On Tue, 20 May 2014 00:39:26 -0000, Chan Kim said:
But still it's confusing. Can two virtual addresses from the "same process" (in init process, one for nocache pool, the other not) point to the same physical address?
I'm not sure what you're trying to say there. In general, the hardware tags like non-cacheable and write-combining are applied to physical addresses, not virtual. And a moment's thought will show that treating the same address (whether it's virtual or physical) as caching in one place and non-caching in another is just *asking* for a stale-data bug when the non-caching reference updates data and the caching reference thinks it's OK to use the non-flushed non-refreshed cached data. It's easy enough to test if two addresses from a single process can point to the same physical address - do something like this: /* just ensure these two map the same thing at different addresses */ foo = mmap(something,yaddayadda); bar = mmap(something,yaddayadda); /* modify via one reference */ *foo = 23; /* you probably want a barrier call here so gcc doesn't screw you */ /* Now dereference it via the other reference */ printf("And what we read back is %d\n", *bar); (Making this work is left as an exercise for the student :) And figuring out why you need a barrier is fundamental to writing bug-free code that uses shared memory. The file Documentation/memory-barriers.txt is a good place to start.