be a benefit also for IA32 and other architectures of course. vmalloc: vmalloc should be outside the first 512GB to keep that space free for the user space. It needs an own pgd to work on in common code. It currently gets an own pgd in the 510th slot of the per CPU PML4. PML4: Each CPU as an own PML4 (=top level of the 4 level page hierarchy). On context switch the first slot is rewritten to the pgd of the new process and CR3 is flushed. Modules: Modules need to be in the same 4GB range as the core kernel. Otherwise a GOT would be needed. Modules are currently at 0xffffffffa0000000 to 0xffffffffafffffff. This is inbetween the kernel text and the vsyscall/fixmap mappings. Vsyscalls: Vsyscalls have a reserved space near the end of user space that is acessible by user space. This address is part of the ABI and cannot be changed. They have ffffffffff600000 to ffffffffffe00000 (but only some small space at the beginning is allocated and known to user space currently). See vsyscall.c for more details. Fixmaps: Fixed mappings set up at boot. Used to access IO APIC and some other hardware. These are at the end of vsyscall space (ffffffffffe00000) downwards, but are not accessible by user space of course. Early mapping: On a 120TB memory system bootmem could use upto 3.5GB of memory for its bootmem bitmap. To avoid having to map 3.5GB by hand for bootmem's purposes the full direct mapping is created before bootmem is initialized. The direct mapping needs some memory for its page tables, these are directly taken from the physical memory after the kernel. To access these pages they need to be mapped, this is done by a temporary mapping with a few spare static 2MB PMD entries. Unsolved issues: 2MB pages for user space - may need to add a highmem zone for that again to avoid fragmentation. Andrea SuSE Andi Kleen SuSE $Id$