On Thu, 15 Feb 2007 23:30:57 -0800 Jeremy Fitzhardinge [off-list ref] wrote:

quoted

If you really need to run atomically, that gets ugly.  Even of one were to
run handle_mm_fault() by hand, it still needs to allocate memory.

Two ugly options might be:

a) touch all the pages, then go atomic, then touch them all again.  If
   one of them faults (ie: you raced with swapout) then go back and try
   again.  Obviously susceptible to livelocking.

b) Do get_user_pages() against all the pages, then go atomic, then do
   put_page() against them all.  Of course, they can immediately get
   swapped out.

But that function's already racy against swapout and I guess it works OK. 
I don't have clue what it is actually trying to do, so I'm guessing madly
here.
  

It's for populating the pagetable in a vmalloc area.  There's magic in
the fault handler to synchronize the vmalloc mappings between different
process's kernel mappings, so if the mapping isn't currently present, it
will fault and create the appropriate mapping.  It's not operating on
swappable user memory, so swapping isn't an issue; but if the fault
handler exits immediately with preempt disabled, then there's a problem.

oh, I see.  The vmalloc fault can run atomically.  In fact it can run at
hard iRQ.  So no probs (apart from the fact that it required an email
dialogue to work this out rather than reading the code, but I do go on).