5.7.5. Allocating Memory

• simplest form kmalloc(size, flags)
  • flags see linux/gfp.h
  • may sleep GFP_KERNEL, or not GFP_ATOMIC
  • fast, continuous, doesn't clear (security? -> kzalloc)
  • free with kfree(ptr)

• more memory needed? allocate pages:
  • get_zeroed_page(flags)
  • __get_free_page(flags) /* non zeroed */
  • __get_free_pages(flags, order)
    • allocate 2^order pages
    • calculate oder from size: get_order(size)
  • freeing
    • free_page(ptr)
    • free_pages(ptr, order)
  • advantages: slightly faster and avoids fragementation

• vmalloc / vfree
  • above functions allocate physically consecutive memory
  • vmalloc allocated memory is only virtually consecutive
  • cons:
    • slower because page tables need to be setup
    • not usable from outside of CPU (DMA)
  • pro:
    • large allocation less likely to fail due to fragementation
  • -> ok for e.g. large software only buffer

• Optimizations (premature opt ...)
  • Situation: driver allocates many objects of the same size over and over again
    • lookaside cache (see slab cache in DocBook "Memory Management in Linux")
    • nice: cache usage information in /proc/slabinfo
  • Sitation: memory allocation is not allowed to fail
    • mempool can be used for reserving memory in emergencies
    • Usually bad: memory is idle. Better deal with allocation failures