Memory Management and RMAP VM of 2.6

1
Memory Management and RMAP VM of 2.6

• By
• A.R.Karthick
• (karthick_r_at_infosys.com)

2
Memory Hierarchies
L1 cache
T I M E
L2 cache
RAM
Hard Disk
3
Page Tables

• Define the virtual to physical mapping
• Page directory,page mid level directory,page
  table entry define the course of translation
• Example
• PGD 10 bits PTE 10 BITS (PMD folded in 32
  bit)
• 0000 0000 00 00 1000 0000 1100 0000 1111 -gt
  (0x00080c0f)
• pgd index(0) pte_index(1 ltlt 7) , pmd is folded to
  pgd

4
Page Table Entry Status Bits (PTE Entry)
PAGE_PRESENT PAGE_RW PAGE_USER PAGE_RESERVED PAGE_
ACCESSED PAGE_DIRTY INTERNAL_STATUS
5
Page Fault

• Processor Exception raised when there is a
  problem mapping the virtual address to physical
  address.
• Handled by do_page_fault in arch/i386/mm/fault.c.
• Write protection faults or COW faults map to
  do_wp_page.
• For pages in swap, do_swap_page is called.
• For pages not found, do_no_page is called that
  either faults in an anonymous zero page or an
  existing page.
• Page faults populate the LRU cache.

6
Page Replacement Algorithms

• Optimal Replacement ? Not possible
• Not Recently Used (NRU) ? Crude hack
• FIFO ? Inefficient
• Second Chance ? Better than above
• Clock Replacement ? Efficient than above

7
Page Replacement Algorithms

• LRU Least Recently used replacement
• NFU Not Frequently Used replacement
• Page Ageing based replacement
• Working Set algorithm based on locality of
  references per process
• Working Set based clock algorithms
• LRU with Ageing and Working Set algorithms are
  efficient to use and are commonly used

8
Page replacement handling in Linux Kernel

• Page Cache
• Pages are added to the Page cache for fast
  lookup.
• Page cache pages are hashed based on their
  address space and page index
• Inode or disk block pages, shared pages and
  anonymous pages form the page cache.
• Swap cached pages also part of the page cache
  represent the swapped pages.
• Anonymous pages enter the swap cache at swap-out
  time and shared pages enter when they become
  dirty.

9
LRU CACHE

• LRU cache is made up of per zone active lists and
  inactive lists.
• Per-CPU lru active and inactive page vectors make
  lru cache additions faster.
• These lists are populated during page faults and
  when page cached pages are accessed or
  referenced.
• kswapd is the page out kernel thread per node
  that balances the LRU cache and trickles out
  pages based on an approximation to LRU algorithm.
• Page stealing is performed on a page vector or
  performed in batches.
• Active state
• Inactive dirty state
• Inactive clean state
• Per-CPU cold pages

10
Zone Balancing

• Kswapd performs zone balancing based on
  pages_high, pages_low and pages_min
• Zone is considered balance with its free pages
  above pages_high
• The page out process takes a page by scanning
  inactive pages in batches.
• Batch page stealing scales well for large
  physical memory.

11
RMAP

• Maintains mapping of a page to a pte/virtual
  address
• Greatly speeds up the page unmap path without
  scanning the process virtual address space
• Unmapping of shared pages is greatly improved
  because of availability of pte mappings for
  shared pages
• Page faults are reduced because pte entries are
  unmapped only when required.
• Reduced search space during page replacement as
  only inactive pages are touched.
• Low overhead involved in adding reverse mapping
  during fork, page fault , mmap and exit paths.

12
RMAP

• struct pte_chain
• unsigned long next_and_idx
• pte_addr_t ptesNRPTE
• ____cachelinealigned
• next_and_idx field contains both the index to the
  next pte in the same chain or a pointer to the
  next pte chain ,thus aiding in fast pte chaining.
• pte chains have free slots at the top or the head
  of the chain and additions happen from the tail.
• process mm_struct pointer is kept in the pages
  address space, that is used during swapout times.

13
VM-Overcommit Policies

• Commit more than available/actual memory space
  which includes the swap space to the process.
• Overcommit policies can be set through sysctl
  vm.overcommit_memory,ratio
• 0 indicates no overcommit
• 1 indicates overcommit totally.
• 2 indicates overcommit with overcommit_ratio on
  total ram pages plus total swap space pages.
• mmap,mprotect, munmap, brk, shared memory,
  affect overcommit.

14
References

• Primarily Linux Kernel Source Code 2.6
• Towards an O(1) VM by Rik Van Riel Proceedings
  of the Linux Symposium Ottawa