1 / 19

Access to High Memory

Access to High Memory. Introduction to the kernel functions ‘kmap()’ and ‘kunmap(). 1GB ram?: too many ‘pages’!. Some hardware needs virtual addresses (e.g., network and video display cards) CPU also needs some virtual addresses (e.g., for interprocessor communication)

butlerb
Download Presentation

Access to High Memory

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Access to High Memory Introduction to the kernel functions ‘kmap()’ and ‘kunmap()

  2. 1GB ram?: too many ‘pages’! • Some hardware needs virtual addresses (e.g., network and video display cards) • CPU also needs some virtual addresses (e.g., for interprocessor communication) • So not enough virtual addresses remain when system has a gigabyte or more of ram

  3. Linux kernel’s ‘solution’ • The kernel-space addresses are ‘shared’ • The high-memory page-frames are made visible to the CPU (i.e., are ‘mapped’) only while access to them is actually needed • Then these page-frames are ‘unmapped’ so that other page-frames can ‘reuse’ the same virtual address • But sharing only affects ‘high memory’

  4. kernel mapping function • ‘kmap()’ can create new page-mappings • It works on a page-by-page basis • ‘kunmap()’ can invalidate these mappings • Prototypes: #include <linux/highmem.h> • void * kmap( struct page * page ); • void kunmap( struct page * page );

  5. Visualizing kernel-space • Memory below 896MB is ‘directly mapped’ • ‘phys_to_virt()’ just adds 0xC0000000 • ‘virt_to_phys()’ subtracts 0xC0000000 • Master Page Directory provides a template user space 0-127 128-255 256-383 384-511 512-639 640-767 768-895 896-1023 kernel space 896 - 1023

  6. Recall ‘mem_map[ ]’ array unsigned long num_physpages; struct page *mem_map; struct page *highmen_start_page; highmem_start_page = &mem_map[ 896 ]; struct page { … ; void * virtual; };

  7. How ‘kmap()’ works void * kmap( struct page * page ) { if ( page < highmem_start_page ) return page->virtual; return kmap_high( page ); }

  8. How does ‘kmap_high()’ work? • Kernel first performs a quick check: • Is the high page is already ‘mapped’? • If so, just return its existing virtual address • If not, find an ‘unused’ virtual address • Set page-table entry that ‘maps’ the page • BUT… … there might not be any ‘unused’ entries!

  9. Details of ‘kmap_high()’ void kmap_high( struct page * page ) { unsigned long vaddr; spin_lock( &kmap_lock ); vaddr = (unsigned long)page->virtual; if ( !vaddr ) vaddr = map_new_virtual( page ); ++pkmap_count[ (vaddr - PKMAP_BASE)>>12 ]; spin_unlock( &kmap_lock ); return (void *)vaddr; }

  10. Task maybe will sleep • If a new kernel mapping cannot be set up (because all addresses are already in use) then the ‘map_new_virtual()’ function will ‘block’ the task (i.e., will put it to sleep)

  11. How process gets blocked DECLARE_WAITQUEUE( wait, current ); current->state = TASK_UNINTERRUPTIBLE; add_wait_queue(&pkmap_map_wait,&wait); spin_unlock( &kmap_lock ); schedule(); remove_wait_queue(&pkmap_map_wait,&wait); spin_lock( &kmap_lock ); if ( page->virtual ) return page->virtual;

  12. Purpose of ‘kunmap()’ • Programming sequence: void * vaddr = kmap( page ); // get address … /* use this virtual address to access the page */ … kunmap( page ); // release the address • Now kernel can awaken ‘blocked’ tasks

  13. Reference counters • Kernel maintains an array of counters static int pkmap_count[ LAST_PKMAP ]; • One counter for each ‘high memory’ page • Counter values are 0, 1, or more than 1 • =0: page is not mapped • =1: page not mapped now, but used to be • =n >1: page was mapped (n-1) times

  14. Flushing the TLB • ‘kunmap( page )’ decrerments refcount • When refcount == 1, mapping isn’t needed • But CPU still has ‘cached’ that mapping • So the mapping must be ‘invalidated’ • With multiple CPUs, all of them must do it

  15. Special CPU instruction • Pentium has a special instruction: invlpg m ; invalidates one TLB entry • But it only affects CPU that executes it. • In SMP systems, aother CPUs may still have ‘stale’ physical-to-virtual address-mapping cached in their TLB • Prevents other CPU from uptodate access

  16. Solution: notify other CPUs Backside bus CPU 0 CPU 1 I/O APIC System RAM Peripheral devices

  17. What about interrupts? • Interrupt-handlers can’t risk using ‘kmap()’ • Yet may need access to high memory data • Data might not be ‘mapped’ when needed! • So an ‘emergency’ mechanism is required

  18. Temporary Mappings • A small number of ‘windows’ are reserved • These are ‘emergency’ page-table entries • Just five or six such entries (per CPU) • Interrupt-handlers can use these safely • A temporary page-mapping is set up fast • No checking for other possible users • No assurance that mapping will persist

  19. ‘kmap_atomic()’ • May be safely called by interrupt-handlers • No need to do ‘unmapping’ • No TLB clashes with other CPUs • You can experiment with ‘kmap_atomic()’ • But use ‘kmap()/kunmap()’ for project #3

More Related