Data (part 3)

0. Pitch

• demo: number of page faults by an application & system-level
• demo: total value of RSS vs. total physical memory used

1. Process Virtual Address Space: Perspectives

• list of zones vs array of page descriptions (PTE - page table entry)

  struct zone *pvas_head;
  struct zone {
        unsigned long start;
        unsigned long size;
        unsigned int permissions;
        struct zone *prev, *next;
  };
  
  struct page page_table[N];
  struct page {
      bool valid;
      unsigned int frame_index;
      unsigned int permissions;
      unsigned int flags;
  }

• demo: pmap $$ how many zones, how many valid PTEs

• diagram: zones, page table, correspondence

• zones: managed by OS

• page table: managed by hardware / MMU / MTU (Memory Management Unit, Memory Translation Unit)

• virtual to physical address translation (done by MMU / MTU): page address + offset, translate page address to frame address, add offset

2. (Virtual) Memory Operations

• virtual memory reservation: only affects the list of zones view
  • reserved explicitly by a memory allocation call (malloc(), mmap())
• diagram: reserve virtual memory
• allocate physical memory: only affects the page table - validity bit is updated
• allocate at access - not explicit
• diagram: allocation of physical memory, mapping frames to pages, validation of PTEs

3. Page Fault

• demo: show page fault metrics of processes
• a page fault is triggered when access is done to a page that's invalid or with the wrong permissions
• flow is: memory access instruction, aim to translate virtual address to physical address (MMU / MTU); if page table entry is invalid, trigger page fault; run page fault handler; check the zones view; allocate physical page and update page table entry
• diagram with flow
• demo: reservation and allocation, see vsz, rss, min_flt, maj_flt metrics
• run kernel space code: syscall or memory access causing a page fault
• page faults: minor page fault, major page fault, invalid page fault (causes seg fault), permissions page fault (generally causes segfault)
• demo: seg faults
• diagram: demand paging
• diagram: swapping

4. Memory Sharing

• demo: resident memory used by 10 processes started from:
  • the same dynamic executable
  • the same static executable
  • 10 different dynamic executables
  • 10 different static executables
• diagram: same physical memory, multiple address spaces accessing it
• implicit: executables, libraries
• explicit: shm... calls, require synchronization - part of App Interaction lectures

5. Storing the Page Table

• page table in memory
• two issues:
  • double memory access: use TLB - a cache of the page table: diagram
  • a lot of space use multi-level paging: diagram

Conclusion and Takeaways

• virtual memory gives the appearance of using more memory than actually being used
• physical memory is allocated at access
• page table stores state of the physical page
• invalid page may mean not part of the VAS (segfault at access), or not yet allocated / mapped (no segfault at access)
• allocation is triggered by access, and by signal sent by the MMU / MTU to the operating system (page fault handler)
• frames not used are swapped on disk (swap space)
• benefit of virtual memory is memory sharing
• downside of virtual memory is space allocated for page table (alleviated with multi-level paging) and double memory access (page table + actual access) (alleviated by TLB)