GFP_ATOMIC or GFP_KERNEL?

本文转载至：http://lists.metaprl.org/pipermail/cs134-labs/2002-October/000025.html

The short (or kinda long) answer is this:

GFP_ATOMIC means roughly "make the allocation operation atomic".  This
means that the kernel will try to find the memory using a pile of free
memory set aside for urgent allocation.  If that pile doesn't have
enough free pages, the operation will fail.  This flag is useful for
allocation within interrupt handlers.

GFP_KERNEL will try a little harder to find memory.  There's a
possibility that the call to kmalloc() will sleep while the kernel is
trying to find memory (thus making it unsuitable for interrupt
handlers).  It's much more rare for an allocation with GFP_KERNEL to
fail than with GFP_ATOMIC.

In all cases, kmalloc() should only be used allocating small amounts of
memory (a few kb).  vmalloc() is better for larger amounts.

Also note that in lab 1 and lab 2, it would have been arguably better to
use GFP_KERNEL instead of GFP_ATOMIC.  GFP_ATOMIC should be saved for
those instances in which a sleep would be totally unacceptable.

This is a fuzzy issue though...there's no absolute right or wrong
answer.